Code for performing Chi Squared test or Fisher's Exact test for a 2X2 table appropriately

[Mengmeng Li]

Hi everyone,

I am working on a project where there is a need to develop a program (codes) to evaluate whether Chi Squared test or Fisher's Exact test is suitable for the tabulated variables (2x2 table).

For example, there are to categorical variable A and B. Both are binary.

I am thinking to first -tab a b, mis col expected- ; second if any expected value for any cell is <5, then do -tab a b, mis col exact-, otherwise do -tab a b, mis col chi-. Although having the thought in mind, I have a hard time to realize it in Stata code and could not find any relevant information on google.

Please accept my sincere appreciation on any thoughts/insights you would like to share with me.

Kindly regards,
Mengmeng

[Tiago Pereira]

It can be done in several ways. It depends on the way your data is coded. A general approach would be to assume a variable called event (1= event, 0 = non-event) and a variable called group (1= exposed/case/treatment,0 =non-exposed/control).

preserve
gene  new_event = event
collapse (count) new_event, by(event  group)
list
sum new_event
local min = r(min)
local k = r(N)
restore

if `min'<5|`k'<4 {
tab event group, exact
}
else {
tab event group, all
}

Even though I find Fisher's exact test one the most counterintuitive and strange test ever invented, I would suggest to use it regardless of your sample size.

[Mengmeng Li]

It can be done in several ways. It depends on the way your data is coded. A general approach would be to assume a variable called event (1= event, 0 = non-event) and a variable called group (1= exposed/case/treatment,0 =non-exposed/control).

preserve
gene new_event = event
collapse (count) new_event, by(event group)
list
sum new_event
local min = r(min)
local k = r(N)
restore

if `min'<5|`k'<4 {
tab event group, exact
}
else {
tab event group, all
}

Even though I find Fisher's exact test one the most counterintuitive and strange test ever invented, I would suggest to use it regardless of your sample size.

Hi Tiago,

Thank you for sharing your ideas. I tried your codes on my dataset, which seemed working well. I want to also have a clear understanding with your codes. For example, `min' is the actual minimal number of "event = 1" cases among four eventXgroup types, but not the minimum of the expected values. Do you know how to capture the minimal expected value instead?

Much appreciated for your help.

Mengmeng

[Tiago Pereira]

Hi, Mengmeng

After the program -collapse-, you can add lines to calculate a new variable with the expected number for each cell. Then, you summarize that new variable and continue from there.

Tiago