NEWS.MD

In almost-matching-exactly/R-FLAME: Interpretable Matching for Causal Inference

FLAME 2.1.1.9000

FLAME 2.1.1

Minor Changes

1. PE computation now based off new xgboost::xgboost() defaults.
2. Introduced nnet as a dependency for compatibility with mice::mice.impute.polyreg() for imputing categorical variables.

FLAME 2.1.0

Major Changes

1. The Dynamic Almost Matching Exactly (DAME) algorithm has been implemented. This algorithm has the same features as FLAME and is more computationally intensive as it is guaranteed to find the best set of covariates to match on for each unit. The n_flame_iters argument in DAME specifies that FLAME should be run for a certain number of iterations to more rapidly eliminate irrelevant covariates, before switching to DAME. See the DAME paper here for more details.
2. The data entry in the list (now of class ame) returned by FLAME or DAME does not contain *'d entries for values that units did not match on. This led to some confusing behavior with DAME when replace = TRUE and also impeded using the matched data for subsequent analysis (e.g. treatment effect estimation via regression adjustment).
3. Support was removed for automatically binning continuous covariates, as this is generally not recommended. All variables passed in to a matching algorithm (DAME or FLAME) will be assumed to be categorical and no longer need to be passed as factors.
4. Whereas users were previously forced to input categorical covariates as factors or numerics with values 0:(n_levels - 1), covariates can now be passed with any values / codings, streamlining the process of running a matching algorithm on one's data and rendering the output more interpretable.
5. Support for fixed weights has been introduced. Users can now pass a vector of covariate importances via the weights argument that will be used to determine dropping order. This means that the holdout set is no longer necessary (it is ignored if passed) because the weights implicitly define a dropping order and there is no need to compute PE. Note that this dramatically speeds up the algorithm.
6. The output of FLAME or DAME is now an S3 object of class ame; print, plot, and summary methods have been introduced.

Minor Changes

1. The vignette has been updated to reflect all changes.
2. It is no longer possible to perform more than 1 round of imputation for the matching data. Users that wish to analyze the results of multiple imputations should wrap calls to FLAME or DAME in a loop.
3. The values that can be passed to missing_data and missing_holdout have been given interpretable names based off their function.
4. Functionality for using user-supplied functions to compute PE has been changed. Now, a single function should be passed to PE_method that does fitting and prediction and the user_* arguments are deprecated and will be removed in a later release. See the documentation and vignette for more details.
5. The arguments early_stop_pe and early_stop_bf have been deprecated and will be removed in a later release.
6. Methods for treatment effect estimation have been changed. The ATE, ATT, ATC functions have been deprecated and will be removed in a later release, as these effect estimates and their variances are computed by summary.ame, albeit slightly differently from before; see the documentation for more details. The CATE function also estimates CATEs slightly differently and returns estimate variances.
7. CATE estimates are no longer automatically returned; this can be toggled with the estimate_CATEs flag, which defaults to FALSE.
8. Small changes in output formats of MG and CATE.
9. Support for multi-class outcomes.
10. Some information about the data and matching process is now returned in the info entry of an ame object.
11. Eliminated dependence on xgboost and mice; now only required if using XGBoost to compute PE or if imputing data.
12. Rownames in the matching data are now unchanged and MG returns matched groups with respect to the rownames of the matching data and not with respect to the index of the unit in the original data frame. Typically, this will also correspond to the indexing of the data (i.e. the third unit has rowname '3'); however, if a separate holdout set was not passed to the matching algorithm or if the original matching data had rownames other than 1:nrow(data) then this is not the case.
    • The index_only argument of MG has therefore been replaced with the more appropriate id_only argument.

FLAME 2.0.0

Version released to CRAN April 15, 2020. Introduced functionality for matching with FLAME.

almost-matching-exactly/R-FLAME documentation built on Jan. 4, 2022, 8:31 a.m.