Mcnemar's test for dichotomous IV vs. dichotomous DV: how pertinent is it?

I intend to study the effect of a given psychological intervention (dichotomous independent variable) on the compliance of a medical treatment (dichotomous dependent variable). For this purpose, the compliance of a number of patients will be “measured” in two different moments (say M1 and M2). The totality of patients will be divided in two subgroups, one of which will be subject of the mentioned intervention (experimental group) while the other will not (control group).

In both groups and in both moments, there will be patients who comply with the medical treatment and other who do not. That is to say that there are 2x2 possibilities for any patient (e.g. complies in M1 and does not comply in M2, complies in both M1 and M2, etc.).

The research questions could, therefore, be: 1) the psychological intervention promotes the compliance of patients to this treatment, and 2) the psychological intervention prevents the abandonment of compliance to this treatment.

Regarding the analysis, I thought of the following two possibilities (although without great conviction with regards to the first one):

1) To divide the totality of patients in the 2x2 groups above mentioned, regardless of the group to which the patients belong (i.e., experimental or control group), and then use a simple chi-square analysis crossing those 4 groups and the independent variable (existence of psychological intervention or not). My goal would be to be able to answer to a question such as this: the subgroup to which the patient belongs (e.g. patients who comply in both moments) is related to the psychological intervention?

2) Using McNemar's test.

Any opinions on these above mentioned analysis? Any alternative that could be more interesting?

Thank you in advance and sorry for my English!

Joao Machado Vaz


TS Contributor
You can use a binary logistic regression model with outcome at t2
as the dependent variable and outcome at t1 as covariate.

Or. if you want to perform something analogous to McNemar's test, then
like in that test you leave out the unchanged subjects and only analyse
those who changed (noncompliance=>compliance versus compliance=>
noncompliance). This means a maybe tremendous loss of statistical power,
of course. It resembles your idea to compare the 4 compliance-types
between groups: if your proposed Chi² analysis became significant, then you
would have to perform pairwise post-hoc comparisons. And in that post-hoc
analyses you would have to adjust for multiple comparisons (=loss of power).
So why not concentrate on the only comparison that is interesting, but
without being forced to adjust.

As an alternative I have seen mentioned generalized estimating equations
(GEE), but personally I don't know much about that analysis.

With kind regards