Propensity Score Matching - variable preicts failure perfectly

[Franziska Jakobs]

Hello,

I have a dataset that conists of roughly 280 observations. Of those, one has received treatment, the others have not. In order to conduct a case-control study, I would like to perform Propensity Score Matching to find the best / closest control cases for my case. In order to do this, i have been using the pscore command as follows:

pscore case sex2 agecat onco, pscore(a) blockid(b) logit

The variables are all stored in float format, and sex2 is stored as byte (1 = female, 0=male). Onco describes a variable of whether or not a patient has an oncological diagnosis (1=true, 0=false).

If i run the pscore command stated above, the output is as follows:

note: sex2 != 0 predicts failure perfectly;
sex2 omitted and 55 obs not used.

note: onco != 0 predicts failure perfectly;
onco omitted and 2 obs not used.


I understand that for the treated observation (case=1), sex2=0 and onco=0 and there are no cases for which this is not true in my dataset. However, i do not understand why these variables are excluded from the propensity score? Is there a way to include them ?

The only option i can think of is to use a penalized regression model (using the firthlogit command).

I would greatly appreciate any help or suggestions!
Many thanks in advance.

[Phil Clayton]

Essentially you're asking Stata to run a logistic regression to predict an event which only happens once in your dataset. When you think about it, how could Stata possibly provide an OR for (eg) sex2, when the only event occurred in a female?

The fundamental problem is that you only have 1 case. There is no getting around that - you don't have enough data to estimate anything.