Ordinal data provides non-dichotomous and multi-levelled outcome. It is widely used in medical scoring systems. However, being "ordinal" means that author can't just simplify outcome result by "average/mean", because the very same ordinal number may have different interpretation between individuals who carried the value.
And I feel this become a problem when calculating sample size.
The standard, widely used "compare two proportion" sample size formula is out of question because it means I have to re-categorize the ordinal results into dichotomous categorical outcome.
And the "compare two means" formula, n = (Zα/2 + Zβ)2 *2*σ2 / (μ2-μ1)2 , while intuitive, requires the input of μ, which is simply not really ordinal.
So, can someone kindly share how to actually calculate sample size for ordinal data "in the correct way?"
Use compare two proportions this is only a guess anyway and you don't have to recategorize your data during your real study so use p=q =.5 which gives the largest n. Calculate the variance by approx. Of your expected numbers and increase that value to take care of extra uncertainty. Best wishes David Booth
Hi,
Tests of proportions are applicable.
Sample Size Calculations in Clinical Research, Third Edition by Shein-Chung Chow and Jun Shao
You are trying to calculate the sample size needed to get a significant test result for some effect size and variability in the data ?
My guess is that the only practical (?) known (?) way would be to run simulations.
But it might be serviceable to treat the ordinal dependent variable as numeric, and use the appropriate formula for the experimental design.
Since you mentioned medical research, I assume this is patient data with unique identifiers. The sample size is the number of unique identifiers. It has nothing to do with whether some of the variables are ordinal.
If you meant to ask a different question, please delete this one and write a new one.
- Badges
- Science topic
- Similar topics
- Papers