I have a dataset which contains 110 data points. This data set is divided in to three categories, say, A,B and C and each categories are sub-categorized into 3 to 5 categories: A1,A2,A3,A4,A5, B1,B2,B3, and C1,C2,C3, and C4. I split this dataset into two subsets. How would I know if each of the categories are well represented by the subsets and if the subsets represent the dataset well?
Thanks
Hello Rueben,
One appoach to comparing the two subsets would be to use a chi-square test of independence (subset membership = one dimension; membership in the 12 categories = the second dimension). The only concern is, that's a 2 x 12 table, and with only a total N of 110 cases, the test (for departures from equal respresentation across the 12 categories) won't have much statistical power.
If you're really concerned about having two subsamples with comparable category representation, you could use a stratified sample. For instance, if your subcategory A3 had 6 cases, you would randomly choose 3 for subsample #1 and assign the remaining 3 to subsample #2. That would assure equal representation.
Good luck with your work!
David Morse ,Thank you for this answer. It really helps.
Reuben -
Do you mean that each category is known, and the known data for each category are split into two groups in each case, and you want to know if either group could be used in each category as a good representative selection for that category? So perhaps each of the two groups in each category should have about the same mean and standard deviation, using close to the same number of data points in each of the paired subsets? I guess it depends on what matters most to you. I assume you want these two groups split in each category to match for some criteria, but it depends on what they are.
Or is that not what you want?
Cheers - Jim
James R Knaub, Thanks for answering my querries. Yes that is exactly what I want. However, all of the categories are not normally distributed, so I guess the mean could not represent the categories well because it was skewed. I was wondering which non-parametric test I should use? Thanks again
Hypothesis tests for such small sample sizes would have low power and not tell you much. It depends on what you want to do with these data, but you know any comparisons would be rather rough. I suppose you just might want some descriptive statistics. Means, medians, and perhaps histograms or some other graphics might be about all you can do. This may not be capable of giving you what you want.
- Badges
- Science topic