Wilcoxon Signed Rank Sum Test- Missing Data

[Staci Ebadirad]

Hello,

I'm trying to run my thesis data analysis and am running into trouble. Here's an overview of my study. I'm trying to see if parenting behavior changes over time in association to parent education. I have 20 dyads, both moms and their child with two time points for data collection, pre and post. I'm running a Wilcoxon signed rank sum test to compare the changes in mean behavior from the pre time point to the post time point.

I noticed that when I ran the Wilcoxon test, it only accounted for the participants with complete data (n=14) and not the whole sample (n=20). I'm not looking at individual changes in behaviors but the overall group changes. Is there a way to include all of my participants in the Wilcoxon signed rank sum test?

Any suggestions or comments are appreciated. Please let me know if I need to clarify any information. THANK YOU!

[Clyde Schechter]

Except for the -mi impute- commands, there is no command I can think of that will include observations that have missing values for the variable(s) they are working with. What value would you have Stata use for those cases?

Also, as an aside, the Wilcoxon rank sum test does not compare means, it compares distributions overall.

Added: OK, the -tabulate- command with the -miss- option will include missing-value observations in the tables it generates. But I still think it would be accurate to say that no command will include missing values in calculations other than just counting them.

[Carlo Lazzaro]

Staci.
welcome to the list.
The practical solution would be to report in the outcome table(s) of your dissertation that you have had 20 observations with 6 missing values: hence, the descriptive statistics refer to a sample of 14 observations (this is the approach I usually follow if dealing with missing data is not included in the study protocol).
Obviously, there's a risk of some challenging question coming from some discussant who may be well interested in missing values mechanisms, patterns, imputation and the like.
Otherwise you may want to try to deal with the missing values that creeped up in your analysis via -mi- (if feasible) and present outcome table(s) with and without missing values dealt properly.
Others may want to sweep that issue under the carpet and pretend that their results do refer to the whole sample (I've seen this methodological mis-behaviour in several instances): however, that attitude has the drawback that, more often than one would imagine, as we say in Italian, "the blanket is too short" to hide the evidence that missing values affect the results of the analysis somehow.