05-04-2016 02:48 PM
Hi All,
I am using proc logistic to estimate propensity scores with the code below:
proc logistic data=anh.anh_sample_sat2;
model tri_both (event='1')=race age bmi parity_1 any_smoke any_diabetes_oshpd htn_disorder non_cchd_sbd;
output out=anh.prop pred=ps;
title "Estimation of propensity score from covariates";
run;
Both my outcome and all covariates are binary (0/1)
Here is the model without the pscore:
proc logistic data=anh.prop;
model cchd (event='1')=tri_both race age bmi parity_1 any_smoke any_diabetes_oshpd htn_disorder non_cchd_sbd ps;
title "Adding propensity score as covariate";
run;
and the model with the p-score added as a covariate:
proc logistic data=anh.prop;
model cchd (event='1')=tri_both race age bmi parity_1 any_smoke any_diabetes_oshpd htn_disorder non_cchd_sbd;
title "Without propensity score as covariate";
run;
When I add the p-score I get output that I do not know how to interpret (see bold text below). Can someone please help me understand this? Thank you in advance!!!
My outcome is rare (298/115,406) and again all variables are binary (0/1)
Odds Ratio Estimates | |||
Effect | Point Estimate | 95% Wald | |
tri_both | 1.410 | 1.080 | 1.840 |
race | 1.018 | 0.671 | 1.545 |
age | 1.264 | 0.599 | 2.666 |
bmi | 0.952 | 0.393 | 2.309 |
parity_1 | 0.823 | 0.218 | 3.107 |
any_smoke | 0.289 | <0.001 | 550.733 |
any_diabetes_OSHPD | 0.706 | 0.006 | 86.076 |
HTN_disorder | 0.502 | <0.001 | >999.999 |
non_cchd_sbd | 7.540 | 0.576 | 98.620 |
ps | >999.999 | <0.001 | >999.999 |
05-04-2016 03:12 PM
The interpretation is that the confidence intervals are noninformative. (That is, you aren't confident at all in the point estimate!) For practical purposes, the Wald confidence interval for the odds ratio is (0, infinity). This can happen in a contingency table if there is a cell count of zero. Your model is too complicated for me to guess what is going on, but there could be a degenerate relationship between the response variable and the ps variable.
It might be interesting to compare the Wald CIs to the profile-likelihood CIs. You can specify
... / CLODDS=BOTH;
on the MODEL statement and see what happens.
05-04-2016 03:24 PM
05-04-2016 04:19 PM
What is the purpose of using posterior probabilites as a predictor in same logistic regression model that generated these probabilites? These results may be as a result of high association between predictived probabilties and observed respose.
05-04-2016 09:33 PM
"Both my outcome and all covariates are binary (0/1)"
HoHo. That is not good for proc logistic . Since they are all category variables, your data more like a contingency table.
So maybe you should use PROC CATMOD to build logistic model .
05-05-2016 11:00 AM - edited 05-05-2016 11:23 AM
You might try a different approach. You can match subjects 1 to 1 having similar propensity scores. With the matched pairs data, you then need only fit a model with your predictor of primary interest omitting the other predictors (model cchd=tri_both; ) .
05-05-2016 11:23 AM - edited 05-05-2016 11:23 AM
Regarding the results that you show, I think you are essentially removing the effects of the secondary predictors twice by including both the propensity score and the predictors in the propensity score model. If you don't do matching, then you should only need to fit the model with the primary predictor and the propensity score (model cchd=tri_both ps; ). Notice that all of the secondary predictors are nonsignficant as well as the redundant propensity score.