I'm working with Nigeria DHS 2008 data (complex survey data) and I'm using the data analysis software available through my school (StataIC 13). I have a large number of 'not-applicable'. I'd initially re-coded them as missing, but then ran into trouble trying to explain why I had such a large amount of missingness and still was not using multiple imputation. Because of this, I'm having a hard time describing my sample. It's also complicated by the fact that of the 33,385 women interviewed, I'm actually only interested in the married women (23,954 women) so I have to specify a sub-population in addition. When I run my logistic regression model, I end up with about 15,449 observations instead of the 23,954 I expect. I'm guessing it's because of the 8021 not-applicables but I don't know how to deal with this. (1) Would 15,449 be my sample size then, instead of the 23,954 married women that I specified? (2) Do I try to address the NAs after I have weighted and specified my subpopulation (svy, sub pop:) or is this something I would need to handle before I weight my data and specify my subpopulation (i.e., married women). (3) How do I deal with the not-applicable since I have now found that re-coding them as missing is not the way to go. (4) Would I still include the NAs as part of my sample size? Thanks in advance for your help
To handle large number of not applicable, first I have the following comments.
1. If the study was piloted, then NA issues could be detected and minimized.
2. The tool should be validated to conform with the culture and setting
3. NA means participants gave answers, thus, this cant be handled as missing
4. only you need to do descriptive statistics for your results, leaving the NA as they came out.
5. if the NA are items in a Likert scale, then it is difficult to calculated total score.
When describing your sample (and sampling technique), just say what you started with and that you will be doing the logistic regression on a group of them. The df for those statistics will (likely) be based just on these. That is fine. Not applicable responses occur in most surveys. You just need to be careful when describe results which groups the analyses have been done on.
Som, are the women reporting married, single, not applicable? Or are some of the outcome variables the participants responded to as 'non-applicable'? The sample set should have sample weights. Usually you apply the sample weights on the entire data set, then filter participants with certain characteristics. Replied everything but 'non-applicable', women, married, for example. your sample size is what is left after applying your exclusion criteria.
Hello,
How about coding your truly missing data as -999, and NA's as -998. Perhaps it would help to distinguish them and while calculating frequencies of missing data, the proportion of these would go radically down.
Hope it helps a bit
I have the following comments:
1. if you did pilot study, then you were able to avoid such NA problem
2.if the NA not withing Likert scale,, then you just report in a descriptive stat. their number and percentage.
3. Not applicable is not a missing value to be replaced. the participants already answered these items.
To handle large number of not applicable, first I have the following comments.
1. If the study was piloted, then NA issues could be detected and minimized.
2. The tool should be validated to conform with the culture and setting
3. NA means participants gave answers, thus, this cant be handled as missing
4. only you need to do descriptive statistics for your results, leaving the NA as they came out.
5. if the NA are items in a Likert scale, then it is difficult to calculated total score.
Depends on why they are not applicable. For example, if the question is "Do you smoke?"If the answer is "no," then the number of cigarettes smoked could be coded "NA" but it could be recoded as "0" (zero). If NA is one of the options, look at the case's characteristics to see if you can see why the NA was used. Otherwise, you may be stuck with lots of missing data. If the variable is of interest in your study, you could compare those who chose "NA" and those who didn't.
You may replace the missing value by the average but I prefer keeping and reporting the missing values as missing. It is common to have no answer or very low valid responses in surveys unless depend and use highly trained interviewers in collecting data using structured interviews . Even though, using highly trained interviewers in collecting data using structured interviews is the best but it is time and money consuming
In sum, good and sound research is time and money consuming
In the case of missing data or not applicable data, termed as missing group, should be compared on rest of the available information with that of non-missing group. If both the groups behave differently than neglecting the missing group data in analysis is not desirable. However, if the missing group is comparable with that of non-missing group then treat the non-missing group data as the representative data and go ahead with the analysis. Considering the missing data as the non-response, I am trying to suggest you to see whether the non-responders are comparable on other data with that of responders. It is generally believed that non-responders, very often, form a different group and hence cannot be neglected in any type of data analysis.
Hi Som,
Not applicable means the question of interest does not apply to all the married women e.g if a question is on children, a married woman who has never given birth is not asked that question. You need to check the questionnaire and identify how the not applicable comes in (by checking the previous skip). Using the example of women above, your study population (sample) should be married women who had given birth not all the married women. There is a useful resource on guide to DHS statistics which shows how NA and missing system should be handled and the denominator (total sample) when handling different indicators. It can be accessed on http://www.dhsprogram.com/publications/publication-dhsg1-dhs-questionnaires-and-manuals.cfm.
I think it'll be good if you shed more light on the question specifically..
If someone never had sex for instance, they will have a not applicable information on age at first....in this case what you can possibly do is recode the NAs to something like never had sex....or better still just drop them...
But as everyone has suggested I think it'll be a good idea if you shed more light on the specific question.
Regards..
So sorry I'm just getting back. I was able to resolve the issue. It turns out it was a variable on age at first marriage and the N/As were those of course who had never been married. Recoding solved the problem. @Emmanuel Olamijuwon - since I was using survey data, I couldn't technically 'drop' the N/As otherwise it would give me a lot of error in estimation of my variances. I just used the subpopulation option instead (essentially "dropping" any variables I'm not interested in, while still keeping them for purposes of computing my standard errors.). My apologies once again for the extremely late update.
- Badges
- Science topic