Hi everyone, I was wondering if someone might have some advice on the best stats test to use? I have body condition scores from individuals from 2 populations and need to know if they are statistically different. (For example is the average body condition of population 1 different to population 2.) I only have weight scores (0-3), not absolute weights.
n for both is over 30.
The data is skewed with fewer individuals in the low weight scores.
So just as an example :
Population 1
Weight Score.......Number of individuals
0........................ 2
1........................ 3
2 ....................... 11
3 ....................... 28
Population 2
0....................... 6
1....................... 4
2....................... 18
3....................... 22
I thought I could use an independant t-test to check if the body condition was statistically different between the two populations, but have been told by a statistician I cannot analyse this data using a t-test (not due to issues with normality of the data, but because it cannot be used for data like this.) Since some groups are less than 5 I don't think a chi square would be useful either. Bit confused how it can be best analysed?
I was also wondering if it is ok to consider using a linear regression where one of the two measurements is a categorical (ordinal) value and the other is continuous? For example looking at a relationship between body burden of a pollutant and weight scores?
Thank you.
Because you are faced with the frequency and number, it is better to use a variety of Contingency tests.
The important point is, what is the difference between your sample A and B? Are they case-control? Has an intervention been performed on one? Are they just a sample that has been measured twice? Are they pre and post?
The answer to the question of what exactly is the conceptual difference between your two samples A and B will be very effective in choosing the correct statistical method (of course, in any case of the contingency type).
The "best stats" (as you asked for) to my knowledge for such data is an ordered logistic model.
For a chi²-test you can unify the two categories with the lowest counts, so that you get a 3x2 table with all expected counts > 5. The p-value is 0.16. If you don't unify (i.e. you use your original 4x2 table), you can bootstrap the p-value, what gives 0.24 (what is in fact the same as using the standard test).
This is not statistically significant, so there is nothing to interpret and so there is no problem with interpretation. But if the result would have been significant, then the interpretation is difficult. You yould then check the residuals to find out which of the categories contributed most (and in what direction).
One alternative is the Mann Whitney U test, which is the nonparametric version of Student's t. It was specifically developed for comparing two groups on an ordinal measure, when the t test cannot be used. The fact that the data are skewed does not matter with the U test.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression