R/reset_coefs.R

In margins: Marginal Effects for Model Objects

# utility function to reset the coefficients in a model object, and return the modified model
reset_coefs <- function(model, coefs) {
    UseMethod("reset_coefs")
}

reset_coefs.default <- function(model, coefs) {
    # in basic model classes coefficients are named vector
    model[["coefficients"]][names(coefs)] <- coefs
    model
}

reset_coefs.lm <- function(model, coefs) {
    # in lm coefficients are named vector
    model[["coefficients"]][names(coefs)] <- coefs
    model
}

reset_coefs.glm <- function(model, coefs) {
    # in glm coefficients are named vector
    model[["coefficients"]][names(coefs)] <- coefs
    ## But, there's an edge case!! When `predict(model, se.fit = TRUE)` is called without `newdata`, `predict.lm()` isn't called.
    ## Instead `model$linear.predictors` is returned directly if `type = "link"` and
    ## `model$fitted.values` is returned directly if `type = "response"`.
    ## `marginal_effects()` for "glm" is always called with `newdata`, so we won't hit this.
    model
}

reset_coefs.betareg <- function(model, coefs) {
    # in betareg, coefficients are a two-element list. We want to substitute the first element!
    model[["coefficients"]]$mean[names(coefs)] <- coefs
    model
}

reset_coefs.merMod <- function(model, coefs) {
    # in 'merMod', predictions work the slot called "beta", which is unnamed
    # `fixef(model)` returns the same thing named
    requireNamespace("methods")
    beta <- methods::slot(model, "beta")
    beta[match(names(coefs), names(lme4::fixef(model)))] <- as.numeric(coefs)
    methods::slot(model, "beta") <- beta
    model
}

reset_coefs.lmerMod <- reset_coefs.merMod