Help with teffects

[Francis Pike]

Hi Guys: I am new to teffects in STATA and am trying to get a causal estimate for survival (Average treatment effect) in patients who had an early brain catheterization compared to those who didn't.
I understand how to enter the model into stata but I don't understand what the coefficient means.
Here is the teffects code: Surv is the outcome var. early_cath is the treatment and the rest are the variables for the propensity score. teffects psmatch (surv) (early_cath age male smoker __epi trop_1 bls cabg_h cad chf cva cxr diabetes htn inc_chol lbbb pcac std vf_vt new_where), nneighbor(1
The output from stata is:

teffects psmatch (surv) (early_cath age male smoker __epi trop_1 bls cabg_h cad chf cva cxr d
> iabetes htn inc_chol lbbb pcac std vf_vt new_where), nneighbor(1) control(0)

Treatment-effects estimation Number of obs = 426
Estimator : propensity-score matching Matches: requested = 1
Outcome model : matching min = 1
Treatment model: logit max = 1
------------------------------------------------------------------------------
| AI Robust
surv | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
ATE |
early_cath |
(1 vs 0) | -.0821596 .0371013 -2.21 0.027 -.1548768 -.0094425
------------------------------------------------------------------------------
Can anyone tell me what a coefficient of -.0821596 means when my outcome variable is (0,1)???
Thanks guys.
Francis

[Marcos Almeida]

Hello, Francis,

This will surely help you:http://www.stata.com/manuals13/te.pdf

Please check the "teffects intro" as well as "teffeccts intro advanced". Also, among other interesting excerpts, page 52, pages 84 to 86 and page 140.

Hopefully that helps.

Best,

Marcos

[Francis Pike]

Thanks Marcos. I downloaded and read through the examples and specifically the pages you mentioned but they are all continuous so its the difference in means. I don't see any example where the outcome is binary. Can I assume the ATE is just an expected difference in proportions ATE= POt-POc. I have ran several other basic logistic and stratification ps adjustments and have ODDS RATIOS that are not significant and thus close to 1 so I need to be sure what this ATE is reporting for matching. The PI has asked to match and report the effect sizes.
Like I said in the original post the ATE for 1:1 PS matching is ~ -0.082. Does this mean the treatment causes a reduction in the outcome of 8%???. Since the outcome is survival it would say that early_catheterization is harmful which tends to go against what the PI supposed. ie It would be beneficial. Thanks again. Francis
PS please excuse my naivety.

[Marcos Almeida]

Hello Francis,

This is an excerpt from "teffects intro advanced" (http://www.stata.com/manuals13/tetef...roadvanced.pdf):

The outcome models can be continuous, binary, count, or nonnegative. Continuous outcomes can be modeled using linear regression; binary outcomes can be modeled using logit, probit, or heteroskedastic probit regression; and count and nonnegative outcomes can be modeled using Poisson regression. The treatment model can be binary or multinomial. Binary treatments can be modeled using logit, probit, or heteroskedastic probit regression, while multinomial outcomes are modeled using multinomial logit regression.

Hopefully that helps.

Now, changing a little bit the subject: as a cardiologist, I wouldn't be much suprised that early catheterization could potentially be associated with worse outcomes. Among the (several reasons), a confounding factor left outside your model, let alone the "burden" of the procedure itself. Also, selection bias: what if the patients who underwent early catheterization were the ones with the worst baseline condition?

Kind regards,

Marcos

[Jeff Wooldridge]

Francis: The output is interpreted as always when your response variable is binary: It is the estimated difference in the probability of seeing a one (survival = 1). In general, all teffects commands, with the default ate option, estimate E(y1) - E(y0), where y1 and y0 are the counterfactuals. Because these are binary in your case, E(y1) = P(y1 = 1), and so on.

I would state your interpretation slightly differently: Early catheterization reduces the probability of survival by 8.2 percentage points (not percent). FWIW, I agree with Marcos: While the invassiveness of the procedure could be causing higher mortality, the unconfoundedness assumption very likely fails. The attending physician has much more information about the status of the patient than is in the controls that you are using. I have experience with a data set on right heart catheterization, which suffers from similar problems. JW

[Francis Pike]

Thanks guys appreciate all the help and advice. The only comment I have for Marcos is that I was under the impression that causal models were invented to specifically counter selection bias and basically form a pseudo randomization which balances all covariates across pscores. Maybe I misunderstand what pscores are doing. I do get that there may be variables we have omitted from the treatment model but I included all the variables the physician felt important. Great learning experience. Appreciate all the help.

[Jeff Wooldridge]

There's nothing magic about matching. If the intervention is based on factors that we don't observe, and that affect the outcome, then matching is biased -- just like regression methods.

[Marcos Almeida]

Hello, Francis,

As Jeff clearly stated, propensity score matching procedures work basically with observable factors. There may be several non-observed variables that may be left outside a model. Also worth remarking, "hidden" (non measured or non measurable) factors may have played a role regarding the physician's decision to indicate early catheterization. Hardly could we add "all covariates" to a model. To end, invasive procedures such as early catheterization shall do some sort of harm and you'll find literature on that so as to back up your study.

Best,

Marcos