SPSS Software | IBM

Hey Phuong Tran

when I tried (and failed) to use multiple imputation in R on the basis of SPSS, I remeber creating a separated data set with all former missing variables (coded: 99 and 98) changed to SYSMIS as you proposed. As always, I am happy to be corrected on the matter by more knowledgeable colleagues.

Mohammed Y. Shakor you answer does not address the question in any way. Please read it before you post something.

Best

Marcel

The don't knows and not applicables are interesting in themselves, and deserve a look. Items where lots of people don't know the answer indicate poor validity of the item – is the item not clear, or do the response categories omit some potential responses. And the not applicable items, again, are interesting because if they are numerous it suggests that they have poor validity in this population setting.

In neither case would I impute the missing value. The values are, in fact, not missing. The person has given you an answer, just not one that you wanted. To impute the data is to ask what the person would have said if they had known the answer, or if the question had applied to them. Both of these strike me as contradictory!

Mohammed Y. Shakor – I don't think you understand the function of questions on RG. If you cannot provide a useful answer, don't spam the question with contributions like that. The questioner already made it clear that they are using SPSS – what were you trying to do in your answer? I hope you weren't trying to inflate your RG score.

Remember, your RG answers are visible to any prospective employer. So far, many of your "answers" don't look all that helpful.

Hi Ronán Michael Conroy

despite my best intentions, I have given poor advice. You are absolutely correct that MI is not warranted because neither "I do not know" nor "Not applicable" are categories of missingness as in someone omitted to answer, but rather missing not completely at random in the case of "Not applicable" and a valid statement in the case of "I do not know".

Phuong Tran I still think that replacing missing values with SYSMIS is the way to go just not with the kind of "missing" answers present in your data set.

Best

Marcel

Thank you Ronán Michael Conroy and Marcel Grieger for taking the time to respond to my question. Your input and advice are much appreciated.

Upon reflection, I agree that multiple imputation would not be an appropriate approach to address these types of "missingness" (they are considered user-missing data in my dataset though...). I am using an existing dataset for my study on fairness perceptions of digital platform workers. I think those who said "I don't know" and "Not applicable" might have provided some valuable insight.

What's I'm currently working on is to find an alternative solution to handle the missing data (both system- and user-missing) and justify it (e.g., using the Little's MCAR test, resulted in p value < .05).

I was considering Listwise deletion, just to simply remove all cases with (system- and user-)missing values. I am concerned however about the possibility of biased parameter estimates produced by listwise deletion as the data is not MCAR (as suggested by the MCAR test result).

Any guidance you could give would be greatly appreciated.

I think that you have a two stage model for the individual data items :

1. Is the item inappliable? The discard the observation for this analysis

2. DId the person know the answer? A logistic regression

3. Conditional on their knowing the answer, what was the answer

The problem is that the errors in the models 2 and 3 have potential linkages between the errors. This is not an easy model. You might look it up under Conditional mixed-process models.

However, I reckon that before you do that, you might pretend there wasn't a problem with the linkage of the two models and estimate them to see if you have something interesting enough to warrant turning the statistical heavy artillery on it! I note that you are on SPSS, which may not have a way of estimating such models (dunno – stopped using SPSS in the 1980s).

Hi Ronán Michael Conroy

Thank you for your advice. I‘ll look into the models and associated SPSS features (if any). I haven’t come across these models in my previous quant classes and readings so far. They sound very interesting.

I have done some deletion of offending individual cases and variables (as you suggested). This seems to have improved the extent of missingness.

Phuong Tran – It has occurred to me several times that there should be someone in your institution that you can talk to about this. While it's been fun on RG discussing your problem (which is an interesting one!) it's not a substitute for you having someone who knows what your research is about, what your data are like and can advise you and explain things to your satisfaction. There really should be someone local who can do this.

This is not to put you off posting problems here – if it starts a discussion it will be of use to others besides yourself. But RG isn't a substitute for in-house stats support.

If I follow Ronán Michael Conroy's suggestion above, another term to search on might be "hurdle model". I think that term is often reserved for the specific situation where the stage 1 model is a binary logit model and the stage 2 model is some kind of count regression model (e.g., Poisson or negative binomial). But I don't see why one could not use some other type of model at stage 2. Did I understand you, Ronan?

Bruce Weaver Thank you.

you should get support from your institution / from locally available statisticians, the tool used for data collection, and your purpose of the study must be consideredd to get the correct solution