Need p-values for pwcompare or contrast using mimrgns

[Jessica Totsky]

Hello,

I am able to produce pairwise comparisons or contrasts with mimrgns with the contrast dy/dx, SE, and CIs, but not the p-values. When I use pwcompare(effects) to obtain p-values, the code stops running after definitions of the _at levels; no contrasts, SEs, or CIs follow. Can you help me to view the p-values for each comparison? Here is the code and output so far:

mi estimate: svy: logistic Alcohol ib0
> .Gender ib0.ParEdu ib0.Race c.Age c.Fa
> mRel##c.FamHx c.FamRel##c.ParAtt_Alc c
> .FamRel##c.FamMan c.FamRel##c.FamConfl

Multiple-imputation estimates Imputations = 10
Survey: Logistic regression Number of obs = 35,198

Number of strata = 65 Population size = 160,214.29
Number of PSUs = 625
Average RVI = 0.0194
Largest FMI = 0.0796
Complete DF = 560
DF adjustment: Small sample DF: min = 381.54
avg = 541.95
max = 557.71
Model F test: Equal FMI F( 26, 557.4) = 95.72
Within VCE type: Linearized Prob > F = 0.0000


Alcohol Coefficient Std. err. t P>t [95% conf. interval]

Gender
Cis male -.1928048 .0964194 -2.00 0.046 -.3821948 -.0034148
Trans .0719635 .3754549 0.19 0.848 -.6655192 .8094461
Other -.0345483 .3089848 -0.11 0.911 -.6414849 .5723883

ParEdu
College -.0596366 .1959337 -0.30 0.761 -.4445058 .3252327
Some College .0235523 .2043461 0.12 0.908 -.3778492 .4249538
High School .0848689 .200519 0.42 0.672 -.3090073 .478745
Some High Sch.. .2030791 .2303097 0.88 0.378 -.2493894 .6555476
Grade School .. .0908952 .2586879 0.35 0.725 -.4172328 .5990233
IDK -.4030745 .2327125 -1.73 0.084 -.8601862 .0540373
N/A -.2749817 .4756616 -0.58 0.564 -1.210228 .6602645

Race
Hispanic .5485 .1195447 4.59 0.000 .3136827 .7833172
Multiracial .1636419 .2024677 0.81 0.419 -.2340536 .5613374
Asian -.370364 .2374845 -1.56 0.119 -.8368873 .0961594
Black -.1600875 .2891246 -0.55 0.580 -.72803 .4078551
Native American -.8731909 .3119547 -2.80 0.005 -1.485941 -.260441
Pacific Islan.. -.0265193 .239738 -0.11 0.912 -.4974197 .4443811

Age .2412244 .0225822 10.68 0.000 .1968664 .2855823
FamRel -.6681113 .0935536 -7.14 0.000 -.8518781 -.4843445
FamHx 1.089896 .2369505 4.60 0.000 .6244371 1.555355

c.FamRel#c.FamHx .3814078 .1126626 3.39 0.001 .1600889 .6027267

FamRel 0 (omitted)
ParAtt_Alc .4816989 .1053049 4.57 0.000 .2748387 .6885591

c.FamRel#
c.ParAtt_Alc .2656229 .0583024 4.56 0.000 .1510997 .3801462

FamRel 0 (omitted)
FamMan .1173456 .0793785 1.48 0.140 -.0385741 .2732652

c.FamRel#
c.FamMan .2659852 .0422492 6.30 0.000 .1829977 .3489727

FamRel 0 (omitted)
FamConfl .2252861 .0783911 2.87 0.004 .0713066 .3792655

c.FamRel#
c.FamConfl -.0882419 .0421145 -2.10 0.037 -.1709645 -.0055192

_cons -6.725947 .3969068 -16.95 0.000 -7.505585 -5.94631

Note: Strata with single sampling unit centered at overall mean.



