Compare paired prediction errors with a year-block bootstrap
Description
Resamples consecutive circular blocks of years, retaining all regions in each sampled year. The interval describes uncertainty in a paired RMSE difference on the supplied evaluation period; it is not a guarantee under nonstationarity. Few observed years give weak uncertainty estimates.
Usage
compare_crop_errors(
predictions,
reference,
repetitions = 2000L,
block_length = 2L,
confidence = 0.95,
seed = 2718L
)
Arguments
predictions |
Data frame with 'Method', 'RS', 'Year', 'Observed', and 'Predicted'. Each method must have the same region/year observations. |
reference |
Reference method name. |
repetitions |
Number of bootstrap resamples, at least 100. |
block_length |
Number of consecutive years per block. |
confidence |
Confidence level between zero and one. |
seed |
Integer random seed; the caller's random state is preserved. |
Value
A data frame with paired RMSE differences (method minus reference), percentile limits, observation/year counts, and bootstrap settings.
Examples
p <- expand.grid(RS = c("A", "B"), Year = 2001:2008,
Method = c("reference", "alternative"))
p$Observed <- 10 + p$Year - 2000
p$Predicted <- p$Observed + ifelse(p$Method == "reference", 2, 1)
compare_crop_errors(p, "reference", repetitions = 100)