Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • #16
    Thanks guys. I am not sure if gsem can be used for continuous variables. I tried both sem and gsem, they gave me slightly different results.

    Suppose I could use svy:regress, but all my models were either sem (for continuous outcomes) or gsem (for binary or categorical outcomes), just want to be consistent so it's easier to explain.

    Comment


    • #17
      -gsem- is fine for continuous variables. Slight differences from -sem- results may be attributable to differences in the maximization algorithms used internally. But the differences should be minor. If they are not, then probably the code for the two models was not the same. In fact, -gsem- can estimate any model that -sem- can. (The main advantage of -sem- is that it is a bit simpler to code, and it offers some fit and other statistics that -gsem- doesn't.)

      Comment


      • #18
        Thank you so much guys! Can I ask one more question, please?

        I am a bit confuse of whether or not there is a significant moderation effect of sp_ss_friend (my moderator) or not. My gsem logistic model showed significant interaction effect but margin showed overlapped 95% CI.


        Code:
        svy:gsem (sm_ptsd_categories <- sp_vvgroup)(sp_prodrink <- i.sm_ptsd_categories##i.sp_ss_friendi.sp_vvgroup##i.sp_ss_friend `control1') , logit
        Code:
        ---------------------------------------------------------------------------------------------
                                    |             Linearized
                                    |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
        ----------------------------+----------------------------------------------------------------
        sm_ptsd_categories          |
                         sp_vvgroup |    2.14388   .1880505    11.40   0.000     1.775102    2.512659
                              _cons |  -2.577016   .1776783   -14.50   0.000    -2.925454   -2.228578
        ----------------------------+----------------------------------------------------------------
        sp_prodrink                 |
                 sm_ptsd_categories |
        PTSD is experienced (51..)  |   .0221354   .3993032     0.06   0.956    -.7609231    .8051938
                                    |
                       sp_ss_friend |
                   at least weekly  |  -1.617157   .7365257    -2.20   0.028    -3.061529   -.1727837
                                    |
                 sm_ptsd_categories#|
                       sp_ss_friend |
        PTSD is experienced (51..) #|
                   at least weekly  |   2.024679   .7222303     2.80   0.005     .6083398    3.441017
                                    |
                         sp_vvgroup |
                                VV  |   .5363737   .3807734     1.41   0.159    -.2103467    1.283094
        As the result above shows (bolded part), both the main effect (-1.617*) of the moderator, and the interaction effect (2.024679**) are significant.

        Can I just report them as significant moderation effect of sp_ss_friend?

        However, the margin effect showed the two 95% CIs are overlapped.

        Code:
        .                                  margins sp_ss_friend, dydx(sm_ptsd_categories) 
        
        Average marginal effects                        Number of obs     =      1,720
        Model VCE    : Linearized
        
        dy/dx w.r.t. : 1.sm_ptsd_categories
        1._predict   : Predicted mean (Created variable - PCL classification), predict(mu
                       outcome(sm_ptsd_categories))
        2._predict   : Predicted mean (RECODE of q24c (Q24C How many standard drinks do you have on
                       a t, predict(mu outcome(sp_prodrink))
        
        ---------------------------------------------------------------------------------------
                              |            Delta-method
                              |      dy/dx   Std. Err.      t    P>|t|     [95% Conf. Interval]
        ----------------------+----------------------------------------------------------------
        0.sm_ptsd_categories  |  (base outcome)
        ----------------------+----------------------------------------------------------------
        1.sm_ptsd_categories  |
        _predict#sp_ss_friend |
                1#less often  |          0  (empty)
           1#at least weekly  |          0  (empty)
           2#less often  |          .0006962  .0125929     0.06   0.956   -.0239993    .0253917
           2#at least weekly  |   .0602262  .0266158     2.26   0.024    .0080309    .1124215
        ---------------------------------------------------------------------------------------
        Note: dy/dx for factor levels is the discrete change from the base level.

        So the margin effect of 0.0006962 is not significantly smaller than 0.0602262.

        In this case, can I still report that sp_ss_friend significantly moderatedthe relationship between PTSD and problem drinking?

        Thank you!!

        Maggie

        Comment


        • #19
          It is the interaction coefficient in the regression output that measures the moderation effect. If you are interested in significance testing, look at the p-value there.

