The magic number of 120 (actually 119) for reference interval studies comes from the numbers needed to determine the 90% confidence intervals of the 2.5th and 97.5th percentiles of a population using non-parametric statistics. A smaller number will not allow you to make this calculation. Thus this number is the smallest number to see how uncertain your estimate is of these percentiles. Whether the accuracy obtained is sufficient  for  your needs is not a statistic question. Some additional factors - with a skewed distribution the 90% CI of the top reference limit can be very wide (but very narrow for the lower reference limit), but will be much narrower for a distribution closer to Gaussian. A larger n is needed for higher centiles (eg for the 99th centile the number is about 600 from memory). Other key statistical issues are outlier removal and transformation if using parametric statistics. Of course this is just the statistical part - the selection of population, whether to partition by age or sex or other,  pre-analytical and analytical quality can also strongly influence the reference interval determined.

Regards,

Graham

Although i am not a statistician but sample size for any study depends on:

1. Acceptable level of significance,  2. Power of the study,  3. Expected effect size 4. Underlying event rate in the population,  5. Standard deviation in the population.

You can read this article...Jennen-Steinmetz C1, Wellek S. A new approach to sample size calculation for reference interval studies. Stat Med. 2005 Oct 30;24(20):3199-212.

We have always used 20 samples to determine a normal range.  In a medical laboratory we already have a good idea of where normal should be.  We only have to develop a local range and mean.  But much like the previous poster said, your research criteria must depend on the listed factors.  20 is what our pathologist has established as a minimum number of samples for correlations, normal ranges, precision studies and quality control ranges.  

The 95% confidence interval is usually applied to analysis of laboratory data e.g  in the determination of reference ranges for biochemical parameters (Ojo, 1995: In manual of calculations in Clin. Chem. Pract). Like the Std deviation, approx 95% of the sample means will lie within 2X S.E.M. And we know that it is SEM that is used to calculate the confidence interval in a population. By implication,  the larger the sample size, the smaller the sample error of mean (SEM). Therefore, few sample size will bring about skewness in distribution and not normal distribution. As long as 120 samples might not be too few, i also think, if Prof. and his team can look at the prevalence of a condition under study, it might help in their choice of their sample size. Take for instance, in deterrmination  of sample size for reference range for glucose level in the population, prevalence of diabetes will be a factor to arrive at an acceptable reference interval for glucose (fasting or postprandial) in that population.  And in that instance,

N = [(1.96)2 pq]/d2  

could be applied.

where, N= sample size , p= prevalence rate (in % ), q=1-p, d = degree of precision ( here for 95% CI, it is set at  5% or 0.05)

NB: P could arbitrarily be set at  0.5 ( i.e 50:50), if this is unknown.

Hope this helps in a way.

Thank you.

 I agree :

1-To to establish

In Teitz Text of Clinical Chemistry  and CAP 120 is the minimum sample size given for Non-Parametric data but statisticians do not agree and say it is very small. Unfortunately they do not give their own formulae, so our projects are delayed.

How to calculate sample size while establishing Reference Values in Clinical Chemistry? - ResearchGate. Available from: https://www.researchgate.net/post/How_to_calculate_sample_size_while_establishing_Reference_Values_in_Clinical_Chemistry [accessed Oct 27, 2016].

2- to confirm 20 samples that's enough

Hope this helps you

Paschoalina Romano

Dear Aamir,

You may find some additional info in this attached paper which we published recently.

Kind regards

Tania

The magic number of 120 (actually 119) for reference interval studies comes from the numbers needed to determine the 90% confidence intervals of the 2.5th and 97.5th percentiles of a population using non-parametric statistics. A smaller number will not allow you to make this calculation. Thus this number is the smallest number to see how uncertain your estimate is of these percentiles. Whether the accuracy obtained is sufficient  for  your needs is not a statistic question. Some additional factors - with a skewed distribution the 90% CI of the top reference limit can be very wide (but very narrow for the lower reference limit), but will be much narrower for a distribution closer to Gaussian. A larger n is needed for higher centiles (eg for the 99th centile the number is about 600 from memory). Other key statistical issues are outlier removal and transformation if using parametric statistics. Of course this is just the statistical part - the selection of population, whether to partition by age or sex or other,  pre-analytical and analytical quality can also strongly influence the reference interval determined.

Regards,

Graham

HI Graham, what about parametric model? How many samples will be needed?

Minimum of 120 male and 120 female population are needed to establishing 95th percentile range (I.e., 2.5th and 97.5th percentiles) of a population using nonparametric tests