The Skeptical Environmentalist, by Bjorn Lomborg
Constantinople
http://www.sfgate.com/cgi-bin/article.cgi?file=/chronicle/archive/2002/03/04/MN77449.DTL
--------------------------------------------------------------------------------
Optimistic researcher draws pessimistic reviews
Critics attack view that life is improving
Keay Davidson, Chronicle Science Writer Monday, March 4, 2002
--------------------------------------------------------------------------------
A controversial new book by a Danish statistician claims that
environmentally speaking, the world is getting better, contrary
to the headline-making scary scenarios of the last few decades.
Some scientists are upset by the book, which they say is a case
study in the perils of the old saw about "lies, damned lies and
statistics." But other experts welcome it as a breath of
optimistic air amid the often alarmist press coverage of the
planetary environment.
Whoever's right, the fuss illustrates the challenge facing
scientifically savvy citizens: How can they decide what is and
isn't "good science" when the disputes are so technically complex
and the atmosphere is so politically charged?
The book is "The Skeptical Environmentalist" by Bjorn Lomborg, a
statistician and associate professor of political science at the
University of Aarhus, Denmark. Published by Cambridge University
Press, the book is a 515- page survey of global trends in
everything from human population and grain production to
illiteracy, working hours and planetary forest cover.
The blonde, sunny-faced, 42-year-old Lomborg says the bottom line
is clear: The world is not going to hell.
FAILED TO DISPROVE ECONOMIST Lomborg was inspired by the writings
of the late Julian Simon, an economist who stirred controversy in
the 1980s by making a similar claim. Lomborg -- who had been a
dues-paying member of the in-your-face environmental group
Greenpeace -- read an interview with Simon in Wired magazine, and
set out to disprove his claims.
But when Lomborg studied Simon's data closely, he concluded that
it wasn't "simple, American right-wing propaganda" after all.
Rather, "a surprisingly large amount of his points stood up to
scrutiny."
In his book, Lomborg takes on what he calls "the litany" of
environmental gloom and doom, and concludes that on every major
point, the claims made by advocacy groups are wrong or
exaggerated. Air pollution is diminishing, not worsening, he
argues; population growth is falling faster than expected; per-
capita food production is improving; and so on.
The book has some prominent champions -- especially the British
business magazine the Economist. A review in the Washington Post
called it a "magnificent achievement." And the noted science
writer Matt Ridley, writing in the London Daily Telegraph, said
it was "probably the most important book on the environment ever
written."
But it has come under heavy attack by others, including the top
scientific journals Science and Nature, which ran unfriendly
reviews of Lomborg's work. Indeed, many scientists and
environmental groups have gone out of their way to criticize and
even ridicule Lomborg.
"This is an old story," says Paul Ehrlich, a famed professor of
conservation biology at Stanford University whose scary book "The
Population Bomb" made him an environmental hero three decades
ago. "Every single review in the scientific literature has
pointed out the many, many, many egregious errors in this
(Lomborg) book. But it's being very heavily promoted for
political purposes."
'EGREGIOUS DISTORTIONS' The brouhaha really heated up in January,
when Scientific American -- an esteemed, 157-year-old publication
-- ran an 11-page attack on Lomborg.
The piece contained articles by four well-known environmental
specialists --
Stephen Schneider of Stanford, who edits the journal Climatic
Change; environmental scientist and energy expert John P. Holdren
of Harvard, formerly of the University of California at Berkeley;
John Bongaarts, a vice president at the Population Council in New
York City; and Thomas Lovejoy, chief biodiversity adviser to the
World Bank.
They assailed Lomborg for "egregious distortions" (Schneider),
for "elementary blunders of quantitative manipulation and
presentation that no self-respecting statistician ought to
commit" (Holdren), and for sections "poorly researched and
presented . . . shallow . . . rife with careless mistakes"
(Lovejoy).
In reply, Lomborg's defenders accuse Scientific American's
authors of overkill, of offering more polemic than substance. The
four short articles cite few specific errors -- for example,
Lomborg's use of the chemical term "catalyzing," when he should
have said "electrolyzing." And whatever his shortcomings, they
say Lomborg hardly deserves such personalized attacks.
His defenders also assail the magazine for not giving Lomborg the
opportunity to present his own views in the same issue of the
magazine.
"Scientific American . . . has stooped so low as to claim to
speak for all scientists by rounding up the usual alarmist
suspects to attack Lomborg," says John Christy, who directs the
Earth System Science Center at the University of Alabama in
Huntsville. He is a leading opponent of the idea that the planet
is experiencing unprecedented global warming.
"Lomborg's real personal achievement was to break out of the
Northern European mind set of an almost religiously held belief
in 'climate calamity' and look at the science," says Christy, a
native of Fresno. "Lomborg hasn't discovered 'new' information --
many of us have been publishing these real numbers for over a
decade."
'NUTS' TO SCIENTIFIC AMERICAN Scientific American chose to air
the story in a very biased way, some critics say. That belief
spurs one detractor, David Wojick of Electricity Daily, an
industry newsletter, to claim the magazine is "an American
institution now apparently gone nuts."
The "nuts" charge draws a laugh from Scientific American Editor
in Chief John Rennie.
"I still don't really feel that we've handled this in a bad way,
notwithstanding the critics," Rennie said in an interview. The
magazine originally considered doing a book review of Lomborg's
tome, but "we started to hear very clearly from scientists
(saying) the book was doing a disservice to their field."
Rennie considered giving Lomborg space to present his views in
the same January issue, alongside the attacks by Schneider,
Holdren, Bongaarts and Lovejoy. This idea was dropped, though.
The reason: Lomborg had already received so much press in other
publications that "we felt it would not be a terrible disservice"
to run just the four critics, while allowing Lomborg to reply in
a future issue. Lomborg's reply is tentatively scheduled to run
in the May issue.
The brouhaha has left Lomborg -- formerly an obscure specialist
in theoretical topics such as game theory -- dizzy over his
sudden ascent to international fame and notoriety. "I thought
initially we would have a couple of weeks of debate and that
would be it and we'd all move on," he said in a phone interview.
"But it just kept on and on and on."
Scientific American "should have focused on giving a balanced
view of the book," Lomborg says. The four critics offered mainly
"a few nitpicking points. . . . (But) on the important issues, it
seems like they (give) lots of very negative adjectives and
fairly little substance."
Lomborg took a 1 1/2-year sabbatical from his university to
debate his critics and promote his book. Last week, the Danish
government announced it was appointing Lomborg to head a new,
small environmental monitoring agency, the Institute for
Environmental Evaluation. It will operate independently of the
Danish version of the U.S. Environmental Protection Agency.
Lomborg says he plans to postpone indefinitely his return to
academia while he runs the new agency, which he expects will have
a staff of 10 and a budget of about $1 million. He says its
mission will be to decide the best ways to spend taxpayer dollars
on environmental remediation. He adds that his appointment has
made "a lot of people very angry" in Denmark.
MAKINGS OF A MARTYR Meanwhile, it seems that the controversy over
his book has helped to drive up sales. If so, it won't be the
first time that part of the scientific community has shot itself
in the foot. History shows that when scientists attack a lone
figure so relentlessly, they risk transforming him into someone's
martyr.
David Wojick, one of Lomborg's defenders, says Lomborg can be
grateful for Scientific American's "incredible arrogance and
bias." The magazine "is an American institution. It is sold in
drugstores and supermarkets across the land. And they spelled
Bjorn's name right.
"So I can imagine millions of four-eyed high school science
nerds, like I once was, now wanting to read 'The Skeptical
Environmentalist.' Thank you, Scientific American. Thank you very
much."
COSTS OF SOLVING ENVIRONMENTAL PROBLEMS
One argument of Bjorn Lomborg's book, "The Skeptical
Environmentalist," is
that it is not cost efficient to spend money on certain environmental
problems.
He cites a Harvard University study comparing the cost efficiency of
various
life-saving initiatives.
An example: regulating radioactive emissions at phosphorous plants
costs $2.
8 million. Doing so would save only one life per decade. The cost per
year of
life saved, the researchers calculated, is $9.2 million. By contrast,
requiring smoke detectors in homes is so cost-efficient that it saves
money.
Here is a selection of the initiatives studied and the
corresponding cost
per year of life saved.
Cost per
Initiative life-year saved
Federal law requiring smoke detectors in homes < $0
Reduced lead content of gasoline from 1.1 gram
to 0.1 gram per leaded gallon < $0
Measles, mumps and rubella immunization < $0
Mandatory seatbelt use laws $69
Influenza vaccination for high risk people $570
Mammography for women age 50 $810
Pneumonia vaccination for people 65 and older $2,000
Chlorination of drinking water $3,100
Screening blood donors for HIV $14,000
Low-cholesterol diet for men age 30 $19,000
Improve basic driver training $20,000
Flashing lights and gates at rail-highway crossings $45,000
National 55 mile per hour speed limit $89,000
Annual mammography for women age 55-64 $110,000
Air bags (versus manual lap belts) $120,000
Seat belts for passengers in school buses $2,800,000
Strengthen buildings in earthquake-prone areas $18,000,000
Arsenic emission control at glass
manufacturing plants $51,000,000
Radiation emission standard for nuclear
power plants $180,000,000
Benzene emission control at rubber tire
manufacturing plants $20,000,000,000
Source: Harvard University Center for Risk Analysis
Madbit
> So what do people think about this book?
"We also note, that Bjørn Lomborg so far has failed to publish his
findings in a relevant international scientific journal with peer
review."
from http://www.au.dk/~cesamat/debate.html
[snip]
> Last week, the Danish
> government announced it was appointing Lomborg to head a new,
> small environmental monitoring agency, the Institute for
> Environmental Evaluation.
The newly elected Danish government which has stated that it does not
trust experts (which must surely be why they chose Lomborg for the job
instead).
The new government which holds it's power due to support from Dansk
Folkeparti (Danish People's Party - socialist-nationalists,
'zi-nas'?). The party is explicitly against human rights ;-> Their
party 'Propagandaminister' Soren Espersen was (or is still?) a member
of DNSB (the Danish nazi movement).
[snip]
> Here is a selection of the initiatives studied and the
> corresponding cost
> per year of life saved.
> Cost per
> Initiative life-year saved
> Federal law requiring smoke detectors in homes < $0
> Reduced lead content of gasoline from 1.1 gram
> to 0.1 gram per leaded gallon < $0
> Measles, mumps and rubella immunization < $0
[snip]
>
> Source: Harvard University Center for Risk Analysis
Less than zero? I have an eerie feeling about what this kind of
statistics would tells us that we ought to do with the handicapped,
the old and the sick...
James A. Donald
On 5 Mar 2002 07:49:19 -0800, constan...@yahoo.com
(Constantinople) wrote:
> So what do people think about this book?
>
> http://www.sfgate.com/cgi-bin/article.cgi?file=/chronicle/archiv
> e/2002/03/04/MN77449.DTL
You can tell a man by the quality of his enemies.
On the internet, we often have debates about statistics, and
always one person cites statistics that allegedly prove one thing,
and another cites statistics that allegedly prove the opposite.
Lomborg was subjected to an attack article by Scientific American,
where a bunch of prominent scientists argued that if he had done
his statistics differently he MIGHT have got a different result.
But for the most part they conspicuously failed to themselves do
the statistics differently and themselves ACTUALLY get a
significantly different result, and when they did so, they were
obviously cherry picking a few particular special statistics in a
contrived fashion.
This debate reminds me of similar quarrels over statistics on
concealed carry, and "the vanishing middle class".
Critics of the statistic "more concealed carry, less crime"
reworked the data in an artificial fashion, and for ONE special and
peculiar contrived way of massaging the data, they managed to get
the result "more concealed carry, no significant effect on crime"
Marx predicted the middle class would vanish, between a few
billionaires on one side, and a mass of unskilled laborers on the
other side. This is an important prediction for those who hope
for anti capitalist revolution, for the forces of capitalist
revolution have always come from the middle class. Commies
therefore are always producing statistics showing that the middle
class is vanishing, for if it does not vanish, revolution will
remain a capitalist phenomenon, and the much promised socialist
revolution will continue to recede into the ever more distant
future. Some of these statistics are simply pulled out of their
asses, some of them force the desired result choosing the right
start and end date for each quantity, for example comparing home
ownership at bottom of the most recent slump with home ownership
at the height of the previous boom
So what did the critics of Lomborg do.
No much. Mostly they said IF we reworked the data, maybe we could
get different results. Well if they could have, they would have.
With the concealed carry data, and the housing data, if you look
at the whole series rather than a couple of cherry picked figures,
it is obvious the critics are cherry picking. The more data
someone provides, the harder it is to cook the statistics.
Lomborg provides lots of data. His critics merely provide glib
sneers.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
AQg8EBRjgxujNCeCDL3OVMvDYyDGAzrgvOTElzVW
4KYv57dXb0LPcbgjhGINtmlmT8WGNG1LFqiZyRLpp
Constantinople
> constan...@yahoo.com (Constantinople) wrote in message news:<e6ee4695.02030...@posting.google.com>...
> > So what do people think about this book?
>
> "We also note, that Bjørn Lomborg so far has failed to publish his
> findings in a relevant international scientific journal with peer
> review."
That "criticism" misrepresents his book. It is not original science.
It makes use of science conducted by others.
> from http://www.au.dk/~cesamat/debate.html
>
> [snip]
>
> > Last week, the Danish
> > government announced it was appointing Lomborg to head a new,
> > small environmental monitoring agency, the Institute for
> > Environmental Evaluation.
>
> The newly elected Danish government which has stated that it does not
> trust experts (which must surely be why they chose Lomborg for the job
> instead).
>
> The new government which holds it's power due to support from Dansk
> Folkeparti (Danish People's Party - socialist-nationalists,
> 'zi-nas'?). The party is explicitly against human rights ;-> Their
> party 'Propagandaminister' Soren Espersen was (or is still?) a member
> of DNSB (the Danish nazi movement).
OK, you started with a misrepresentation, and you continue with an
ad hominem.
> [snip]
> > Here is a selection of the initiatives studied and the
> > corresponding cost
> > per year of life saved.
> > Cost per
> > Initiative life-year saved
> > Federal law requiring smoke detectors in homes < $0
> > Reduced lead content of gasoline from 1.1 gram
> > to 0.1 gram per leaded gallon < $0
> > Measles, mumps and rubella immunization < $0
> [snip]
> >
> > Source: Harvard University Center for Risk Analysis
>
> Less than zero? I have an eerie feeling about what this kind of
> statistics would tells us that we ought to do with the handicapped,
> the old and the sick...
Now you latch onto a detail of notation.
I wonder if anyone has something real to say about Lomborg's book.
David Friedman
constan...@yahoo.com (Constantinople) wrote:
> > [snip]
> > > Here is a selection of the initiatives studied and the
> > > corresponding cost
> > > per year of life saved.
> > > Cost per
> > > Initiative life-year saved
> > > Federal law requiring smoke detectors in homes < $0
> > > Reduced lead content of gasoline from 1.1 gram
> > > to 0.1 gram per leaded gallon < $0
> > > Measles, mumps and rubella immunization < $0
> > [snip]
> > >
> > > Source: Harvard University Center for Risk Analysis
> >
> > Less than zero? I have an eerie feeling about what this kind of
> > statistics would tells us that we ought to do with the handicapped,
> > the old and the sick...
>
> Now you latch onto a detail of notation.
1. If I followed the post correctly, the table is quoted in Lomborg's
book, but the source is a study by someone at Harvard.
2. Cost per life less than zero is a perfectly straightforward concept.
Suppose some precaution (immunization, say) costs $100,000, saves
$200,000 in lost time at work (or anything else other than mortality or
consequences of mortality), and saves 10,000 lives. The net cost os
saving those lives is -$100,000, or -$10/life saved.
--
David Friedman
www.daviddfriedman.com/
Madbit
http://www.gristmagazine.com/grist/books/lomborg121201.asp
Mike Hammock
> Critics of the statistic "more concealed carry, less crime"
> reworked the data in an artificial fashion, and for ONE special and
> peculiar contrived way of massaging the data, they managed to get
> the result "more concealed carry, no significant effect on crime"
To what paper do you refer here?
Mike Hammock
mham4...@aol.com
mha...@emory.edu
mha...@learnlink.emory.edu
Constantinople
news:aec8d815.0203...@posting.google.com:
It's easy enough to find negative reviews for a book that claims to explode
many of the myths of environmentalist activists, regardless of whether or not
it is a good book. The most painful negative reviews come from the magazines
Science, Nature, and Scientific American. Lomborg's website has some quite
devastating replies to Scientific American and Nature, penned not by him but
by scientists who take to task other scientists for letting their political
passions get the better of their scientific competence.
The essence of the approach of the scientists who attacked Lomborg so
hysterically and sloppily seems to be captured well in a reply to one of
these replies, written by another scientist, a meteorologist at MIT.
http://www.ps.au.dk/vip/lomborg/files/LindzenSciAmerLomborg.pdf
Thus, at one fell swoop, Schneider misrepresents both the book he
is attacking and the science that he is allegedly representing.
This goes well beyond Schneider’s infamous admission in 1989: "...
we are not scientists but human beings as well. And like most
people we’d like to see the world a better place, which in this
context translates into our working to reduce the risk of
potentially disastrous climate change. To do that we need to
get some broad based support, to capture the public’s imagination.
That, of course, entails getting loads of media coverage. So we
have to offer up scary scenarios, make simplified, dramatic
statements, and make little mention of any doubts we might have."
But, Schneider’s approach proved a slippery slope indeed .
Richard S. Lindzen, Alfred P. Sloan Professor of Meteorology at MIT
Constantinople
read.news.verio.net:
You're right! I had assumed it was a peculiarity of notation, but it is
perfectly possible for a precaution to pay off in other gains on top of the
lives saved.
Thanks, that blew off a conceptual block.
Tim Lambert
> jam...@echeque.com (James A. Donald) wrote in message news:<3c7f0a40...@east.usenetserver.com>...
>
> > Critics of the statistic "more concealed carry, less crime"
> > reworked the data in an artificial fashion, and for ONE special and
> > peculiar contrived way of massaging the data, they managed to get
> > the result "more concealed carry, no significant effect on crime"
>
> To what paper do you refer here?
I imagine that he is referring to Black and Nagin's critique. One of
their points is that excluding Florida causes the results for murder
and rape to no longer be significant. Why James considers this to be
artificial peculiar and contrived, I don't know.
You might be interested in Ted Goertzel's article in the Jan 2002
Skeptical Inquirer. It starts:
"Do you believe that every time a prisoner is executed in the United
States, eight future murders are deterred? Do you believe that a 1%
increase in the number of citizens licensed to carry concealed weapons
causes a 3.3% decrease in the state's murder rate? Do you believe that
10 to 20% of the decline in crime in the 1990s was caused by an
increase in abortions in the 1970s? Or that the murder rate would have
increased by 250% since 1974 if the United States had not built so
many new prisons?"
and goes on to explain why he thinks the studies these beliefs are
based on are junk science.
http://www.crab.rutgers.edu/~goertzel/mythsofmurder.htm
Tim
James A. Donald
> > > Critics of the statistic "more concealed carry, less crime"
> > > reworked the data in an artificial fashion, and for ONE special and
> > > peculiar contrived way of massaging the data, they managed to get
> > > the result "more concealed carry, no significant effect on crime"
> > To what paper do you refer here?
Tim Lambert
> I imagine that he is referring to Black and Nagin's critique. One of
> their points is that excluding Florida causes the results for murder
> and rape to no longer be significant. Why James considers this to be
> artificial peculiar and contrived, I don't know.
They also do lots of other things with dates, aggregation, and
disaggregation. to make the numbers come out right, of which simply
throwing away one major state is merely the most conspicuous.
The other things they do are not so outrageous, but they are merely
one special and way of looking at the data, out of a large number of
possible ways.
