NEWS

RobinCar 0.3.1 (2024-11-07)

When the following things happened simultaneously, the variance results were incorrect: (1) when the system locale was not set to "C", (2) the user did not specify their treatment variable as a factor, and (3) when inconsistent cases were used in naming treatment levels (e.g., "a", "B"). We have fixed this issue.

Simplified user experience for robincar_glm: removed robincar_glm, and renamed robincar_glm2 to robincar_glm. Same with robincar_linear and robincar_linear2. In effect, the older versions of robincar_linear and robincar_glm have been deprecated.

Previously, the factor level order passed by the user for the treatment variable was ignored. We have fixed this, so all estimates will be presented in the order of treatment levels specified by user.

Added a NEWS.md file to track changes to the package.
Added two new functions: robincar_linear2 and robincar_glm2. These are wrappers for robincar_linear and robincar_glm, respectively, but they give the user more full control over their covariate adjustment settings. The differences are listed below:
- The difference between robincar_linear and robincar_linear2 is that in robincar_linear2, if you want to include strata variables as covariates in the working model, you need to add those strata variables to the covariate_cols as well as the car_strata_cols.
- robincar_glm2 is exactly robincar_glm but only with the formula working model functionality.

Fixed two issues with the formula argument in robincar_glm. The first is if you had overlapping names for your treatment/response/covariate/strata variables (e.g., "A" for treatment col, and "A2" for covariate col), the formula would parse incorrectly. The second is if you used parentheses (e.g., "A*(x1+x2)"), the formula would also parse incorrectly. This would have given breaking errors to the user, and now the behavior is fixed.

Updated variance estimator for covariate-adjusted hazard ratio so that it uses the un-adjusted HR estimate rather than the adjusted HR estimate. Both result in consistent variance estimators but using un-adjusted theta is the expected behavior for the user.

Support for negative binomial glm (with known or unknown dispersion parameter)
Beta version of SuperLearner working model
Updated variance calculation that fixes negative variance when all/most outcome observations are 0

Initial Release