I am developing single use sensors to test a concentration of an analyte in a sample. I am interested in the reproducibility of the individual sensor batches and the subsequent probability to find results of single sensors of this batch within an specific range.


I am producing a batch of sensors (~2000) and then test 5 sensors. The sample results give me a mean and a sample variance. Now I am interested to calculate the probability of any single sensors measured in the future to give me a result with a deviation larger then 10% form the expected value.

Calculation approach

I estimate the population variance from the n=5 sample and then assuming a normal distribution for any future measurements. I confirmed the normality by a shapiro wilk test of the sample residues.

Next I calculate the Z value of my 10% boundary from the mean and the estimate of the population variance from the sample and derive the corresponding probability. I have particular troubles how to deal with the situation that I want to predict the result of just one sensor, no average of several sensors which would trigger me to use a T statistic given the small sample size.

I also considered to use the Chi Squared upper 95% CI limit for the sample variance as estimate for the worst case population variance for calculation of the Z statistic, which results obviously in a higher probability to find a sensor outside of the 10% boundary.


  1. Is it correct to assume normal distribution of future single measurement and hence to use a Z statistic?
  2. Is the estimate for a population variance from a n=5 sample suitable for calculating the Z statistic or do i need to correct somehow for the small sample size?
  3. Is a Chi Squared derived upper 95% CI limit for the sample variance more precise in estimating the population variance when i am interested in the worst case scenario?
  4. Is there an alternative approach to the problem

I hope I have described the problem detailed enough and I am really looking forward to your input.


1 Answer 1


Welcome to the site!

Without more information, yes, a normal distribution sounds fine. There are scenarios where other models work better, but this is if nothing else a good, standard starting place.

I'd recommend what you mentioned in #3, correcting for population size as you are constructing the Gaussian model; Use the worst mean/stddev for a given % chance occurrence. (Don't forget to pessimistically estimate mean too, using that same fudge factor.) Then compare how bad those "worst 0.x% of items" are expected to be vs requirements.

  • $\begingroup$ Dear Andrew, thanks for having me and confirming my approach. If I understand you correctly I should derive a fudge factor, like the 95 CI limit from a t statistic on the mean, right? To give some additional information: I use the same sample set to calibrate the sensor batch so the mean of this sample will be equivalent to the requirement. I am not sure how to integrate a pessimist mean. I have thought about using the t test derived 95CI of the sample mean as a measure of spread of the population. I just not sure how to work with this in predicting p of future results. $\endgroup$
    – Fons
    May 13, 2020 at 18:29
  • $\begingroup$ Correct. The overall idea is whatever method you use to "be pessimistic about" your variance should be consistent with how you estimate the mean. $\endgroup$ May 13, 2020 at 19:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.