For my dissertation, I'm using the Parenting Stress Index, which, based on 36 items, generates four scores--one for each factor (difficult child, parent stress, environment) and one overall stress score. The problem is that there are a few cases where the participant missed one item, so then when the syntax runs to generate those four scores, the category that item falls into gets no score, and the participant also doesn't get an overall score.
I've been looking into some options for missing data. I am aware of the option of simple mean imputation, but I'm reading that using it can weaken results. My understanding is that regression imputation has the reverse problem, and I'd like to avoid that too. I have read a little bit about multiple imputation, and so far conceptually it reminds me of bootstrapping and it sounds like it might be the most precise option, but I am unclear on exactly how it works and how to get SPSS to do it. Do you have any thoughts on methods to deal with this issue? With my small sample size, I can't really afford to lose cases.
Please let me know if you have had any experience with this kind of thing and if so, what you might recommend. Thank you!
Its common, happens a lot. The way you will deal with it may be more important for your formation than to analyse a full dataset. Just make clear in your manuscript if missing data is random or related to any other variable. More on the subject:
http://www.emgo.nl/kc/Analysis/9%20Handling%20Missing%20Data.html
multiple imputation you need when you are interested in variances of your estimates. You may first visualise the missing data pattern, e.g. using the VIMGUI R package.
Regression imputation is surely a good option, best when it is based on some randomness (e.g. Bayesian regression or just to add random noise to your imputed data based on the standard deviation of your residuals). Software is available for it (eg. packages mi or VIM in R) but I don't know if there exists something in SPSS as well.
As you write that it's really about participants not responding to single items, how about you use mean scores for your four outcomes instead of sums (and then multiply by the numer of items to be comparable to sum scores)? I guess this would be equivalent to replacing each missing score with the mean of the respective subscale for each participant and you will not have the problem with missing overall scores with SPSS.
Wow, what a wonderful community of helpful folks! Thanks so much, everybody. Lots to chew on here.
The National Academy of Sciences, Food and Drug Administration, and National Research Council has issued expert panel guidelines for handling missing data in clinical trials, which is worth looking at in detail. The lead author (Rod Little) has a 2012 New England Journal of Medicine article summarizing their recommendations.
I found a copy of the accepted manuscript of the Little et al. (2012) article. Here it is.
There are two ways to look at the issue: missingness at the scale score level vs. missingness at the item level. If you impute at the scale score level, you lose information from the non-missing item level data.
Take a look at Schafer & Graham's paper. Not only is it a great description of old and newer methods (including MI), but it discusses exactly your situation: missing items within an otherwise complete set of responses to a scale. In this case Schafer & Graham suggest MI is useful, but averaging over the non-missing items may be just fine (given sufficient item intercorrelation and a single underlying scale for the summed items).
Schafer, J. L., & Graham, J. W. (2002). Missing data: our view of the state of the art. Psychological methods, 7(2), 147.
Your research data is typical of many data sets in survey research (Hair, Anderson, Tatham, & Black, 1998), in that they contain a number of missing data spread among both respondents and variables. To address the issue of missing data, as well as the potential bias such missing data might have on the analysis, you should consider the interdependency of variables in analysis, as you seek and employ a remedy.
Following the argument that “the simplest and most direct approach for dealing with missing data is to include only those [responses] with complete data” (Hair, Anderson, Tatham, & Black, 1998, p. 51), you may excluded all cases with missing data. This strategy will reduce the number of available responses for analysis but will allow you to submit only cases having complete information. Since your sample size was small to begin with, you might end up with a sample size that will not allow results that are generalizable to the population. The issue of missing data sets is one reason for researchers to engage an ample sample size for their studies.
Reference:
Hair Jr., J. F., Anderson, R. E., Tatham, R. L., & Black, W. C. (1998). Multivariate data analysis (5th Ed.). Upper Saddle River, NJ: Prentice Hall.
p.s. I'm sure there is a more recent version of Hair, et al's work.
- Badges
- Science topic