1
$\begingroup$

What is the proper formula for estimating the Average Treatment Effect with a simple main effects generalized linear model?

My first pass at defining the Average Treatment Effect for a GLM with a log link:

$$ ATE = \biggl[\frac{\mathbb{E}(Y|A=1,\mathbf{X})}{\mathbb{E}(Y|A=0,\mathbf{X})}\biggr]\\ = \biggl[\frac{e^{\beta_0+\beta_A\cdot 1+\mathbf{\beta}\mathbf{X}}}{e^{\beta_0+\beta_A\cdot0+\mathbf{\beta}\mathbf{X}}}\biggr] \\ =e^{\beta_A}$$

where the 95% confidence interval for the ATE estimate would be $e^{\beta_A}\pm e^{1.96 \times SE(\beta_{A})}$.

For a project at work I'm doing a difference-in-differences case-control healthcare analysis to estimate the % change in episode-level medical costs from period=0 to period=1 for treatment=1 relative to treatment=0, using pooled cross-sectional observational data. (i.e., not longitudinal records but independent observations in each period $\times$ treatment group). Expanding the above ATE formula:

$$ ATE = \biggl[\frac{\mathbb{E}(Y|P=1,A=1,\mathbf{X})/\mathbb{E}(Y|P=0,A=1,\mathbf{X})}{\mathbb{E}(Y|P=1,A=0,\mathbf{X})/\mathbb{E}(Y|P=0,A=0,\mathbf{X})}\biggr]\\ = \biggl[\frac{e^{\beta_0+\beta_P\cdot1+\beta_A\cdot1+\beta_{PA}\cdot1+\mathbf{\beta}\mathbf{X}}/e^{\beta_0+\beta_P\cdot0+\beta_A\cdot1+\beta_{PA}\cdot0+\mathbf{\beta}\mathbf{X}}}{e^{\beta_0+\beta_P\cdot1+\beta_A\cdot0+\beta_{PA}\cdot0+\mathbf{\beta}\mathbf{X}}/e^{\beta_0+\beta_P\cdot0+\beta_A\cdot0+\beta_{PA}\cdot0+\mathbf{\beta}\mathbf{X}}}\biggr] \\ =(e^{\beta_P +\beta_{PA}}/e^{\beta_P})\\ =e^{\beta_{PA}}$$

where the 95% confidence interval for the ATE estimate would be $e^{\beta_{PA}}\pm e^{1.96 \times SE(\beta_{PA})}$.

If I want to use inverse propensity score weighting, doubly robust methods, etc. for estimating the ATE, observed records would fall in the "treatment" group for the above difference-in-differences study if P=1 and A=1, with the "control" group containing records where P=0 and A=1, P=1 and A=0, or P=0 and A=0.

Is my reasoning on-track here?

$\endgroup$
0
$\begingroup$

I believe there is a parameteric solution to estimating the Average Treatment Effect for a difference-in-differences study using the $g$-computation method:

1) Fit a GLM to the data, including terms for the time period and treatment group plus interactions:

$$ \mathbb{E}[Y|P,T,\mathbf{X}]=e^{\beta_0+\beta_P\cdot P+\beta_T\cdot T+\beta_{PT}\cdot P\cdot T+. . .}$$

2) For each record in the data compute the three counterfactual expected values of $Y$ (in addition to the existing expected value in the original data) by substituting alternative values for the the time period and treatment variables into the regression formula: $$\mathbb{E}[Y_i|P_i=0, T_i=0, \mathbf{X}_i]$$$$\mathbb{E}[Y_i|P_i=0, T_i=1,\mathbf{X}_i]$$$$\mathbb{E}[Y_i|P_i=1, T_i=0,\mathbf{X}_i]$$$$\mathbb{E}[Y_i|P_i=1, T_i=1,\mathbf{X}_i]$$

3) Compute the Average Treatment Effect by taking the geometric mean of the individual treatment effects:

$$ ATE = \biggl[\prod_{i=1}^n\frac{\mathbb{E}(Y_i|P_i=1,A_i=1,\mathbf{X}_i)/\mathbb{E}(Y_i|P_i=0,A_i=1,\mathbf{X}_i)}{\mathbb{E}(Y_i|P_i=1,A_i=0,\mathbf{X}_i)/\mathbb{E}(Y_i|P_i=0,A_i=0,\mathbf{X}_i)}\biggr]^{\frac{1}{n}}$$

The 95% confidence interval for the $ATE$ estimate can be estimated with bootstrapping, computer resources permitting.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.