The developers of fitdistrplus got back to me already with a code fix. The new version (1.0-12) is available from R-forge using
install.packages("fitdistrplus", repos="http://R-Forge.R-project.org")So now with fitdistrplus loaded in global.dcf I can run
mydata <- rnorm(20, 0, 10)
mydata_cens <- data.frame(left = floor(mydata), right = ceiling(mydata))
dmynorm <- function(x, mean, sd, log = FALSE) dnorm(x, mean, sd, log)
pmynorm <- function(q, mean, sd, lower.tail = TRUE, log.p = FALSE) pnorm(q, mean, sd, lower.tail, log.p)
fitdistcens(mydata_cens, "mynorm", start = list(mean = 0, sd = 10))Fitting of the distribution ' mynorm ' on censored data by maximum likelihood
Parameters:
estimate
mean -0.0008706237
sd 9.6356247206They also explained that for “standard” distributions such as “norm”, the default mle start matches the mean and sd of the data, so that is the source of the apparent discrepancy.