One important manipulation of their data is to conceal the fact that
concealed carry laws tended to be passed in response to a rise in
violent crime. They aggregated the data so that rises in violent
crime preceding concealed carry laws were counted as violent crime
following concealed carry laws, an aggregation that comes completely
out of the blue with no more justification than excluding Florida.
The more you process the data, more stuff that you add, subtract,
factor, multiply, and divide, the more opportunities you have to make
arbitrary choices to make the data come out right. They cooked the
data a lot, making lots and lots of choices. Because the data was so
thoroughly massaged, they had thousands of possible equally legitimate
ways of doing it. Presumably they tried thousands of ways, until they
found the one way that gave ONE number close to those they wanted.
Those who lie with statistics give a magick process that yields ONE
magic number. Those who tell the truth with statistics give tables
and chart, providing lots of data, rather than a single magick number
that has been conjured up by a complicated and incompletely explained
procedure.
>
> You might be interested in Ted Goertzel's article in the Jan 2002
> Skeptical Inquirer. It starts:
>
> "Do you believe that every time a prisoner is executed in the United
> States, eight future murders are deterred? Do you believe that a 1%
> increase in the number of citizens licensed to carry concealed weapons
> causes a 3.3% decrease in the state's murder rate
The number "3.3%" would certainly be junk science, if anyone had given
it. That a small increase in concealed carry causes a large decrease
in violent crime is not junk science, but simple fact.
Mike Hammock
> > To what paper do you refer here?
>
> I imagine that he is referring to Black and Nagin's critique. One of
> their points is that excluding Florida causes the results for murder
> and rape to no longer be significant. Why James considers this to be
> artificial peculiar and contrived, I don't know.
I thought perhaps he was referring to:
Rubin, Paul, and Dezhbakhsh, Hashem, "Lives Saved or Lives Lost: The
Effect of Concealed Handgun Laws on Crime," American Economic Review
Papers and Proceedings, May, 1998, 468-474.
They use Lott's own data set, with some enhancements, and suggest that
he doesn't include some important cross-county variation. Including
them appears to make concealed-carry laws have little to no effect on
crime (and in some cases, seems to increase it--although I think the
strongest statement one can make is that the laws have no clear
effect).
> You might be interested in Ted Goertzel's article in the Jan 2002
> Skeptical Inquirer. It starts:
>
> "Do you believe that every time a prisoner is executed in the United
> States, eight future murders are deterred? Do you believe that a 1%
> increase in the number of citizens licensed to carry concealed weapons
> causes a 3.3% decrease in the state's murder rate? Do you believe that
> 10 to 20% of the decline in crime in the 1990s was caused by an
> increase in abortions in the 1970s? Or that the murder rate would have
> increased by 250% since 1974 if the United States had not built so
> many new prisons?"
It doesn't to be the case that every execution deters eight murders;
the statistics suggested that on _average_, eight murders were
deterred. There's a paper coming out sometime soon in J. Law and Econ
suggesting that it's more like between 10 and 20-something murders
deterred per execution. I don't see how one could have strong a
priori beliefs about the impact of a 1% increase in concealed-carry
permits--I didn't have much more than a guess myself. That's what the
data is for. Is Goertzel suggesting that we do statistics, and then
ignore the results, going instead by intuition? It's better to
examine and criticize statistical techniques than to say "The results
don't jive with my intuition, so I disregard them."
For the last bits of data--murder, abortions, and prisons--I'm not
familiar with the empirical work, and am therefore not qualified to
say anything. But they don't strike me as obviously wrong.
> and goes on to explain why he thinks the studies these beliefs are
> based on are junk science.
It seems to me that this is the important part. Thanks for the link.
Tim Lambert
> James A. Donald
> > > > Critics of the statistic "more concealed carry, less crime"
> > > > reworked the data in an artificial fashion, and for ONE special and
> > > > peculiar contrived way of massaging the data, they managed to get
> > > > the result "more concealed carry, no significant effect on crime"
>
> > > To what paper do you refer here?
>
> Tim Lambert
> > I imagine that he is referring to Black and Nagin's critique. One of
> > their points is that excluding Florida causes the results for murder
> > and rape to no longer be significant. Why James considers this to be
> > artificial peculiar and contrived, I don't know.
>
> They also do lots of other things with dates, aggregation, and
> disaggregation. to make the numbers come out right, of which simply
> throwing away one major state is merely the most conspicuous.
In fact, they only have to make one change to Lott's model (exclude
Florida) to change the results for murder and rape. They also try
other things (such as just looking at counties with a large
population) but that does not make much difference.
> The other things they do are not so outrageous, but they are merely
> one special and way of looking at the data, out of a large number of
> possible ways.
There are ten different states that could have been excluded. This
means that the significance levels have to be adjusted to account for
this choice. Even after making this adjustment, the changes for
murder and rape were not significant.
> One important manipulation of their data is to conceal the fact that
> concealed carry laws tended to be passed in response to a rise in
> violent crime. They aggregated the data so that rises in violent
> crime preceding concealed carry laws were counted as violent crime
> following concealed carry laws, an aggregation that comes completely
> out of the blue with no more justification than excluding Florida.
What has come completely out of the blue is your claim. We are
referring to:
Black, Dan. and Daniel Nagin 1998. Do right-to-carry
laws deter violent crime? Journal of Legal Studies 27: 209-219.
Right? Where do they do this aggregation you claim? Please be
specific.
> The more you process the data, more stuff that you add, subtract,
> factor, multiply, and divide, the more opportunities you have to make
> arbitrary choices to make the data come out right. They cooked the
> data a lot, making lots and lots of choices. Because the data was so
> thoroughly massaged, they had thousands of possible equally legitimate
> ways of doing it. Presumably they tried thousands of ways, until they
> found the one way that gave ONE number close to those they wanted.
Ten states they could have excluded. Ten is a much smaller number
than thousands. Other researchers have looked at the data with
different models and found different results to Lott.
(For example Dezhbakhsh and Rubin, Ludwig, Ayres and Donahue, etc)
> Those who lie with statistics give a magick process that yields ONE
> magic number. Those who tell the truth with statistics give tables
> and chart, providing lots of data, rather than a single magick number
> that has been conjured up by a complicated and incompletely explained
> procedure.
Umm, you did actually read Black and Nagin's paper? Are you claiming
that they just give a single magic number???
> > You might be interested in Ted Goertzel's article in the Jan 2002
> > Skeptical Inquirer. It starts:
> >
> > "Do you believe that every time a prisoner is executed in the United
> > States, eight future murders are deterred? Do you believe that a 1%
> > increase in the number of citizens licensed to carry concealed weapons
> > causes a 3.3% decrease in the state's murder rate
>
> The number "3.3%" would certainly be junk science, if anyone had given
> it. That a small increase in concealed carry causes a large decrease
> in violent crime is not junk science, but simple fact.
For a summary of many more problems with this claim, see:
http://www.cse.unsw.edu.au/~lambert/guns/lott/
Tim
James A. Donald
> They use Lott's own data set, with some enhancements, and suggest that
> he doesn't include some important cross-county variation. Including
> them appears to make concealed-carry laws have little to no effect on
> crime
If one "enhances" any set of data sufficiently, one can always get any
result one wants.
Juggling numbers in cross county comparisons can get any result one
wishes for cross county comparisons, but could not make Lott's
scattergrams within countys for countys over time go away.
The correct approach is to construct scattergrams of violent crime
against concealed carry permits for various data sets, and let the
data speak for itself, as Lott, for the most part, did, rather than
endlessly manipulate and re-manipulate the data.
Similar counties with higher crime rates are more likely to have high
rates of concealed carry permits, because citizens are afraid. By
adding and subtracting numbers reflecting such facts, one can get any
result one wishes. If, for example one defines county A as similar to
county B, even though county B has a large criminal underclass, and in
consequence a high rate of concealed carry, and county B is composed
primarily of white quakers, one can readily demonstrate that concealed
carry causes crimes.
It is hard to play smart games with a scattergram. Lott uses
scattergrams, not magick numbers. His opponents pretend he presented
magic numbers, and then cook up their own magic numbers, which show
whatever they want.
You cannot cannot rebut a scattergram with a mere number, and anyone
who pretends to do so is pulling the wool over your eyes.
David Friedman
Tim Lambert <lam...@cse.unsw.edu.au> wrote:
> In fact, they only have to make one change to Lott's model (exclude
> Florida) to change the results for murder and rape.
As best I recall, they also have to drop Lott's column showing the
combined effect for all violent crime, since that one is still
significant after they omit Florida. Am I mistaken?
Obviously that doesn't contradict what you just said, but it does
contradict the implication--that the only selection they engaged in was
eliminating Florida.
I don't currently have a link up to Black and Nagin's paper, and the
link to your page of links on my page appears to be broken. My memory
was that in order to make murder and rape insignificant they also had to
eliminate all counties of less than 100,000.
Note also that "change the results" means "make the result no longer
statistically significant." Even after their changes, three of the four
results still have the right sign.
--
David Friedman
www.daviddfriedman.com/
James A. Donald
Tim Lambert
> > > I imagine that he is referring to Black and Nagin's
> > > critique. One of their points is that excluding Florida
> > > causes the results for murder and rape to no longer be
> > > significant. Why James considers this to be artificial
> > > peculiar and contrived, I don't know.
James A. Donald:
> > They also do lots of other things with dates, aggregation, and
> > disaggregation. to make the numbers come out right, of which
> > simply throwing away one major state is merely the most
> > conspicuous.
Tim Lambert
> In fact, they only have to make one change to Lott's model
> (exclude Florida) to change the results for murder and rape.
Untrue:
Lott gives a scatter diagram, they give a single magic number.
You cannot turn a scatter diagram into a single magic number
without lots of drastic measures. It is like rewriting a novel
into a thirty second movie trailer.
To turn a scatter diagram into a single magic number is a major
rewrite. You have to aggregate together lots of information that
the scatter diagram breaks out and disaggregates.
This gives you lots of opportunities to make lots of arbitrary
decisions, to aggregate data in one fashion rather than another
fashion.
To illustrate the creativity and imagination of the methods they
used to get their magic number, imagine the following hypothetical:
a state where the violent crime rate rose by twenty percent every
year, until, after eight years, the alarmed citizens voted for
concealed carry, whereupon the violent crime rate fell by ten
percent a year every year for eight years, falling in the same
proportion as concealed carry permits were issued.
Under their magic number manipulation, this hypothetical story
about a lot of numbers would be reduced to a single magic number
--the magic number being that violent crime rates almost doubled
AFTER concealed carry was introduced.
> There are ten different states that could have been excluded.
A scatter diagram is a big complicated picture. There are a
thousand different ways one can condense that big picture into
something completely different, and not very closely related, into
a single number. Obviously they must have tried every possible
way, before they got so desperate they had to throw out an entire
state.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
/k7cUWVfDW460R2jcHbRTTkn+mEx1YDt5zgne5xI
4TZSiwXWPqad0ObnYc1IuNakVpp5eDHiW5yL9BV1k
Constantinople
read.news.verio.net:
> Note also that "change the results" means "make the result no longer
> statistically significant."
Just curious: how does removing enough data to make the result not
statistically significant constitute a valid scientific attack on a result?
Tim Lambert
If removing a single state (Florida) eliminates the significance of a
result, then the result is unstable and, although it is evidence, it
isn't very strong evidence.
The key point is that we just have to eliminate one state and not a
whole bunch.
Tim
Tim Lambert
> --
> Tim Lambert
> > > > I imagine that he is referring to Black and Nagin's
> > > > critique. One of their points is that excluding Florida
> > > > causes the results for murder and rape to no longer be
> > > > significant. Why James considers this to be artificial
> > > > peculiar and contrived, I don't know.
>
> James A. Donald:
> > > They also do lots of other things with dates, aggregation, and
> > > disaggregation. to make the numbers come out right, of which
> > > simply throwing away one major state is merely the most
> > > conspicuous.
>
> Tim Lambert
> > In fact, they only have to make one change to Lott's model
> > (exclude Florida) to change the results for murder and rape.
>
> Untrue:
>
> Lott gives a scatter diagram,
No he doesn't. Lott's entire book does not contain a single scatter
diagram. Here's a simple challenge for you: give the page number of
any scater diagram in Lott's book.
> they give a single magic number.
Both Lott and Black and Nagin present their results as "magic numbers"
(fitted parameters in the multivariate regression if you want to get
technical.)
> You cannot turn a scatter diagram into a single magic number
> without lots of drastic measures. It is like rewriting a novel
> into a thirty second movie trailer.
So you think Lott is wrong when he does this?
> To turn a scatter diagram into a single magic number is a major
> rewrite. You have to aggregate together lots of information that
> the scatter diagram breaks out and disaggregates.
>
> This gives you lots of opportunities to make lots of arbitrary
> decisions, to aggregate data in one fashion rather than another
> fashion.
>
> To illustrate the creativity and imagination of the methods they
> used to get their magic number, imagine the following hypothetical:
> a state where the violent crime rate rose by twenty percent every
> year, until, after eight years, the alarmed citizens voted for
> concealed carry, whereupon the violent crime rate fell by ten
> percent a year every year for eight years, falling in the same
> proportion as concealed carry permits were issued.
>
> Under their magic number manipulation, this hypothetical story
> about a lot of numbers would be reduced to a single magic number
> --the magic number being that violent crime rates almost doubled
> AFTER concealed carry was introduced.
Again, this is what Lott did, so if things had happened as in your
hypothetical example, he would have found that violent crime went up
after concealed carry.
Tim
James A. Donald
James A. Donald:
> > You cannot turn a scatter diagram into a single magic number
> > without lots of drastic measures. It is like rewriting a
> > novel into a thirty second movie trailer.
Tim Lambert
> So you think Lott is wrong when he does this?
John Lott presents a wide range of data, not a single magic
number. Here is his reply to the study you cite:
John Lott:
: : Lott: When you drop out counties with fewer than
: : 100,000 people, if anything it actually increases the
: : size of the effect. What [the reviewer is] saying is
: : that if you not only drop out counties with fewer than
: : 100,000 people--which is 86 percent of the counties in
: : the sample, so it's not just a few small counties that
: : we're talking about--but also drop out Florida, then
: : the changes in two of the violent crime categories,
: : when you're just looking at the simple
: : before-and-after averages, aren't statistically
: : significant. But the results still imply a drop, and
: : for robberies and aggravated assaults you still get a
: : drop that's statistically significant.
: :
: : I think it's somewhat misleading to look only at the
: : simple before-and-after averages. Take the case where
: : violent crime rates are rising right up to the point
: : when the law goes into effect and falling afterward,
: : and let's say it was a perfectly symmetrical inverted
: : V. If I were to take the average crime rate before the
: : law goes into effect and the average afterward, where
: : the point of the V is when the law changed, they're
: : going to be the same. Does that mean the law had no
: : impact? When you drop Florida from the sample, [the
: : results] look more like this inverted V than they do
: : when Florida is in there. So I would argue that it
: : strengthens the results, if what you care about is the
: : change in direction.
: :
: : In any case, the bottom line to me is this: I wanted
: : all the data that were available....I didn't pick and
: : choose, and when somebody drops out 86 percent of the
: : counties along with Florida, you know they must have
: : tried all sorts of combinations. This wasn't the first
: : obvious combination that sprang to mind. And it's the
: : only combination they report....If, after doing all
: : these gymnastics, and recording only one type of
: : specification, dealing with before-and-after averages
: : that are biased against finding a benefit, they still
In short, his reply is that he gives lots of numbers, not just one
magic number, and these guys extremely drastic manipulations only
make one of these many numbers go away. They do not make the
numbers for in county comparisons go away, they do not make
numbers for robberies and assualts go away, and the do not make
the numbers for changes over time go away.
To make each of Lott's many, many, numbers go away, one requires a
different manipulation for each number.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
IExLDCLj+PIiXEY0UdUAeqTsfhmlhWJDwB9wIAUJ
48kSckIheUjiMsNuGev2djvz5k8oWjOyRM8tsEzWb
James A. Donald
On 08 Mar 2002 00:15:05 +1100, Tim Lambert
<lam...@cse.unsw.EDU.AU> wrote:
> If removing a single state (Florida) eliminates the significance
> of a result, then the result is unstable
But they did not just remove a single state. They removed 86% of
the countys, and a state, and several categories of crime, and
aggregated time data into before and after.
In short, they removed the great majority of the data, partly by
leaving it out, and partly by aggregation.
If they had to do all that, the result is not unstable.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
s0arvlwqCtoYob+h1MZFQBTdQj3pxSbGzUTARzLw
4EQE52biRf7IIUuaisHYgpZCxVYTrCkr+5hI0TaZk
Mike Hammock
> Tim Lambert
> > They use Lott's own data set, with some enhancements, and suggest that
> > he doesn't include some important cross-county variation. Including
> > them appears to make concealed-carry laws have little to no effect on
> > crime
>
> If one "enhances" any set of data sufficiently, one can always get any
> result one wants.
But Rubin and Dezhbakhsh are not randomly adding variables or
attempting to cook the results. They identify problems with Lott and
Mustard's assumptions, then estimate a model intended to address these
problems. From the paper:
"We believe Lott and Mustard's findings are suspect, mainly because of
the way they parameterize and measure the effect of permissive handgun
laws on crime. They model the effect as a shift in the intercept of
the linear crime equation they estimate at the county level. This
approach is predicated on two assumptions: (i) all behavioral
(response) parameters of this equation (slope coefficients) are fixed
(unaffected by the law), and (ii) the effect of the law on crime is
identical across counties. Obviously, if the law affects the behavior
of the criminals or of citizens, then these parameters should change,
and not only the intercept term. Moreover, it seems highly unlikely
that the maginitude of the effects such laws may have on crime rates
in ac ounty would be independent of economic and demographic
characteristics of the county," (Page 468-469).
Then they reestimate the model, allowing the parameters to change.
The paper is available on Jstor, if you have access to it. A
different version of the paper is available here:
http://userwww.service.emory.edu/~cozden/dezhbakhsh_99_03_paper.pdf
I haven't read this version of the paper, though.
> Juggling numbers in cross county comparisons can get any result one
> wishes for cross county comparisons, but could not make Lott's
> scattergrams within countys for countys over time go away.
>
> The correct approach is to construct scattergrams of violent crime
> against concealed carry permits for various data sets, and let the
> data speak for itself, as Lott, for the most part, did, rather than
> endlessly manipulate and re-manipulate the data.
The only correct approach is scatterplots? Isn't that like saying one
should always do simple regression instead of multiple regression?
Scatterplots can't control for variation in other relevant variables.
Furthermore, you're understating the contribution that Lott and
Mustard made. Their panel data estimation was clearly a step in the
right direction, as compared to scatterplots.
I'm looking at the "Crime, Deterrence, and Right-to-Carry Concealed
Handguns" in the Journal of Legal Studies. I don't see any
scatterplots. Perhaps they're in his book instead; I don't have it on
hand at the moment.
> Similar counties with higher crime rates are more likely to have high
> rates of concealed carry permits, because citizens are afraid. By
> adding and subtracting numbers reflecting such facts, one can get any
> result one wishes. If, for example one defines county A as similar to
> county B, even though county B has a large criminal underclass, and in
> consequence a high rate of concealed carry, and county B is composed
> primarily of white quakers, one can readily demonstrate that concealed
> carry causes crimes.
That's why one should use county-level data describing the
characteristics of the inhabitants, in order to control for such
variation. The solution is to include relevant information--not to
exclude it.
> It is hard to play smart games with a scattergram. Lott uses
> scattergrams, not magick numbers. His opponents pretend he presented
> magic numbers, and then cook up their own magic numbers, which show
> whatever they want.
Lott uses Cross-Sectional Time Series, or Panel, Data. At least, he
does in the paper. That's where his numbers come from. It's a
rigorous approach.
I also don't like what you're implying here--As if anyone who
challenges Lott and Mustard's results is cooking the numbers in a
desparate attempt to make them look wrong. "His opponents pretend",
and then "cook up their own magic numbers, which show whatever they
want". Have you considered the possibility that Lott and Mustard may
have made some mistakes, and that subsequent studies might correct
those mistakes?