mimrgns, dydx(FamRel) at (FamMan=(0(1)3)) predict(pr) vce(unconditional) cmdmargi
> ns

Multiple-imputation estimates Imputations = 10
Average marginal effects Number of obs = 35,198

Number of strata = 65 Population size = 160,214.29
Number of PSUs = 625
Average RVI = 0.0056
Largest FMI = 0.0083
Complete DF = 560
DF adjustment: Small sample DF: min = 551.06
avg = 552.73
Within VCE type: Linearized max = 554.32

Expression : Pr(Alcohol), predict(pr)
dy/dx w.r.t. : FamRel

1._at : FamMan = 0

2._at : FamMan = 1

3._at : FamMan = 2

4._at : FamMan = 3


dy/dx Std. err. t P>t [95% conf. interval]

_at
1 -.0141557 .002657 -5.33 0.000 -.0193748 -.0089366
2 -.0108982 .002422 -4.50 0.000 -.0156556 -.0061407
3 -.0006164 .0039758 -0.16 0.877 -.008426 .0071932
4 .0226229 .010087 2.24 0.025 .0028094 .0424364


mimrgns, dydx(FamRel) at (FamMan=(0(1)3)) predict(pr) vce(unconditional) pwcompar
> e cmdmargins

Multiple-imputation estimates
Pairwise comparisons of average marginal effects

Imputations = 10
Number of obs = 35,198
Number of strata = 65 Population size = 160,214.29
Number of PSUs = 625
Average RVI = 0.0056
Largest FMI = 0.0083
Complete DF = 560
DF adjustment: Small sample DF: min = 551.06
avg = 552.73
Within VCE type: Linearized max = 554.32

Expression : Pr(Alcohol), predict(pr)
dy/dx w.r.t. : FamRel

1._at : FamMan = 0

2._at : FamMan = 1

3._at : FamMan = 2

4._at : FamMan = 3


Contrast
dy/dx Std. err. [95% conf. interval]

_at
2 vs 1 .0032575 .001418 .0004723 .0060427
3 vs 1 .0135393 .0044723 .0047546 .022324
4 vs 1 .0367786 .011111 .0149537 .0586034
3 vs 2 .0102818 .0031301 .0041335 .0164301
4 vs 2 .0335211 .0098788 .0141166 .0529256
4 vs 3 .0232393 .0068077 .0098673 .0366113


mimrgns, dydx(FamRel) at (FamMan=(0(1)3)) predict(pr) vce(unconditional) pwcompar
> e(effects) cmdmargins

Multiple-imputation estimates
Pairwise comparisons of average marginal effects

Imputations = 10
Number of obs = 35,198
Number of strata = 65 Population size = 160,214.29
Number of PSUs = 625
Average RVI = 0.0056
Largest FMI = 0.0083
Complete DF = 560
DF adjustment: Small sample DF: min = 551.06
avg = 552.73
Within VCE type: Linearized max = 554.32

Expression : Pr(Alcohol), predict(pr)
dy/dx w.r.t. : FamRel

1._at : FamMan = 0

2._at : FamMan = 1

3._at : FamMan = 2

4._at : FamMan = 3

[daniel klein]

mimrgns is probably from SSC.

It's a bug that I have fixed about a year ago but somehow the update on the SSC server failed. Thanks to Kit Baum, the most recent version of mimrgns (*! version 4.0.7 24jul2022 daniel klein) is now available from the SSC. Updating mimrgns should solve the problem.

[Jessica Totsky]

Thank you! Also, is it theoretically incorrect to calculate marginal effects for non-significant coefficients in a regression, and likewise, contrast non-significant marginal effects? From my ouput, I see that it's possible, but I'm trying to make sense of why and if it's important to present. daniel klein

[daniel klein]

I am sorry, I cannot generally answer the broader follow-up question. What is considered significant and what is not varies across fields. The same is probably true for what is and is not "theoretically" correct and almost certainly true for what is "important to present".