Testing IC estimates

[nhejazi]

• use only intercept in design matrix; otherwise, unclear what the t-test is actually doing...
• using more than just the intercept term after applying the IC transform would result in a test that is not necessarily well-defined

[nhejazi]

119de84 implements a change from limmatmle to the new function modtest_ic, which now accepts as input a design vector corresponding to the treatment/exposure of interest.

Remaining questions include:

• how does one split up a sample (to test the parameter) when the treatment/exposure of interest has more than 2 levels?

[nhejazi]

Currently, the target causal parameter is similar to the ATE -- that is, it takes the form $\Psi = E[E[Y \mid A = max(A), W] - E[Y \mid A = min(A), W]]$, for some vector A with categorical entries.

[nhejazi]

This should be restricted to a design vector that corresponds to an intercept term, as the group-level difference is captured in the influence curve representation of the ATE, reducing the required (moderated) t-test to that required for a 1-sample problem.

[nhejazi]

fixed in 3e7dc3b