> You cannot cannot rebut a scattergram with a mere number, and anyone
> who pretends to do so is pulling the wool over your eyes.
You cannot rebut serious econometric work by dismissing it as "mere
numbers".
I suggest you reread the Lott and Mustard paper and then read the
Rubin and Dezhbakhsh paper, if you can get access to it.
David Friedman
mham4...@aol.com (Mike Hammock) wrote (responding to James):
> I also don't like what you're implying here--As if anyone who
> challenges Lott and Mustard's results is cooking the numbers in a
> desparate attempt to make them look wrong. "His opponents pretend",
> and then "cook up their own magic numbers, which show whatever they
> want". Have you considered the possibility that Lott and Mustard may
> have made some mistakes, and that subsequent studies might correct
> those mistakes?
My impression is that there were two generations of critics. The first
negative responses to Lott and Mustard's work, the ones that inspired me
to set up a web page to cover the controversy, were I think to varying
degrees dishonest or incompetent--for details see my page. At least some
later ones--I've just skimmed the Rubin and Dezhbakhsh paper--are honest
attempts to redo the calculations while changing the details in ways the
authors think improve the results.
Two things struck me as odd about the Rubin and Dezhbakhsh paper. The
first is that they report the estimated effect if states that didn't
have shall issue laws had adopted them, but not the other way around. If
we believe, optimistically, that the states where the laws work best are
the first to adopt them, then their result (ambiguous consequences if
the remaining states adopted them, with some crime rates in some states
going up, some crime rates in some states going down) is consistent with
Lott's conclusion that shall issue laws significantly reduce
confrontational crime. They're looking at the states where the laws work
least well, and concluding that in those states they sometime work and
sometimes don't.
The other odd things was that they write as if they think the details of
their results are solid enough to base state level policy on--to decide
that Illinois probably should pass the law and Maryland shouldn't. That
strikes me as optimistic, given how complicated these issues are.
In any case, thanks for the cite--I've added a link to my page.
--
David Friedman
www.daviddfriedman.com/
David Friedman
Constantinople <constan...@yahoo.com> wrote:
That depends on the reason you remove it.
One common reason is that you are testing how robust the results are.
Suppose you report that out of a thousand coin flips, six hundred were
heads. That's very strong evidence that the coin is biased. Then someone
asks you how you did a thousand flips in only ten minutes. You explain
that, to save time, you flipped the coin only ten times, and multiplied
each result by a hundred. It suddenly becomes a lot weaker evidence.
Now suppose I show that something is true for 600 out of a thousand
counties, where by chance it should happen half the time in a county.
Again it sounds impressive. But what if the counties were in ten states,
a hundred counties each, and the nature of what is happening is such
that it either happens or doesn't happen for almost all the counties in
a state. The evidence is much weaker.
Now suppose the argument is "100 counties all had something in common,
which might have effected the result for all of them." You remove those
counties and see if the result is still significant. If it is, you
conclude that that particular problem didn't generate the significant
result. If it isn't, you conclude that it might have.
Basically you are checking subsets of the data to see if the result is
an artifact produced by a special feature of parts of the data set. In
the case of Black and Nagin, they argued that Florida had other things
happening which drove the results. Lott, in his rebuttal, argued that in
fact those things didn't effect the results in the way Black and Nagin
asserted, hence there wasn't any good reason to eliminate Florida.
--
David Friedman
www.daviddfriedman.com/
Constantinople
> Constantinople <constan...@yahoo.com> writes:
>
> > David Friedman <dd...@best.com> wrote in news:ddfr-F6EEF1.17485006032002@ord-
> > read.news.verio.net:
> >
> > > Note also that "change the results" means "make the result no longer
> > > statistically significant."
> >
> > Just curious: how does removing enough data to make the result not
> > statistically significant constitute a valid scientific attack on a result?
>
> If removing a single state (Florida) eliminates the significance of a
> result, then the result is unstable and, although it is evidence, it
> isn't very strong evidence.
That doesn't follow. The mere fact of the data being on the edge of
statistical significance does not mean that the evidence is poor. The
whole point of setting the level of statistical significance to some
value, say .05, is that this establishes an acceptable limit. To
say that it is acceptable is to say that data that goes below that
level, *by however small an amount*, is acceptable. What you argue
here is that if it is only slightly below that level then it is
"unstable" and "not strong evidence", i.e., unacceptable. So your
argument is really a problem with the chosen level of statistical
significance.
Or is it? Is your problem really that you don't know what statistical
significance is?
David Friedman
> > If removing a single state (Florida) eliminates the significance of a
> > result, then the result is unstable and, although it is evidence, it
> > isn't very strong evidence.
> That doesn't follow. The mere fact of the data being on the edge of
> statistical significance does not mean that the evidence is poor. The
> whole point of setting the level of statistical significance to some
> value, say .05, is that this establishes an acceptable limit. To
> say that it is acceptable is to say that data that goes below that
> level, *by however small an amount*, is acceptable. What you argue
> here is that if it is only slightly below that level then it is
> "unstable" and "not strong evidence", i.e., unacceptable. So your
> argument is really a problem with the chosen level of statistical
> significance.
Suppose you got a significance at the .05 level, then checked at the .04
level and it didn't make it. That would be much weaker evidence than if
it checked all the way to .01. So eliminating a little data and lowering
significance from .05 to .06 would be evidence that your result was
weaker than if eliminating the data didn't have that effect.
But the interesting case isn't that. The interesting case is where the
initial result is significant down to .01 (say), you remove a little
data, and it is no longer even significant to .05. That suggests that
something odd is going on--that your significance is an artifact of some
very unusual data.
Consider two scatter plots. One is a random cloud of points, plus one
point way out in the upper righthand corner, far beyond everything else
in both dimensions. The other is a clould of points running roughly
along a diagonal line from 0,0 to 5,5.
Fit both of them to Y=A+BX. In each case you get A=0, B=1. In each case
the relation is significant at the .05 level. Which do you believe?
--
David Friedman
www.daviddfriedman.com/
Constantinople
read.news.verio.net:
> In article <e6ee4695.02030...@posting.google.com>,
> constan...@yahoo.com (Constantinople) wrote:
>
>> > If removing a single state (Florida) eliminates the significance of a
>> > result, then the result is unstable and, although it is evidence, it
>> > isn't very strong evidence.
>
>> That doesn't follow. The mere fact of the data being on the edge of
>> statistical significance does not mean that the evidence is poor. The
>> whole point of setting the level of statistical significance to some
>> value, say .05, is that this establishes an acceptable limit. To
>> say that it is acceptable is to say that data that goes below that
>> level, *by however small an amount*, is acceptable. What you argue
>> here is that if it is only slightly below that level then it is
>> "unstable" and "not strong evidence", i.e., unacceptable. So your
>> argument is really a problem with the chosen level of statistical
>> significance.
>
> Suppose you got a significance at the .05 level, then checked at the .04
> level and it didn't make it. That would be much weaker evidence than if
> it checked all the way to .01.
But you don't have to eliminate a little data to see that it's at the .05
level but not at the .04 level. You can just recheck the same data. So this
seems to be beside the whole issue of eliminating data.
> So eliminating a little data and lowering
> significance from .05 to .06 would be evidence that your result was
> weaker than if eliminating the data didn't have that effect.
"Weaker than" does not mean "weak". My point stands. We can do the same thing
for .01 and .005, for .001 and .0005, etc., always showing that one test is
weaker than another test. The issue is not whether the evidence is weaker
than some possible evidence, but whether it is weak to the point that we
should disregard it. If barely passing the .05 mark is to be considered to be
so weak that it should be disregarded, as the other poster is arguing, then
his argument is that .05 wasn't a strong enough test.
> But the interesting case isn't that. The interesting case is where the
> initial result is significant down to .01 (say), you remove a little
> data, and it is no longer even significant to .05. That suggests that
> something odd is going on--that your significance is an artifact of some
> very unusual data.
Yes I know this, because I discussed the content of my post with two
statisticians before posting it, and they brought this up, after first making
the major point I made above. For example an extreme outlier can really skew
the data. The fact remains that until you show that something unusual is
going on, then you can't argue from the assumption that something odd is
going on.
> Consider two scatter plots. One is a random cloud of points, plus one
> point way out in the upper righthand corner,
That is pretty much what was pointed out to me before I posted my argument. I
didn't think it affected the argument, because the person I was talking to
was not arguing from the observation that a little change in data caused a
big change in significance. All he said was that it moved it from
significance to nonsignificance, which is compatible with an infinitesimally
small move.
I know what you're saying and insofar as you have just brought up some
points, I don't find any problem. But I question whether any of these points
affect my argument. I didn't think they did before I posted the argument, and
I still don't.
Constantinople
OK - this argument about Florida being special was not clear to me
previously, and this changes matters with respect to my other post in this
thread replying to yours. I had misconstrued the argument because it was not
clear to me from the posts I read that there was anything more to it than
that a reduction of data would render the result insignificant, and Mr.
Lambert's reply reinforced that impression. I think my argument addresses Mr.
Lambert's argument, but his argument was not the same as Black and Nagin's.
Constantinople
> Basically you are checking subsets of the data to see if the result is
> an artifact produced by a special feature of parts of the data set. In
> the case of Black and Nagin, they argued that Florida had other things
> happening which drove the results. Lott, in his rebuttal, argued that in
> fact those things didn't effect the results in the way Black and Nagin
> asserted, hence there wasn't any good reason to eliminate Florida.
By the way, thanks for making it clear why data might need to be removed,
something that I did not know the answer to (though now I find it hard to
imagine not understanding this).
David Friedman
Constantinople <constan...@yahoo.com> wrote:
> I had misconstrued the argument because it was not
> clear to me from the posts I read that there was anything more to it than
> that a reduction of data would render the result insignificant, and Mr.
> Lambert's reply reinforced that impression. I think my argument addresses Mr.
> Lambert's argument, but his argument was not the same as Black and Nagin's.
A further element is the question of small counties. There were some
statistical problems which Lott and Mustard noted having to do with
small counties (I no longer remember the details, but it may have been
cases where some variables were zero for some years). One possible
solution was to eliminate the small counties from the sample. I'm pretty
sure that Black and Nagin did that for one of their reruns of the data.
My memory is that it took that plus removing Florida to eliminate the
significance of the result for two categories of crime, but Tim thinks
it only took removing Florida, and I haven't checked to see if he is
right.
--
David Friedman
www.daviddfriedman.com/
James A. Donald
James A. Donald:
> > If one "enhances" any set of data sufficiently, one can always get any
> > result one wants.
Mike Hammock
> But Rubin and Dezhbakhsh are not randomly adding variables or
> attempting to cook the results. They identify problems with Lott and
> Mustard's assumptions, then estimate a model intended to address these
> problems. From the paper:
>
> "We believe Lott and Mustard's findings are suspect, mainly because of
> the way they parameterize and measure the effect of permissive handgun
> laws on crime. They model the effect as a shift in the intercept of
> the linear crime equation they estimate at the county level. This
> approach is predicated on two assumptions: (i) all behavioral
> (response) parameters of this equation (slope coefficients) are fixed
> (unaffected by the law), and (ii) the effect of the law on crime is
> identical across counties. Obviously, if the law affects the behavior
> of the criminals or of citizens, then these parameters should change,
> and not only the intercept term. Moreover, it seems highly unlikely
> that the maginitude of the effects such laws may have on crime rates
> in ac ounty would be independent of economic and demographic
> characteristics of the county," (Page 468-469).
Firstly: This argument is an argument for overfitting the data -- to
introduce more variables than the data can possibly determine, a
measure that guarantees a nonsense outcome. that guarantees that many
of the variables will drop below signficance. This approach could
equally be used to prove that food intake has no statistically
significant relationship to obesity.
It is a little like the uncertainty principle. The more you assume
you can know about one variable, the more you allow it to be adjusted
to fit, the less you can know about the other variable. Thus by
assuming that criminals in different counties have very different
responses to to the threat of violent death, one can then interpret
the statistics as providing accurate information on the variation of
criminals concern for violent death, and thus as not providing
accurate information on extent that concealed carry increases the
perceived risk of criminal conduct resulting in violent death.
Secondly, the effect of the law will result in an increase in the
danger of committing confrontational crimes proportional to the number
of concealed carry holders in the county, which will indeed be largely
independent of the economic and demographic characteristics of the
county. Black criminals, white criminals, rich criminals and poor
criminals will be near enough equally deterred by a similar prospect
of violent death. They are arbitrarily and artificially assuming
large variablility in something that does not seem likely to vary
measurably.
To get results significantly different from those of Lott, it is
necessary to introduce numbers that implicitly suppose high
variablility in the average response of criminals in a county to the
threat of violent death, and that the statistician can accurately
predict this large variation in response from the county
characteristics. Neither assumption seems likely.
This is related to a general problem in fitting theory to experiment.
One can always save the theory by multiplying hypotheses, by
introducing more hidden variables.
Large differences between counties in criminal's concern for violent
death is just a hidden variable, introduced ad hoc to to obscure other
relationships visible in the data.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
jaP7ZDKYlHhPseMHax8ak8o8XY/sp+RqredSIk1a
4zF2H7sUCScl7TheCdwgjXf2Q96Ap0sAbmaRnZ022
Constantinople
read.news.verio.net:
> Consider two scatter plots. One is a random cloud of points, plus one
> point way out in the upper righthand corner, far beyond everything else
> in both dimensions. The other is a clould of points running roughly
> along a diagonal line from 0,0 to 5,5.
>
> Fit both of them to Y=A+BX. In each case you get A=0, B=1. In each case
> the relation is significant at the .05 level. Which do you believe?
By the way, I think this is a good example of one of the points James was
making, which is that it is harder to go wrong with a scatterplot, than it is
to go wrong with a number derived from the scatterplot.
The human visual system, evidently, performs a more informative and reliable
analysis of some scatterplots than do statistical calculations. This may be
because the probabilistic model of the null hypothesis (i.e., the model of
how the scatterplot should look if the null hypothesis was true) is overly
simple, whereas the human visual system has a richer and more realistic set
of expectations assuming that the null hypothesis is true. Specifically, the
human visual system expects outliers, which turns out to be a more reasonable
expectation than expecting clean, bell-curve-like data. In fact, I even
imagine that the *absence* of significant outliers might sometimes set off
alarms, suggesting that the data has been "preprocessed" or possibly even
faked.
Tim Lambert
> In article <Xns91CAD6C434F...@140.99.99.130>,
> Constantinople <constan...@yahoo.com> wrote:
>
> > I had misconstrued the argument because it was not
> > clear to me from the posts I read that there was anything more to it than
> > that a reduction of data would render the result insignificant, and Mr.
> > Lambert's reply reinforced that impression. I think my argument addresses Mr.
> > Lambert's argument, but his argument was not the same as Black and Nagin's.
Huh? Why do you think that?
> A further element is the question of small counties. There were some
> statistical problems which Lott and Mustard noted having to do with
> small counties (I no longer remember the details, but it may have been
> cases where some variables were zero for some years).
Their model includes the arrest rate, which is undefined if the crime
rate is zero. Consequently L&M have to elimate all observations with
a zero crime rate. This potentially biases the results.
> One possible
> solution was to eliminate the small counties from the sample. I'm pretty
> sure that Black and Nagin did that for one of their reruns of the
> data.
As did L&M.
> My memory is that it took that plus removing Florida to eliminate the
> significance of the result for two categories of crime, but Tim thinks
> it only took removing Florida, and I haven't checked to see if he is
> right.
Black and Nagin wrote: "Nor is this result a function of our use of
the large-county sample. Without Florida in the sample, the
estimation of Lott and Mustard's model, which is given by equation
(1), for all counties provides no evidence of an impact of RTC laws on
homicide and rape."
Tim
James A. Donald
On 7 Mar 2002 09:05:05 -0800, mham4...@aol.com (Mike Hammock)
wrote:
> But Rubin and Dezhbakhsh are not randomly adding variables or
> attempting to cook the results. They identify problems with
> Lott and Mustard's assumptions, then estimate a model intended
> to address these problems. From the paper:
>
> "We believe Lott and Mustard's findings are suspect, mainly
> because of the way they parameterize and measure the effect of
> permissive handgun laws on crime. They model the effect as a
> shift in the intercept of the linear crime equation they
> estimate at the county level. This approach is predicated on
> two assumptions: (i) all behavioral (response) parameters of
> this equation (slope coefficients) are fixed (unaffected by the
> law), and (ii) the effect of the law on crime is identical
> across counties. Obviously, if the law affects the behavior of
> the criminals or of citizens, then these parameters should
> change, and not only the intercept term. Moreover, it seems
> highly unlikely that the maginitude of the effects such laws may
> have on crime rates in ac ounty would be independent of economic
> and demographic characteristics of the county," (Page 468-469).
What Rubin and Dezhbakhsh are up to is simply the logical inverse
of the trick that Black and Nagin did. Black and Nagin
arbitrarily threw out all the most inconvenient data, about 90% of
the data (most of the counties, most of the population, and half
the crimes, aggregated the rest, and announced that the result
was, surprise, surprise, no longer statistically significant.
Black and Nagin used deletion
Rubin and Dezhbakhsh used dilution.
Their technique is slightly trickier, and I will tell a little
fairy tale to illustrate the fallacy:
Fairy tale about lying with statistics:
Scientist Ann says "I believe that if you disarm the public it has
no effect on confrontational crime, and that if you throw a pig
out of a ten story window, it will not harm the pig."
Scientist Bob calls up lab supply for a pig. "OK" says Bob. "Here
is a pig, and this conference room is ten stories up." He throws
the pig out. It goes splat.
"Observe", says scientist Bob. "The pig died."
To which Scientist Ann replies "Not statistically significant.
How do you know it would not have died anyway?.
"OK" says Scientist Bob. "I will get in seventy five more pigs."
Bob throws every second pig out the window. They all go splat.
"Observe", says Bob. "The thirty eight pigs that I threw out the
window are all dead. The thirty eight pigs I left are alive and
well."
To which Scientist Ann replies "I believe Bob's findings are
suspect, mainly because of> the way he parameterizes and measures
the effect of falling out of a ten story window. Bob ignores the
possibility that the fate of the pig might be effected by other
factors, such as its age, ancestry, weight, and umm, ah, umm,
errh, many, many, many other factors. I will now reanalyze the
data taking these factors into account."
Scientist Ann: "observe, when we introduce all these factors, the
data is no longer statistically significant.""
Bob is startled. " How many factors did you use?
Ann: "My modeling employs 76 factors, and to my vast surprise,
not one turned out to be statistically significant".
The point of this story is that the more factors one introduces,
the more data one needs to determine all the factors with
statistical significance. If what Black and Nagin did was throw
out all the unwanted data, until what remained became
statistically insignificant, Rubin burdened down the data with the
task of discovering all sorts of hypothetical variations in the
criminal population, until the data became statistically
insignificant.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
YKPMz5zV2DOLKFiscjiDz9EtqINZJlTRF00nv5f6
40QdY+GDHFNCoaq4MjqFjRRDl9E6RwtTYiE0QnqQg
James A. Donald
On 09 Mar 2002 01:21:45 +1100, Tim Lambert <lam...@cse.unsw.EDU.AU>
wrote:
> Black and Nagin wrote: "Nor is this result a function of our use of
> the large-county sample. Without Florida in the sample, the
> estimation of Lott and Mustard's model, which is given by equation
> (1), for all counties provides no evidence of an impact of RTC laws on
> homicide and rape."
Throwing out all small counties "the large county sample" means
throwing away more than two thirds of the data. Removing all Florida
counties from the few remaining cases county sample means throwing
out most of the population.
Then they also threw out most of the crimes.
Then they aggregated what cases were left, aggregating like with
unlike.
