heplots (version 1.8.5)

stdmodel: Standardize a Fitted Linear Model

Description

stdmodel() refits a fitted lm/mlm object on standardized (mean 0, SD 1) versions of its response variable(s) and its numeric predictors, leaving factor-coded predictors on their original (raw 0/1 dummy) scale. This is the engine behind stdcoef() for multivariate models, and behind coefplot.mlm(..., std = TRUE).

Since the result is just an ordinary refit "lm"/"mlm" object, it also works with other tools that expect one, e.g. lmtest::coeftest() and broom::tidy.coeftest() on its result.

Usage

stdmodel(object, ...)
# S3 method for lm
stdmodel(object, ...)

Value

The refit model, of the same class as object

Arguments

object

A fitted "lm" or "mlm" object

...

Additional arguments. Not used.

Author

Michael Friendly

Details

Only the response(s) and numeric predictors are standardized; factor predictors are left as-is, since "one SD" isn't a meaningful unit for a 0/1 dummy variable. This still gives an interpretable, commonly-used effect size for factor predictors: "SD units of y per unit change in the predictor."

Interactions and main effects of raw variables (e.g. SES * (n + s)) are handled correctly, since stats::model.frame() reduces a formula down to its base variables before any interaction terms are built -- refitting on the standardized base variables reconstructs the interaction terms correctly.

Computed/transformed predictor terms (e.g. log(x), poly(x, 2)) are not supported and raise an error, since standardizing the already computed column would not do what a user expects once the model is refit and the original expression is re-evaluated. Similarly, a computed univariate response (e.g. log(y) ~ ...) is not supported. Pre-compute these if you want to use such variables.

subset=/weights=/na.action= in the original fitting call are not specially handled -- stdmodel() is designed for models fit directly via lm(formula, data = ...).

See Also

stdcoef(), coefplot.mlm(), lmtest::coeftest(), broom::tidy.coeftest()

Other multivariate linear models: coefplot(), glance.mlm(), stdcoef()

Examples

rohwer.mod <- lm(cbind(SAT, PPVT, Raven) ~ SES + n + s + ns + na + ss, data = Rohwer)
coef(rohwer.mod)

rohwer.std <- stdmodel(rohwer.mod)
coef(rohwer.std)