A metric of a numeric response and its predictions

The statistics biomod2 reads an abundance model by. With e = y - p:

Usage

regression_metric(y, p, metric = names(.regression_metrics))

Arguments

y

Observed values.

p

Predictions for the same units, in the same order.

metric

One of the names in the table.

Value

One number, or NA where the cell defines none.

Details

metric	reads
"r_squared"	1 - sum(e^2) / sum((y - mean(y))^2), NA where y is constant
"pearson"	the correlation of y and p, NA where either is constant
"rmse"	sqrt(mean(e^2))
"mse"	mean(e^2)
"mae"	mean(abs(e))
"max_error"	max(abs(e))
"poisson_deviance"	mean(2 (y log(y / p) - (y - p))), the logarithm taken as zero at y = 0; NA where a prediction is negative or is zero beside a count above zero

A comparison across candidates reads the highest score as the best, so the five errors are registered under the names neg_rmse, neg_mse, neg_mae, neg_max_error and neg_poisson_deviance with their sign reversed, and r_squared and pearson under their own names. A cell holding a prediction that is not a number scores NA.

Examples

y <- c(1, 2, 3, 4)
p <- c(1.5, 2, 2.5, 5)
regression_metric(y, p, "rmse")
regression_metric(y, p, "r_squared")