
Enforce monotone tariff relativities for numeric bands or ordered categories
Source:R/model_refinement_monotonicity.R
add_monotonicity.RdProject the current level effects of a standalone categorical factor onto a
non-decreasing or non-increasing sequence. This is a post-estimation
adjustment, not likelihood-constrained GLM estimation. Unlike
add_smoothing(), it needs no continuous source variable, draws no curve
between levels and does not impose smoothness.
Arguments
- model
A refinement specification created by
prepare_refinement().- model_variable
Name of a standalone factor in a log-link GLM with an intercept. Numeric labels (e.g.
"100","200","1000") and numeric intervals (e.g. levels created bycut()) are ordered numerically, not alphabetically, including for ordered factors. A conflicting ordered-factor level order produces a warning; numeric order is used for the projection. Intervals must not overlap, even for ordered factors. For non-numeric labels, declare the order withordered(x, levels = ...)before fitting; that explicit order is retained.- direction
Desired order of relativities in the numeric or declared level order. Both directions permit ties.
- method
"isotonic"minimizes the weighted sum of squared changes to log-relativities by pooling adjacent violating levels. Alternatively,"increase_only"raises only levels that must rise;"decrease_only"lowers only levels that must fall.- weights
Optional name of a non-negative numeric volume column. Weights are summed by level; each total must be finite and positive.
NULLgives each level equal weight, irrespective of its row count. Weights are recorded but do not affect the two one-sided methods.- output_variable
Name of the new numeric tariff column used as a fixed offset at refit and displayed in
rating_table(). Defaults to the model variable name followed by_monotone, with a unique suffix (_1,_2, ...) if needed. An explicitly supplied name must be new.
Value
A rating_refinement with an appended step; no GLM is fitted.
The step's values table contains the ordered levels, input and adjusted
relativities and level weights. Use ggplot2::autoplot() for a preview.
Details
For grouped insured amount, reconstruction value or similar
variables, pass the existing factor used in the GLM, just as the
model_variable of add_smoothing(). No source_variable or new breaks
are needed: existing categories are retained and only their relativities
are adjusted. A continuous numeric slope is not supported.
A treatment-coded reference level enters the projection with its
current relativity (initially one); it can move. For polynomial contrasts
the input is the exponentiated fitted level contribution, not an
individual contrast coefficient. Projected values are not renormalized.
refit() fixes their ratios and estimates the intercept and remaining
free coefficients. Later explicit refinements can change these ratios.
Repeated calls for the same model_variable append separate steps and
use the current relativities, including any intervening refinements.
Earlier steps remain in the history; the latest projection replaces the
preceding effect at refit, rather than multiplying it again. Applying the
same direction to an already monotone effect leaves its values unchanged.
Examples
portfolio <- data.frame(
insured_amount = c(50000, 150000, 750000, 1500000, 3000000),
claims = c(100, 115, 110, 135, 130), exposure = rep(1000, 5)
)
portfolio$amount_band <- cut(portfolio$insured_amount,
breaks = c(0, 100000, 500000, 1000000, 2000000, Inf))
# cut() creates an ordinary factor; its numeric intervals provide the order.
fit <- glm(claims ~ amount_band + offset(log(exposure)), poisson(), portfolio)
spec <- prepare_refinement(fit) |>
add_monotonicity("amount_band", weights = "exposure")
spec$steps[[1]]$values
#> level original_relativity weight adjusted_relativity
#> 1 (0,1e+05] 1.00 1000 1.000000
#> 2 (1e+05,5e+05] 1.15 1000 1.124722
#> 3 (5e+05,1e+06] 1.10 1000 1.124722
#> 4 (1e+06,2e+06] 1.35 1000 1.324764
#> 5 (2e+06,Inf] 1.30 1000 1.324764
rating_table(refit(spec))
#> risk_factor level est_model
#> 1 (Intercept) (Intercept) 0.1000174
#> 2 amount_band_monotone (0,1e+05] 1.0000000
#> 3 amount_band_monotone (1e+05,5e+05] 1.1247222
#> 4 amount_band_monotone (5e+05,1e+06] 1.1247222
#> 5 amount_band_monotone (1e+06,2e+06] 1.3247641
#> 6 amount_band_monotone (2e+06,Inf] 1.3247641
# Rechecking appends a step without overwriting the earlier specification.
rechecked <- spec |>
add_monotonicity("amount_band", weights = "exposure")
vapply(rechecked$steps, function(step) step$output_variable, character(1))
#> [1] "amount_band_monotone" "amount_band_monotone_1"