I can't tell whether Simon is using single imputation or multiple imputation. If it is some form of single imputation then it is valid only with small numbers of missing values since, as Deborah's quote points, it artificially shrinks variation. Multiple imputation is theoretically sound even for large percentages of missing values (say

What percentage of your total N have > 50% missing data? Do your participants with a lot of missing data look different from the rest of the participants in your database? If not, I would exclude them completely and report it in your methods section.

Thanks Krista. There are quite a few - I haven't counted yet, but a quick eye-ball of the data I'd say maybe 4-5% of the total sample. We too were leaning toward exclusion.

Hello Mr/ Dr Hunter,

this is from a book online called statsoft, I don't know if this is of any use.

Pairwise Deletion of Missing Data vs. Mean Substitution. Another common method to avoid loosing data due to casewise deletion is the so-called mean substitution of missing data (replacing all missing data in a variable by the mean of that variable). Mean substitution offers some advantages and some disadvantages as compared to pairwise deletion. Its main advantage is that it produces "internally consistent" sets of results ("true" correlation matrices). The main disadvantages are:

1.Mean substitution artificially decreases the variation of scores, and this decrease in individual variables is proportional to the number of missing data (i.e., the more missing data, the more "perfectly average scores" will be artificially added to the data set).

2.Because it substitutes missing data with artificially created "average" data points, mean substitution may considerably change the values of correlations.

I tend to agree with exclusion as per comment above for those with lots of missing data, unless you can just use the pairwise for individual analysis and then you may see a trend which will help you understand why or maybe it is obvious already

Thanks Deborah. I definitely would not use mean substitution. I would use a multiple imputation procedure as per the current recommendations.

I can't tell whether Simon is using single imputation or multiple imputation. If it is some form of single imputation then it is valid only with small numbers of missing values since, as Deborah's quote points, it artificially shrinks variation. Multiple imputation is theoretically sound even for large percentages of missing values (say

Thanks David - I was indeed referring to % of variables in a given case. I'm using multiple imputation, but feel that this is not wise when there are such large numbers of data missing on cases.

Simon raises a good common sense point about grossly incomplete observations, but what to do about these observations is not so clear and depends heavily on the specific research questions being pursued, preconceptions about which variables are critical to these questions (aka domain expertise), and which variables are just nice to have available. It also depends on the ability of the non-missing variables to predict the values of the missing variables (since that drives the quality of the imputation). Perhaps these considerations will help Simon think through the specifics of the missing value issues in his study.

Thanks again David. As you say, how we resolve this issue may be as much about being clear about what we are / are not doing as it is about doing the 'right thing' since the 'right thing' seems unclear. It may be that the complete data cam be used for some analyses, and that the problematic cases are omitted for other analyses.

In your situation, I'd run the data with and without the "incomplete" cases and see if the results change any. Also, if possible I'd test to see if the "inccomplete" cases were in any way different from the rest of the population. If you don't have demographic data to compare, can you see if they tend to different answers on the questions they *did* answer?

Hello Dr Hunter,

it's been quite a long time since you've asked this question. So I cautiously assume that you have already made a decision regarding your grossly incomplete data. Could you tell me how did you treat this feature and what was the reasoning behind it? Or maybe you could refer me to the relative literature?

Thanks

In an article I found " ..used the EM algorithm to impute missing values. Scales with 4–5 items allowed two missing values to be imputed. Scales with six items allowed for three missing values to be imputed. If more responses among the scales items were missing, the scale score would not be computed."

Does this mean we can impute missing values upto 50% if any variable in an observation is missing?