Mi impute pmm nearest neighbor matching

[Yujin Kwon]

Hello,
I am using mi impute pmm command and I have a question regarding nearest neighbor matching.
I want to fill in the missing data with the observation from one nearest neighbor, hence using the knn(1) option.
However I have noticed that my outcomes vary each time I run the code, and I don't quite understand why there is randomness to the imputed values when I am almost certain that there must be only one observation with the nearest predicted value.
Any help/advice would be very much appreciated.

foreach y in 2013 2014 2015 2016 2017 2018 2019{  
use "data.dta",clear
keep if year==`y'|year==`y'+1
mi set fl
mi register imputed lnhw
mi xtset pid year
mi impute pmm lnhw lnimphw age edudum1 edudum2 edudum3 marital_simp head i.jobhourt i.jobkind_simp i.jobarea_simp i.wsize regular sex region i.howwage if year==`y', replace knn(1) by(year) add(1) force
save "impute_`y'.dta",replace
}

[daniel klein]

Randomness in the imputed values is a core property of MI; it is essential for getting statistical inference right. With the pmm model, there are two sources of randomness: first, the linear prediction is obtained using randomly drawn parameters from the underlying regression model; second, the final imputed value is chosen randomly from the set of nearest-neighbors. Setting knn(1) (which used to be the default in earlier releases of Stata) removes the randomness from step 2 but not from step 1. Note that you do not want to completely remove the randomness from any step (see Paul Allison's discussion of the subject).

There seem to be a couple of other potential problems with your code that I lack the time to address in detail. I will leave it at bullet points:

- use set seed to get reproducible results
- you only register two variables as imputed but then go on and list a series of additional variables; you do not want that
- you use the force option; you do not want that
- you only create one complete dataset; this will lead to incorrect inference (significance tests, p-values, confidence intervals) in the analyses because there is no variation in the (one) imputed value

[Yujin Kwon]

Randomness in the imputed values is a core property of MI; it is essential for getting statistical inference right. With the pmm model, there are two sources of randomness: first, the linear prediction is obtained using randomly drawn parameters from the underlying regression model; second, the final imputed value is chosen randomly from the set of nearest-neighbors. Setting knn(1) (which used to be the default in earlier releases of Stata) removes the randomness from step 2 but not from step 1. Note that you do not want to completely remove the randomness from any step (see Paul Allison's discussion of the subject).

There seem to be a couple of other potential problems with your code that I lack the time to address in detail. I will leave it at bullet points:

- use set seed to get reproducible results
- you only register two variables as imputed but then go on and list a series of additional variables; you do not want that
- you use the force option; you do not want that
- you only create one complete dataset; this will lead to incorrect inference (significance tests, p-values, confidence intervals) in the analyses because there is no variation in the (one) imputed value

Thank you so much for your helpful comment. I completely missed out on step 1 but now the idea became much more clear after reading the article you attached.
I have edited my codes, But can you tell my why will using the force option will lead to misleading results?The variables I used in the regression model has some missing values here and there so using the force option seemed to be the only choice...

[daniel klein]

But can you tell my why will using the force option will lead to misleading results?The variables I used in the regression model has some missing values here and there so using the force option seemed to be the only choice...

You have answered your first question yourself: your predictors have missing values which causes missing imputed values. Instead of forcing Stata to ignore this, why not impute the missing values in the predictor variables as well?