0

Regarding data, there are three binary variables: trt(0,1), fail(0,1), and female(0,1).

For sub-group analysis (male vs. female), I am running typical 2x2 treatment vs. failure tables separately for males and females. These tables yield gender-specific failure ORs, which are

Male: OR = 0.75(.46,1.24)
      p-value 0.26

Female: OR = 0.38(.16,.92)
        p-value 0.03

So, one of the sub-groups was identified to show a significant treatment effect.

If the male and female univariate models and interaction model are run using a logit model, you can see that the interaction term is not signif (P=0.184), but if the constant is left out the interaction term is significant - however, that interaction is biased by the constant term - obviously. Several collaborators like the last logit model without the constant term, since it yields a significant interaction term -- like the female sub-group analysis. However, I believe that it's an erroneous assumption to assume that the interaction p-value would be significant as long as one of the sub-groups has a significant treatment effect? I actually favor the interaction model with the constant term, since the slope difference between the male and female treatment effects (univariate models) can be discerned.

. logit fail trt if male==1



    fail |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]



-------------+----------------------------------------------------------------
         trt |   -.285727   .2535624    -1.13   0.260    -.7827001     .211246
       _cons |   .4192584   .1858278     2.26   0.024     .0550427    .7834742




. logit fail trt if female==1




    fail |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]



-------------+----------------------------------------------------------------
         trt |  -.9751312   .4522192    -2.16   0.031    -1.861464    -.088798
       _cons |   .7339692   .3511885     2.09   0.037     .0456524    1.422286




. logit fail trt female trtfem




    fail |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]



-------------+----------------------------------------------------------------
         trt |   -.285727   .2535624    -1.13   0.260    -.7827001     .211246(Male trt)
      female |   .3147107   .3973227     0.79   0.428    -.4640274    1.093449(Female const - Male const)
      trtfem |  -.6894042   .5184554    -1.33   0.184    -1.705558    .3267498(Female trt - Male trt)-->Interaction
       _cons |   .4192584   .1858278     2.26   0.024     .0550427    .7834742(Male const)




. logit fail trt female trtfem, nocon




    fail |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]



-------------+----------------------------------------------------------------
         trt |   .1335314   .1725164     0.77   0.439    -.2045945    .4716573(Const + trt, 0.419 - 0.286)
      female |   .7339692   .3511885     2.09   0.037     .0456524    1.422286(Const + Female trt, 0.419 + 0.315)
      trtfem |  -1.108663   .4840083    -2.29   0.022    -2.057302   -.1600237(Interaction - const, -0.689 - 0.419)
  • 2
    Just because the effect for males is nonsignificant and the effect for females is doesn't mean that the treatment has a larger effect for females than for males. To draw that conclusion, one must accept the null that the true effect for males = 0 (odds ratio of 1). The interaction is testing whether the difference between the log of .75 differs from the log of .38. It does not. You cannot conclude that there is a differential treatment effect. Removing the intercept does not make any sense. It needs to be in the model. – dbwilson Jul 14 '20 at 20:36
  • 1
    Interactions in non-linear models can be tricky. See this post for the continuous variable example. Is there any reason not to use robust regression here since the model is saturated? – dimitriy Jul 14 '20 at 22:24
  • I don't really follow the logic behind why the constant leads to bias. – dimitriy Jul 14 '20 at 22:25
  • @Dimitry - the Male-only model intercept value is subtracted from the true interaction value when the intercept is dropped from the interaction model. That is, it's not a true delta of slopes between males and females -- it's biased. But your point is well taken about non-linearity. –  Jul 14 '20 at 22:41
  • 2
    Agreed. Another way to think about it is that with no intercept, the intercept is forced to zero. In this model, that translates into forcing the logit to zero for the no treatment and male condition (value of zero for the independent variables). A logit of zero is a failure rate of .5. Thus, without the intercept, you are fixing the failure rate for the male/no treatment group at .5 no matter what its observed value is. – dbwilson Jul 14 '20 at 23:38

1 Answers1

0

I find it really hard to think about coefficients on the index function scale and translate that to things that I ultimately care about, like probabilities or ORs. This is especially true for interactions. So I would do something like this, which leads to identical conclusions any way you do it (as long as the model stays fully saturated). Perhaps you will find it useful. I've omitted explanations since everything is so similar to the linear case, and I am just doing the equivalent comparison for the nonlinear logit models. I start with OLS, then combined logit, and then subsample logits.

