6

I have a sample of measurements from a real life device which misses all the measurements that are less than some threshold (given device is not precise enough).

From theory and also measurements from more precise devices I know that the data distributed very closely to Gumbel distribution.

I found parameters of a Gumbel distribution for a given sample with MLE and now I want to generate "perfect" missing values and add them to the sample.

I am not sure what is the easiest way to do that in R. I have my sample and sample size, obviously. I want to create a copy of it and then add some values using theoretical PDF.

Xi'an
  • 90,397
  • 9
  • 157
  • 575
Igor Korotkov
  • 63
  • 3

1 Answers1

4

I'm assuming that the values are truncated below the threshold, $t$ rather than censored below $t$ (that is, you don't know how many there are below the threshold).

Let the number of points observed above the truncation point be $n_o$.

A simple approach could go as follows:

  1. Estimate $p$, the proportion of the distribution below the threshold from the fitted parameters and the truncation point (this ignores the uncertainty in the parameter estimates; however if your sample is large that may not be a terrible approximation)

  2. Simulate $N_u$, the number of missing values from the negative binomial $NB(n_o,p)$, obtaining the specific value $n_u$.

  3. Simulate $n_u$ values from the truncated Gumbel on (0,$t$); for example you might consider some form of accept-reject.

Glen_b
  • 257,508
  • 32
  • 553
  • 939