I'd like to know if there's a suitable epidemiological test that could
use data in this format in order to determine whether or not the rate
of increase in health spending slows down in the years after smoking
legislation has passed (the Law Year on the left of the table)...

On top of that, I'd also like to know if (and then how) I would need
to account for the population size of each state...

St.     Law Year        1997    1998    1999    2000    2001    2002    2003    2004    2005    2006    2007
1       None            0%      3%      3%      4%      1%      1%      2%      1%      4%      3%      2%
2       1998            2%      4%      5%      1%      1%      3%      1%      2%      1%      4%      2%
3       None            1%      3%      3%      0%      3%      4%      3%      2%      4%      1%      1%
4       None            2%      3%      2%      4%      5%      1%      4%      4%      0%      4%      3%
5       2003            3%      4%      3%      3%      1%      1%      2%      3%      2%      4%      5%
6       None            1%      3%      4%      4%      1%      4%      2%      0%      3%      5%      4%
7       2005            1%      2%      3%      2%      2%      5%      0%      4%      3%      5%      1%
8       2006            0%      4%      2%      1%      4%      3%      1%      3%      2%      3%      5%
9       None            4%      3%      1%      1%      3%      5%      1%      2%      3%      5%      2%
10      2001            1%      3%      0%      4%      3%      0%      4%      1%      5%      3%      2%

Any advice on how I might begin to analyze data like this would be
tremendously appreciated.

Thanks,
  Anthony Damico

Your outcome is the % increase.  Theoretically, I suppose costs could
decrease, but I'm guessing it's pretty rare.  Anyway, suffice to say
that you might consider a log transform or other sort of transform.
You also might try to see if you can get numbers, rather than %
increase per year.  And if you get those, then you might also do a log
transformm.  Anyway, all I'm suggesting is that whatever outcome you
regress (or do whatever) on, you should do some serious EDA to see if
the assumptions of regression (or whatever model) are met.

Having said that, let's say you nailed down a continuous outcome
variable and you're calling it outc.  I'd start by getting the data
long, like this:

st      lawyr           year            outc
AK      none            2001            0.01
AK      none            2002            0.03
AK      none            2003            0.02
AL      2004            2001            0.06
AL      2004            2002            0.05

etc.  We have one record per state per year, so with 8 years of data
you'd have 400 records.

Next I'd generate a variable "law" which is 1 if and only if smoking
legislation is in effect that year:

st      lawyr           year            outc            law
AK      none            2001            0.01            0
AK      none            2002            0.03            0
AL      2004            2003            0.06            0
AL      2004            2004            0.05            0
AL      2004            2005            0.05            1
AL      2004            2006            0.02            1

etc.

Next, I'd do regression, accounting for clustering at the state level,
with indicator variables for the year.  If you're using Stata, it
might look like this:

xi: glm outc i.year law, vce(cluster st) family(whatever) link
(whatever)

[By "whatever" I don't mean that you should just put whatever you
feel, but you should decide based on the distribution of the outc
variable.]

You'll get back output with a coefficient for each year* (indicating
the overall change in "outc" that year from the previous year in a
state with no smoking law), and also a coefficient for law (indicating
the effect of a smoking law on outc per year).  If the coefficient for
"law" is statistically significantly negative (or < 1 if you're using
some sort of multiplicative model), then smoking laws are good.  If
the coefficient is statistically significantly positive, then they're
bad.

(*Actually I think you won't get a constant for the first year.  The
constant term will represent year-on-year increase for the first year,
and coefficients for years will be the difference between the first
year and each other year.  If you add the option "noconst" you might
get around this.)

An advantage of this method is that it doesn't assume anything about
the overall distribution of cost increases.  For example, you can
allow costs to shoot up from 2003 to 2004, then stay constant from
2004 to 2005.

On Sep 14, 12:10 pm, Anthony <ajdam...@gmail.com> wrote: