mhodds for 2 x 5 table

[Laura Myles]

Dear Listers,

I have data on cancer characteristics. I am interested in exploring whether patients who take aspirin (yes vs. no) have a better cancer stage at diagnosis and whether this depends by age group (5 age groups) to check if this association is stronger in younger patients.

I am wondering if I could use the homogeneity of test produced as part of the mhodds command to address this question by using:

mhodds risk stage, by(age5)

Is there a better solution?

Thanks!

[Carlo Lazzaro]

Laura:
can't you consider -ologit-?

ologit <cancer_stage> i.aspirin##i.age_group

[Clyde Schechter]

Well, this is one possible approach to the question, and might be the best starting point for exploring the issue.

It has several limitations that can be overcome with other approaches, however. The homogeneity test result will give you a test of the null hypothesis that the odds ratio (cancer stage: ASA exposure) is the same for all age groups. But it will not give you any sense of whether the differences are slight or great, nor in what direction they go. The -mhodds- command, by design, produces a single odds ratio estimate. If your homogeneity hypothesis is rejected, then further analysis will necessarily take you out of the scope of -mhodds-.

It isn't clear how your stage variable is defined. Is it some sort of early stage vs late stage dichotomy? Dichotomous outcomes are the main strength of -mhodds-. But in most cancers multiple stages are recognized, and there are more flexible analyses that can handle multi-level outcomes better.

-mhodds- can only adjust for confounders through stratification, which requires that the confounders be discrete variables, such as your five-year age groups. But age is an intrinsically continuous variable, and imposing arbitrary categorizations on it discards information and can also lead to biased results. Again, there are other techniques that work better with continuous outcomes.

I think if I were approaching this, my first step would be to get the crude odds ratio, and then age-group specific odds ratios. A Mantel-Haenszel odds ratio, such as you propose here would be next. But then I would probably advance to either an ordinal or multinomial logistic model that included a risk by (continuous) age interaction.

Evidently this is contingent on the data you have being suitable for that. If all you have about age is 5-year groups, then analyzing age continuously is out of the question. If your data set is small, it may not be able to support fine-grained analysis. How good are your measures of risk and stage? And so on. The usual questions apply.

Added: Crossed with #2 which recommends one of the possibilities I have advocated here.

[Laura Myles]

Thank you both for your replies.

Stage is a categorical variable with 4 levels and, unfortunately, I only have age broken down into 5-year age groups.

I will explore your suggestions so I will run a regression which contains the interaction term to ascertain of there is an actual effect of age on stage and subsequently separate regressions for each age group.