Mirror Splits from Partial Dependence

For ech split, we return an array of dimension n_taxa x 2 fits x n_lags. Each entry contains the estimated counterfactual effect for that taxon and lag combination across the two models (each fit on a different random sample of data).

Usage

pd_splits(ts, w0, w1, tr_fun, n_splits = 20, ...)

Arguments

ts

An object of class ts_inter containing the time-varying microbiome community, environmental interventions, and static host features data. The columns for each element of the values matrix are expected to be consecutive timepoints in an changing community. ts_from_dfs() is a helper to create these objects from data.frames with the relevant information.

w0

One of the counterfactuals with which to compute partial dependence profiles. See steps or pulses for helpers in generating these counterfactuals. The procedure concatenates these counterfactuals to the end of the series and computes the difference in the forecasts.

w1

One of the counterfactuals with which to compute partial dependence profiles. See steps or pulses for helpers in generating these counterfactuals. The procedure concatenates these counterfactuals to the end of the series and computes the difference in the forecasts.

tr_fun

A function that can be used to train over random splits. In the examples in this package, we use an anonymous function that fills our chosen hyperparameters in mbtransfer. For example \(x) mbtransfer(x, P = 1, Q = 1) will fit a lag-1 transfer function model on all the random splits.

Examples

data(sim_ts)
w0 <- cbind(sim_ts[[1]]@interventions, matrix(0, nrow = 1, ncol = 3))
w1 <- cbind(sim_ts[[1]]@interventions, matrix(1, nrow = 1, ncol = 3))
pd_result <- pd_splits(sim_ts, w0, w1, \(x) mbtransfer(x, 1, 1, nrounds = 10), n_splits = 1)
#> Training models for split 1/1