upsis

upsis updates fitted brms models after modest additions or removals of data using Pareto-smoothed importance sampling (PSIS). Updating with PSIS gives an effecient way to "upsize" the dataset used for model fitting, without a full re-fit.

The package computes the change in log likelihood implied by the requested data update, smooths the resulting importance weights with loo::psis(), and stratified-resamples posterior draws back into the fitted brmsfit object. The updated model can then be used with familiar brms post-processing tools.

library(brms)
library(upsis)

fit <- brm(y ~ x, data = old_data)
result <- upsis(fit, data_add = new_data)

summary(result$updated_model)
result$pareto_k

upsis is an approximation to refitting the model on the changed data. It is best suited to small, weakly influential updates that do not change the model structure.

Prefer a full refit when:

• the Pareto k diagnostic is high;
• the new data are highly influential or much larger than the original data;
• the update changes data-dependent priors, spline knots, basis expansions, or other model components chosen during the original fit;
• the update introduces unsupported new grouping levels or otherwise requires new parameters.

Run the lightweight test suite with:

devtools::test()

The optional brms integration test is skipped by default. To run it locally, set RUN_BRMS_TESTS=true before testing.