. #delimit;
delimiter now ;
. sysuse auto, clear;
(1978 Automobile Data)

. gen high_mpg = mpg>22;


. gen high_price = price>6000;


. reg foreign i.high_mpg##i.high_price, robust;


Linear regression                               Number of obs     =         74
                                                F(3, 70)          =       8.78
                                                Prob > F          =     0.0001
                                                R-squared         =     0.2495
                                                Root MSE          =     .40711




                |               Robust
        foreign |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]



--------------------+----------------------------------------------------------------
         1.high_mpg |   .4032258   .1272598     3.17   0.002     .1494142    .6570374
       1.high_price |   .1385199   .1190367     1.16   0.249    -.0988913    .3759312
                    |
high_mpg#high_price |
               1 1  |   .1948134   .2277168     0.86   0.395    -.2593534    .6489802
                    |
              _cons |   .0967742   .0545964     1.77   0.081    -.0121149    .2056633




. margins high_mpg#high_price;


Adjusted predictions                            Number of obs     =         74
Model VCE    : Robust


Expression   : Linear prediction, predict()




                |            Delta-method
                |     Margin   Std. Err.      t    P>|t|     [95% Conf. Interval]



--------------------+----------------------------------------------------------------
high_mpg#high_price |
               0 0  |   .0967742   .0545964     1.77   0.081    -.0121149    .2056633
               0 1  |   .2352941   .1057779     2.22   0.029     .0243267    .4462616
               1 0  |         .5   .1149534     4.35   0.000     .2707327    .7292673
               1 1  |   .8333333   .1564318     5.33   0.000       .52134    1.145327




. margins high_price, dydx(high_mpg);


Conditional marginal effects                    Number of obs     =         74
Model VCE    : Robust


Expression   : Linear prediction, predict()
dy/dx w.r.t. : 1.high_mpg




         |            Delta-method
         |      dy/dx   Std. Err.      t    P>|t|     [95% Conf. Interval]



-------------+----------------------------------------------------------------
0.high_mpg   |  (base outcome)
-------------+----------------------------------------------------------------
1.high_mpg   |
  high_price |
          0  |   .4032258   .1272598     3.17   0.002     .1494142    .6570374
          1  |   .5980392   .1888382     3.17   0.002     .2214133    .9746652




Note: dy/dx for factor levels is the discrete change from the base level.


. margins r.high_price, dydx(high_mpg);


Contrasts of conditional marginal effects       Number of obs     =         74
Model VCE    : Robust


Expression   : Linear prediction, predict()
dy/dx w.r.t. : 1.high_mpg




         |         df           F        P>F



-------------+----------------------------------
0b.high_mpg  |
  high_price |  (not testable)
-------------+----------------------------------
1.high_mpg   |
  high_price |          1        0.73     0.3952
             |
 Denominator |         70






         |   Contrast Delta-method
         |      dy/dx   Std. Err.     [95% Conf. Interval]



-------------+------------------------------------------------
0.high_mpg   |  (base outcome)
-------------+------------------------------------------------
1.high_mpg   |
  high_price |
   (1 vs 0)  |   .1948134   .2277168     -.2593534    .6489802




Note: dy/dx for factor levels is the discrete change from the
      base level.


. logit foreign i.high_mpg##i.high_price, nolog;


Logistic regression                             Number of obs     =         74
                                                LR chi2(3)        =      18.67
                                                Prob > chi2       =     0.0003
Log likelihood = -35.697459                     Pseudo R2         =     0.2073




        foreign |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]



--------------------+----------------------------------------------------------------
         1.high_mpg |   2.233592   .7543524     2.96   0.003     .7550886    3.712096
       1.high_price |   1.054937   .8342486     1.26   0.206      -.58016    2.690034
                    |
high_mpg#high_price |
               1 1  |   .5545007   1.447747     0.38   0.702    -2.283031    3.392032
                    |
              _cons |  -2.233592   .6074929    -3.68   0.000    -3.424256   -1.042928




. margins high_mpg#high_price;


Adjusted predictions                            Number of obs     =         74
Model VCE    : OIM


Expression   : Pr(foreign), predict()




                |            Delta-method
                |     Margin   Std. Err.      z    P>|z|     [95% Conf. Interval]



--------------------+----------------------------------------------------------------
high_mpg#high_price |
               0 0  |   .0967742   .0531003     1.82   0.068    -.0073005    .2008489
               0 1  |   .2352941   .1028794     2.29   0.022     .0336543     .436934
               1 0  |         .5   .1118034     4.47   0.000     .2808694    .7191306
               1 1  |   .8333333   .1521452     5.48   0.000     .5351343    1.131532




. margins high_price, dydx(high_mpg);


Conditional marginal effects                    Number of obs     =         74
Model VCE    : OIM


Expression   : Pr(foreign), predict()
dy/dx w.r.t. : 1.high_mpg




         |            Delta-method
         |      dy/dx   Std. Err.      z    P>|z|     [95% Conf. Interval]



-------------+----------------------------------------------------------------
0.high_mpg   |  (base outcome)
-------------+----------------------------------------------------------------
1.high_mpg   |
  high_price |
          0  |   .4032258   .1237725     3.26   0.001     .1606361    .6458155
          1  |   .5980392   .1836636     3.26   0.001     .2380652    .9580132




Note: dy/dx for factor levels is the discrete change from the base level.