Then, surprise surprise, the data was no longer statistically
significant.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
dACNUnDa8lorjIRz/+tp+MuBl+OMwka/sZuIQzLM
4aYmdWucesPHN2BwIj/cYFNjo67XpnJezVmpE/x7F
Mike Hammock
> Two things struck me as odd about the Rubin and Dezhbakhsh paper. The
> first is that they report the estimated effect if states that didn't
> have shall issue laws had adopted them, but not the other way around. If
> we believe, optimistically, that the states where the laws work best are
> the first to adopt them, then their result (ambiguous consequences if
> the remaining states adopted them, with some crime rates in some states
> going up, some crime rates in some states going down) is consistent with
> Lott's conclusion that shall issue laws significantly reduce
> confrontational crime. They're looking at the states where the laws work
> least well, and concluding that in those states they sometime work and
> sometimes don't.
Good point; it seems to me that would have been a fairly obvious thing
to do. I'll ask Prof. Rubin about it; perhaps they did try it but
didn't report the results (and if so, I'll try to find out why).
> The other odd things was that they write as if they think the details of
> their results are solid enough to base state level policy on--to decide
> that Illinois probably should pass the law and Maryland shouldn't. That
> strikes me as optimistic, given how complicated these issues are.
It's funny that you say that. I've asked Prof. Rubin about what he
thinks policy makers ought to do. His response was something like
"Given that the results don't show a strong effect one way or the
other, I lean toward passing concealed-carry laws,". They don't seem
to do much good, but they don't do much harm, either, but since he
prefers more freedom to less, he'd prefer to see concealed carry laws
passed.
> In any case, thanks for the cite--I've added a link to my page.
You're welcome.
Constantinople
> David Friedman <dd...@best.com> writes:
>
> > In article <Xns91CAD6C434F...@140.99.99.130>,
> > Constantinople <constan...@yahoo.com> wrote:
> >
> > > I had misconstrued the argument because it was not
> > > clear to me from the posts I read that there was anything more to it than
> > > that a reduction of data would render the result insignificant, and Mr.
> > > Lambert's reply reinforced that impression. I think my argument addresses Mr.
> > > Lambert's argument, but his argument was not the same as Black and Nagin's.
>
> Huh? Why do you think that?
Here is your argument:
[begin]
If removing a single state (Florida) eliminates the significance of a
result, then the result is unstable and, although it is evidence, it
isn't very strong evidence.
The key point is that we just have to eliminate one state and not a
whole bunch.
[end]
Just for starters, you make no mention here about why Florida should
be removed. Black and Nagin do. That by itself makes their argument
different from yours. There are other differences, but this is
sufficient to answer your question.
Mike Hammock
> Ann: "My modeling employs 76 factors, and to my vast surprise,
> not one turned out to be statistically significant".
Here's the really funny thing: She wouldn't get that result. There
would be a perfect correlation between the dependent variable--pig
death--and one of the explanatory variables (pig throwing). Any other
variables would show up as irrelevant. If Ann excluded the "pig
throwing" variable, then it's possible that she might get strange
results. But Rubin and Dezhbakhsh are not excluding Lott's variables.
> The point of this story is that the more factors one introduces,
> the more data one needs to determine all the factors with
> statistical significance. If what Black and Nagin did was throw
> out all the unwanted data, until what remained became
> statistically insignificant, Rubin burdened down the data with the
> task of discovering all sorts of hypothetical variations in the
> criminal population, until the data became statistically
> insignificant.
What you're talking about are degrees of freedom. You're right that
we can't estimate too many parameters, or we run out of degrees of
freedom. However, Lott (and therefore Rubin and Dezhbakhsh) wisely
uses data from 3054 counties over a fifteen year period (subject to
the caveat that some of the series only go from 1982 to 1992)--that's
over 45,000 observations.
What's also interesting is that they used a Wald test to investigate
whether Lott's specification--with fixed slope coefficients--is
subject to misspecification bias, and were able to reject the
hypothesis that the coefficients are constant across legal rules for
all crime types, with p-values close to zero. Lott and Mustard appear
to have a misspecification problem.
Finally, the results didn't become statistically insignificant--it
became mixed. Look at table 1 (in the published version). Some signs
are positive, some are negative, some are mixed across counties, and
some are insignificant.
James A. Donald
James A. Donald
> > Ann: "My modeling employs 76 factors, and to my vast
> > surprise, not one turned out to be statistically significant".
On 8 Mar 2002 14:14:41 -0800, mham4...@aol.com (Mike Hammock)
wrote:
> Here's the really funny thing: She wouldn't get that result.
> There would be a perfect correlation between the dependent
> variable--pig death--and one of the explanatory variables (pig
> throwing). Any other variables would show up as irrelevant.
Only if she uses the "unsophisticated" method of data analysis
used by Lott.
If we use the "sophisticated" method employed by Rubin and
Dezhbakhsh any unwanted data goes away.
You are assuming that the probability of the pig dying is
approximated by P = a*x + b*y + c *z .... .... Where x is how fat
the pig is, y, is whether the pig was thrown through a ten story
window, z is the age of the pig, and so on and so forth.
This is equivalent to Lott's "unsophisticated" model for county
levels of confrontational crime.
With this system, one can introduce any number of irrelevant
variables and it will make no difference.
The "sophisticated" model employed by Rubin and Dezhbakhsh makes
the rather startling assumptions that the likelihood that black
people are deterred from confrontational crime by concealed gun
carry does not tell us anything about the likelihood that white
people are deterred from confrontational crime by concealed gun
carry, and similarly for rich and poor, urban and rural, and so
and so forth.
With Rubin's system, one can force any desired result by
introducing appropriate variables. Any unwanted data can be
explained away by some special case variable, or rendered
irrelevant by another variable. One's outcome is completely model
dependent.
Similarly, Ann is assuming that death of one pig after it is
thrown out of a ten story window does not tell us anything about
the likelihood that a slightly different type of pig will die
after being thrown out of a ten story window.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
Q99v6zerOzvE+y/WjdpEY4YvwPXgA1AryCUWkok8
4/sX6ITDxlm/UTMyT7BDi+7TIu0FX+7lz+wbpJXle
Tim Lambert
> --
> James A. Donald:
> > > You cannot turn a scatter diagram into a single magic number
> > > without lots of drastic measures. It is like rewriting a
> > > novel into a thirty second movie trailer.
>
> Tim Lambert
> > So you think Lott is wrong when he does this?
>
> John Lott presents a wide range of data, not a single magic
> number.
No, he presents lots of "magic numbers", just as Black and Nagin do.
You claimed that Lott presents scatter diagrams. Are you now
retracting your claim? Hove you even read his book?
Tim
Tim Lambert
> Two things struck me as odd about the Rubin and Dezhbakhsh paper. The
> first is that they report the estimated effect if states that didn't
> have shall issue laws had adopted them, but not the other way around. If
> we believe, optimistically, that the states where the laws work best are
> the first to adopt them,
But is there any reason at all to believe this?
> then their result (ambiguous consequences if
> the remaining states adopted them, with some crime rates in some states
> going up, some crime rates in some states going down) is consistent with
> Lott's conclusion that shall issue laws significantly reduce
> confrontational crime. They're looking at the states where the laws work
> least well, and concluding that in those states they sometime work and
> sometimes don't.
>
> The other odd things was that they write as if they think the details of
> their results are solid enough to base state level policy on--to decide
> that Illinois probably should pass the law and Maryland shouldn't. That
> strikes me as optimistic, given how complicated these issues are.
That seems to be common attitude among economists who have published
on this matter. Lott certainly hasn't been shy about making public
policy recommendations.
Tim
Tim Lambert
> Secondly, the effect of the law will result in an increase in the
> danger of committing confrontational crimes proportional to the number
> of concealed carry holders in the county, which will indeed be largely
> independent of the economic and demographic characteristics of the
> county. Black criminals, white criminals, rich criminals and poor
> criminals will be near enough equally deterred by a similar prospect
> of violent death.
Nonsense. By the same reasoning crime rates should be the same
everywhere since criminals should be equally deterred by the sanctions
of the criminal justice system. Furthermore, you are assuming that
all counties will have the same percentage of concealed carry
holders. This is false. There is a large variation in this
percentage.
> They are arbitrarily and artificially assuming
> large variablility in something that does not seem likely to vary
> measurably.
No they are not. They assume that there is some variability (and
present evidence that there is such variability). They let the data
determine the amount of that variability.
Here is Dezhbakhsh's summary of their findings:
"My work with Emory economist Paul Rubin-described in "Lives Saved or
Lives Lost: The Effect of Concealed Handgun Laws on Crime," published
in American Economic Review in 1998-criticizes Lott on simple but
fundamental grounds. We show that Lott's work is erroneous not only in
theory but also in its arithmetic.
"Lott's finding relies on the assumption that the effect of permissive
handgun laws on crime is identical across all counties and
independent of any county characteristics. This assumption is flatly
contradicted by conventional wisdom. Such laws would not have the
same effect in crime-ridden urban areas as they would in remote rural
counties or affluent suburbs. Some of Lott's results also assume that
the number of arrests made by police does not depend on the number of
crimes committed! So rural counties with very few crimes may
presumably have more police arrests than urban counties with very
large crime rates.
"Moreover, Lott's central results are invalid because of errors in
computing expected arrest rates: he obtains mostly negative numbers
for arrests. For example, more than 19,000 of approximately 33,000
county-level auto theft arrests are "negative"; the number of
negative arrest rates for aggravated assault and property crimes are,
respectively, 9,900 and 13,500. What does a negative arrest rate
mean? Obviously, the number of individuals arrested for crimes can
only be zero or positive.
"Once we correct for these errors, the more-guns-less-crime claim
disintegrates. In fact, we show not only that Lott's strong
crime-reducing effect does not materialize, but also that concealed
handguns lead to a higher robbery rate.
"The peer examination process usually exposes flawed research
quickly. The ideologically intoxicating finding of Lott-advocating a
governmental hands-off policy-seems to have impaired this healthy
process. Endorsing Lott's book, the arch-conservative Milton Friedman
of the Hoover Institute exults, "Lott has done us a service by his
thorough, thoughtful, scholarly approach to a highly controversial
issue." Friedman, obviously, is prepared to overlook all the
documented inadequacies in Lott's work to embrace his claim.
"The academic survival of a flawed study may not be of much
consequence. But, unfortunately, the ill-effects of a bad policy,
influenced by flawed research, may hurt generations. It would be
tragic for lawmakers to base any gun laws on Professor Lott's More
Guns, Less Crime claim."
Tim Lambert
> --
> On 09 Mar 2002 01:21:45 +1100, Tim Lambert <lam...@cse.unsw.EDU.AU>
> wrote:
> > Black and Nagin wrote: "Nor is this result a function of our use of
> > the large-county sample. Without Florida in the sample, the
> > estimation of Lott and Mustard's model, which is given by equation
> > (1), for all counties provides no evidence of an impact of RTC laws on
> > homicide and rape."
>
> Throwing out all small counties "the large county sample" means
> throwing away more than two thirds of the data.
You didn't actually read what you quoted did you? Here it is again
(my emphasis):
"Without Florida in the sample, the estimation of Lott and Mustard's
model, which is given by equation (1), FOR ALL COUNTIES provides no
evidence of an impact of RTC laws on homicide and rape"
That's "FOR ALL COUNTIES", you know, like including all the small
counties.
Tim
Tim Lambert
> Tim Lambert <lam...@cse.unsw.EDU.AU> wrote in message news:<tnzo1kc...@oolong.orchestra.cse.unsw.EDU.AU>...
> > Constantinople <constan...@yahoo.com> writes:
> >
> > > David Friedman <dd...@best.com> wrote in news:ddfr-F6EEF1.17485006032002@ord-
> > > read.news.verio.net:
> > >
> > > > Note also that "change the results" means "make the result no longer
> > > > statistically significant."
> > >
> > > Just curious: how does removing enough data to make the result not
> > > statistically significant constitute a valid scientific attack on a result?
> >
> > If removing a single state (Florida) eliminates the significance of a
> > result, then the result is unstable and, although it is evidence, it
> > isn't very strong evidence.
>
> That doesn't follow. The mere fact of the data being on the edge of
> statistical significance does not mean that the evidence is poor.
1. The data is not on the edge of statistical significance. In the
case of murder, with Florida there is a 9% reduction, significant at
the 1% level, without Florida there is a 1% reduction, not even
significant at the 50% level.
2. If a result is on the edge of statistical significance then it is
only a little better than one that hust misses out on being
statistically significant. You seem to believe that all statistically
significant results are equal to each other.
Tim
Tim Lambert
> --
> On 08 Mar 2002 00:15:05 +1100, Tim Lambert
> <lam...@cse.unsw.EDU.AU> wrote:
> > If removing a single state (Florida) eliminates the significance
> > of a result, then the result is unstable
>
> But they did not just remove a single state. They removed 86% of
> the countys,
Not so. Include all counties except Florida and the results for
murder and rape are not significant.
> and a state, and several categories of crime,
Oh, really? What categories of crime did they remove?
> and aggregated time data into before and after.
That aggregation was done by Lott. They did not change it.
Tim
Tim Lambert
No it doesn't. I'm summarizing and explaining their argument. The
key is what happens to the results when one state is excluded. The
reasons for excluding that state are less important.
Tim
James A. Donald
James A. Donald:
> > John Lott presents a wide range of data, not a single magic
> > number.
Tim Lambert
> No, he presents lots of "magic numbers", just as Black and Nagin do.
No Black and Nagin do not.
Lott presents a lot of numbers, each of which independently of all of
the others, shows that concealed carry has a large effect, comparable
to that of arrest rate, and usually greater than that of arrest rate.
You can take any one aspect of the data, for example changes in the
rape rate as the number of concealed carry permits increase over time
within an area, and get a significant result, and another independent
aspect of the data, for example murders compared between different
areas with different numbers of concealed carrry holders, and again
get an independent significant result.
Black and Nagin show that for ONE of the many numbers Lott produces,
if you selectively throw out the great majority of the data on which
it is based, that ONE NUMBER ceases to be statistically significant.
Presumably all the others remained significant no matter what special
subset of the data Black and Nagin chose.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
3ipCJ3SYgUcGUcU8IiV8YvgizO0qvE7wdezJqYH+
4YIfNxu0SiSBUCQM/9mF9JoIJvhX1FO1J2ezBZejf
James A. Donald
Tim Lambert
> > > If removing a single state (Florida) eliminates the
> > > significance of a result, then the result is unstable
James A. Donald:
> > But they did not just remove a single state. They removed 86%
> > of the counties,
Tim Lambert
> Not so.
Is so. They restricted the analysis to counties of more than one
hundred thousand people. This gives you a rather small number of
counties, in some states only a single county, and thus a much
less significant result, with the result that then throwing out
one selected state, the state that gives us a very large share of
the few remaining counties, can bias the results sufficiently, and
that some particular states give fluky data because of particular
events in particular towns of that state.
> Include all counties except Florida and the results for
> murder and rape are not significant.
Is significant. That is why they had to throw out all counties of
less than one hundred thousand people. This threw out 86% of the
data, and reduced some states to a single county, thus
automatically rendering the results for many states below the
level of significance
The justification for throwing out these counties was that there
statistical problems in a few very small counties, but they could
have used a much smaller cut off. The problems only occurred in a
very few tiny counties, yet they threw out the vast majority of
counties.
> > and several categories of crime,
>
> Oh, really? What categories of crime did they remove?
To be precise, even if you first throw out 86% of the counties,
then throw out Florida, and then add aggregate ALL categories of
confrontational crime, the results are still statistically
significant, but if you take any one category of confrontational
crime and ignore all the other categories, then the results with
most of the data thrown out, and florida thrown out, and ignoring
all the other categories of confrontational crime, are not
statistically significant, merely because so much data is thrown
out, and because the counties that showed the greatest effect were
preferentially thrown out. but if you add up all the categories of
confrontational crime, the result is still significant, even with
only a few counties retained, and all florida counties thrown out.
James A. Donald:
> > and aggregated time data into before and after.
Tim Lambert
> That aggregation was done by Lott. They did not change it.
Lott does that aggregation, merely in order to explicitly point
out it is not an accurate or realistic way of treating the data,
and he then proceed to break out the data, treating the data in a
way that is clearly more accurate and realistic, His point in
aggregating the data in that unrealistic fashion, is that even
such aggregates still produce a significant result In original
book he deprecated that number -- it was merely part of the
process of showing that no matter what way you mangle the data,
the effect is still visible. He was just trying to show that he
was not pick and choosing just those ways of grouping the data
that give the desired result, that no matter how one processed the
data, the result was still apparent. It was just a devils
advocate statistic, presented to show he was bending over
backwards to give the null hypothesis more than a fair shake.
In any case, it is one thing to do a drastic aggregation on the
complete data set. It is a very different thing to do a drastic
aggregation after throwing out almost all the data. Either way
you lose a lot of statistical significance, but if you have just
thrown out most of the data, then the additional lost of
significance due to using time aggregation rather than time series is
going to hurt.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
iWiiTm607bl1znG00g2BgcgiCoEp/+7yS2liKtl/
4F3HntOU1C0Jp/CWSyCuwBq90oSDNg74xVVXm80xE
James A. Donald
On 10 Mar 2002 01:34:53 +1100, Tim Lambert <lam...@cse.unsw.edu.au>
wrote:
> 1. The data is not on the edge of statistical significance. In the
> case of murder, with Florida there is a 9% reduction, significant at
> the 1% level, without Florida there is a 1% reduction, not even
> significant at the 50% level.
Untrue:
Without florida, the results for murder are only slightly changed.
Without the florida counties, AND with smallest 86% of the remaining
counties thrown out, the reduction for murder is a statistically
insignificant 1%, but if we add all confrontational crimes together,
the result remains statistically significant, though considerably
smaller.,
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
Bcww25esYzzP+uKXeY0OCxsPm1OBEQxmxnZ/Eci4
4+sH1K6WSGIicLOx+AAdn0dLvErIgbE8qT4/DaEmT
David Friedman
Tim Lambert <lam...@cse.unsw.EDU.AU> wrote:
> David Friedman <dd...@best.com> writes:
>
> > Two things struck me as odd about the Rubin and Dezhbakhsh paper. The
> > first is that they report the estimated effect if states that didn't
> > have shall issue laws had adopted them, but not the other way around. If
> > we believe, optimistically, that the states where the laws work best are
> > the first to adopt them,
>
> But is there any reason at all to believe this?
Any reason at all? Yes. Even political institutions have some tendency
towards efficient outcomes, just not much (for details see the chapter
on the economics of government in my webbed _Price Theory_). Enough
reason to make it odd that they reported only half the results implied
by their analysis.
> > The other odd things was that they write as if they think the details of
> > their results are solid enough to base state level policy on--to decide
> > that Illinois probably should pass the law and Maryland shouldn't. That
> > strikes me as optimistic, given how complicated these issues are.
>
> That seems to be common attitude among economists who have published
> on this matter. Lott certainly hasn't been shy about making public
> policy recommendations.
There's a considerable difference in level of detail.
--
David Friedman
www.daviddfriedman.com/
James A. Donald
On 10 Mar 2002 00:51:12 +1100, Tim Lambert
I do not know if they are telling the truth when they say that it
is not a function of the large county sample, but I do know they
are not telling the truth when they say "the estimation of Lott
and Mustard's model." Lott's model attempts to control for the
variation in concealed carry holders over time. Black and Nagin
do not, with the result that their model, which they deceptively
attribute to Lott, generates all sorts of implausible
correlations. If you do stupid things with Lott's data and use
some of his methods in stupid ways, you can get stupid results.
