Psm and cox regression

#1
Dear members
I am new at this forum. I thank you all in advance for your kind help.
I ran a propensity score matching in order to evaluate a rehabilitation program.
We published this paper due to a siginificant finding that showed that the program is effective in reducing recidivism rates.
Now, I am writing a new paper and I am using the matches samples from the study. I am checking whether the interaction between ethnicity and program participation is significant in reducing recidivism rates.
I actually find significant findings.
I did the analysis in 2 ways: using Kaplan Meier curves and also using Cox regression.
I the cox regression I controlled again for the variables I already controlled in the PSM. Is that ok?
Thanks a lot
All the best
Noam
 
#2
Dear members
I am new at this forum. I thank you all in advance for your kind help.
I ran a propensity score matching in order to evaluate a rehabilitation program.
We published this paper due to a siginificant finding that showed that the program is effective in reducing recidivism rates.
Now, I am writing a new paper and I am using the matches samples from the study. I am checking whether the interaction between ethnicity and program participation is significant in reducing recidivism rates.
I actually find significant findings.
I did the analysis in 2 ways: using Kaplan Meier curves and also using Cox regression.
I the cox regression I controlled again for the variables I already controlled in the PSM. Is that ok?
Thanks a lot
All the best
Noam
By this I mean that I did a double adjustment
 

hlsmith

Omega Contributor
#3
Look up augmented inverse propensity scores. It may be fine, I would also take into account sample size and proportion of subjects with or without the event.

Welcome to the forum.

PS, I forgot you said you matched. Same logic holds, but you should be able to compare distributions of potential confounder between groups.
 
Last edited:
#5
One more question please.
I read few of the articles regarding the Aipw.
I used In my previous paper the psmatch2 command - can I still assume that it is fine to control for few of the variables again - for robustness?
Thanks again
Noam
 

hlsmith

Omega Contributor
#6
But my follow-up question is do you need to? Did you match on a lot of variables? And again, I am not sure of your sample size and breakdown of events. Controlling for them a second time may over saturate the final model if your sample size isn't large. I believe you should examine the distribution of potential confounders after matching. If there aren't major differences you are probably fine. You can always model with and without them and see if estimate on variable interest changes, but ideally you set you r analysis plan a priori and stick to it, otherwise you risk false discovery.
 
#7
Thanks a lot
It really helps
I will ran the analysis one time only with the main and interaction effects and with all the variables and will see I something change.
My sample size is1094 - 547 treatment and 547 control group.
The cox regression assumptions are met.
Because this is a new paper and I am using a matched sample from previous study I thought a double robustness would be better.
Again, your help is much appreciated
All the best
Noam
 

hlsmith

Omega Contributor
#8
Glad to hear i could be of help. If the matching process was successful i would imagine controlling for the potential confounders in the model won't be necessary.

A side issue, propensity scores shouldn't include instrumental variables or effects of the variable of interest and outcome. If you accidentally included them in creating scores that can be an issue, but including them twice (in scores and outcome model) would double down on the issue. Matching on scores, functions to balance baseline covariates but shrinks the sample size. Now including them again adds sparsity to the data in the model, so you shrunk sample then included more terms in final model. Not ideal if matching worked well.

My other question I keep asking is what proportion of subjects had the event? Just like in logistic regression you look at the smaller of the group's (event: yes/no) and get an idea how many variables you have per total number of events.

Lastly, was the dataset originally created to answer a comparable question, so you have all relevant potential confounders collected? When disseminating results you must disclose this is a secondary analysis and some people may opt to use a smaller alpha value for significance for this reason.
 
Last edited:
#9
In the propensity score matching process there were 29 variables controlled for. The variables included criminal history, socio-demographic and incarceration characteristics of the prisoner.
In this paper we evaluated the effectiveness of a prigram running within prison walls for five years.
The matching went well.
I took the file after matching and ran a Cox regression. Follow up period is 11 years and about 40 percent has an event. Less events of course in the treatment group.
I added 2 variables to the regression that were not included in the psm - e didn’t have this variable originally, but I fixed some of the variables and added them again and also checked for multicolliniarity and assumptions.
Before running the cox regression I ran a Kaplan Meier test and found out that the reduction of crime for Jews is 30% and for Arab only 15%.
This is the reason I thought of running the regression with the interaction.