 Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.

# st: Re: interpret dydx as probabilities after logistic or xtlogit

 From "Seyi Soremekun" To Subject st: Re: interpret dydx as probabilities after logistic or xtlogit Date Mon, 11 Jul 2011 12:25:42 +0100

```Hi All,

I have a model with a categorical (binary) by continuous interaction, which I estimate using logistic regression or RE logistic regression thus:

logistic infacility i.intvn##c.quintile
xtlogit infacility i.intvn##c.quintile, re i(zone) or

I would like to estimate the change in probability (not odds) of my outcome "infacility" for a unit change in quintile, at intvn=0 and intvn=1, i.e. explain the interaction by looking at the actual estimated slopes of quintile, separately for intvn=0 and 1

Are the slopes obtained via a simple post-estimation margins dydx calculation?:
margins, dydx(quintile) over(intvn) continuous post
or
margins, dydx(quintile) over(intvn) predict(pu0) continuous post
...where one then exponentiates the results to show the risk ratio for the unit change?

The reason I ask is that from graphing the raw proportions of infacility for each level of quintile separately for intvn=0 and 1 gives a very different picture to the results obtained from the above methods, even without the adjustment for clustering (i.e. via logistic model):

1. Logistic model
odds ratio for quintile (when intvn==0) = 0.4446739
dydx for quintile (over intvn when =0) = -0.1383194, exponentiated to 0.87

2. Random effects model
odds ratio for quintile (when intvn==0) = 0.6044537
dydx for quintile (over intvn when =0) = -.0948134, exponentiated to .91

3. raw data when intvn=0
change in probability (% reduction) of infacility at:
quintile 2/1 = 0.73
quintile 3/2 = 0.68
quintile 4/3 = 0.59
quintile 5/4 = 0.38

One can see that the average pecentage reduction (~0.60) in the probability of infacility from the raw data (3.) is much higher than either logistic model.
So:
1. Am I interpreting the dydx results from the logistic models in the correct fashion - as a percentage reduction in probability for a unit change in quintile (i.e. averaged risk ratio)?
2. If so any ideas why the model results look so different to the raw data - even in the simple logistic model without clustering adjustment?

thanks, Seyi

Seyi Soremekun
Faculty of Epidemiology and Public Health
London School of Hygiene and Tropical Medicine
London WC1E 7HT
+442079272464

*
*   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/
```