I am reminded of a famous Ripleyism (his was in response to stepwise regression)
“The only reason to do this is to show why it is wrong”
But maybe one of us will come up with some exception!
Peter
--
--
To post a new thread to MedStats, send email to MedS...@googlegroups.com .
MedStats' home page is http://groups.google.com/group/MedStats .
Rules: http://groups.google.com/group/MedStats/web/medstats-rules
---
You received this message because you are subscribed to the Google Groups "MedStats" group.
To unsubscribe from this group and stop receiving emails from it, send an email to medstats+u...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.
Confidentiality Statement:
This email message, including any attachments, is for the sole use of the intended recipient(s) and may contain confidential and privileged information. Any unauthorized use, disclosure or distribution is prohibited. If you are not the intended recipient, please contact the sender by reply email and destroy all copies of the original message.
Diana,
Floor effects are not the problem here. Z-transforms are not appropriate for ratios. The correct approach, usually, is to take logs of ratios and to operate on those, then later anti-log to get fold change.
Arithmetic means can be appropriate for proportions, not for ratios in general. If you know of cases where arithmetic means of two ratios is appropriate, where the ratios are not proportions and none of the numerators and denominators is a fixed literal constant, please post it.
Frank
On Friday, May 10, 2013 9:17:49 AM UTC-5, d.e.ko...@herts.ac.uk wrote:
No reason why you shouldn’t take arithmetic mean of ratios or proportions.
BUT if the reason you are doing it is to predict what another similar group of entities will do next time, it may be more accurate to 1st transform to z-scores, take mean of z-scores & then back trasnofr to the proportion
WHY?
Because ratios have floor and ceiling effect as they approach 0 or 1.
It is much easier to increase % correct in an exam from 505%ot 55% than from 90% to 95%
It is much easier to increase % buing product after viewing ad from 505% to 55% than from 90% to 95%
Etc.
Best
diana
On 10/05/2013 15:05, "John Sorkin" <jso...@grecc.umaryland.edu <http://jso...@grecc.umaryland.edu> > wrote:
Prof. Harrell,
At the risk of appearing unknowledgeable, can you give me an example, or reason why it is not legitimate to compute the arithmetic mean of ratios or percent change? Additionally, what is the correct method to use to get a summary measure of these statistics? I don't want to do that which is not correct.
John
P.S. In asking this question, I take courage from the aphorism that the only stupid question is the unasked question,
>>> Frank Harrell <harr...@gmail.com <http://harr...@gmail.com> > 5/10/2013 9:40 AM >>>
Hi Steve,
I'm looking for examples where the elements of the ratios are not fixed constants. No one has thought of one yet - I hope to get some more responses.
Frank
On Friday, May 10, 2013 8:26:49 AM UTC-5, SteveDrD wrote:
And I can think of exactly one trivial example (although it may fall into the sum of proportions caveat). Suppose one of the ratios was equal to zero. Would the sum be meaningful?
Steve Denham
On Friday, May 3, 2013 2:16:57 PM UTC-4, plf515 wrote:
I am reminded of a famous Ripleyism (his was in response to stepwise regression)
“The only reason to do this is to show why it is wrong”
But maybe one of us will come up with some exception!
Peter
From: meds...@googlegroups.com <http://meds...@googlegroups.com> [mailto:meds...@googlegroups.com] On Behalf Of Frank Harrell
Sent: Friday, May 03, 2013 1:08 PM
Subject: {MEDSTATS} Example where adding ratios is correct?
I am having an argument with a journal that has published several papers where authors have created a prognostic score by anti-logging Cox proportional hazards model regression coefficients and adding the hazard ratios. This is wrong on so many levels that it is difficult to fathom. Ewout Steyerberg has pointed out that with this scheme, a protective factor will be scored as harmful upon anti-logging its negative regression coefficient.
