Questions on mcartest

[Alyssa Beavers]

Hello,

I am trying to test if my data are missing completely at random (MCAR) by using the mcartest command, a user generated command. I have read the article Little’s test of missing completely at random (By Cheng Li, The Stata Journal (2013) 13, Number 4, pp. 795–809).

I do have a few questions on how to use the mcartest:

1. In the article, it says the syntax is:

mcartest depvars

However, my impression is that when it says "depvars" it is NOT referring to the variables that will be dependent variables in your subsequent analysis, but actually the variables that you are interested in testing are MCAR. In my case, I only have missing values for my independent variables so should I put those where depvars go in the syntax?

2. For covariate dependent missingness, it says to use the syntax,

mcartest depvars = indepvars

My understanding is that the independent variables for this command are your covariates that do NOT have missing values.

3. Can you only test if continuous variables are MCAR using Little's test?

Many thanks,
Alyssa

[Carlo Lazzaro]

Alyssa:
1) Correct:

mcartest <yourindependentvariables>

2) Correct:

use bloodtest
mcartest chol trig diasbp sysbp = weight height exercise i.age i.female i.alcohol i.smoking

Little's CDM test

Number of obs       = 371
Chi-square distance = 141.1465   
Degrees of freedom  = 140
Prob > chi-square   = 0.4569

3) No, as you can see from the following toy-example:

use "C:\Program Files (x86)\Stata15\ado\base\a\auto.dta"
clonevar foreign_1=foreign
replace foreign_1=. in 2/7
mcartest foreign_1 rep78
note: 2 observations omitted from EM estimation because of all imputation variables missing

Little's MCAR test

Number of obs       = 72
Chi-square distance = 0.1074     
Degrees of freedom  = 2
Prob > chi-square   = 0.9477

[Alice Ellyson]

Carlo, Can you help me understand the validity behind using Little's test on binary and categorical variables?

I see that mcartest can be applied to variables that are binaries (foreign) or categorical (if, for example, you do not declare a categorical variables as one using i.make), but the Stata Journal article that introduces mcartest states that "When the normality is not satisfied, Little’s test still works in the asymptotic sense for quantitative random vectors yi’s but is not suitable for categorical variables (Little 1988)." https://journals.sagepub.com/doi/pdf...867X1301300407

Is there a way to implement Little's test for categorical missing variables, or an alternate approach for categorical variables in Stata that similarly attempts to help a researcher understand whether making a MAR assumption is wildly inappropriate?

[Carlo Lazzaro]

Alice:
https://www.crcpress.com/Flexible-Im.../9781138588318 provides some hints on imputing categorical variables at pages 87-91.