I have a large data set containing children's scores on arithmetic tasks with a lot of missing values possibly due to the age of the children? My hypothesis is that the missingness is not completely at random but at random. I need to do a CFA and want to know whether I need to condition on age or not? I know SPSS 17 has a missing values analysis but my data is in R and is quite large. I would like to do a missing values analysis in R but have not been able to find a package that can do this. Does anyone know?
Add a comment
|
3 Answers
$\begingroup$
$\endgroup$
3
As @Dirk Eddelbuettel already mentioned, your questions is not very clear. In fact, I think you are asking two questions. The first question is related to your M(C)AR assumption. The second question is about (an) appropriate R package(s).
(1) "Testing" for MAR
To test if age has an effect on the missingness of your score variable, you could run a simple logistic regression model with age as a predictor variable. Your response variable is 0: score is not missing, 1: score is missing (see also @mbq's answer and @Macro's comment). Given the assumption that younger children are more likely to not report math scores, we expect to see a significant negative effect of age .
## Make up some data
set.seed(2)
## Younger children are more likely to not report math scores,
## so I use a Poisson distribution to model that behaviour
missData <- rpois(10000, 10)
dfr <- data.frame(score=rnorm(100), age=sample(6:15, 100, replace=TRUE))
dfr <- dfr[order(dfr$age), ]
dfr$agemiss <- sort(sample(missData, 100, replace=TRUE))
dfr$miss <- ifelse(dfr$agemiss == dfr$age, 1, 0)
## Run the logistic regression with age as predictor
> summary(glm(miss ~ age, data=dfr, family=binomial))
[...]
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 5.9729 1.4946 3.996 6.43e-05 ***
age -0.7997 0.1760 -4.544 5.53e-06 ***
---
[...]
(2) (Some) Missing data related R packages
Some of these packages also have functions to explore patterns of missingness (e.g., missing.pattern.plot() in the mi package).
-
1$\begingroup$ Interesting approach to
MCARtesting! But what do you think about alternatives: Little's test forMCAR(jstor.org/discover/10.2307/2290157), implemented inBaylorEdPsychR package and tests by Jamshidian and Jalal (jstatsoft.org/v56/i06/paper), implemented inMissMechR package? $\endgroup$ Commented Aug 23, 2014 at 17:01 -
$\begingroup$ @AleksandrBlekh Thanks, I wasn't aware of those tests! $\endgroup$ Commented Aug 24, 2014 at 5:34
-
$\begingroup$
mi::missing.pattern.plot()looks like not available anymore. $\endgroup$– zx8754Commented May 16, 2019 at 10:11
$\begingroup$
$\endgroup$
2
As far as I understand your question, you want to investigate if missing values in your data appear due to some pattern. In this case, you don't need any "missing value analysis" -- this is the same problem as checking whether the score is bigger than 0.7 or whatever. Just convert your dataset into two-class factor (missing, not-missing) and look for correlations.
-
$\begingroup$ BTW this is the method how people artificially proliferate functions in stuff like SPSS or SAS. $\endgroup$– user88Commented Jun 16, 2011 at 16:32
-
1$\begingroup$ +1. Consider "missing" as a binary outcome and see if it is associated with any of the predictor variables using, for example, logistic regression. Not sure why CFA would be necessary. Like hammering a nail with a sledgehammer. $\endgroup$– MacroCommented Jun 16, 2011 at 16:46
$\begingroup$
$\endgroup$
2
Your question is a little difficult to decipher. One approach for dealing with missing data is imputation -- and there is a substantial literature on this and an already large and growing set of packages at CRAN so you may want to start there.
answered Jun 16, 2011 at 15:27
-
$\begingroup$ Thank you for all your responses. I see that my question was indeed not very clear. What I meant to ask was, Is the missing data completely at random (MCAR) and how do I do this in R. I t $\endgroup$– RobertCommented Jun 17, 2011 at 12:21
-
$\begingroup$ The logistic regression seems to be an adequate method to test this. The reason that I won't to know the nature of the missing is because I'm performing a CFA on the complete data set (with missings) with FIML. However, if the missing is not MCAR but MAR then I would need to condition the latent variables on age (have age predict the latent variable arithmetic ability). Anyway, my problem is solved, thanks again. Cheers $\endgroup$– RobertCommented Jun 17, 2011 at 12:28