Announcement

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

  • Mediation using sgmediation

    Dear all,

    I am conducting a fe panel regression in stata and want to check for a mediation
    I am using the sgmediation package and I am running the following code:

    sgmediation dv, iv(independent variable) mv(mediator) cv(controls) -> the variable names are in brackets
    I am getting significant effects of the mediator and the IV (for the direct and indirect effect)

    After that I want to bootstrap the results

    bootstrap r(ind_eff) r(dir_eff), reps(5000): sgmediation dv, iv(independent variable) mv(mediator) cv(controls)

    But now the indirect effect _bs_1 has a p value of 0.3 and thereby is not significant anymore

    Could you please help me how I can interpret the results?
    Is the model with a mediation not working for me?

    when running estat bootstrap, percentile the interval value is not 0

  • #2
    I found some information on the command here:
    http://ederosia.byu.edu/misc/example_stata_output2.pdf
    There some information is available on how to read the output. What I take from your description (next time maybe post your entire output) is that some coefficients are significant, but these are not the same as the indirect effect. Otherwise the bootstrapping would be useless since you already have your p-values. This is not the case. The individual coefficients can be significant but the indirect effect is something different. This means that in your specific sample, the indepvar and the mediator might be significantly related to your dependent var, but the inference for the indirect effect shows that there is probably no effect for this statistic in the population, thus, you cannot generalize your finding, it might be random.

    Best wishes

    Stata 18.0 MP | ORCID | Google Scholar

    Comment


    • #3
      Thank you so much. That is a great example.
      But in this example the authors state in the end of page 2 "To be ignored for our purposes"
      So it seems that they do not really look at the results and just interpret the ones on page 3 (
      "This confidence interval corresponds to the Preacher and Hayes SPSS confidence interval. If this interval does not include zero, the indirect effect is statistically significant.")

      Please find below the results from the sgmediation command:
      Sobel-Goodman Mediation Tests

      Coef Std Err Z P>|Z|
      Sobel .57509944 .26681823 1.877 .06207487
      Goodman-1 (Aroian) .57509944 .28686939 1.833 .06396984
      Goodman-2 .57509944 .29475229 1.854 .06218608

      Coef Std Err Z P>|Z|
      a coefficient = .790312 .41393 1.90929 .057225
      b coefficient = .674542 .083017 8.12536 4.3e-16
      Indirect effect = .533099 .286818 1.85866 .06396984
      Direct effect = 3.82448 1.05055 3.64045 .000272
      Total effect = 4.35757 1.0845 4.01806 .000059

      Proportion of total effect that is mediated: .12333836
      Ratio of indirect to direct effect: .13839125
      Ratio of total to direct effect: 1.1383913


      After bootstrapping the indirect effect _bs_1 has a p value of 0.12 and thereby is not significant anymore

      But if I get it right the authors state in the end of page 2 that this is "To be ignored for our purposes" and that it is only important that the interval does not include zero. Then the indirect effect is statistically significant

      Comment


      • #4
        What I take from the PDF is that you should ignore the inference of the indirect effect, that is, SE, Z and p-value. You first use conduct the analysis with the actual data, note the point estimate of the indirect effect, and then conduct the bootstrapping to get the correct p-value for the coefficient. I assume that is what they mean with ignoring the p-value from the actual command since this is probably not valid.
        Best wishes

        Stata 18.0 MP | ORCID | Google Scholar

        Comment


        • #5
          Thank you!
          So to get it right: Even if after bootstrapping the indirect effect _bs_1 has a p value of 0.12 I can still move on with the mediation (assuming that the interval does not include zero)?

          Comment


          • #6
            Well the conclusion would be that there is no significant indirect in the population that is different from zero, that is, there is no mediation effect visible in your analysis.
            Best wishes

            Stata 18.0 MP | ORCID | Google Scholar

            Comment

            Working...
            X