Interpretation of ATT using a Binary Outcome Variable

[Juan Venegas]

Dear all,

I am applying PSM having as a dependent variable a dummy variable that state if the student dropped out or not from higher education and as independent variable if the student participated or not in a counseling program.

This is my command : teffects nnmatch (dropout x2 x3 x4 ... (program_participation), atet

I obtain as a coefficient -0.074. How do I interpret this coefficient? Is it correct to say that students participating in the counseling program have on average 7,4% less odds of dropping out from higher education than non participatnts? Or should I talk about probabilities? or should I estimate the exp of -0.074 like logit analysis?

Thank you!!

Juan

[Joshua D Merfeld]

Dear Juan,

With PSM, we are constructing two groups that we believe are similar (on observables) except for the treatment variable. Since you have not specified caliper(), the difference between att and atet is that att finds a match for every observation in the data. In other words, it takes every treated observation and matches it to an untreated observation, with matches taking place based on propensity score. It also takes every untreated observation and matches a treated observation to it. atet, on the other hand, only finds a match for the treated observations. The treatment effect is a difference in means between the two groups. So, that is a 7.4 percentage point difference. If the assumptions for PSM hold, then the program reduces the probability of dropping out by 7.4 percentage points.

I actually haven't used the tseffects command myself, so if my understanding is incorrect, I know someone out there will correct me!

HTH,

Josh

[Juan Venegas]

Thank you Joshua.

I understand the difference between ATT and ATET. My concern was specifically about how yo interpret ATT having a binary dependent variable. With logistic model you have to estimate the exp of your coefficient to estimate the OR and then substantially interpret your results. In this case, I am not sure if I have to estimate the exp or if I get a coefficient of -0,07 I can say that treated have 7% less probability of dropping out.

[Gayatri Krishna]

Hello all
Even I have the same query that Juan Venegas had. The command I used was: teffects psmatch (lung_cancer) (exposure age gender education marital_status mpce_quintile caste religion residence), atet. My outcome variable here is lung cancer which is binary with responses as yes/no and treatment variable is exposure (exposed v/s unexposed). I got an ATT of -0.031. How can I interpret them?

[Weiwen Ng]

You interpret this as you would a linear probability model, I.e. one where your DV is 0/1 and you fit a linear model to it. You have probably heard that logistic models are advantageous, and they are. But nothing will explode if you fit a linear probability model to a binary outcome, although you do need to be aware of possible edge cases (e.g. predictions are allowed to go outside 0 or 1).