CanonMan –
Your organizational procedures should set some form of standard for this. If your company has not established this, you should solve that problem before you begin worrying about a specific process. Otherwise, you may end up with lots of different criteria applied to different processes that can leave your validation program in a questionable state (it eliminates the temptation to pick an arbitrary sample size and then attempt to justify it after the fact).
An example of such a system might be as follows:
For critical defects/characteristics: 95% confidence/99% reliability
For major defects/characteristics: 95% confidence/95% reliability
For minor defects/characteristics: 95% confidence/90% reliability
Once you have established these guidelines, there is a table in Montgomery that can be used to determine a k factor and sample size for continuous data (assuming a normal distribution) and you can utilize a binomial table for pass/fail data (it is a good idea to copy these into your validation procedures so that everyone in your organization is given the same information).
An alternative to the above, comes from Taylor (http://www.variation.com ; visit the tech library, article “Selecting Statistically Valid Sample Sizes”). If your organization has already established AQLs for your defects/characteristics (i.e. critical – .065%; major – 1.0%; minor – 2.5%, etc.) he advises using your AQL as your LTPD for the first three lots and then transitioning to a less severe sampling plan after validation is completed. He presents examples using attribute (binomial) data, but the same concept is easily applied to contiuous data.
The final method your organization may consider is to set Cpk goals for your validations (again these are typically proportional to the criticality of the characteristic) and then use an alpha to beta ratio to determine the sample size necessary to claim you have achieved it as a minimum. This procedure is also outlined in Montgomery.
Any of the above approaches would work and can be statistically justified. Just remember when dealing with continuous data you need to check your distributional assumptions to ensure the model (typically a normal distribution) you used when determining your acceptance criteria and sample sizes were correct.
.
