BYOB
(Bring your own Bias)
From SP4.27
Goal:
Determine smallest sample size likely to result in a normal-distribution tolerance interval narrow
2011
enough to fit within the specifications
Typical Usage: Process Validation Activities (e.g. OQ/PQ)
Approach: Iterative approach to arrive at minimum sample size for demonstrating a level of process
performance at a specified confidence level using tolerance intervals. For PQ sample size applies to each of the
3 runs separately. The user must enter an initial guess for sample size (start at 10 if you dont know where to
start). Software is calculating tolerance bound(s). User tries different sample sizes until the tolerance bound is
just inside the defined specification limit(s). This is the minimum sample size necessary.
Needed Before Determining Sample Size:
Justification for using normal distribution assumption (e.g. probability plot of relevant historical data, scientific
rationale for distribution shape)
Estimate for mean and standard deviation (e.g. pull from historical data)
Desired Confidence Level (site requirement typically 95%)
Process Performance you need to demonstrate (often called Reliability Level)
Specification limits
Risk of wrong answer: Too small sample size causes unnecessary test failure. Too large sample size wastes
resources.
Be conservative: A larger sample size can protect you from failing unnecessarily. Boost estimated standard
deviation estimate by 25% to 50% and move the mean closer to the specification to use as the input values.
This will lead to a larger sample size, decreasing the width of the computed tolerance interval.
Questions: Contact your local Subject Matter Expert or LeRoy Mattson
Example (one sided spec limit): What sample size (N) is needed to prove 95% confidence than at least
90% within tolerance? Preliminary data showed data was normally distributed with sample mean (xbar) of 8
and sample standard deviation (s) of 1.5. Process has a lower specification limit of 3.
Example
Input 1 = 10 (Initial
Guess)
Input 2 = 8
Input 3 = 1.5
Input 4 = 95%
Input 5 = 90%
Input 6 = lower spec
only
Size
=6
Example (two-sided spec limits): What sample size (N) is needed to prove 95% confidence than at least
90% within tolerance? Preliminary data showed data was normally distributed with sample mean (xbar) of 9
and sample standard deviation (s) of 1.5. Process has a lower specification limit of 3 and upper
specification limit of 13. (note: mean is closer to UCL than LCL)
Example
Input 1 = 10 (Initial
Guess)
Input 2 = 9
Input 3 = 1.5
Input 4 = 95%
Input 5 = 90%
Input 6 = Two-sided
Size = 13
N=12 is still above upper spec limit of 13
Initial Settings
Try it with 10 Samples
Result
Not so good! With that sample size
there is 95% confidence that the
lower bound will be .227 months!