What test can we use to compare the sample proportion of multiple dependent samples (> 2) for non-dichotomous data (> 2 categories)?

28 Views Asked by At

I am currently studying hypothesis testing for dependent two-sample (proportion). The crux for my question is this, what test does one use to compare the proportion of multiple samples for non-dichotomous data (i.e., multiple categories)? For example, a 1 to 5 star rating by 50 critics for 5 restaurants.

The context in which my question arises from is as follows. For comparing the sample proportions ($H_0:\pi_1=\pi_2$) of dependent two-samples for dichotomous data, we use the McNemar test. Going with an example, let us say we are comparing the effectiveness of 2 drugs, drug (1) and drug (2). We would employ the McNemar test if we have data like this,

Effective Not effective
Effective $n_{11}$ $n_{12}$
Not effective $n_{21}$ $n_{22}$

Continuing with this example, if we wanted to compare the effectiveness of these 2 drugs on multiple categories (i.e., non-dichotomous), we cannot use the McNemar test. Instead, we would employ the Bhapkar test or Stuart-Maxwell test for data like this,

Effective Somewhat effective Not effective
Effective $n_{11}$ $n_{12}$ $n_{13}$
Somewhat effective $n_{21}$ $n_{22}$ $n_{23}$
Not effective $n_{31}$ $n_{32}$ $n_{33}$

If we had wanted to test for multiple samples (> 2 drugs; $H_0:\pi_1=\pi_2=\pi_3$...) on dichotomous data, we also cannot employ the McNemar test. Instead, we would employ the Cochran Q test for data like this,

drug (1) drug (2) drug (3)
person (1) $0/1$ $0/1$ $0/1$
person (2) $0/1$ $0/1$ $0/1$
person (3) $0/1$ $0/1$ $0/1$
person (4) $0/1$ $0/1$ $0/1$
person (5) $0/1$ $0/1$ $0/1$

Where,

  • 0 = Effective
  • 1 = Not effective

To summarise the properties of the 3 tests,

No. of samples No. of categories
McNemar 2 2
Bhapkar / Stuart-Maxwell 2 > 2
Cochran's Q > 2 2
? > 2 > 2

This goes back to my question, what test can we use for comparing the proportion of multiple samples with multiple categories?

p.s. I know that we can employ Cohen/Fleiss Kappa to analyse "agreement" for multiple samples/categories. What I am asking is for hypothesis testing of sample proportion equality (i.e., $H_0:\pi_1=\pi_2=\pi_3$...).

Please correct me if my understanding of the subject matter above is wrong. I am very much in the process of trying to grasp this subject and greatly appreciate any further insight on the matter. What I have stated in this post is based on my personal reading of the subject thus far.