Please help: comparing three average treatment effects (ATE) coefficients to determine if they differ significantly

[Josip Borovac]

Dear colleagues,

I am running an average treatment effects analysis (teffects) for the binary outcome and I have designated three subgroups in the propensity-score matching estimator (under by/if/in command - repeat commands by groups).

The output gives me three different ATE coefficients with 95% CI, z-value, and p>z value and standard error for all three subgroups.

My question is: which command do I use and how do I compare these three ATE coefficients to see if they significantly differ?
Is there a way to do this in STATA?

Those three ATE coefficients are clearly different numerically, however, I would wish to see if they differ significantly, therefore, would appreciate it if you could let me know on the ways how to formally test this?

I would highly appreciate your help since this is the last part I need to finish for my analysis. I apologize if similar content was reported previously, however, I am a novice STATA user and have used SPSS almost exclusively for the purpose of my research, however, this analysis is not available in SPSS and for this instance, I need to use STATA.

Thank you very much for your help!

[Phil Bromiley]

You didn't get a quick answer. You will increase your chances of useful answer by following the FAQ on asking questions – provide Stata code in code delimiters, readable Stata output, and sample data using dataex. Being able to replicate your problem is often essential to helping you.

While you talk about average treatment effects, you don't tell us precisely what estimator you used or exactly how you set these estimates of these average treatment effects up. Providing the exact code in code delimiters will solve this problem. Without such information, it is almost impossible to help you.

Depending on what you have used to estimate this, suest might get all of the parameters into one estimate in which case you can simply use the test procedure. After any estimate, if you issue
estimate_name, coefl it will tell you how to refer to the coefficients in subsequent activities.

[Josip Borovac]

You didn't get a quick answer. You will increase your chances of useful answer by following the FAQ on asking questions – provide Stata code in code delimiters, readable Stata output, and sample data using dataex. Being able to replicate your problem is often essential to helping you.

While you talk about average treatment effects, you don't tell us precisely what estimator you used or exactly how you set these estimates of these average treatment effects up. Providing the exact code in code delimiters will solve this problem. Without such information, it is almost impossible to help you.

Depending on what you have used to estimate this, suest might get all of the parameters into one estimate in which case you can simply use the test procedure. After any estimate, if you issue
estimate_name, coefl it will tell you how to refer to the coefficients in subsequent activities.

Dear Phil, many thanks for your reply.

I am apologizing for perhaps not being able to provide you with sufficient details.

Basically, I am running STATA 14 with the following analysis used: treatment effects --> binary outcomes --> propensity-score matching.

The dependent outcome is binary - mortality (0-no, 1-yes) while we test treatment 1 vs. treatment 2 in the ATE analysis.

Both treatment groups are propensity score-matched for various baseline characteristics and caliper size of 0.2 was used in the neighbor matching algorithm.

The output that I get is attached in the PDF file.

Basically, I've run this analysis in three different risk groups to see if the effects of treatment 1 are consistent vs. treatment 2 across all groups.

As you can see, I've got that across all three risk groups there is an average treatment effect that is characterized by the reduction of dependent endpoint of death.

Three ATE coefficients are provided for each of these analyses (marked in bold). Now, I only need to compare these three coefficients between themselves to see if they differ significantly.

i.e. ATE of high risk group (3) vs. ATE of low risk group (1), ATE of high risk group (3) vs. ATE of intermediate risk group (2), etc.

I hope I made this more clear.

Many thanks for your help.

With best wishes,
Josip

[samian sulaiman]

I have similar question, suest did not work for me. it gives me an error message because the teffects psmatch used robust vce "was estimated with a nonstandard vce (robust)". when I changed vce to iid, it again gave me a different error message "unable to generate scores for model ***
suest requires that predict allow the score option", any advise?