Big deal.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
5ovunXANhGJRBaUC4OxIBI70MvDZFHXXs1/CqJ5/
4qdJezjVpJtCXzttub4qoKJHaUe0Ne6N9MUxPb/J+
Mike Hammock
way around--that is, if he tried predicting whether crime rates would
have gone up or down in states that did pass concealed carry laws if
they had not passed such laws. His recollection was that they had
indeed done this, although the results are not reported in the AER
version of the paper (nor, it appears, in the other version to which I
provided a link). He said there might be another version of the paper
out there with the results in it. His recollection is that the
results were similar to the reported results--no clear direction of
the effect emerges. That is, it seems that the states that did pass
the laws wouldn't have experienced significantly different crime rates
if they hadn't passed the laws.
If you'd like, I could pester him a bit more and see if I can get him
to either give me the results, or let me try the estimation myself. I
don't know if he'd lend me the data, though. It's a rather important
data set, which they've milked for several papers now (including a
rather interesting one on capital punishment which has gotten some
press).
Mike Hammock
mham4...@aol.com
mha...@emory.edu
mha...@emory.learnlink.edu
James A. Donald
James A. Donald:
> > Secondly, the effect of the law will result in an increase in
> > the danger of committing confrontational crimes proportional
> > to the number of concealed carry holders in the county, which
> > will indeed be largely independent of the economic and
> > demographic characteristics of the county. Black criminals,
> > white criminals, rich criminals and poor criminals will be
> > near enough equally deterred by a similar prospect of violent
> > death.
Tim Lambert:
> Nonsense. By the same reasoning crime rates should be the same
> everywhere since criminals should be equally deterred by the
> sanctions of the criminal justice system.
The sanctions vary everywhere.
Obviously arrest rates vary greatly, just as the number of
conealed carriers vary greatly. For example it is well known,
especially well known among blacks, that a black person committing
a crime against another black person faces a far lower risk of
arrest than a white person committing a crime against another
white person, which doubtless explains some substantial part of
the high crime rate among blacks.
It is well known that courts and police, both black police and
white police, have little regard for black property rights, and
little concern for enforcing them. This directly leads to a high
rate of property crime among blacks, which is one important cause
of the higher rate of violent confrontations between blacks.
Similarly poor people will mug little old ladies, and rich people
will not mug little old ladies, since the risk and financial
reward is the same for both, but the marginal utility of that
financial reward is far greater for poor people than for rich
people.
It seems to me that you are implying that black people have higher
crime rates because they are intrinsically stupid and nasty.
Irrespective of whether this is true, it seems highly improbable
that black criminals could be so stupid that the prospect of
finding their intended victim armed had a significantly different
effect on them than it would have on a white criminal.
> Furthermore, you are assuming that all counties will have the
> same percentage of concealed carry holders.
No I am not. Rubin and Dezhbakhsh are assuming that the *effect*
of concealed carry is intrinsically different in different
counties -- that the apparent fluctuation in the numbers is not
noise, but rather is real. This gives them an unreasonable and
counterfactual model that they employ to match the data, which in
turn leads to nonsense correlations.
> Here is Dezhbakhsh's summary of their findings: "My work with
> Emory economist Paul Rubin-described in "Lives Saved or
> Lives Lost: The Effect of Concealed Handgun Laws on Crime,"
> published in American Economic Review in 1998-criticizes Lott
> on simple but fundamental grounds. We show that Lott's work is
> erroneous not only in theory but also in its arithmetic.
>
> "Lott's finding relies on the assumption that the effect of
> permissive
> handgun laws on crime is identical across all counties and
> independent of any county characteristics. This assumption is
> flatly contradicted by conventional wisdom. Such laws would not
> have the same effect in crime-ridden urban areas as they would
> in remote rural counties or affluent suburbs. Some of Lott's
> results also assume that the number of arrests made by police
> does not depend on the number of crimes committed! So rural
> counties with very few crimes may presumably have more police
> arrests than urban counties with very large crime rates.
Dezhbakhsh lies.
Lott, quite obviously, does not make that assumption. After all
the whole point of a multivariate analysis is precisely to take
account of such variations.
The question at issue is what kind of variations are likely
between counties. Dezhbakhsh arbitrarily assumes complicated
variations that have no basis in human nature. Lott assumes
simple linear variation, which makes perfect sense from our
intuitive knowledge of human anture.
Complicated variations offer more scope in cooking the books.
Unrealistic variations offer scope to force improbable results.
The realistic assumption is the one that Lott makes: That a
criminal will be deterred, and the rate of confrontational crime
will drop, proportionally to number of concealed carry holders,
and this drop will be approximately uniform for all classes of
criminals, and all classes of victim under circumstances affected
by concealed carry laws. There will be far fewer muggers in a
rural county, but they will be deterred in proportion to the
amount of concealed carry in that county, and the mugger in an
urban ghetto will be deterred in similar proportion according to
the proportion of concealed carry holders in that ghetto.
Any othe assumption is artificial, and concocted for the purpose
of cooking the books.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
f/+FG4s3h5/CgFAYzycrxa1z6FLvTL/RWJaiCFBa
4OhNLUS2fIATLLlK3JQ/2JrsQzi4Nf6uUqmgR8eyW
James A. Donald
> the case of murder, with Florida there is a 9% reduction,
> significant at the 1% level, without Florida there is a 1%
> reduction, not even significant at the 50% level.
This is all a pack of lies. The various critics start out with a
bunch of very complicated half truths, and when these half truths
are simplified to simple punchy statements such as the above, they
cease to be complicated half truths, and become simple llies.
Lott massaged his data in complicated ways. His critics quite
correctly pointed out that this massage was a little too
complicated. a bit too clever. He tried to predict what could not
be predicted.
To which Lott responds that even when the data is minimally massaged,
the truth sticks out like dog's balls: See for example his
time chart:
http://www.daviddfriedman.com/Lott_v_Teret/Response_to_Teret.html
Then, to force the right results, the critics proceeded to massage
the data in ways considerably more complicated, and
correspondingly less realistic, than the arguably already overly
complicated and excessively sophisticated methods used by Lott,
attempting to predict more things, less easily measured, or even
defined, than that which Lott attempted, Dezhbakhsh and Rubin
being the most extreme example of this.
Then, when they summarized their complicated manipulations in
simple punchy ways that make a good headline, by simplifiying,
they unavoidably found themselves lying.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
erIB1X7WgKGbOBqzrPpVVMsOLDl3AjScvzHNo5gc
4qSiXVwmRAyJ/9A06619MtV3Ue6HZ6ENIqlN/6U2y
James A. Donald
http://bingweb.binghamton.edu/~fplass/gun.pdf gives review of the
statistical questions at issue, and reanalyzes Lott's data taking
the various criticisms into account.
They get pretty much the same resuls as Lott did.
To summarize their results: The critics were correct to point out
various defects in Lotts models and methods, but the critics
committed all the same offenses and far worse offenses, and when
the data is reanalyzed in a way that takes account of the
criticisms, Lott's result gets stronger rather than weaker.
Conclusion: Concealed carry laws result in a level of concealed
carry that has a very large effect in deterring criminals. This
effect is significant not only if we take all the states together,
but in half the states it is significant taking that state by
itself.
In other words, yes, we can leave out Florida, and it does not
make any difference.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
RxN9OJl8RtUGazUVNQ3FJEhKzNQnFTUb14TD4rzn
4/fqtakN+BNbeqIBzNCLuMNQeykgEezZU60X6iUvD
James A. Donald
On 10 Mar 2002 01:08:07 +1100, Tim Lambert
<lam...@cse.unsw.EDU.AU> wrote:
> Not so. Include all counties except Florida and the results for
> murder and rape are not significant.
http://bingweb.binghamton.edu/~fplass/gun.pdf
does a state by state analysis, and find that for half the states
each state is statistically significant by itself. You do not
need Florida to get the conclusion that shall-issue laws cause a
statistically significant fall in confrontational crime -- neither in
Lott's original analysis, nor in Plassmann's reanalysis.
Much of the data available is contaminated by plain and simple
barefaced lying. Thus the Universty of Maryland issue a study
that Florida was a disaster, and Nagel issues a study that Florida
was a fluke in having such a good outcome, and so should not be
counted. Obviously one or both of these very respectable seeming
studies have to be lying. They cannot both be telling the truth.
The Dezhbaksh paper
http://userwww.service.emory.edu/~cozden/dezhbakhsh_99_03_paper.pdf
has an abstract which does not appear to me to be consistent
with its body. Perhaps someone more sophisticated than me might
interpret that summary in a way consistent with the body of the
paper, but to me it apears untrue, deceptive, a lie. The
conclusion that the effect is more mixed is not a conclusion, but
merely a premise. They modelled the data on the basis of that
premise, got noise results, and correctly but deceptively
concluded that noise results are consistent with the premise.
However, results substantially larger than noise would provide
rather more support for the premise. The fact that their model
gives noise results in no way casts doubt on models that give
better than noise results, unless one starts from the unsupported
theoretical presupposition that their model is superior apriori to
other models.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
yiPjsDyC4WOg6LrAt3Ov3laMqlVglfEPgDn6l3G9
4g2iNHmXsRdHexT6mLsJRbGYNmSvrQF+sTmx1CwX+
James A. Donald
Mike Hammock
> I've asked professor Dezhbakhsh if he tried the prediction the
> other way around--that is, if he tried predicting whether crime
> rates would have gone up or down in states that did pass
> concealed carry laws if they had not passed such laws. His
> recollection was that they had indeed done this, although the
> results are not reported in the AER version of the paper (nor,
> it appears, in the other version to which I provided a link)
I have just read the paper:
http://userwww.service.emory.edu/~cozden/dezhbakhsh_99_03_paper.pdf
They assume, a priori, large variability in the effect of
concealed carry in deterring crime.
They attempt to measure that variability, but get no very strong
results. Their results do not strongly contradict the null
hypothesis that there is no large variability in the effect of
concealed carry in deterring confrontational crime.
They misleadingly represent this as evidence for no large effect
of concealed carry in deterring confrontational crime, when in
fact their methods, assumptions, and measures are tangential to
that issue. In particular the conclusion in their abstract "such
effects appear to be much smaller and more mixed than Lott and
Mustard suggest" is not directly supported by the data they
present, or their method of analysis.
Lott and Mustard present many estimates for the effect of "shall
issue" laws in counties in states that introduced such rules for
concealed carry permits. Dezhbakhsh and Rubin do not, and do not
attempt to do so, thus the conclusion in their abstract has no
direct immediate connection to any specific numbers in their
paper. They attempt to predict effects for various places, yet we
can find no numbers in their paper that can be directly compared
with equivalent numbers appearing in Lott and Mustard's paper.
They attempt to predict, what Lott and Mustard attempt to measure.
The obvious thing then is to list a bunch of counties and states
with Lott and Mustard's measures in one column, and Dezhbakhsh and
Rubin predictions for what should have happened in the other
column, and see if they are consistent with common sense and the
typical crime news of that place, thereby not only providing
evidence for the claims made in the abstract, but also providing a
basic sanity check on the plausibility of their model. This is
what one would expect from the abstract, and yet such information
just is not present in the body of the paper that the abstract
purports to summarize.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
OKYRUJn8Id4F1GOqjEPaL6NgOl6MIIvY+k8XWwwp
4E1WsNObnQeESkjxlfDP5E+CxXnS6WEks4tLwcVWV
James A. Donald
Tim Lambert
> I'm summarizing and explaining their argument.
The trouble is that we have here a bunch of very complicated
arguments, many of them half truths, containing very complicated
criticism of overly complicated models, models whose main defect
is that they are overly complicated, and therefore potentially
subject to creative manipulation.
Any summary and explanation of a complicated half truth is apt to
unintentionally turn a complicated half truth into a simple lie.
I do not think you are intentionally lying, but you have said many
things that I do not believe are true, and I doubt you have any
good reason to believe that they are true.
All of the critics of Lott objected to his use of a model that
assumed continuous variability, when some of his variables, such
as arrest rate, could not go below zero, and were frequently
bounded by zero. Nagin used this defect as justification for
reworking his numbers in drastic and unreasonable ways. Yet
almost all of the critics who made a big deal out of this, who
complained a lordly manner how absurd this was, and how it
contaminated his results, proceeded to use models with the same
defect. This shows that at least some of the criticism of Lott is
dishonest. Nagin found the boundedness of such variables good
reason for throwing away any data he found inconvenient, but found
the problem no reason to throw away data he found convenient, yet
the problem has solutions that are considerably more respectable
and widely used than throwing away data.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
hStDZOMhBCa38mLNPcr7YMHS6NKWgEvcFxLr/lLR
4M5il+thwzJiYP8DnqgGru0g+Xmr06Dcd/uymN9ba
David Friedman
mham4...@aol.com (Mike Hammock) wrote:
> If you'd like, I could pester him a bit more and see if I can get him
> to either give me the results, or let me try the estimation myself. I
> don't know if he'd lend me the data, though. It's a rather important
> data set, which they've milked for several papers now (including a
> rather interesting one on capital punishment which has gotten some
> press).
Considering that they are using Lott's data, which he made available for
free, it would seem somewhat discourteous if they are now locking up
their data set. The relevant norms may not be as strong in the academic
version of open source as in the software version, but they're basically
similar.
--
David Friedman
www.daviddfriedman.com/
Constantinople
David Friedman's version of their argument makes their argument
defensible. Yours doesn't. You eliminated too much from your "summary".
Mike Hammock
Hmm. I wasn't aware that Lott had made the data available for free; do
you mean that it's available for anonymous download from a web page,
or that he'll e-mail it to anyone that requests it? In any case, if
it really is available to anyone, you might consider investigating the
results yourself if you're curious about what the Rubin and Dezhbakhsh
model predicts for counties that did pass concealed-carry laws.
Tim Lambert
> --
> James A. Donald:
> > > John Lott presents a wide range of data, not a single magic
> > > number.
>
> Tim Lambert
> > No, he presents lots of "magic numbers", just as Black and Nagin do.
>
> No Black and Nagin do not.
You haven't read their paper have you?
> Lott presents a lot of numbers,
You claimed that Lott presents scatter diagrams. Are you now
retracting your claim?
Tim
Tim Lambert
> In article <tnk7slr...@oolong.orchestra.cse.unsw.EDU.AU>,
> Tim Lambert <lam...@cse.unsw.EDU.AU> wrote:
>
> > David Friedman <dd...@best.com> writes:
> >
> > > Two things struck me as odd about the Rubin and Dezhbakhsh paper. The
> > > first is that they report the estimated effect if states that didn't
> > > have shall issue laws had adopted them, but not the other way around. If
> > > we believe, optimistically, that the states where the laws work best are
> > > the first to adopt them,
> >
> > But is there any reason at all to believe this?
>
> Any reason at all? Yes. Even political institutions have some tendency
> towards efficient outcomes, just not much (for details see the chapter
> on the economics of government in my webbed _Price Theory_). Enough
> reason to make it odd that they reported only half the results implied
> by their analysis.
But there is no evidence that the laws work better in those states.
If there was such evidence, then it possible that influenced the
lawmakers, but absent such evidence, there is no reason to believe
that the law worked better in those states.
> > > The other odd things was that they write as if they think the details of
> > > their results are solid enough to base state level policy on--to decide
> > > that Illinois probably should pass the law and Maryland shouldn't. That
> > > strikes me as optimistic, given how complicated these issues are.
> >
> > That seems to be common attitude among economists who have published
> > on this matter. Lott certainly hasn't been shy about making public
> > policy recommendations.
>
> There's a considerable difference in level of detail.
From the abstract of Lott and Mustard:
"If those states without right-to-carry concealed gun provisions had
adopted them in 1992, county- and state-level data indicate that
approximately 1,500 murders would have been avoided
yearly. Similarly, we predict that rapes would have declined by over
4,000, robbery by over 11,000, and aggravated assaults by over
60,000."
"The estimated annual gain
from all remaining states adopting these laws was at least $5.74
billion in 1992.
Seems pretty detailed.
Tim
Tim Lambert
> All of the critics of Lott objected to his use of a model that
> assumed continuous variability, when some of his variables, such
> as arrest rate, could not go below zero, and were frequently
> bounded by zero.
Not true. This is the objection of Plassmann and Tideman. Black and
Nagin and Dezhbakhsh and Rubin do not make this objection. And if you
want to claim that they did, you'd better provide an actual quote from
them.
> Nagin used this defect as justification for
> reworking his numbers in drastic and unreasonable ways.
Not true. I already explained the reason why they did some analyses
just on counties with more than 100,000 people. And if it was so
drastic and unreasonable, why did Lott do it too?
Page 35 of Lott and Mustard: "We reran all the regressions in this
section first by limiting the sample to those counties over 10,000,
100,000, and then 200,000 people."
> Yet
> almost all of the critics who made a big deal out of this, who
> complained a lordly manner how absurd this was, and how it
> contaminated his results, proceeded to use models with the same
> defect.
Not true. Plassmann and Tideman make this objection, but they then go
on to use a different model.
> This shows that at least some of the criticism of Lott is
> dishonest.
No, it shows that you don't know what you are talking about. Do you
want to tell us about those scatter diagrams in Lott again?
Tim
Tim Lambert
> --
> On 10 Mar 2002 00:51:12 +1100, Tim Lambert
> <lam...@cse.unsw.EDU.AU> wrote:
> > You didn't actually read what you quoted did you? Here it is
> > again (my emphasis):
> >
> > "Without Florida in the sample, the estimation of Lott and
> > Mustard's
> > model, which is given by equation (1), FOR ALL COUNTIES
> > provides no evidence of an impact of RTC laws on homicide and
> > rape"
> >
> > That's "FOR ALL COUNTIES", you know, like including all the
> > small counties.
>
> I do not know if they are telling the truth when they say that it
> is not a function of the large county sample, but I do know they
> are not telling the truth when they say "the estimation of Lott
> and Mustard's model." Lott's model attempts to control for the
> variation in concealed carry holders over time. Black and Nagin
> do not, with the result that their model, which they deceptively
> attribute to Lott, generates all sorts of implausible
> correlations. If you do stupid things with Lott's data and use
> some of his methods in stupid ways, you can get stupid results.
> Big deal.
At this point I can only conclude that:
1. You haven't read the paper by Black & Nagin that you are attacking
so enthusiastically.
and
2. You haven't read the paper by Lott & Mustard that you are defending
so enthusiastically.
Lott & Mustard summarize their results in their abstract:
"If those states without right-to-carry concealed gun provisions had
adopted them in 1992, county- and state-level data indicate that
approximately 1,500 murders would have been avoided
yearly. Similarly, we predict that rapes would have declined by over
4,000, robbery by over 11,000, and aggravated assaults by over
60,000. We also find criminals substituting into property crimes
involving stealth, where the probability of contact between the
criminal and the victim is minimal."
The regressions that this is based on are listed in table 3 of their
paper. Taking those regressions and rerunning them without Florida
causes the results for murder and rape no longer to be statisitically
significant.
Tim
David Friedman
Tim Lambert <lam...@cse.unsw.EDU.AU> wrote:
Less detailed than separate results by state.
--
David Friedman
www.daviddfriedman.com/
James A. Donald
James A. Donald:
> > All of the critics of Lott objected to his use of a model that
> > assumed continuous variability, when some of his variables,
> > such as arrest rate, could not go below zero, and were
> > frequently bounded by zero.
Tim Lambert
> Not true. This is the objection of Plassmann and Tideman. Black
> and Nagin and Dezhbakhsh and Rubin do not make this objection.
Liar.
The bounded by zero problem is Black and Nagin's justification for
throwing out great gobs of data that they do not want.
> And if it was so drastic and unreasonable, why did Lott do it
> too? Page 35 of Lott and Mustard: "We reran all the regressions
> in this section first by limiting the sample to those counties
> over 10,000, 100,000, and then 200,000 people."
Out of context. Lott's purpose is to show that these counties do
not bias the results, to answer in advance the criticism that
Black and Nagle dishonestly made.
James A. Donald:
> Yet
> > almost all of the critics who made a big deal out of this, who
> > complained a lordly manner how absurd this was, and how it
> > contaminated his results, proceeded to use models with the
> > same defect.