One thing I'd like to say to the journal is that there is no example where adding ratios is appropriate when (1) the ratios do not represent the ratios of parts to a whole (i.e., proportions) and (2) the numerators and denominators are both variables (i.e., not fixed constants). It is OK to add proportions, and we have variance formulas where you take 1/m + 1/n, but the numerators are fixed constants in the latter.
Can anyone think of an example in biology, medicine, statistics, physics, or any other field (subject to the two exclusions above) where adding ratios actually works? I note in passing that it is not legitimate to compute the arithmetic mean of ratios or percent change - that's why we have geometric means and why we average logs then anti-log to get fold change.
Frank
Emeritus Professor Diana Kornbrot
Assume we have two subjects one with an initial systolic blood pressure of 200 mmHg. which after treatment decreases to 150 mmHg, a 25% decrease. The second subject has an initial systolic blood pressure of 100 mmHg and after treatment has a systolic blood pressure of 75 mmHg, also a 25% decrease. Other than being concerned that I have lowered the second subject's systolic blood pressure too much, and I not correct to say that the average percentage decrease brought about by treatment was 25%?
John
>>> Frank Harrell <harr...@gmail.com> 5/10/2013 10:57 AM >>>
Let me turn it around and ask you for an example where it is valid to do so, with specifics.
Details about problems caused by percent change are at http://biostat.mc.vanderbilt.edu/ManuscriptChecklist - look for Measures of Change.
For Cox and logistic models the simple answer is that the regression coefficients were derived using maximum likelihood for a model in which the effects were additive on the log scale, not multiplicative on the log scale. So adding antilogs gives the wrong estimate of risk and is inconsistent with how the betas were estimated.
Frank
On Friday, May 10, 2013 9:05:32 AM UTC-5, John Sorkin wrote:
Prof. Harrell,
At the risk of appearing unknowledgeable, can you give me an example, or reason why it is not legitimate to compute the arithmetic mean of ratios or percent change? Additionally, what is the correct method to use to get a summary measure of these statistics? I don't want to do that which is not correct.
John
P.S. In asking this question, I take courage from the aphorism that the only stupid question is the unasked question,
>>> Frank Harrell <harr...@gmail.com <javascript:> > 5/10/2013 9:40 AM >>>
Hi Steve,
I'm looking for examples where the elements of the ratios are not fixed constants. No one has thought of one yet - I hope to get some more responses.
Frank
On Friday, May 10, 2013 8:26:49 AM UTC-5, SteveDrD wrote:
And I can think of exactly one trivial example (although it may fall into the sum of proportions caveat). Suppose one of the ratios was equal to zero. Would the sum be meaningful?
Steve Denham
On Friday, May 3, 2013 2:16:57 PM UTC-4, plf515 wrote:
I am reminded of a famous Ripleyism (his was in response to stepwise regression)
“The only reason to do this is to show why it is wrong”
But maybe one of us will come up with some exception!
Peter
From: meds...@googlegroups.com [mailto:meds...@googlegroups.com] On Behalf Of Frank Harrell
Sent: Friday, May 03, 2013 1:08 PM
To: meds...@googlegroups.com
Subject: {MEDSTATS} Example where adding ratios is correct?
I am having an argument with a journal that has published several papers where authors have created a prognostic score by anti-logging Cox proportional hazards model regression coefficients and adding the hazard ratios. This is wrong on so many levels that it is difficult to fathom. Ewout Steyerberg has pointed out that with this scheme, a protective factor will be scored as harmful upon anti-logging its negative regression coefficient.
One thing I'd like to say to the journal is that there is no example where adding ratios is appropriate when (1) the ratios do not represent the ratios of parts to a whole (i.e., proportions) and (2) the numerators and denominators are both variables (i.e., not fixed constants). It is OK to add proportions, and we have variance formulas where you take 1/m + 1/n, but the numerators are fixed constants in the latter.
Can anyone think of an example in biology, medicine, statistics, physics, or any other field (subject to the two exclusions above) where adding ratios actually works? I note in passing that it is not legitimate to compute the arithmetic mean of ratios or percent change - that's why we have geometric means and why we average logs then anti-log to get fold change.
