margins, pwcompare

[Liliani Costari]

I have written a program that exports tables from margins. I was hoping to replicate the results from pwcompare with the margins, pwcompar command. In most cases the results are similar, however, when using the option pveffects, the results differ. Am I missing something. Here is an example:

webuse nhanes2
logistic highbp sex##agegrp##c.bmi

. margins agegrp, pwcompare(pv)

Pairwise comparisons of predictive margins              Number of obs = 10,351
Model VCE: OIM

Expression: Pr(highbp), predict()

--------------------------------------------------------
                |            Delta-method    Unadjusted
                |   Contrast   std. err.      z    P>|z|
----------------+---------------------------------------
         agegrp |
30-39 vs 20-29  |   .0798159    .013507     5.91   0.000
40-49 vs 20-29  |   .1738605   .0155477    11.18   0.000
50-59 vs 20-29  |   .3122507   .0161705    19.31   0.000
60-69 vs 20-29  |   .3610133   .0126272    28.59   0.000
  70+ vs 20-29  |   .4504747   .0174674    25.79   0.000
40-49 vs 30-39  |   .0940445   .0164988     5.70   0.000
50-59 vs 30-39  |   .2324348    .017087    13.60   0.000
60-69 vs 30-39  |   .2811974   .0137814    20.40   0.000
  70+ vs 30-39  |   .3706588   .0183192    20.23   0.000
50-59 vs 40-49  |   .1383903   .0187419     7.38   0.000
60-69 vs 40-49  |   .1871529   .0157866    11.86   0.000
  70+ vs 40-49  |   .2766143   .0198717    13.92   0.000
60-69 vs 50-59  |   .0487626   .0164004     2.97   0.003
  70+ vs 50-59  |    .138224   .0203626     6.79   0.000
  70+ vs 60-69  |   .0894614   .0176804     5.06   0.000
--------------------------------------------------------

. pwcompare agegrp, pv

Pairwise comparisons of marginal linear predictions

Margins: asbalanced

--------------------------------------------------------
                |                            Unadjusted
                |   Contrast   Std. err.      z    P>|z|
----------------+---------------------------------------
highbp          |
         agegrp |
30-39 vs 20-29  |  -.0248978   .5020654    -0.05   0.960
40-49 vs 20-29  |   1.217738    .522244     2.33   0.020
50-59 vs 20-29  |   2.427516    .506242     4.80   0.000
60-69 vs 20-29  |   3.083275    .419672     7.35   0.000
  70+ vs 20-29  |   4.267813   .5528408     7.72   0.000
40-49 vs 30-39  |   1.242636   .5468977     2.27   0.023
50-59 vs 30-39  |   2.452414   .5316383     4.61   0.000
60-69 vs 30-39  |   3.108173    .449981     6.91   0.000
  70+ vs 30-39  |   4.292711   .5761868     7.45   0.000
50-59 vs 40-49  |   1.209778   .5507345     2.20   0.028
60-69 vs 40-49  |   1.865538   .4723898     3.95   0.000
  70+ vs 40-49  |   3.050076   .5938522     5.14   0.000
60-69 vs 50-59  |   .6557598   .4546364     1.44   0.149
  70+ vs 50-59  |   1.840298   .5798298     3.17   0.002
  70+ vs 60-69  |   1.184538   .5060101     2.34   0.019
--------------------------------------------------------

[Andrew Musau]

As stated in the output, pwcompare calculates marginal linear predictions, which are conventional linear predictions (predicted values of the outcome holding all other variables in the model constant). You can certainly compute linear predictions using margins.

clear all
webuse nhanes2
logit highbp sex##agegrp##c.bmi
pwcompare agegrp, pv
margins, over(agegrp) expression(predict(xb)) at(bmi=0) asbalanced pwcompare(pv)

Res.:

. pwcompare agegrp, pv

Pairwise comparisons of marginal linear predictions

Margins: asbalanced

--------------------------------------------------------
                |                            Unadjusted
                |   Contrast   Std. err.      z    P>|z|
----------------+---------------------------------------
highbp          |
         agegrp |
30–39 vs 20–29  |  -.0248978   .5020654    -0.05   0.960
40–49 vs 20–29  |   1.217738    .522244     2.33   0.020
50–59 vs 20–29  |   2.427516    .506242     4.80   0.000
60–69 vs 20–29  |   3.083275    .419672     7.35   0.000
  70+ vs 20–29  |   4.267813   .5528408     7.72   0.000
40–49 vs 30–39  |   1.242636   .5468977     2.27   0.023
50–59 vs 30–39  |   2.452414   .5316383     4.61   0.000
60–69 vs 30–39  |   3.108173    .449981     6.91   0.000
  70+ vs 30–39  |   4.292711   .5761868     7.45   0.000
50–59 vs 40–49  |   1.209778   .5507345     2.20   0.028
60–69 vs 40–49  |   1.865538   .4723898     3.95   0.000
  70+ vs 40–49  |   3.050076   .5938522     5.14   0.000
60–69 vs 50–59  |   .6557598   .4546364     1.44   0.149
  70+ vs 50–59  |   1.840298   .5798298     3.17   0.002
  70+ vs 60–69  |   1.184538   .5060101     2.34   0.019
--------------------------------------------------------

. 
. margins, over(agegrp) expression(predict(xb)) at(bmi=0) asbalanced pwcompare(pv)

Pairwise comparisons of adjusted predictions            Number of obs = 10,351
Model VCE: OIM

Expression: Linear prediction (log odds), predict(xb)
Over:       agegrp
At: 1.agegrp
        sex      (asbalanced)
        agegrp   (asbalanced)
        bmi    = 0
    2.agegrp
        sex      (asbalanced)
        agegrp   (asbalanced)
        bmi    = 0
    3.agegrp
        sex      (asbalanced)
        agegrp   (asbalanced)
        bmi    = 0
    4.agegrp
        sex      (asbalanced)
        agegrp   (asbalanced)
        bmi    = 0
    5.agegrp
        sex      (asbalanced)
        agegrp   (asbalanced)
        bmi    = 0
    6.agegrp
        sex      (asbalanced)
        agegrp   (asbalanced)
        bmi    = 0

--------------------------------------------------------
                |            Delta-method    Unadjusted
                |   Contrast   std. err.      z    P>|z|
----------------+---------------------------------------
         agegrp |
30–39 vs 20–29  |  -.0248978   .5020654    -0.05   0.960
40–49 vs 20–29  |   1.217738    .522244     2.33   0.020
50–59 vs 20–29  |   2.427516    .506242     4.80   0.000
60–69 vs 20–29  |   3.083275    .419672     7.35   0.000
  70+ vs 20–29  |   4.267813   .5528408     7.72   0.000
40–49 vs 30–39  |   1.242636   .5468977     2.27   0.023
50–59 vs 30–39  |   2.452414   .5316383     4.61   0.000
60–69 vs 30–39  |   3.108173    .449981     6.91   0.000
  70+ vs 30–39  |   4.292711   .5761868     7.45   0.000
50–59 vs 40–49  |   1.209778   .5507345     2.20   0.028
60–69 vs 40–49  |   1.865538   .4723898     3.95   0.000
  70+ vs 40–49  |   3.050076   .5938522     5.14   0.000
60–69 vs 50–59  |   .6557598   .4546364     1.44   0.149
  70+ vs 50–59  |   1.840298   .5798298     3.17   0.002
  70+ vs 60–69  |   1.184538   .5060101     2.34   0.019
--------------------------------------------------------

.