categorical age or year-age?

Hi John: I decided to use 3-level categorical age: children, younger adolescents, older adolescents. I also included year-age in the stratified analysis. Both year-age and categorical age significantly correlated with treatment adherence (N = 300), however when they were stratified into 3 age groups (each n is approx. 100), year-age was no longer significant. This is done by using logistic regression. So, I concluded that categorical age was better than year-age in this study. Am I right or am I too hasty to jump into conclusion. There is a possibility that I did not have enought people in each group? Thank you for your help.


TS Contributor
The reason why year-age was no longer significant when you stratified them into age-categories is because the correlation coefficient is affected by the range of the independent variable.

In effect, you shortened the year-age variable to the "range" of the particular categories, thus reducing the correlation coefficient (you are analyzing year-age within each category, or in other words, you are analyzing year-age while controlling for age-category).

Pick either year-age or age-category and go with it - putting both of them in the model will only confuse the issue - they are both measuring the same thing, in a different way....
I understand, John. But I have to justify it... Let's say I used only a year-age or continuous age and found Pearson corr significant in relation to treatment adherence. After that I tried to see how it is by using only categorical, and I saw the corr to be more significant. I also wanted to find out that within each group if one year increase made a difference, and it did not except in older adolescence. So, this tells me, categorical age was more appropriate when I investigate children and younger adolescents. Within each age group, age does not make any difference. So, categorical age was appropriate... this is my conclusion. One reviewer said that I could not make such a conclusion because the sample size 300 for all is much bigger than the size in each group; so I did not see year-age was significant perhaps because I did not have enough sample size. But I remember you saying that Pearson corr can be used in smaller number like 100. Thank you, and I am sorry that I am not able to explain in a succinct way...


TS Contributor
Not a problem - it appears that you've done what you can with the data you have - I really can't do anything about objections from reviewers, and their comments on sample size may be valid....

It's certainly possible that you don't have enough samples to detect (if it in fact exists) the year-age effect within certain categories....