Frank
Assume we have two subjects one with an initial systolic blood pressure of 200 mmHg. which after treatment decreases to 150 mmHg, a 25% decrease. The second subject has an initial systolic blood pressure of 100 mmHg and after treatment has a systolic blood pressure of 75 mmHg, also a 25% decrease. Other than being concerned that I have lowered the second subject's systolic blood pressure too much, and I not correct to say that the average percentage decrease brought about by treatment was 25%?
The numbers used in my example were chosen simply for simplicity of exposition. They were not chosen to represent a true experiment.
My question remains, am I wrong in stating that the drug, on average, results in a 25% decrease in blood pressure?
I wonder if the definition of “ratio” or lack thereof may be causing some confusion here.
When I first read Frank’s question my thought was that a ratio just meant a fraction and I thought of a few different examples where adding proportions (with equal, fixed denominators) made sense. But reading further I believe that Frank is talking about epidemiologic ratios which are a division of 2 rates or odds and important from a statistical view is that at least one piece in each of the numerator and denominator is a random variable. For example, the proportion of smokers who develop lung cancer is a rate (and a fraction and a proportion) but is not a ratio. The proportion of smokers who develop lung cancer divided by the proportion of non-smokers who develop lung cancer is a ratio.
I don’t think that the change in blood pressure example (and other examples so far) meet that definition of ratio that Frank is interested in.
From the statistics/random variable approach it make sense to work with ratios on the log scale because taking the log turns those pesky divisions into subtractions and dealing with the difference between 2 random variables is simple, the ratio of 2 random variables is not so simple.
If I remember my elementary school math correctly, you can only add 2 fractions when the denominator is exactly equal. So adding 2 rates from the same population (therefore with the same, fixed denominator) can be meaningful. But if there is a random variable in the denominator then can the denominators ever be considered to be equal?
Even if we have a case where numerically the denominators are equal, for example we have an odds ratio (or hazard ratio, or risk ratio) for heavy smokers developing lung cancer compared to non-smokers and we also have an odds ratio for light smokers developing lung cancer compared to non-smokers then we have the rate of lung cancer in non-smokers as the denominator in both cases. Can we add those 2 odds ratios together? Would that be the odds ratio smokers (either heavy or light) developing lung cancer compared to non-smokers? It would be paradoxical if smokers (undetermined amount) had a higher odds ratio than either heavy or light smokers. It could be considered the risk of being both a light smoker and a heavy smoker at the same time (even though that is impossible) but that would still require additivity (no interaction) on the non-logged scale. Averaging might make sense here, but there is still the main issue of weights in the averaging.
So my 2 cents worth is that I cannot think of a meaningful case of adding ratios (epidemiology definition) on the non-logged scale.
--
Gregory (Greg) L. Snow Ph.D.
Statistical Data Center
Intermountain Healthcare
--
I don’t know about “best” but it seems to me that you would get good information from a Cronbach’s alpha, or, rather, 4 alphas – one for each study
Peter
Peter Flom
My web site: http://www.statisticalanalysisconsulting.com/
Linked in: http://www.linkedin.com/in/peterflom
Twitter: @PeterFlomStat
From: meds...@googlegroups.com [mailto:meds...@googlegroups.com] On Behalf Of John Sorkin
Sent: Monday, June 03, 2013 8:32 AM
To: meds...@googlegroups.com
Subject: {MEDSTATS} assessing agreement with multiple observers, data on ratio scale
Colleagues,
--
--
--
But one comment – I doubt that you have “ratio” level data. You may have a “rational” zero, but that does not make it ratio-level data.
| ----------------------------------------------------------------------- Confidentiality Notice: This e-mail message, including any attachments, |
is for the sole use of the intended recipient(s) and may contain |
| privileged and confidential information. Any unauthorized review, use, |
disclosure or distribution is prohibited. If you are not the intended |
| recipient, please contact the sender by reply e-mail and destroy |