Can you provide some more details on why you want to discard one option and count it as missing? I think it might become difficult to argue why you want to impute a value (1 or 2) where participants' answer was clearly "3".

The correct answer to this depends on how big your sample size is and what the research question you're trying to answer is. Imputation of categorical variables is difficult to do with any reasonable accuracy, and the accuracy also depends on how large the sample size is.

If you can retain sufficient power while excluding the participants with missing data (listwise deletion), then that is generally reasonable. You will generally bias your results far less by excluding 25% of your sample size than you will by performing multiple imputation of multiple categorical variables, which are less amenable to imputation.

If you do decide to impute the missing variables, the best way to do this depends mostly on the prevalence of positive findings in the response variable in your data. As a general rule, you can/should simply use linear imputation to impute a continuous substitute if the prevalence is high (say ≥20%). If the prevalence is lower, consider imputation using logistic regression or linear discriminant analysis instead.

Hello Sven Greving ,

3 of the 4 variables ask whether a participant reports being investigated for a disease or not (e.g., I have had my blood FSH and LH levels tested = yes/ no/ I don't know). Each question asks about a specific test, some tests might be easily recalled (e.g., having an ultrasound; 2% missing) while others are more difficult (e.g., did I get my FSH or my DHEA level tested last year?).

My thought process was as follows: The "I don't know" participants clearly were either tested or not, there is no in between, but they probably just do not recall if this exact test was administered, so I don't believe they constitute a different group (although I might be wrong).

I have thought of combining the three question in one variable (got tested and didn't get tested; this will result in very few missing data), but although the tests are used for the same disease, patient experience with them is not the same, so I didn't combine them and I thought of predicting them instead.

That's my thought process, and I am by no means an expert at this field and any guidance is appreciated.

Hello Jack Henry ,

My sample size is large enough so that even after deleting about 25% of it, it is going to be a good representative of my target population, but the problem is that I believe most of those who answered "I don't know" are actually "yes" answers (i.e., they know they got investigated but they don't recall the exact test), so deleting them would introduce some level of bias.

Would pairwise deletion introduce less bias than listwise in this case?

Husam Aldean Mahdi Abuhayyeh There is no way to account for missing data without introducing any bias, but what you're describing does not sound like it precludes listwise deletion. Issues with recall are inherent to self-reported data even when it is complete. Without evidence, it is probably not reasonable to infer that a majority of the "I don't know" answers are in fact "Yes".

Pairwise deletion is reasonable and the degree to which it introduces bias depends on the involved variables, namely whether they are correlated. If you choose to perform pairwise deletion, you should first complete the analysis using listwise deletion so that you have baseline plausible results to compare to when you then perform pairwise deletion to ensure that separate analyses on non-overlapping datasets do not yield radically different results.

It is also reasonable to include the "I don't know" category as a separate category as Sven Greving has suggested because it is plausible that the participants not knowing whether they had the investigations has implications in and of itself. Either retaining it as a separate category or performing deletion is probably more reasonable than imputing values.

Apart from the good points made by Jack Henry, I thought of an additional issue you could consider.

Check whether answering "I dont know" occurs (completely) at random or if it is correlated with another variable. Finding that answering with this option is correlated with something else, might make you reconsider this category

Dear Jack Henry and Sven Greving ,

I have just analyzed the "I don't know" as a separate group in my regression model, and it turns out to be significantly different from "yes" group!, it is not significantly different from "no" for every variable, however. I really didn't expect that! I should have done this step way earlier. Now I guess I will include it in the final analysis as a separate group, as it seems to occur not completely at random (I think).

I am really thankful for pointing out my mistakes, and for your detailed explanations that you provided. It was very helpful to me and I hope others find it helpful as well.

create dummy varriables for every category and the problem will be solved.

right now you have one column with 3 categories (yes, no, i dont know)

create dummy variables for each category, so that you will have 3 new columns

the 1st new column will be titled "yes" and the values will be 0 or 1

the 2nd new column will be titled "no" and the values will be 0 or 1

the 3rd new column will be titled "i_dont_know" and the values will be 0 or 1

ofc if one of these columns has 1 in it, naturally the two other columns will have zero (if a participant answered "yes", the "yes" column will = 1, the two other columns will = 0)

now to correctly do analysis with dummies, the rule is that:

[number of dummy variables should be = number of categories minus 1]

(i.e. # dummies= # categories - 1)

why minus 1 ?

when the "yes" and "no"columns are both 0, the software will understand it as "i_dont_know" = 1 (because these are logically mutually exclusive events, and this is the default natural conclusion)

so to apply the [minus 1] rule; you should delete the "i_dont_know" column, and keep the "yes" and "no" columns. you will end up with 2 dummy variables: "yes" and "no"

finally as a last step... you should delete the original column that had 3 categories (which you used to create your dummies) because categorical variables >2 cannot be used in regression and this is why we created those dummy variables

and now you are set to do the regression analysis :D

* it does not make a lot of sense to assume people who answered "i dont know" to be missing values, this decreases your sample for no good reason. and also it makes you miss any statistical finding that these people may have that is different from others who answred "yes" or "no". so they are a population on their own. but if that question was acutally left unanswered; then it makes perfect sense to treat it as missing data and consider imputing it or ignoring that row

I have to recommendations for you.

The first paper deployed several imputation techniques for comparison.

The second paper implemented a novel approach for both Missing assessment and treatment of Missing values  .

I think the size of Missing values are discussed very well. . .

Knn in row is a new imputation technique with a wide application.

https://ieeexplore.ieee.org/abstract/document/8859963

https://www.researchgate.net/publication/349384200_Investigating_the_impact_of_missing_data_imputation_techniques_on_battery_energy_management_system?_sg=T89qlVyc07IPNPI2AGCHIStg_maiCNHj8UZq6nO7fwgxVSZ7fyEW2Ao8owGbUxSVFnDOBLQkxps3Hzk_hraC1-OJ5wXs9YThAVJy4ZBk.G6mSFbsYwPE2eZv0huCWE92yDYRU78bzWOat5T1tTFmYILtMiYkWLZdw4Bq4o6-pfmGboQtsu9IKbnSTjO80Ug