Calculate partial effects based on a model with interactions

Usage

tl_interaction_effects(model, var, by_var, at_values = NULL, intervals = TRUE)

Arguments

model: A tidylearn model object
var: Variable to calculate effects for
by_var: Variable to calculate effects by (interaction variable)
at_values: Named list of values at which to hold other variables
intervals: Logical; whether to add 95\ lower and upper. This needs a model whose underlying fit is an lm or glm; for any other fit a message is shown and only point estimates are returned. For a glm the interval is built on the link scale and transformed to the response scale.

Value

For numeric var: a list with effects (data frame of predicted values across the variable range for each value of by_var) and slopes (data frame with the slope of var at each value of by_var). For categorical var: a data frame of predicted values at each factor level for each level of by_var. A numeric by_var is evaluated at its quartiles; quartiles that tie are evaluated once, with a label naming each quartile they stand for, such as "Q0/Q25".

fit, lower, upper and slope are on the response scale whatever intervals is set to: predicted probabilities for a logistic model. slope is the slope of a straight line fitted to fit across the range of var, so for a non-linear link it is an average rate of change over that range.

slopes$slope_se is the standard error of a straight line fitted to the prediction grid, not the sampling uncertainty of the marginal effect. For a linear model the grid is exactly linear in var, so this is near zero by construction and should not be read as a precise estimate. Use summary(model$fit) for inference on the interaction coefficient itself.

Examples

# \donttest{
model <- tl_model(mtcars, mpg ~ wt * hp, method = "linear")

# How the effect of weight changes across horsepower
effects <- tl_interaction_effects(model, var = "wt", by_var = "hp")
head(effects$effects)
#>         wt hp      fit    lower    upper by_value by_label
#> 1 1.513000 52 33.32234 30.78023 35.86445       52       Q0
#> 2 1.552505 52 33.05495 30.57550 35.53440       52       Q0
#> 3 1.592010 52 32.78756 30.37004 35.20508       52       Q0
#> 4 1.631515 52 32.52017 30.16379 34.87655       52       Q0
#> 5 1.671020 52 32.25278 29.95669 34.54887       52       Q0
#> 6 1.710525 52 31.98539 29.74867 34.22211       52       Q0
effects$slopes
#>      by_value by_label     slope     slope_se
#> Q0       52.0       Q0 -6.768521 5.695734e-16
#> Q25      96.5      Q25 -5.529278 4.387243e-16
#> Q50     123.0      Q50 -4.791302 6.071491e-16
#> Q75     180.0      Q75 -3.203958 4.018576e-16
#> Q100    335.0     Q100  1.112505 4.116338e-16

# slopes$slope_se describes the fitted grid, not the sampling
# uncertainty of the marginal effect -- for that, read the coefficient
summary(model$fit)$coefficients["wt:hp", ]
#>     Estimate   Std. Error      t value     Pr(>|t|) 
#> 0.0278481483 0.0074195805 3.7533319407 0.0008108307 
# }