--- Christy McKinney <[email protected]> wrote:
> I am examining the association between binge drinking and
> neighborhood poverty using survey data. Because binge drinking is
> highly prevalent I would like to use glm to estimates relative risks
> (instead of logistic regression & odds ratios).
Why is high prevalence an argument against logistic regression? The
only argument I have heard is that in those cases the odds ratios no
longer closely resemble risk ratios, but that is a fallacy: Odds ratios
are intended to measure the ratio of odds, and should be interpreted
that way. The fact that they no longer approximate something they
weren't suppose to approximate anyhow is not a valid argument. Odds
ratios are easy to interpret once you stop trying to interpret them in
metrics they were not made for, like risk differences or risk ratios.
The odds, the expected number successes for every failure, is a measure
for the likelihood of an event that is different, but equally easy, as
the probability. The ratio of odds is a simple way of showing how much
the likelihood of an even differs between groups. That is all you need
when interpreting odds ratios, any reference to risk ratios will only
complicate things.
Do you have multiple observations in the same neighborhood? In that
case observations are nested within neighborhood and you need to take
this multilevel structure into account.
> Past year binge drinking is categorized into 3 groups: no binge
> (referent); binge <1 month; and binge >=1 month. In one glm model, I
> compare binge <1 month (binge2lt) to no binge drinking. In a separate
> glm model, I compare binge >=1 month (binge2gt) to no binge
> drinking.
A more appropriate model would take into account the ordinal nature of
your data. Take a look at -help ologit- and -ssc describe gologit2-. If
you have multilevel data you should probably use -gllamm-, see -ssc
describe gllamm- and www.gllamm.org .
> I want to evaluate whether the risk of binge drinking associated with
> neighborhood poverty is different for men and women (e.g. does sex
> modify the relation between binge drinking and neighborhood
> poverty?). I am new to the suest command and am not sure I am using
> properly. Can I use the suest command to combine the two separate
> models and test the interaction jointly across binge drinking
> categories?
You can, but you shouldn't. Just add the interaction like you would in
any other situation, create the interaction term (e.g. -gen povXfem =
poverty*female -) and add the interaction term into your analysis.
There is a big problem with this, as is discussed here:
http://www.nd.edu/~rwilliam/oglm/index.html . However, I have not seen
a truly convincing solution to it, most solutions are very smart but
way too fragile for my taste.
Hope this helps,
Maarten
-----------------------------------------
Maarten L. Buis
Department of Social Research Methodology
Vrije Universiteit Amsterdam
Boelelaan 1081
1081 HV Amsterdam
The Netherlands
visiting address:
Buitenveldertselaan 3 (Metropolitan), room N515
+31 20 5986715
http://home.fsw.vu.nl/m.buis/
-----------------------------------------
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