. margins r.high_price, dydx(high_mpg);


Contrasts of conditional marginal effects       Number of obs     =         74
Model VCE    : OIM


Expression   : Pr(foreign), predict()
dy/dx w.r.t. : 1.high_mpg




         |         df        chi2     P>chi2



-------------+----------------------------------
0b.high_mpg  |
  high_price |  (omitted)
-------------+----------------------------------
1.high_mpg   |
  high_price |          1        0.77     0.3791






         |   Contrast Delta-method
         |      dy/dx   Std. Err.     [95% Conf. Interval]



-------------+------------------------------------------------
0.high_mpg   |  (base outcome)
-------------+------------------------------------------------
1.high_mpg   |
  high_price |
   (1 vs 0)  |   .1948134   .2214768     -.2392731    .6288999




Note: dy/dx for factor levels is the discrete change from the
      base level.


. logit foreign i.high_mpg if high_price == 0, nolog;


Logistic regression                             Number of obs     =         51
                                                LR chi2(1)        =      10.46
                                                Prob > chi2       =     0.0012
Log likelihood = -23.718984                     Pseudo R2         =     0.1807




 foreign |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]



-------------+----------------------------------------------------------------
  1.high_mpg |   2.233592   .7543524     2.96   0.003     .7550885    3.712096
       _cons |  -2.233592   .6074929    -3.68   0.000    -3.424256   -1.042928




. margins high_mpg;


Adjusted predictions                            Number of obs     =         51
Model VCE    : OIM


Expression   : Pr(foreign), predict()




         |            Delta-method
         |     Margin   Std. Err.      z    P>|z|     [95% Conf. Interval]



-------------+----------------------------------------------------------------
    high_mpg |
          0  |   .0967742   .0531003     1.82   0.068    -.0073005    .2008489
          1  |         .5   .1118034     4.47   0.000     .2808694    .7191306




. margins, dydx(high_mpg);


Conditional marginal effects                    Number of obs     =         51
Model VCE    : OIM


Expression   : Pr(foreign), predict()
dy/dx w.r.t. : 1.high_mpg




         |            Delta-method
         |      dy/dx   Std. Err.      z    P>|z|     [95% Conf. Interval]



-------------+----------------------------------------------------------------


1.high_mpg |   .4032258   .1237725     3.26   0.001     .1606361    .6458155


Note: dy/dx for factor levels is the discrete change from the base level.


. logit foreign i.high_mpg if high_price == 1, nolog;


Logistic regression                             Number of obs     =         23
                                                LR chi2(1)        =       6.83
                                                Prob > chi2       =     0.0090
Log likelihood = -11.978475                     Pseudo R2         =     0.2219




 foreign |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]



-------------+----------------------------------------------------------------
  1.high_mpg |   2.788093   1.235687     2.26   0.024     .3661903    5.209995
       _cons |  -1.178655   .5717719    -2.06   0.039    -2.299307   -.0580027




. margins high_mpg;


Adjusted predictions                            Number of obs     =         23
Model VCE    : OIM


Expression   : Pr(foreign), predict()




         |            Delta-method
         |     Margin   Std. Err.      z    P>|z|     [95% Conf. Interval]



-------------+----------------------------------------------------------------
    high_mpg |
          0  |   .2352941   .1028794     2.29   0.022     .0336543     .436934
          1  |   .8333333   .1521452     5.48   0.000     .5351343    1.131532




. margins, dydx(high_mpg);


Conditional marginal effects                    Number of obs     =         23
Model VCE    : OIM


Expression   : Pr(foreign), predict()
dy/dx w.r.t. : 1.high_mpg




         |            Delta-method
         |      dy/dx   Std. Err.      z    P>|z|     [95% Conf. Interval]



-------------+----------------------------------------------------------------


1.high_mpg |   .5980392   .1836636     3.26   0.001     .2380652    .9580132


Note: dy/dx for factor levels is the discrete change from the base level.

I think there might be a way to use suest to combine the two subsample logit models and compare the cross-equation marginal effects, but I am not sure how to do that immediately.

You can also get OR results like this:

. logit foreign i.high_mpg##i.high_price, or nolog;

Logistic regression                             Number of obs     =         74
                                                LR chi2(3)        =      18.67
                                                Prob > chi2       =     0.0003
Log likelihood = -35.697459                     Pseudo R2         =     0.2073




        foreign | Odds Ratio   Std. Err.      z    P>|z|     [95% Conf. Interval]



--------------------+----------------------------------------------------------------
         1.high_mpg |   9.333333   7.040623     2.96   0.003       2.1278    40.93952
       1.high_price |   2.871795   2.395791     1.26   0.206     .5598088    14.73218
                    |
high_mpg#high_price |
               1 1  |   1.741071   2.520631     0.38   0.702     .1019747     29.7263
                    |
              _cons |   .1071429   .0650885    -3.68   0.000     .0325735    .3524213




Note: _cons estimates baseline odds.

All these models indicate that the effect of high MPG is not moderated by heaviness to a significant degree.

dimitriy
  • 35,430