I want to divide a group of 20 participants into 2 subgroups based on their median performance value on a continuous variable, and then run t-test. Is this all right? If not, what is the problem, and how can can I devise a better mode of analysis? Do I have to turn to logistic regression?
Artificially creating two groups with a median split is rarely a good idea b/c you are giving up information and thus losing statistical power (e.g., someone 1 point out of 50 above the median is in the same group as someone who has a max score of 50 and is much higher on this variable).
You would get more out of your data doing continuous analyses, such as correlations between this continuous variable and other study variables, rather than a t-test.
I agree with Michael, and here are two references on this topic.
"Dichotomizing continuous predictors in multiple regression: a bad idea" by Royston et al. (2006).
"On the practice of dichotomization of continuous variables" by MacCallum et al. (2002).
Instead, I would start by doing a correlation between your two variables, as well as examining a scatterplot to make sure that you do not have a non-linear relationship.
try this article
Article TO DETERMINE SKEWNESS, MEAN AND DEVIATION WITH A NEW APPROAC...
I agree with Michael. Median split is rarely a good idea. Correlations, or regression analysis, are much better alternatives to describe the relationship between two continuous variables.
Completely agree with Francis as The test has low power (efficiency) for moderate to large sample sizes. Correlations, or linear regression is better idea to describe the relationship between two continuous variables.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics