Model coefficients as a tibble

Returns the coefficients of a fitted tidylearn model as a tibble, with standard errors, test statistics, p-values and – on request – confidence intervals. Available for "linear", "polynomial", "logistic", "ridge", "lasso" and "elastic_net". Other methods have no coefficients; use tl_table_importance for those.

Usage

tl_coefficients(
  model,
  conf_int = FALSE,
  level = 0.95,
  exponentiate = FALSE,
  lambda = "1se"
)

Arguments

model

A tidylearn supervised model object from tl_model.

conf_int

Whether to add conf_low and conf_high columns (default FALSE). Not available for regularised methods.

level

Confidence level for the interval (default 0.95). Ignored unless conf_int = TRUE.

exponentiate

Whether to report estimate and the interval on the odds scale rather than the log-odds scale (default FALSE). Only meaningful for a classification model, whose coefficients are log odds. The standard error stays on the log-odds scale it was computed on and is renamed std_error_log to say so.

lambda

For regularised methods: "1se" (default), "min", or a numeric penalty value.

Value

A tibble, one row per model term. For "linear", "polynomial" and "logistic": term, estimate, std_error, statistic, p_value, plus conf_low and conf_high when conf_int = TRUE. For regularised methods: term, estimate and the lambda the estimate came from – glmnet reports no standard errors, so there is nothing to test or bound. A multiclass regularised model has one set of coefficients per class, so its rows are led by a class column; exponentiate is not available for it, because glmnet's multinomial coefficients are not relative to a reference class.

Details

Intervals are Wald intervals, computed from the same standard errors as the statistic and p_value columns beside them, so the interval and the p-value always agree about whether zero is excluded. For "linear" and "polynomial" that means t quantiles on the residual degrees of freedom, which is exactly what confint returns for an lm. For "logistic" it means z quantiles, which is what the reported z statistic implies but not what confint() gives – that profiles the likelihood, which is the better interval when the sample is small or a class is nearly separated. Call stats::confint(model$fit) when you want it.

A rank-deficient fit – two perfectly collinear predictors, or an interaction of factors with a combination no row has – cannot estimate every term. Those terms are returned with an NA estimate rather than dropped, so a term named in the formula never disappears from the output without saying so.

See also

tl_table_coefficients for the same numbers as a formatted table.

Examples

model <- tl_model(mtcars, mpg ~ wt + hp, method = "linear")
tl_coefficients(model)
#> # A tibble: 3 × 5
#>   term        estimate std_error statistic  p_value
#>   <chr>          <dbl>     <dbl>     <dbl>    <dbl>
#> 1 (Intercept)  37.2      1.60        23.3  2.57e-20
#> 2 wt           -3.88     0.633       -6.13 1.12e- 6
#> 3 hp           -0.0318   0.00903     -3.52 1.45e- 3
tl_coefficients(model, conf_int = TRUE)
#> # A tibble: 3 × 7
#>   term        estimate std_error conf_low conf_high statistic  p_value
#>   <chr>          <dbl>     <dbl>    <dbl>     <dbl>     <dbl>    <dbl>
#> 1 (Intercept)  37.2      1.60     34.0      40.5        23.3  2.57e-20
#> 2 wt           -3.88     0.633    -5.17     -2.58       -6.13 1.12e- 6
#> 3 hp           -0.0318   0.00903  -0.0502   -0.0133     -3.52 1.45e- 3

# Odds ratios from a logistic fit
am_data <- transform(mtcars, am = factor(am))
model <- tl_model(am_data, am ~ wt, method = "logistic")
tl_coefficients(model, conf_int = TRUE, exponentiate = TRUE)
#> # A tibble: 2 × 7
#>   term           estimate std_error_log conf_low conf_high statistic p_value
#>   <chr>             <dbl>         <dbl>    <dbl>     <dbl>     <dbl>   <dbl>
#> 1 (Intercept) 169460.              4.51 24.6       1.17e+9      2.67 0.00759
#> 2 wt               0.0179          1.44  0.00107   2.99e-1     -2.80 0.00509