Tim Lambert
> Not true. Plassmann and Tideman make this objection, but they
> then go on to use a different model.
Plassmann and Tideman are not "critics who make a big deal out of
this", since their results confirm those of Lott, drawing the
conclusion that treating data bounded by zero as if it was not so
bounded does not cause any big problem in this data set -- the
same conclusion that Lott and Mustard have already supported by
less rigorous means.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
Esa27/RMiQcaVszNal1BFRKmqU0s/Ynwy4kJQtxZ
4Xi+4YHcmFSeuKsUtsilMEbAvfbx6/s3i+dXa+N4o
James A. Donald
David Friedman
> > > > Two things struck me as odd about the Rubin and Dezhbakhsh
> > > > paper. The first is that they report the estimated effect
> > > > if states that didn't have shall issue laws had adopted
> > > > them, but not the other way around. If we believe,
> > > > optimistically, that the states where the laws work best
> > > > are the first to adopt them,
Tim Lambert:
> > > But is there any reason at all to believe this?
David Friedman:
> > Any reason at all? Yes. Even political institutions have some
> > tendency towards efficient outcomes, just not much (for
> > details see the chapter on the economics of government in my
> > webbed _Price Theory_). Enough reason to make it odd that they
> > reported only half the results implied by their analysis.
Tim Lambert:
> But there is no evidence that the laws work better in those
> states.
It was a curious and striking omission that Rubin failed to
compare his predictions with what in fact actually happened.
I, like you, and unlike David Friedman, tend to suspect that
people are liars. However, I suspect a different set of people
as being liars.
When I compare Plassmann's paper with Rubin's paper I am struck by
the fact that Plassmann went to great lengths to construct numbers
that could be compared with real world events, and compared with
other people's numbers, whereas Rubin et al criticize other
people's numbers as if they are going to produce different and
conflicting numbers, yet then sneak off quietly, and do not do
what they announce.
Rubin et al's paper is not in fact evidence that contradicts
Lott's paper, though the abstract announces that it is. Rather,
the conclusion that should be drawn from Rubin's paper is that if
we analyze the data in a certain way, it does not tell us much.
This is not in any way evidence against the claim that if we
analyze data in a different way, it tells us a great deal.
David Friedman and yourself are taking the abstract of Rubin et
al's paper at face value. I do not think that abstract is an
accurate account of the contents of the paper.
For example, even after the drastic measures applied by Black et
al to their data, concealed carry still shows up as a large and
statistically significant deterrent to robbery in those states
were concealed carry laws went into effect. Suppose Rubin's
predictions showed that there should be no such effect. Then
obviously people would doubt his numbers. Suppose they showed
that there would be such an effect. Then people would doubt his
conclusions. Either way, these numbers would be bad for him, and
the fact that they are absent should raise a red flag.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
7VU3DvVAlFTzRoyZhfjzf3Azt5bQ9sFA+5fBprB8
4rsdtRvnHWuCixG9moa/h/cWMpnLbGyY0lSJl0DLB
James A. Donald
On 10 Mar 2002 00:11:18 +1100, Tim Lambert
<lam...@cse.unsw.EDU.AU> wrote:
> That seems to be common attitude among economists who have
> published on this matter. Lott certainly hasn't been shy about
> making public policy recommendations.
But the Rubin paper contains ONLY recommendations. It tells us
what would have happened, if their model is accurate, and public
officials did things differently from what they actually did. If
fails to tell us what their model predicts should have happened
when public officials actually did what they did.
Surely that is a rather odd omission, and one that suggests
conclusions rather more cynical than those that David Friedman is
implying.
It is also an omission that makes it very difficult to compare
the Rubin model with reality and with other paper's analyzing
those events.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
VZlQ75EhJB3huYhjepfWZzL+kFHmyAj/CbwD6h54
43gMsdANXwfGADutu1L96XqRaTtTnGiRuEeFYtu28
James A. Donald
> their paper. Taking those regressions and rerunning them
> without Florida causes the results for murder and rape no longer
> to be statisitically significant.
By omitting to mention robbery, you uttered a deceptive half
truth. By failing to mention time distribution, implying that
this regression was the primary evidence that Lott relied on, you
uttered not a half truth, but a lie.
On reviewing your data, I see that you are technically right and I
was wrong -- but the data is still suggestive, and if we add in
robbery, it is statistically significant even without Florida.
You are correct in the sense that if we ignore Florida AND robbery
AND ignore the time distribution of declines in confrontational
crime, the evidence that concealed carry is effective in reducing
confrontational crime could be a mere statistical fluctuation.
So phrased, phrased honestly instead of deceptively, your
statement, though technically true, becomes considerably less
persuasive.
Your statement is deceptive, so deceptive as to amount to a lie,
in that the regression in question is not the primary evidence
that Lott relies on, so in fact if one merely throws out Florida
and robbery, the results Lott and Mustard primarily rely on REMAIN
statistically significant.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
/fUMpa+luOVgebzpCyNqiMmpxPVwapOTYcqRcA4B
4MdSKVbIo/qmRbwmh2e5vHOgzt1YJIIl0Z8upr6B5
James A. Donald
> > Lott presents a lot of numbers,
Tim Lambert
> You claimed that Lott presents scatter diagrams. Are you now
> retracting your claim?
Yes. The diagram I misremembered was a time series, not a scatter
plot.
The averaging that goes into constructing a time series can
conceal lots of stuff.
Still, the fact remains that he presents a lot more data than his
critics, that he does not massage his numbers in the extensive and
creative fashion that his critics do. It is a lot easier to rig
one magic number, as his critics do, than a large collection of
numbers.
The paper http://bingweb.binghamton.edu/~fplass/gun.pdf should
settle the controversy between Lott's book, and Black's attack on
it.
At each step of their analysis, they compare their data and
results with those of Lott and Mustard, and those of Black and
Nagin.
Plassmann and Tidernan deal with the problem of values bounded by
zero by a less drastic method than throwing out most of the data,
and deal with the problem of Florida allegedly being an outlier
state by addressing each state separately.
Their results are almost the same as Lott's
For murder, Florida is an outlier in that shall issue appeared to
have a much greater impact than it did in the other states, but
several other states show significant, and statistically
significant, impacts. The impacts for murder in the other states
are similar to those reported by Black when they exclude Florida,
but with the big difference that because Plassmann does not throw
out most of the data, these impacts remain statistically
significant for the most part, state by state.
For rape, Florida is not an outlier. It shows the highest impact
of any state, but not by very much. Georgia shows almost
identical numbers. Idaho, Missisippi, and Pennsylvania show
similar numbers. (All of them statistically significant.)
For robbery, Florida is not an outlier, it shows less impact than
most other states.
Now in a sense, both Lott and Black were telling the truth, in
that the results in body of Lott's paper, and the results in the
body of Black's paper, were both consistent with the results in
Plassmann's paper.
But when they and their respective reporters condensed these
results down to short snappy slogans, Lott's condensation
"Concealed carry stops crime", was true, and Black's condensation
"Florida is a outlier, exclude Florida and there is no significant
correlation" was a lie.
Exclude Florida, and the results for rape and robbery do not
change by much. The results for murder change a lot, but are
still pretty good. The loss of statistical significance in
Black's results was wholly an artificat of throwing out or
aggregating large amounts of unwanted data, and even throwing out
lots of data and Florida could not conceal the impact on robbery,
so they simply ignored that unwanted result.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
efRwJCvI69Qz4+hIsoNGDJ7NUBDbwwnsaodKBkBA
4zV90MVlL0u4W4nQ5wHWs/kZMQwhYOyyqaI6xwgzJ
David Friedman
jam...@echeque.com (James A. Donald) wrote:
> Plassmann and Tidernan deal with the problem of values bounded by
> zero by a less drastic method than throwing out most of the data,
> and deal with the problem of Florida allegedly being an outlier
> state by addressing each state separately.
That's Tideman--T. Nicolaus Tideman to be precise.
Thanks for the reference. I've added it to my page on the controversy.
--
David Friedman
www.daviddfriedman.com/
David Friedman
David Friedman <dd...@best.com> wrote:
Oops--the risk of working late. It was already there--probably thanks to
Tim's help.
--
David Friedman
www.daviddfriedman.com/
Tim Lambert
> --
> Tim Lambert <lam...@cse.unsw.EDU.AU> wrote in message
> > The regressions that this is based on are listed in table 3 of
> > their paper. Taking those regressions and rerunning them
> > without Florida causes the results for murder and rape no longer
> > to be statistically significant.
>
> By omitting to mention robbery, you uttered a deceptive half
> truth.
Let's put my paragraph back in its original context:
--start restored text--
Lott & Mustard summarize their results in their abstract:
"If those states without right-to-carry concealed gun provisions had
adopted them in 1992, county- and state-level data indicate that
approximately 1,500 murders would have been avoided
yearly. Similarly, we predict that rapes would have declined by over
4,000, robbery by over 11,000, and aggravated assaults by over
60,000. We also find criminals substituting into property crimes
involving stealth, where the probability of contact between the
criminal and the victim is minimal."
The regressions that this is based on are listed in table 3 of their
paper. Taking those regressions and rerunning them without Florida
causes the results for murder and rape no longer to be statistically
significant.
--end restored text--
We see that James first cleverly removed the mention of robbery and
then turned around and accused my of deception for not mentioning robbery.
> By failing to mention time distribution, implying that
> this regression was the primary evidence that Lott relied on, you
> uttered not a half truth, but a lie.
Well, it WAS the primary evidence that Lott relied on. The quote from
the abstract of Lott and Mustard proves that. Next time you could
read it instead of snipping it.
> On reviewing your data, I see that you are technically right and I
> was wrong -- but the data is still suggestive, and if we add in
> robbery, it is statistically significant even without Florida.
Well, no actually. Robbery was not statistically significant even
with Florida. Excluding Florida actually increases the significance
for robbery so that it is not quite significant at the 5% level.
Perhaps you were confused when Lott wrote: "By any standards I know, a
t-statistic of 1.9 for robberies is still statistically significant at
better than the 5 percent level" Lott should check his statistical
tables :-)
Tim
Tim Lambert
> --
> James A. Donald:
> > > All of the critics of Lott objected to his use of a model that
> > > assumed continuous variability, when some of his variables,
> > > such as arrest rate, could not go below zero, and were
> > > frequently bounded by zero.
>
> Tim Lambert
> > Not true. This is the objection of Plassmann and Tideman. Black
> > and Nagin and Dezhbakhsh and Rubin do not make this objection.
>
> Liar.
>
> The bounded by zero problem is Black and Nagin's justification for
> throwing out great gobs of data that they do not want.
That is not true. Here, again, is their justification:
" In the regression whose results are shown in Table 3 of the L&M
article, one of the independent variables is the rate of
arrests. This creates a problem, as Lott and Mustard point out,
because for some counties for some years for some classes of offense,
there are no offenses and no arrests, giving an arrest rate of
0/0. Such cases are omitted in the regression. This could conceivably
bias the results, since it is omitting a non-random set of data--ones
for counties with very low crime rates.
In order to eliminate this problem, Black and Nagin rerun the
regressions, eliminating all counties with fewer than 100,000 people
in them--and thus most of the counties for which this problem arises."
Are you seriously claiming that the quote above is a lie?
Tim
Tim Lambert
> I, like you, and unlike David Friedman, tend to suspect that
> people are liars. However, I suspect a different set of people
> as being liars.
You have repeatedly made false statements about what Black and Nagin
and what they gave as their reasons. This is despite being corrected
over and over again. I suppose it would not be correct to call your
statements "lies", since you haven't read Black and Nagin's paper and
so you could not be absolutely certain that what you wrote was false.
The best description I can think of for your behaviour is "writing
with a wilfull disregard for the truth". Perhaps some reader could
come up with a more succint description?
Tim
Tim Lambert
> --
> > > Lott presents a lot of numbers,
>
> Tim Lambert
> > You claimed that Lott presents scatter diagrams. Are you now
> > retracting your claim?
>
> Yes. The diagram I misremembered was a time series, not a scatter
> plot.
Nope. The diagrams you are thinking of show perfectly smooth curves.
If they were time series they would show some noise. What Lott did
was compute some of those "magic numbers" you are complaining about
and then made a diagram using the magic numbers. It gets worse. He
than argued that since his diagrams showed something happening at the
time of the laws, whatever it was that caused the change in crime
rates must have happened at exactly the time of the laws. This is an
error.
It's as if he took a scatter plot showing a cloud of data, fitted a
straight line to it and then presented a diagram showing the straight
line and not the original data AND THEN argued that diagram showed
that there was a linear relationship.
Tim
Tim Lambert
> Tim Lambert <lam...@cse.unsw.edu.au> wrote in message news:<m3vgc9z...@cycloid.localdomain>...
>
> > > To what paper do you refer here?
> >
> > I imagine that he is referring to Black and Nagin's critique. One of
> > their points is that excluding Florida causes the results for murder
> > and rape to no longer be significant. Why James considers this to be
> > artificial peculiar and contrived, I don't know.
>
> I thought perhaps he was referring to:
>
> Rubin, Paul, and Dezhbakhsh, Hashem, "Lives Saved or Lives Lost: The
> Effect of Concealed Handgun Laws on Crime," American Economic Review
> Papers and Proceedings, May, 1998, 468-474.
>
> They use Lott's own data set, with some enhancements, and suggest that
> he doesn't include some important cross-county variation. Including
> them appears to make concealed-carry laws have little to no effect on
> crime (and in some cases, seems to increase it--although I think the
> strongest statement one can make is that the laws have no clear
> effect).
>
> > You might be interested in Ted Goertzel's article in the Jan 2002
> > Skeptical Inquirer. It starts:
> >
> > "Do you believe that every time a prisoner is executed in the United
> > States, eight future murders are deterred? Do you believe that a 1%
> > increase in the number of citizens licensed to carry concealed weapons
> > causes a 3.3% decrease in the state's murder rate? Do you believe that
> > 10 to 20% of the decline in crime in the 1990s was caused by an
> > increase in abortions in the 1970s? Or that the murder rate would have
> > increased by 250% since 1974 if the United States had not built so
> > many new prisons?"
>
> It doesn't to be the case that every execution deters eight murders;
> the statistics suggested that on _average_, eight murders were
> deterred.
I think that's what he means.
> There's a paper coming out sometime soon in J. Law and Econ
> suggesting that it's more like between 10 and 20-something murders
> deterred per execution.
I think that he is referring to Ehrlich's work with that eight figure.
That new paper is also by Dezhbakhsh and Rubin, right?
Looks like they're trying to annoy both sides of politics:
Paper 1: Concealed carry doesn't deter crime.
Paper 2: Executions do deter crime.
Tim
James A. Donald
Tim Lambert:
> Nope. The diagrams you are thinking of show perfectly smooth
> curves. If they were time series they would show some noise.
I have Lott and Mustard's book, second edition, open in front of
me. I flipped at random, till I came to one of his many time
series. These ones happened to be on page 138 and 139 They show
noise.
> What Lott did was compute some of those "magic numbers" you are
> complaining about and then made a diagram using the magic
> numbers.
I flipped a little more. His preferred graphs, page 78, show near
zero noise. According to his commentary, not because of magic
number manipulation, but because they employ more data.
Similarly, his graphs on page 94, which employ a smaller subset of
the data, show a little noise, more than on page 78, less than on
138.
Like almost everything you have said about Lott and Mustard, and
their data, your claims are simply untrue. Either you are lying,
or you are indifferent to what the facts actually are.
> It gets worse. He
> than argued that since his diagrams showed something happening
> at the time of the laws, whatever it was that caused the change
> in crime rates must have happened at exactly the time of the
> laws. This is an error.
The crime rates declined when "shall issue" introduced, but "shall
issue was permitted at different times in different states. This
offers convincing evidence that the decline was the result of
concealed carry.
The decline is a robust change. It does not matter what dates you
throw out, what states you throw out, what counties you throw out.
This is the convincing and decisive evidence that permitting
people to defend themselves stops crimes.
In essence, Lott presents decisive evidence for the effect on
concealed carry laws, and you conclude he must be lying. Yet
other people look at the same data, and fail to call out his lie.
Lott presents lots of extremely decisive evidence. His critics
response has been to look at the weakest evidence that he has
presented, and argue that it is weaker than it looks, while simply
ignoring the strongest evidence that he presented -- county level
variations in the decline of crime rate according to their policy
on concealed carry before the statewide change.
From the lying and weasily character of their response, we know
that his argument and evidence must be truthful and decisive.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
evi8s/yh/MYD5B3smne/4P5+OQFfpfnWHpZbnT2N
4fCxEWSPGUZQVGikd+5TBkto/W8RJrIYdKs6GUk8o
David Friedman
Tim Lambert <lam...@cse.unsw.EDU.AU> wrote:
> I think that he is referring to Ehrlich's work with that eight figure.
> That new paper is also by Dezhbakhsh and Rubin, right?
>
> Looks like they're trying to annoy both sides of politics:
A good policy.
I have a vague memory of some piece by Rubin arguing that government
expenditure was determined by what the government could get its hands
on, not what it "needed"--a claim appealing to only a few extremists
such as myself, Lott, and the late C. Northcote Parkinson--but I may be
misremembering. I know he's contributed to the literature on why one
would expect the common law to be economically efficient.
> Paper 1: Concealed carry doesn't deter crime.
> Paper 2: Executions do deter crime.
Paper 1 doesn't say that concealed carry doesn't deter crime. It says
that concealed carry sometimes deters some crimes--and if the state
legislatures would just check with the authors before deciding whether
to legalize it, national crime rates would fall.
--
David Friedman
www.daviddfriedman.com/
James A. Donald
James A. Donald:
> > I, like you, and unlike David Friedman, tend to suspect that
> > people are liars. However, I suspect a different set of
> > people as being liars.
Tim Lambert:
> You have repeatedly made false statements about what Black and
> Nagin and what they gave as their reasons. This is despite
> being corrected over and over again.
I have made numerous minor errors. However, on the key facts in
dispute, the important issues that matter, you are lying, and I am
telling the truth.
Your account of what Lott did is incompatible with what Lott says
he did. Either Lott is lying, or you are. When we examine the
weasel worded way in which is critics attack Lott, the implication
of their forked tongued words is that Lott is speaking the truth.
If you were telling the truth, and Lott was lying, his critics
would not have to express themselves in such a serpentine evasive
fashion.
You said in plain and simple lies, what Black and Nagle evasively
implied with weasel worded equivocal half truths.
Your claim is that if you throw out Florida, Lott's result's go
away. They don't. You can check this from Lott's book, where he
discusses sensitivity to particular states at great length, and
from people such as Plassmann who reanalyses the same data with
different methods, or merely by noting the carefully evasive
phrasing used by Black and Nagle where they say something that
sounds as if they are saying the data goes away, when in fact they
do not actually say that, merely try to sound as if they are
saying that.
Lott never relied on the aggregations that Black and Nagin
criticized, and even if he relied on those aggregations, which are
among the weakest evidence he presented, leaving out Florida does
not make those aggregations go away, though it weakens them
significantly.
But even if leaving out Florida did make that particular item of
evidence go away, that item was the least impressive of the
statistics presented by Lott.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
1DpDuJgwj94pne/qtd8TYOGNqOzctm7tuP8s1pdQ
48yH09i4VyxQ0zSP2nHkhKxaIG8gatJP3aWzomNvi
James A. Donald
On 12 Mar 2002 17:57:26 +1100, Tim Lambert
> Nope. The diagrams you are thinking of show perfectly smooth
> curves. If they were time series they would show some noise.
> What Lott did was compute some of those "magic numbers" you are
> complaining about and then made a diagram using the magic
> numbers.
You are, as usual, lying. Black and Nagin do not make that
accusation when they criticize those curves.
Instead they say: "Each data point is an average for states that
liberalized their gun carrying laws but the states that make up
the average are not the same each year [...] for many of the
states studied, data was only available one to three years after
the laws were implemented."