          The two marginal effects you estimated with -margins- do have overlapping confidence intervals, but that proves nothing. In any context, not just marginal effects or interactions, the presence of overlapping confidence intervals for statistics A and B does not exclude the possibility of the difference between A and B being statistically significant. In particular, your conclusion "So the margin effect of 0.0006962 is not significantly smaller than 0.0602262." is incorrect. The overlap of the confidence intervals does not imply this.

          (It does, however, work in the other direction: if the confidence intervals do not overlap, then the difference is statistically significant.)

          Comment


          • #20
            Thank you so much Clyde!

            Many of us in the office are having problem interpreting the results of moderation effect. We thought the difference between the two marginal effects have to be significantly different in order to say there is a significant moderation effect.

            Is the p value of the interaction coefficient in the regression output the ONLY measure of the significance of the moderation effect?

            I was told that BOTH the (1) regression coefficient of the moderator and the (2) interaction coefficient of the interaction term have to be significant in order to claim a significant moderation effect.

            for example, the interaction coefficient was significant (p = 0.008), but the main effect of the moderator was not significant (p = 0.216). In this case, can I say there is a significant moderation effect of sp_ss_family (interaction term of sp_ss_family#ptsd)?

            Code:
            .                 svy:sem (sm_ptsd_categories <- sp_vvgroup)(sp_mentsum_sf <- sm_ptsd_categories sp
            > _vvgroup sp_ss_family sp_famss_vvgp sp_famss_ptsd `control5b') , method(mlmv) standardized     
            (running sem on estimation sample)
            
            Survey: Structural equation model               Number of obs     =      2,211
            Number of strata     =         1                Population size   = 4,451.1958
            Number of PSUs       =     2,211                Design df         =      2,210
            
            --------------------------------------------------------------------------------------------------
                                             |             Linearized
                                Standardized |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
            ---------------------------------+----------------------------------------------------------------
            Structural                       |
              sm_ptsd_categories             |
                                  sp_vvgroup |   .3870005   .0225092    17.19   0.000     .3428591    .4311419
                                       _cons |   .1702186   .0264601     6.43   0.000     .1183294    .2221078
              -------------------------------+----------------------------------------------------------------
              sp_mentsum_sf                  |
                          sm_ptsd_categories |  -.1284914   .0288805    -4.45   0.000    -.1851271   -.0718556
                                  sp_vvgroup |  -.0699664   .0299362    -2.34   0.020    -.1286724   -.0112604
                                sp_ss_family |   .0398745   .0322037     1.24   0.216    -.0232781    .1030271
                               sp_famss_vvgp |  -.0242583   .0351974    -0.69   0.491    -.0932819    .0447652
                               sp_famss_ptsd |   .0812741   .0308544     2.63   0.008     .0207675    .1417807
            Thank you!

            Comment


            • #21
              I was told that BOTH the (1) regression coefficient of the moderator and the (2) interaction coefficient of the interaction term have to be significant in order to claim a significant moderation effect.
              That is quite wrong. The coefficient of the moderator itself is nothing but the marginal effect of the moderator when the other variable is zero. The coefficient of the moderator is often, as a result, a marginal effect conditional on a condition that never actually happens, or, if it does happen, is often of no interest. In fact, unless you have centered the other variable so that the value of zero is actually of some interest, the coefficient of the moderator alone should just be ignored--it is meaningless.

              In the example you show, yes, there is a statistically significant moderation effect.

              Comment


              • #22
                I see. Do we interpreting other main effects in the model at all? the coefficient of IV (VVgroup) and MV (PTSD)?

                Or should we only report the interaction coefficient?

                Thank you!

                Comment


                • #23
                  If a variable is part of an interaction, then its "main" effect is actually the effect of that variable conditional on the other variable(s) of the interaction being 0. That is often an impossible condition, and usually an uninteresting one. So these coefficients are usually ignored. Again, there is an exception if the variable has been "centered" so that 0 is a meaningful and interesting value.

                  For other variables, however, that are not part of any interaction, their coefficients are interpreted just as they would be in any regression that has no interactions.

                  Comment


                  • #24
                    Thank you so much for your help with moderation analyses Clyde!

                    Comment

                    Working...
                    X