Announcement

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

  • Mediation analysis confidence intervals versus p-values

    Hello,

    I am running a mediation analysis with a binary predictor, a binary mediator and a binary outcome. I used paramed on stata to run this. I got the following output, but am confused as the confidence intervals suggest each of the effects should be significant, but the p values do not

    | Estimate Std Err P>|z| [95% Conf Interval]
    --------------------------------------------------------------------
    cde | .68979196 .11338915 (0.001) [.5523309, .86146358]
    nde | .68979196 .2267783 (0.102) [.44226294, 1.0758599]
    nie | 1.1819007 .05215623 (0.001) [1.0670499, 1.3091135]
    mte | .81526564 .2319472 (0.379) [.51744202, 1.2845073]

    cde: controlled direct effect, nde: natural direct effect, nie: natural indirect effect, mte: marginal total effect
    I am unsure how to interpret the results in this case. I would be grateful if anyone has any insight into this. Could it be related to the fact that my variables are binary?

    Thanks for your help

  • #2
    These are odds ratios (to be fair, I dont see it mentioned in the help file that the estimates are ORs for binary vars). So you get a significant p-value when the CI does not include 1 . This might help: https://stats.idre.ucla.edu/stata/fa...ic-regression/
    Eric A. Booth
    Senior Research Scientist, Gibson Consulting Group

    eric.a.booth@gmail.com | http://www.eric-a-booth.com

    Comment


    • #3
      Originally posted by eric_a_booth View Post
      These are odds ratios (to be fair, I dont see it mentioned in the help file that the estimates are ORs for binary vars). So you get a significant p-value when the CI does not include 1 . This might help: https://stats.idre.ucla.edu/stata/fa...ic-regression/
      Great thanks that makes a lot more sense now.

      I'm also a little confused about how these confidence intervals have been calculated. I thought CI's for 95% level were the coefficient +/- 1.96 x SE, but that doesn't seem to be the case here. Is some form of transformation needed in the case of binary variables as it is a logistic regression?

      Comment


      • #4
        You'd want to get your critical value from the t-distribution via invttail() in lieu of just using 1.96. Here's a quick example:




        Code:
        sysuse auto, clear
        
        regress mpg price turn
         di _b[price] - invttail(71,0.025) * _se[price]
        
         
         paramed mpg, avar(for) mvar(price) a0(0) a1(1) m(1200) yreg(linear) mreg(linear)
          di  312.2587  - invttail(72,0.025) * 754.4488
        Eric A. Booth
        Senior Research Scientist, Gibson Consulting Group

        eric.a.booth@gmail.com | http://www.eric-a-booth.com

        Comment

        Working...
        X