A Tweedie family for stats::glm(), for responses with a mass at zero and a
continuous positive part, such as the loss cost (claims amount per unit of exposure) of
insurance policies, where most policies have no claim. Its variance is
\(\phi \mu^p\) for a power \(1 < p < 2\): a compound Poisson distribution of Gamma
claim sizes, the Poisson frequency and Gamma severity of a pure premium in one model.
Value
A family object, as stats::poisson() returns, with the family name
"Tweedie" and the power as variance_power.
Details
The family is quasi-likelihood: the coefficients and their standard errors are those of
the Tweedie distribution, with the dispersion \(\phi\) estimated from the Pearson
residuals, but its density has no closed form, so AIC() is NA. Compare
Tweedie models by their deviance, ideally on rows they were not fitted to. The unit
deviance is
$$2 \left( \frac{y^{2-p}}{(1-p)(2-p)} - \frac{y \mu^{1-p}}{1-p} + \frac{\mu^{2-p}}{2-p} \right).$$
See also
run_shiny_glm_fitting_tool, which fits it from a form.
Examples
set.seed(1)
n <- 2000
age <- sample(18:80, n, replace = TRUE)
claims <- rpois(n, exp(-2 + 0.02 * age))
loss <- vapply(claims, function(k) sum(rgamma(k, shape = 2, scale = 500)), numeric(1))
fit <- glm(loss ~ age, family = tweedie_family(1.5))
exp(coef(fit))
#> (Intercept) age
#> 165.849246 1.017522
summary(fit)$dispersion
#> [1] 80.73336
