Currently as one of our KPI's we monitor the DOA (dead on arrival) and ELF (early life failure) rates. Currently we monitor these as percentages using the calculation:

No. of reported ELF's or DOA's/Total number of units shipped that month.

We have what a control level set for what we consider the maximum acceptable failure rate for these two items.

Some months a relativley small number of units may be shipped meaning that even 1 or 2 units being reported with either of these failure rates can send us over our target without there necessarily being a "real" issue.

To try and improve this, i.e. so when we go over target we can be sure it's not just 1 or 2 units skewing the percentage in a particular month, we are looking at changing to using a 3 month moving average, we typically have a spike in the number of units shipped at the end of each quarter for financial reasons so a 3 month moving average seems a sensible choice IMO.

My question is; is the 3 month moving average method I intend to use the best or even good method for doing this? Are there other methods that may be better ? and what are they?

Thought it would be better to get the opinion of some that are more knowledgable than myself before going to management with it.

Thanks in advance.

A rolling average is a good way to "even out" percentage spikes that can arise due to low volume. Of course as you've certainly seen, the risk is that taking too long a period could mask what you might otherwise call an issue.

For a couple indicators we settled on a slight variation on what you are considering: a rolling average over a given volume instead "per time period". The advantage is that it minimizes percentage fluctuation due to volume changes when there are long periods with low volume. The disadvantage is that you need to be careful choosing the volume that is right for your operation. Too large, you risk not seeing small systematic increases for a while. Too small, you still risk "false" spikes.

Some groups have just stuck with the "old way", but always plot volume with the indicator.

If you want to get geeky, you can simulate some confidence interval estimations to help you decide where to set your volume or time period. Based on a hypothetical percentage of defects found, you can determine how wide your CI will be for any given sample size and decide if that is sufficient for your needs.

Cheers.

use a p chart with variable control limits based on 'sample size' (number of shipments per month. that is what the chart is for. It works. no need for the other 'fancy' and yet misleading approaches.

you might want to search for an article that Steve Prevette posted on alternate analysis methods. I believe it's called: "Case study of differing analysis techniques on the same data" or something like that. It shows the dangers of rolling averages etc.

I cannot remember the details, but there is an alternate method for analyzing very low frequency failures that involves analysis of the time period between failures to detect changes rather than the failure rate.

I cannot remember the details, but there is an alternate method for analyzing very low frequency failures that involves analysis of the time period between failures to detect changes rather than the failure rate.

One way is documented in Dr. Wheeler's book - Understanding Variation, and may be in a few of his others. I have posted the way I do it at Hanford using Dr. Wheeler's method at http://www.hanford.gov/rl/uploadfiles/VPP_TrendLow-Rate.pdf

another similar approach is covered in the thread
http://elsmar.com/Forums/showthread.php?t=18249&highlight=geometric+distribution
on "cumulative count" charts for very low frequency. However, whether or not you use Wheeler's approach or the approach I've referenced be sure your situation applies. if the failure rate averages less than 1 in a measured period then you would use the charts for rare events, otherwise use the p chart.