On page 247 of "More guns, less crime, second edition". Lott
discuses how those curves are generated. No smoothing. Black
and Nagin do not accuse him of smoothing, and the procedure he
describes involves no smoothing.
Those curves are smooth, because Lott aggregates a lot of data.
Where he aggregates less data, they are less smooth. He presents
many such charts, and some are noisier than others, depending on
how much data they average together.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
EH9J4F6yE68/zFA/VPW1Lgk46SG0tqMinsXwPB+Q
4HfWOP5tzgaTkvPA90GYBvCTZjHaz4WrRomyVIscZ
Tim Lambert
From the abstract:
"Our results show that the expected effect of the law on crime varies
across the counties and states and depends on county-specific
characteristics. Such effects appear to be much smaller and more mixed
than Lott and Mustard suggest, and are not crime-reducing in most
cases."
From the paper:
"For murder, for example, there is only a small
reduction. For robbery, many states experience increases in crime. For
other crimes, results are ambiguous, with some states showing
predicted increases and some predicted decreases."
Dezhbakhsh's comments on their results:
"My work with Emory economist Paul Rubin-described in "Lives Saved or
Lives Lost: The Effect of Concealed Handgun Laws on Crime," published
in American Economic Review in 1998-criticizes Lott on simple but
fundamental grounds. We show that Lott's work is erroneous not only
in theory but also in its arithmetic.
"Lott's finding relies on the assumption that the effect of permissive
handgun laws on crime is identical across all counties and
independent of any county characteristics. This assumption is flatly
contradicted by conventional wisdom. Such laws would not have the
same effect in crime-ridden urban areas as they would in remote rural
counties or affluent suburbs. Some of Lott's results also assume that
the number of arrests made by police does not depend on the number of
crimes committed! So rural counties with very few crimes may
presumably have more police arrests than urban counties with very
large crime rates.
"Moreover, Lott's central results are invalid because of errors in
computing expected arrest rates: he obtains mostly negative numbers
for arrests. For example, more than 19,000 of approximately 33,000
county-level auto theft arrests are "negative"; the number of
negative arrest rates for aggravated assault and property crimes are,
respectively, 9,900 and 13,500. What does a negative arrest rate
mean? Obviously, the number of individuals arrested for crimes can
only be zero or positive. Once we correct for these errors, the
more-guns-less-crime claim disintegrates. In fact, we show not only
that Lott's strong crime-reducing effect does not materialize, but
also that concealed handguns lead to a higher robbery rate."
Tim
Tim Lambert
> --
> Tim Lambert:
> > Nope. The diagrams you are thinking of show perfectly smooth
> > curves. If they were time series they would show some noise.
>
> I have Lott and Mustard's book, second edition, open in front of
> me. I flipped at random, till I came to one of his many time
> series. These ones happened to be on page 138 and 139 They show
> noise.
These aren't the main graphs that he relies on, and yes, page 139
shows a time series. Notice how there's noise, unlike Lott's other
graphs?
> > What Lott did was compute some of those "magic numbers" you are
> > complaining about and then made a diagram using the magic
> > numbers.
>
> I flipped a little more. His preferred graphs, page 78, show near
> zero noise.
Not near zero noise, but zero noise. The points all fall exactly on
two parabolic arcs. I scanned one of the graphs:
http://www.cse.unsw.edu.au/~lambert/guns/lott/figure4.7.gif
Real data does not look like that. For example, see:
http://cdpsweb.state.co.us/ors/grpharc2.htm
This shows a time series of US homicide rates. Notice how it doesn't
lie on perfect parabolic arcs.
> According to his commentary, not because of magic
> number manipulation, but because they employ more data.
You're making stuff up again. He describes his graphs as "a simple
time trend and time trend squared for the number of years before and
after the concealed-handgun laws". That means that the curve he
fitted has the functional form a+bt+ct^2 ie a parabola. Do you see
why his "time series" are parabolas?
And no, using more data does not eliminate noise completely. Look at
http://cdpsweb.state.co.us/ors/grpharc2.htm again to see what a time
series using even more data than Lott looks like.
> Similarly, his graphs on page 94, which employ a smaller subset of
> the data, show a little noise, more than on page 78, less than on
> 138.
Wrong. The points all fall exactly on parabolic arcs again. They've
just been connected with straight lines so it's less obvious. Have a
look at graphs on pages 64 to 67. 36 graphs and they're all
monotonic. Real data does not look like that.
> Like almost everything you have said about Lott and Mustard, and
> their data, your claims are simply untrue. Either you are lying,
> or you are indifferent to what the facts actually are.
James, if you are determined not to believe me, then I suggest you
consult someone else with statistical expertise. Once you've done
that, you should post a retraction and an apology.
Tim
James A. Donald
On 13 Mar 2002 21:59:02 +1100, Tim Lambert
<lam...@cse.unsw.edu.au> wrote:
> From the abstract: "Our results show that the expected effect of
> the law on crime varies
> across the counties and states and depends on county-specific
> characteristics. Such effects appear to be much smaller and
> more mixed than Lott and Mustard suggest, and are not
> crime-reducing in most cases."
The problem is that the abstract does not conform to the body.
The abstract promises a comparison between their numbers, and
Lott's numbers, which are precisely the numbers that are curiously
missing from the body.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
DzXfDBlaTug3VL0xAMm1R7yQtBhEHPY8t9BOwwQB
4lv5Z1XnSeUR58/T1gSLd//c7D0M/6c1IKmP8wby2
Tim Lambert
> --
> James A. Donald:
> > > I, like you, and unlike David Friedman, tend to suspect that
> > > people are liars. However, I suspect a different set of
> > > people as being liars.
>
> Tim Lambert:
> > You have repeatedly made false statements about what Black and
> > Nagin and what they gave as their reasons. This is despite
> > being corrected over and over again.
>
> I have made numerous minor errors. However, on the key facts in
> dispute, the important issues that matter, you are lying, and I am
> telling the truth.
You claimed:
> The bounded by zero problem is Black and Nagin's justification for
> throwing out great gobs of data that they do not want.
That is not true. Here, again, is their justification:
" In the regression whose results are shown in Table 3 of the L&M
article, one of the independent variables is the rate of
arrests. This creates a problem, as Lott and Mustard point out,
because for some counties for some years for some classes of offense,
there are no offenses and no arrests, giving an arrest rate of
0/0. Such cases are omitted in the regression. This could conceivably
bias the results, since it is omitting a non-random set of data--ones
for counties with very low crime rates.
In order to eliminate this problem, Black and Nagin rerun the
regressions, eliminating all counties with fewer than 100,000 people
in them--and thus most of the counties for which this problem
arises."
I ask again: Are you seriously claiming that the quote above
describing their justification is a lie?
>
> Your claim is that if you throw out Florida, Lott's result's go
> away.
You misrepresent my position. Here it is again:
Lott & Mustard summarize their results in their abstract:
"If those states without right-to-carry concealed gun provisions had
adopted them in 1992, county- and state-level data indicate that
approximately 1,500 murders would have been avoided
yearly. Similarly, we predict that rapes would have declined by over
4,000, robbery by over 11,000, and aggravated assaults by over
60,000. We also find criminals substituting into property crimes
involving stealth, where the probability of contact between the
criminal and the victim is minimal."
The regressions that this is based on are listed in table 3 of their
paper. Taking those regressions and rerunning them without Florida
causes the results for murder and rape no longer to be statistically
significant.
> Lott never relied on the aggregations that Black and Nagin
> criticized,
Yes he did. Look at the abstract again. Those numbers are based on
the aggregations that Black and Nagin criticized.
Tim
David Friedman
Tim Lambert <lam...@cse.unsw.edu.au> wrote:
> David Friedman <dd...@best.com> writes:
> > Paper 1 doesn't say that concealed carry doesn't deter crime. It says
> > that concealed carry sometimes deters some crimes--and if the state
> > legislatures would just check with the authors before deciding whether
> > to legalize it, national crime rates would fall.
> From the abstract:
> "Our results show that the expected effect of the law on crime varies
> across the counties and states and depends on county-specific
> characteristics. Such effects appear to be much smaller and more mixed
> than Lott and Mustard suggest, and are not crime-reducing in most
> cases."
> From the paper:
> "For murder, for example, there is only a small
> reduction. For robbery, many states experience increases in crime. For
> other crimes, results are ambiguous, with some states showing
> predicted increases and some predicted decreases."
From the paper:
"We identify states (Illinois, Kansas, Minnesota) that might benefit
from passage of these laws, and states (Maryland, New Mexico, and Iowa)
that probably would not.
...
The sort of analysis developed here could be used to enable policy
makers to more carefully tailor laws to particular conditions in a
jurisdiction."
> Dezhbakhsh's comments on their results:
> "My work with Emory economist Paul Rubin-described in "Lives Saved or
> Lives Lost: The Effect of Concealed Handgun Laws on Crime," published
> in American Economic Review in 1998-criticizes Lott on simple but
> fundamental grounds. We show that Lott's work is erroneous not only
> in theory but also in its arithmetic.
Perhaps you could point at where in the article he actually does so? The
point of the article is not that Lott makes any mistakes in arithmetic
but that he ought to have estimated a more elaborate model.
> Some of Lott's results also assume that
> the number of arrests made by police does not depend on the number of
> crimes committed! So rural counties with very few crimes may
> presumably have more police arrests than urban counties with very
> large crime rates.
> "Moreover, Lott's central results are invalid because of errors in
> computing expected arrest rates: he obtains mostly negative numbers
> for arrests. For example, more than 19,000 of approximately 33,000
> county-level auto theft arrests are "negative"; the number of
> negative arrest rates for aggravated assault and property crimes are,
> respectively, 9,900 and 13,500. What does a negative arrest rate
> mean? Obviously, the number of individuals arrested for crimes can
> only be zero or positive.
Perhaps you can point to where in the article these claims are
supported? I could not find it. This seems to be simply one of the
authors of the article offering his own criticisms, not in a peer
reviewed article, not with any endorsement by the other author, not
supported by the published work. You might try contrasting the tone of
that passage, which sounds like a wholly one sided diatribe, with the
article, which makes a very much more moderate and reasonable set of
claims.
Note the following from the article:
"In regression analysis an intercept-shifting dummy variable is often
used to estimate the effect of an institutional change. The statistical
and conceptual ramifications of this practice is seldom examined,
particularly when the empirical analysis is not predicated on economic
theory. To better motivate our procedure which is intended to overcome
the shortcomings of this approach, we elaborate on this issue using Lott
and Mustard's (1997) highly publicized study on the effect of concealed
handgun laws on crime."
Or in other words, what the authors say is that Lott is using a
conventional approach to making the estimation, that there are serious
problems with that approach, and they will demonstrate that by providing
a more sophisticated analysis. That isn't exactly the impression you get
from the Deshbakhsh quote.
I'm curious as to how much of the quote Paul Rubin would agree with.
Incidentally, it's worth making a general point that can get lost in the
details of these arguments--the tradeoff between unrealistic
simplification in modelling on the one hand and permitting specification
searches on the other.
Suppose there is one obvious way of doing the statistics--but it is
unrealistic in ways that make its results less reliable than some other
and much more sophisticated way. On the face of it, you should obviously
use the latter.
But now suppose that one can plausibly argue for any of a hundred more
sophisticated ways. The first researcher uses the simple approach and
reports result X. A critic later publishes, using one more sophisticated
result, and reports not-X. Whom do you believe?
The answer depends on whether you think the critic engaged in a
specification search, either explicitly (try all hundred versions and
report the only one that yields not-X) or implicitly (a hundred critics
try one version each and only the not-X result is worth publishing).
To some extent, Lott dealt with this problem by offering both the simple
version and a variety of more sophisticated versions, some (if I
remember correctly) using data not available for the whole sample. The
critics seem to be each doing one sophisticated analysis and reporting
its result. Some (Plassman and Tideman) confirm Lott's general result,
some (this paper) don't.
--
David Friedman
www.daviddfriedman.com/
Mary Rosh
> jam...@echeque.com (James A. Donald) writes:
> You claimed:
> > The bounded by zero problem is Black and Nagin's justification for
> > throwing out great gobs of data that they do not want.
>
> That is not true. Here, again, is their justification:
>
> " In the regression whose results are shown in Table 3 of the L&M
> article, one of the independent variables is the rate of
> arrests. This creates a problem, as Lott and Mustard point out,
> because for some counties for some years for some classes of offense,
> there are no offenses and no arrests, giving an arrest rate of
> 0/0. Such cases are omitted in the regression. This could conceivably
> bias the results, since it is omitting a non-random set of data--ones
> for counties with very low crime rates.
>
> In order to eliminate this problem, Black and Nagin rerun the
> regressions, eliminating all counties with fewer than 100,000 people
> in them--and thus most of the counties for which this problem
> arises."
The problem that James is referring to is that when you throw out the
arrest rate you have many counties with zero crime rates. That
creates a truncation problem and biases the results towards zero.
Lott goes through and shows that there are still a nontrivial number
of zero crime rate counties even with populations over 100,000.
>
> >
> > Your claim is that if you throw out Florida, Lott's result's go
> > away.
>
> You misrepresent my position. Here it is again:
> Lott & Mustard summarize their results in their abstract:
>
>
> "If those states without right-to-carry concealed gun provisions had
> adopted them in 1992, county- and state-level data indicate that
> approximately 1,500 murders would have been avoided
> yearly. Similarly, we predict that rapes would have declined by over
> 4,000, robbery by over 11,000, and aggravated assaults by over
> 60,000. We also find criminals substituting into property crimes
> involving stealth, where the probability of contact between the
> criminal and the victim is minimal."
>
> The regressions that this is based on are listed in table 3 of their
> paper. Taking those regressions and rerunning them without Florida
> causes the results for murder and rape no longer to be statistically
> significant.
>
>
As usual you completely misrepresent Lott & Mustard. First, the
results in Table 3 were biased against finding a drop in crime. Lott
& Mustard never claimed that those were the most believeable results
in fact they went on in many ways to explain the problems with them in
the original paper. Lott quotes David Friedman as agreeing with this
point in his book. Second, as Lott has pointed out in both his
response to Black and Nagin and his book, dropping out Florida and
counties with fewer than 100,000 people weakens the results for muder
and rape when you compare the before and after averages, but doing so
actually strengthens the results when you are looking at the before
and after trends. Lott points out that dropping out these
observations make the data pattern look like even more of what he
calls an inverted "V" and that is why you are less likely to get
statistically significant differences in the before and after
averages. Do you understand why looking at the before and after
averages can be misleading when the crime rates are rising before the
law and falling there after?
James A. Donald
On 14 Mar 2002 01:05:53 +1100, Tim Lambert <lam...@cse.unsw.EDU.AU>
wrote:
> jam...@echeque.com (James A. Donald) writes:
>
> > --
> > Tim Lambert:
> > > Nope. The diagrams you are thinking of show perfectly smooth
> > > curves. If they were time series they would show some noise.
> >
> > I have Lott and Mustard's book, second edition, open in front of
> > me. I flipped at random, till I came to one of his many time
> > series. These ones happened to be on page 138 and 139 They show
> > noise.
>
> These aren't the main graphs that he relies on, and yes, page 139
> shows a time series. Notice how there's noise, unlike Lott's other
> graphs?
>
> > > What Lott did was compute some of those "magic numbers" you are
> > > complaining about and then made a diagram using the magic
> > > numbers.
> >
> > I flipped a little more. His preferred graphs, page 78, show near
> > zero noise.
>
> Not near zero noise, but zero noise. The points all fall exactly on
> two parabolic arcs. I scanned one of the graphs:
> http://www.cse.unsw.edu.au/~lambert/guns/lott/figure4.7.gif
They do indeed look suspiciously parabolic. Almost unbelievably
parabolic. So I checked one of his graphs, the right hand side of
graph 4.7 on page 78, "more guns, less crime, second edition", the one
you post in the above link.f
The points were 169, 164, 160.5, 157.5, 155.75
Taking first order differences, we get 5, 3.5,, 3, 1.75
Taking second order differences we get 1.5, 0.5, 1.25
If the graphs were parabolic, the second order differences would be
constant.
The graphs are close enough to parabolic that you cannot see the
noise, but the noise is there. Almost any smooth curve is going to
look parabolic if you eyeball it.
However the most compelling graph is figure 5.1, on page 192, of mass
public shootings. Because the dataset is small, (listed in on the
preceding page) the data has quite visible noise, yet even so, the
graph is fairly smooth, and demonstrates the dramatic effect of
concealed carry on mass public shootings, the kind of crime for which
we would expect it to be most effective.
Surely it is unsurprising that graphs for much larger data sets should
be correspondingly smooth.
If we look at the graphs for individual states, for example florida
page 109, we see the curve is almost parabolic with one odd point.
Obviously if you average all the states together, odd points like that
one are going to average out, drop below visibility.
Lott presents lots of graphs. The ones that average lots of data look
perfectly smooth, but are not perfectly smooth when you measure them.
The ones that average moderate amounts of number look almost perfectly
smooth, but not quite. The ones that average sparse data, look noisy.
I would say this is just the well known large number effect -- that
any statistical curve is going to smooth out when you average over
enough disparate cases.
> Real data does not look like that. For example, see:
> http://cdpsweb.state.co.us/ors/grpharc2.htm
> This shows a time series of US homicide rates. Notice how it doesn't
> lie on perfect parabolic arcs.
His graph is not a time series, nor does it lie on a perfect parabolic
arc.. It is the average of several time series each with a different
starting date. When he gives a single time series with a single
starting date, it has about as much noise as that graph,
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
bcbG6i5hK4x1d+shTrqTRt+hnqDf7bjmNGXfEkIG
4k3UyY6rKMxntzam9CeGjTpTjg6UR0VYKFEdEwE4L
If we look at the graphs for individual states, for example florida
page 109, we see the curve is almost parabolic with one odd point.
Obviously if you average all the states together, odd points like that
one are going to average out, drop below visibility.
Lott presents lots of graphs. The ones that average lots of data look
perfectly smooth, but are not perfectly smooth when you measure them.
The ones that average moderate amounts of number look almost perfectly
smooth, but not quite. The ones that average sparse data, look noisy.
I would say this is just the well known large number effect -- that
any statistical curve is going to smooth out when you average over
enough disparate cases.
> Real data does not look like that. For example, see:
> http://cdpsweb.state.co.us/ors/grpharc2.htm
> This shows a time series of US homicide rates. Notice how it doesn't
> lie on perfect parabolic arcs.
His graph is not a time series, nor does it lie on a perfect parabolic
arc.. It is the average of several time series each with a different
starting date. When he gives a single time series with a single
starting date, it has about as much noise as that graph,
James A. Donald
James A. Donald:
> > The bounded by zero problem is Black and Nagin's justification
> > for throwing out great gobs of data that they do not want.
Tim Lambert
> That is not true. Here, again, is their justification:
>
> " In the regression whose results are shown in Table 3 of the
> L&M
> article, one of the independent variables is the rate of
> arrests. This creates a problem, as Lott and Mustard point out,
> because for some counties for some years for some classes of
> offense, there are no offenses and no arrests, giving an arrest
> rate of 0/0. Such cases are omitted in the regression. This
> could conceivably bias the results, since it is omitting a
> non-random set of data--ones for counties with very low crime
> rates.
If you understood statistics, you would understand that that is
merely a different way of saying what I just said.
> In order to eliminate this problem, Black and Nagin rerun the
> regressions, eliminating all counties with fewer than 100,000
> people in them--and thus most of the counties for which this
> problem
> arises."
But that is a curiously drastic way of elminating the problem, and
in principle does not eliminate.the problem, since one is still
calculating probabilities as if arrest rates could be negative,
but by sheer chance never happen to be negative.
There are lots of more conventional methods for addressing this
problem. It is most strange that Black and Nagin chose this
rather unusual method for dealing with this quite common problem.
> I ask again: Are you seriously claiming that the quote above
> describing their justification is a lie?
