MLE Hotwire Article

[Aurora Tiffany-Davis]

In Issue 9, November 2001 of Hotwire, it is stated...

"...our recommendation is to use rank regression techniques when the sample sizes are small and without heavy censoring. When heavy or uneven censoring is present, when a high proportion of interval data points are present and/or when the sample size is sufficient, MLE should be preferred."

I have read elsewhere that "sufficient" sample size for MLE ranges between 30 and 100 samples.

In my application, all sample sizes will be between 5 - 15. Therefore I am concerned about using MLE. My question is - how much censoring is "heavy censoring"? And, what does "uneven censoring" refer to?

Just one example would be a life test with 6 samples with 5 (small) interval failures and 1 suspension. Which method would be preferable in this case? What if I had 3 suspensions out of 6 samples?

Thank you.

[Tarik El-Azzouzi]

Here is what I heard. Statisticians say that sample sizes about 30 is a large size. Engineers are more brave, they say above 18 is a large size.

Uneven censoring is when the censor times are spread across the data of failure times. Ex: F at 200, S at 230, F at 340, S at 379, F at 580, S at 600…

As for your example, if the intervals are significantly narrow compared to the precision of the results you are trying to get, in other words if the granuality of the data is coarser then the desired results. For example, the intervals are of a few hours, whereas the desired results are in month.

So if you have 5 failure times (small intervals) and one failure, you are probably better off with the Regression method.

If however you have a sample size of 6 and 3 suspensions, I think that MLE is better.

Note that you can compare MLE and Regression results.

[Aurora Tiffany-Davis]

Thanks for the