Compare predicted probabilities?

[Jaycob Applegate]

Hi all,

Say I run a logistic regression with a 3x3 categorical by categorical interaction. I then run

margins var1#var2 , predict(pr)

to get the predicted probability by each category. I am given confidence intervals for each prediction. However, I can I statistically test to see if predicted values are statistically different? Would I use average marginal effects?

[Clyde Schechter]

margins var1#var2, pwcompare

and see the help for options to customize the output.

[Jaycob Applegate]

Thank you for this! Two follow-up questions:
1. Is this pairwise comparison different from checking average marginal effects? For example, does the following code tell me something different?

 margins var1, dydx (var2) predict(pr)

2. Can this be executed with the 'mimrgns' package, as I have multiply imputed data? I have taken a stab at it, but get the following error:

nothing found where name expected
st_matrixcolstripe_term_size():  3300  argument out of range
  _b_stat::fill_mcinfo():     -  function returned error
      _b_stat::compute():     -  function returned error
           _coef_table():     -  function returned error
                 <istmt>:     -  function returned error

[Clyde Schechter]

If var2 is a dichotomous variable, -margins var2, pwcompare(effects)- will produce the exact same results as -margins, dydx(var2)-.

I can't comment on -mimrgns- as I have only minimal experience with it. I know it does not have a -pwcompare- option. Nor does it have a -dydx()- option. But I believe there is a way to get contrasts of multiple imputation results for logistic regression coefficients in some other way. But I have not done that myself and don't remember what it is. Others on this Forum are more familiar with -mimrgns- and can probably provide a more helpful answer.

[Jaycob Applegate]

Thanks for all the help! I have used dydx successfully with mimrgns I should note. I welcome anyone else with experience using -mimrgns-!