Conflict results from "cc" command?

[Zihan Dong]

Hello everyone,

I was running cc command, and got the output below:

The test of homogeneity suggests that the 3 ORs are homogeneous; but the test of combined OR=1 suggests that there is decreasing trend of OR with age increasing.

Any insight about this? Thank you.

[Leonardo Guizzetti]

You have misinterpreted the test that the combined OR=1 as a test of trend, which it is not. Each stratum-specific odds ratio is very roughly 3.0, which the M-H combined estimate tells you, and that they are compatible with the hypothesis that they are homogeneous. However, the OR is still greater than 1 in its point estimate, and the second test tells you is statistically different from 1.

[Zihan Dong]

You have misinterpreted the test that the combined OR=1 as a test of trend, which it is not. Each stratum-specific odds ratio is very roughly 3.0, which the M-H combined estimate tells you, and that they are compatible with the hypothesis that they are homogeneous. However, the OR is still greater than 1 in its point estimate, and the second test tells you is statistically different from 1.

Thank you for your explanation Leonardo! Now i understand the meaning of the two tests.

But when I run tabodds, there do shows a test of trend which is still significant...

So I feel conflict still exists. Would you mind also explain this question? Thank you.

[Leonardo Guizzetti]

I believe that -tabodds- as used here gives you the test of trend for the numerical levels of diabetes, adjusting for age groups, and weighting those age groups using M-H weights. This would mean the test is with respect to diabetes, not age group, but with two groups, amounts to a test of difference from OR=1. These yield similar results with -cc- as the score test and 95% CI show.

[Zihan Dong]

I believe that -tabodds- as used here gives you the test of trend for the numerical levels of diabetes, adjusting for age groups, and weighting those age groups using M-H weights. This would mean the test is with respect to diabetes, not age group, but with two groups, amounts to a test of difference from OR=1. These yield similar results with -cc- as the score test and 95% CI show.

thanks, I got it. but how could I test the trend of OR with respect to age group using stata?

[Leonardo Guizzetti]

I suppose you could reverse age and diabetes in the -tabodds- command. If you have individual level data, you could model it directly:

logit cvd i.diabetes agegroup

Note that in this model, I am explicitly modeling age group as a continuous variable, whose slope is the estimate of the linear trend in age group.