Yes: I think the real justification was to get the errror bars up
high enough so that TWO of the many, many, many, numbers provided
by Lott would drop out of statistical significance, so that lots
people could then make statements, either phrased to be
misleading, or phrased to be barefaced lies, that implied or
outright proclaimed that ALL of the numbers provided by Lott drop
out of statistical significance.
James A. Donald:
> > Your claim is that if you throw out Florida, Lott's result's
> > go away.
Tim Lambert
> You misrepresent my position. Here it is again: Lott & Mustard
> summarize their results in their abstract:
>
> "If those states without right-to-carry concealed gun provisions
> had
> adopted them in 1992, county- and state-level data indicate
> that approximately 1,500 murders would have been avoided
> yearly. Similarly, we predict that rapes would have declined by
> over 4,000, robbery by over 11,000, and aggravated assaults by
> over 60,000. We also find criminals substituting into property
> crimes involving stealth, where the probability of contact
> between the criminal and the victim is minimal."
>
> The regressions that this is based on are listed in table 3 of
> their paper. Taking those regressions and rerunning them
> without Florida causes the results for murder and rape no longer
> to be statistically significant.
Your full statement is equivalent to saying "If we ignore Florida,
and ignore assaults and robberies, and ignore that fact that crime
was rising before concealed carry laws and fell after concealed
carry laws, then Lott's results are not statistically
significant." So phrased, that statement is hilarious. Anyone
would understand it as really meaning "Lott's results are
undeniable, no matter how much I want to deny them."
When, however, you cut it down to a sound bite, as you so often
do, "Without florida, Lott's results for murder and rape are no
longer significant", it sounds very impressive, but is so
deceptive as to constitute a lie.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
ruDUab8Hl80jCefHf8klRVKB56X0pkE+v20dCgs4
4pbLgq11U4XIoME0xJ+hKpaII0VOF18Vk6NuJ8z4V
Tim Lambert
> In article <m3k7sgy...@cycloid.localdomain>,
> Tim Lambert <lam...@cse.unsw.edu.au> wrote:
>
> > David Friedman <dd...@best.com> writes:
> > Dezhbakhsh's comments on their results:
> > "My work with Emory economist Paul Rubin-described in "Lives Saved or
> > Lives Lost: The Effect of Concealed Handgun Laws on Crime," published
> > in American Economic Review in 1998-criticizes Lott on simple but
> > fundamental grounds. We show that Lott's work is erroneous not only
> > in theory but also in its arithmetic.
>
> Perhaps you could point at where in the article he actually does so? The
> point of the article is not that Lott makes any mistakes in arithmetic
> but that he ought to have estimated a more elaborate model.
I think that's what he meant. I.e. not that Lott add two and three
and got four, but that he added two and three when he should have been
multiplying two and three.
I got the impression from the quote that Deshbakhsh thought that there
were serious problems with Lott's approach.
> I'm curious as to how much of the quote Paul Rubin would agree with.
In an earlier posting in this thread, Mike HAmmock wrote:
] I've asked Prof. Rubin about what he
] thinks policy makers ought to do. His response was something like
] "Given that the results don't show a strong effect one way or the
] other, I lean toward passing concealed-carry laws,". They don't seem
] to do much good, but they don't do much harm, either, but since he
] prefers more freedom to less, he'd prefer to see concealed carry laws
] passed.
> Incidentally, it's worth making a general point that can get lost in the
> details of these arguments--the tradeoff between unrealistic
> simplification in modelling on the one hand and permitting specification
> searches on the other.
>
> Suppose there is one obvious way of doing the statistics--but it is
> unrealistic in ways that make its results less reliable than some other
> and much more sophisticated way. On the face of it, you should obviously
> use the latter.
>
> But now suppose that one can plausibly argue for any of a hundred more
> sophisticated ways. The first researcher uses the simple approach and
> reports result X. A critic later publishes, using one more sophisticated
> result, and reports not-X. Whom do you believe?
>
> The answer depends on whether you think the critic engaged in a
> specification search, either explicitly (try all hundred versions and
> report the only one that yields not-X) or implicitly (a hundred critics
> try one version each and only the not-X result is worth publishing).
>
> To some extent, Lott dealt with this problem by offering both the simple
> version and a variety of more sophisticated versions, some (if I
> remember correctly) using data not available for the whole sample. The
> critics seem to be each doing one sophisticated analysis and reporting
> its result. Some (Plassman and Tideman) confirm Lott's general result,
> some (this paper) don't.
Looking at Plassman and Tideman's paper I notice that they found a
larger decrease for murder, a similar result for rape, and, instead of
a decrease, a significant *increase* in robberies. The state by state
results are also mixed--more decreased than increased, but some
experienced significant increases. They write:
"While this ambiguous result is somewhat discouraging, because it
indicates that a right to carry concealed handguns is unlikely to be
the ultimate weapon against crime, it is not very
surprising. Whenever the theoretically possible and in-practice
plausible effects of public policy are ambiguous, it can be expected
that the effects of such a policy will differ across localities that
are different from each other. It is rather remarkable that these
effects differ so clearly, and that our analysis produced so many
statistically significant and consistent results."
Looking at the state by state numbers also produces some troubling
questions. Oregon experienced a 37% decrease in robbery, while
Mississipi had a 28% increase. Pennsylvania had 15% increase in rape,
while Florida had a 15% decrease. For murder, Florida had a 21%
decrease while Virginia had an 8% increase. Nor are these large
differences just noise -- for most of the estimates the standard error
is just 1 or 2 percentage points.
At the very least some other unknown factor or factors must be
operating here in conjunction with the carry laws and it may be that
if that factor is identified why find that the laws had no effect.
Tim
Tim Lambert
And they are constant, once you allow for rounding error in your
measurements. If your measurements are accurate to the nearest 1/4,
then your first order differences are accurate to the nearest 1/2, and
your second order differences to the nearest 1. All your second order
differences are within 1 of each other.
> The graphs are close enough to parabolic that you cannot see the
> noise, but the noise is there. Almost any smooth curve is going to
> look parabolic if you eyeball it.
Have a look at:
http://www.cse.unsw.edu.au/~lambert/guns/lott/figure4.7.html
(Your browser needs to support Java.) I have superimposed two red
parabolic arcs on Lott's diagram. Notice how perfectly the red
parabolas match Lott's curve. Take a random smooth curve and I
guarantee that you will not be able to get a parabola to fit to it
like that. If you could get parabolas to match as well as that
everyone would use parabolas instead of NURBs for geometric
modelling.
What is really interesting is the one spot where Lott's curve deviates
from a parabola. The parabola turns upwards at the end -- showing
rape rates increasing seven years after the laws. However, Lott has
been naughty and fudged his graph so that his curve merely levels off
after seven years.
Tim
David Friedman
Tim Lambert <lam...@cse.unsw.EDU.AU> wrote:
> > Perhaps you could point at where in the article he actually does so? The
> > point of the article is not that Lott makes any mistakes in arithmetic
> > but that he ought to have estimated a more elaborate model.
> I think that's what he meant. I.e. not that Lott add two and three
> and got four, but that he added two and three when he should have been
> multiplying two and three.
But that isn't what he says. All you have to do is read the whole
passage you were quoting it from and compare it to the published article
to see that the author of the passage is engaged in demagoguery, not
trying to give an accurate report of what the article actually did.
That, of course, is why Paul Rubin's name isn't on the passage.
Do you disagree? Presumably you have read both.
And that he was pretending that he had demonstrated mistakes by Lott in
the published article that he did not demonstrate. Hence that the quote
is worthless, save as evidence that Deshbakhsh is not to be trusted.
Do you disagree?
Again, is that quote consistent with what I quoted above from the
article? Can an honest man say that someone has made lots of mistakes in
arithmetic when what he means is "done the problem in a conventional
fashion when I believe there is a better way of doing it"?
> > I'm curious as to how much of the quote Paul Rubin would agree with.
> In an earlier posting in this thread, Mike HAmmock wrote:
> ] I've asked Prof. Rubin about what he
> ] thinks policy makers ought to do. His response was something like
> ] "Given that the results don't show a strong effect one way or the
> ] other, I lean toward passing concealed-carry laws,". They don't seem
> ] to do much good, but they don't do much harm, either, but since he
> ] prefers more freedom to less, he'd prefer to see concealed carry laws
> ] passed.
The simplest reason to believe Rubin doesn't agree with the webbed
comment is that his name isn't on it. It's particularly striking given
that the page is on the Emory site, Rubin is at Emory, Deshbakhsh is not.
> At the very least some other unknown factor or factors must be
> operating here in conjunction with the carry laws and it may be that
> if that factor is identified why find that the laws had no effect.
I would assume that lots of factors are working here. Nobody in his
right mind claims that the specification of a model like this (Lott's,
or T&T or D&R) is a full and accurate description of what is going on.
The hope is only that it gets close enough so that doing statistics on
it yields at least some information.
It's easy enough to imagine local circumstances that would make such a
law much more or much less effective.
--
David Friedman
www.daviddfriedman.com/
Mary Rosh
I don't think that this discussion of the Plassman and Tideman piece
is at all remotely accurate. In Table 6, the year zero effect is .19
and by year 5 it is -.27. In all the years after year zero, the
coefficients values are lower after year zero.
Table 9 shows drops for Florida, Georgia, Idaho, MT, and Oregon and
increases for Mississippi and Virginia. Yet, Lott keeps on claiming
that one shouldn't look at the before and after averages and it looks
to me that these numbers confirm his claims. For all the states, the
robbery rate falls after the law is in effect. Even for Mississippi
and Virginia the drops are large: Ms. .21 to .07 and Va. .07 to -.07.
The increases in these two states are only in the before and after
averages.
James A. Donald
David Friedman <dd...@best.com> writes:
> > Perhaps you could point at where in the article he actually
> > does so? The point of the article is not that Lott makes any
> > mistakes in arithmetic but that he ought to have estimated a
> > more elaborate model.
On 14 Mar 2002 18:07:00 +1100, Tim Lambert
<lam...@cse.unsw.EDU.AU> wrote:
> I think that's what he meant. I.e. not that Lott add two and
> three and got four, but that he added two and three when he
> should have been multiplying two and three.
Yet, multiplying two and three is a most unusual procedure with
this kind of data, and leads to results that defy common sense.
Let us suppose we want to test whether throwing pigs out a ten
story window causes increased risk of sudden pig death. We throw
36 pigs out the window, and keep another 36. We find that the 36
that went through the window all died instantly, and the 36 that
remain all live. If we use Dezhakhsh's procedure, and consider
enough additional factors, we get the surprising result that there
is no significant correlation between falling ten stories, and
sudden death.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
2frPynpwL+B83Zuw1HJuuLGafIZVKdIZlSTLfI32
4PpRN56ugmSwnIax4hAOA4BSgqxkF/46nrWTMvU1V
James A. Donald
On 15 Mar 2002 02:23:59 +1100, Tim Lambert <lam...@cse.unsw.EDU.AU>
wrote:
You are correct. So I pasted the image into paintbrush, and measured
the locations to the exact pixel
The sequence is
219
98
317 23
75
392 18
57
449 17
40
489 19
21
510 15
6
516 5
1
517
Suspiciously close to a parabola, but I still see more than
measurement noise. Of course that might be inaccurate drawing. The
noise level, though not zero, is curiously low.
On other pages we see lots and lots of graphs with obvious measurement
noise, and they still tell the same story just as decisively.
If you are correct, these graphs are misleading, being a mere
dramatization of some magic numbers, not the eyeball confirmation of
these numbers that they purport to be.
But is the graph on page 102 misleading?
Tables 7, 8, and 9 of http://bingweb.binghamton.edu/~fplass/gun.pdf
tell the same story as these graphs. If these graphs are based on
magic numbers, the numbers are true nonetheless.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
fn3c0xfg1ZiXslo+ojNqVNU4N94NQcIgGTGZnHN+
4EnG6AdZe+OTwz6gU9vaUYtaOhRt52np24XrOWO7i
Tim Lambert
> In article <tny9gv4...@oolong.orchestra.cse.unsw.EDU.AU>,
> Tim Lambert <lam...@cse.unsw.EDU.AU> wrote:
>
> > > Perhaps you could point at where in the article he actually does so? The
> > > point of the article is not that Lott makes any mistakes in arithmetic
> > > but that he ought to have estimated a more elaborate model.
>
> > I think that's what he meant. I.e. not that Lott add two and three
> > and got four, but that he added two and three when he should have been
> > multiplying two and three.
>
> But that isn't what he says.
Adding when you should have multiplied is an error in arithmetic.
> All you have to do is read the whole
> passage you were quoting it from and compare it to the published article
> to see that the author of the passage is engaged in demagoguery, not
> trying to give an accurate report of what the article actually did.
I have read the whole passage. I cannot agree with you at all.
Dezhbaksh makes his purpose in writing the passage quite clear. He writes:
"Scholars, however, have pointed out several technical flaws in Lott's
work, but these criticisms, appearing as an academic squabble to
nonexperts, have not received much publicity."
He is trying to explain what is wrong with Lott's paper to
non-experts, so he is trying to avoid technical terms like "2SLS".
Now, let's go through it, and I'll try to translate back into jargon.
"Some of Lott's results also assume that
the number of arrests made by police does not depend on the number of
crimes committed! So rural counties with very few crimes may
presumably have more police arrests than urban counties with very
large crime rates."
Translation: Some of Lott's results use OLS instead of 2SLS.
"Moreover, Lott's central results are invalid because of errors in
computing expected arrest rates: he obtains mostly negative numbers
for arrests. For example, more than 19,000 of approximately 33,000
county-level auto theft arrests are "negative"; the number of
negative arrest rates for aggravated assault and property crimes are,
respectively, 9,900 and 13,500. What does a negative arrest rate
mean? Obviously, the number of individuals arrested for crimes can
only be zero or positive.
In the article D&R write:
]Their predicted arrest rates also include a large number of negative
]values (i.e., over 19,000 of about 33,000 observation for auto theft
]are negative; the number of negative arrest rates for aggravated
]assault and property crimes are, respectively, 9,900 and 13,500.
> > > Perhaps you can point to where in the article these claims are
> > > supported? I could not find it.
Done.
By the way, have you looked at
http://www.cse.unsw.edu.au/~lambert/guns/lott/figure4.7.html ?
The red curve shows what Lott's graph should have shown: an upturn in
rapes at the end. The black curve is what he actually published --
rapes level off at the end.
Do you agree that Lott has fudged his graph?
> > > I'm curious as to how much of the quote Paul Rubin would agree with.
>
> > In an earlier posting in this thread, Mike HAmmock wrote:
> > ] I've asked Prof. Rubin about what he
> > ] thinks policy makers ought to do. His response was something like
> > ] "Given that the results don't show a strong effect one way or the
> > ] other, I lean toward passing concealed-carry laws,". They don't seem
> > ] to do much good, but they don't do much harm, either, but since he
> > ] prefers more freedom to less, he'd prefer to see concealed carry laws
> > ] passed.
>
> The simplest reason to believe Rubin doesn't agree with the webbed
> comment is that his name isn't on it. It's particularly striking given
> that the page is on the Emory site, Rubin is at Emory, Deshbakhsh is not.
It seems to me that he does not agree with your claim that their paper
"says that concealed carry sometimes deters some crimes".
> > At the very least some other unknown factor or factors must be
> > operating here in conjunction with the carry laws and it may be that
> > if that factor is identified why find that the laws had no effect.
>
> I would assume that lots of factors are working here. Nobody in his
> right mind claims that the specification of a model like this (Lott's,
> or T&T or D&R) is a full and accurate description of what is going on.
> The hope is only that it gets close enough so that doing statistics on
> it yields at least some information.
>
> It's easy enough to imagine local circumstances that would make such a
> law much more or much less effective.
But there was a 37% decrease in robbery in one state and a 28%
increase in another. It's not just that it is less effective in some
places, it's that apparently it can cause either a large increase or a
large decrease.
Tim
Tim Lambert
> --
> David Friedman <dd...@best.com> writes:
> > > Perhaps you could point at where in the article he actually
> > > does so? The point of the article is not that Lott makes any
> > > mistakes in arithmetic but that he ought to have estimated a
> > > more elaborate model.
>
> On 14 Mar 2002 18:07:00 +1100, Tim Lambert
> <lam...@cse.unsw.EDU.AU> wrote:
> > I think that's what he meant. I.e. not that Lott add two and
> > three and got four, but that he added two and three when he
> > should have been multiplying two and three.
>
> Yet, multiplying two and three is a most unusual procedure with
> this kind of data, and leads to results that defy common sense.
>
> Let us suppose we want to test whether throwing pigs out a ten
> story window causes increased risk of sudden pig death. We throw
> 36 pigs out the window, and keep another 36. We find that the 36
> that went through the window all died instantly, and the 36 that
> remain all live. If we use Dezhakhsh's procedure, and consider
> enough additional factors, we get the surprising result that there
> is no significant correlation between falling ten stories, and
> sudden death.
Not true at all. An analogous procedure would be to consider if the
death rate for fat pigs was the same as for thin pigs. In your
example, it would be, and the Lott and Dezhakhsh approaches give the
same answer.
What is more interesting is if we are throwing pigs out of a lower
window and some survive, with the death rate for fat pigs thrown out
being higher. The Dezhakhsh approach gives the right answer, while
the Lott approach predicts that fat pigs who are NOT thrown out are
more likely to die.
Tim
Tim Lambert
> --
> On 15 Mar 2002 02:23:59 +1100, Tim Lambert <lam...@cse.unsw.EDU.AU>
> wrote:
>
> > jam...@echeque.com (James A. Donald) writes:
> >
> > > --
> > > On 14 Mar 2002 01:05:53 +1100, Tim Lambert <lam...@cse.unsw.EDU.AU>
> > > wrote:
> > >
> > > > jam...@echeque.com (James A. Donald) writes:
> > > >
> > > > > --
> > > > > Tim Lambert:
If your measurement error is +-1 pixel, then the second differences
are +-4. All (except the last one) are within 4 of 19.
The last one is different because Lott fudged his graph so that it
wouldn't show rapes increasing at the end.
which sows the almost perfect fit to parabola of Lott's graph.
>
> If you are correct, these graphs are misleading, being a mere
> dramatization of some magic numbers, not the eyeball confirmation of
> these numbers that they purport to be.
If one of Lott's critics had done this you would be jumping up and down
and accusing him of dishonesty. Why aren't you doing the same for Lott?
Tim
James A. Donald
Tim Lambert:
> > <lam...@cse.unsw.EDU.AU> wrote:
> > > I think that's what he meant. I.e. not that Lott add two
> > > and three and got four, but that he added two and three when
> > > he should have been multiplying two and three.
James A. Donald:
> > Yet, multiplying two and three is a most unusual procedure
> > with this kind of data, and leads to results that defy common
> > sense.
> >
> > Let us suppose we want to test whether throwing pigs out a ten
> > story window causes increased risk of sudden pig death. We
> > throw 36 pigs out the window, and keep another 36. We find
> > that the 36 that went through the window all died instantly,
> > and the 36 that remain all live. If we use Dezhakhsh's
> > procedure, and consider enough additional factors, we get the
> > surprising result that there is no significant correlation
> > between falling ten stories, and sudden death.
Tim Lambert:
> Not true at all. An analogous procedure would be to consider if
> the death rate for fat pigs was the same as for thin pigs. In
> your example, it would be, and the Lott and Dezhakhsh approaches
> give the same answer.
If, however, we consider 76 other attributes in parallel with
fatness, the MAINSTREAM approach used by Lott, and the unusual and
idiosyncratic approach used by Dezhakhsh give very different
answers.
--digsig
James A. Donald
6YeGpsZR+nOTh/cGwvITnSR3TdzclVpR0+pr3YYQdkG
CyEWQO83JdC0nui6lrpNfJagkERok21ltHTQdhYg
4iAkJiVi30UxRTBUwVJTc/bNdQ5w3cIuRaf7l/FA0