binreg vs logistic different p-values

[Rannveig Mowinckel]

Hello. I have a dataset where I look at associations - the dependent variable is binary, and the independent variables are all categorical. The depentens variable measure prevalence of musculoskeletal disorders, and have a prevalence of 31%. I have therefor chosen to use the binreg command as it reports risk ratio rather than odd ratio due to the difference in OR/RR at low/high prevalences. BUT why do I get a different result in regard to the p-values when using the binreg vs the logistic command? I am particulary interested in the p-value for females in age group 25. - see under:
I am most grateful for any thoughts on this topic.
Best regards
Rannveig


. by sex01, sort : binreg prevalence i.age_groups, rr

--------------------------------------------------------------------------------------
-> sex01 = male
------------------------------------------------------------------------------
| EIM
prevalence | Risk ratio std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
age_groups |
18 | 1 (base)
25 | 1.445646 .3400905 1.57 0.117 .9116231 2.292495
35 | 1.253019 .3377062 0.84 0.403 .7388343 2.125047
45 | 1.673663 .3978536 2.17 0.030 1.050328 2.666924
|
_cons | .2168675 .0452351 -7.33 0.000 .1440945 .3263935
-> sex01 = female
-----------------------------------------------------------------------------
| EIM
prevalence | Risk ratio std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
age_groups |
18 | 1 (base)
25 | 1.942857 .6841564 1.89 0.059 .9743149 3.874203
35 | .5 .3644339 -0.95 0.342 .1198275 2.086332
45 | 2 .9354143 1.48 0.138 .7996813 5.001993
|
_cons | .25 .0765465 -4.53 0.000 .1371873 .4555816
------------------------------------------------------------------------------

by sex01, sort : logistic prevalence i.age_groups

-----------------------------------------------------------------------------
-> sex01 = male
------------------------------------------------------------------------------
prevalence | Odds ratio Std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
age_groups |
18 | 1 (base)
25 | 1.649169 .5111241 1.61 0.106 .8983667 3.027447
35 | 1.347429 .4780347 0.84 0.401 .6722344 2.700793
45 | 2.057493 .6602478 2.25 0.025 1.096955 3.85912
|
_cons | .2769231 .0737573 -4.82 0.000 .1643028 .4667381
------------------------------------------------------------------------------
-> sex01 = female
-----------------------------------------------------------------------------
prevalence | Odds ratio Std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
age_groups |
18 | 1 (base)
25 | 2.833333 1.502056 1.96 0.049 1.002406 8.008506
35 | .4285714 .3681963 -0.99 0.324 .0795674 2.308402
45 | 3 2.44949 1.35 0.178 .6055055 14.86361
|
_cons | .3333333 .1360828 -2.69 0.007 .1497536 .7419597
------------------------------------------------------------------------------

[Andrew Musau]

You estimate a log-binomial model to obtain the risk ratios. Both the coefficients and standard errors will be smaller in logistic regression and the p-values will be similar to log-binomial regression, but not exact. In any case, 0.049 and 0.059 are not that different.

[Leonardo Guizzetti]

As a note, -binreg- is nothing more than a wrapper for -glm- with some convenient defaults, so it's not the only way to estimate the same log-binomial model.

[Rannveig Mowinckel]

Thank you ever so much! I appreciate your kind responses!