Clips predicted retention times by fitting a log-normal distribution to the observed training RTs and bounding predictions to the central 99.99% interval. All observed RTs must be positive to estimate the distribution. If the estimated lower bound would be negative, it is replaced by 1% of the observed minimum RT instead.
Examples
# Draw only a few samples (10) and clip based on these. The allowed range will
# be much bigger than the observed range.
set.seed(42)
y <- rlnorm(n = 1000, meanlog = 2, sdlog = 0.1)
yhat <- y
yhat[1] <- -100 # way too low to be realistic
yhat[2] <- 1000 # way too high to be realistic
yhat <- clip_predictions(yhat, y)
range(y) # [ 6.18, 8.93]
#> [1] 5.274195 10.480647
yhat[1:2] # [ 4.96, 10.61] # Limited by theoretical bounds
#> [1] 5.274195 10.480647
# Draw more samples (1000) and clip based on these. The allowed range will
# be almost identical to the observed range.
set.seed(42)
y <- rnorm(n = 100, mean = 100, sd = 5)
yhat <- y
yhat[1] <- -100
yhat[2] <- 1000
yhat <- clip_predictions(yhat, y)
range(y) # 83.14, 117.47
#> [1] 85.03455 111.43323
yhat[1:2] # 83.14, 117.72
#> [1] 84.06834 119.01241