Dear Semra,

Assuming you have used repeated measures, and the assumptions were met, you may use the mixed models approach. It is elaborated on the following link.

https://www.uvm.edu/~statdhtx/StatPages/Mixed-Models-Repeated/Mixed-Models-for-Repeated-Measures1.html

For more insights on various techniques, see the following studies.

Xi, W., Pennell, M. L., Andridge, R. R., & Paskett, E. D. (2018). Comparison of intent-to-treat analysis strategies for pre-post studies with loss to follow-up. Contemporary Clinical Trials Communications, 11, 20–29. https://doi.org/10.1016/j.conctc.2018.05.008

Anderson, S. F. (2018). Missing data in randomized pretest posttest studies. Multivariate Behavioral Research, 53(1), 125–126. https://doi.org/10.1080/00273171.2017.1404893

Good luck,

I don't know of a general consensus about when too much missing data from a participant's measure means you should treat it as entirely missing (vs, say, averaging the completed items). Maybe because the reason for missing responses may be very different depending on your measure. For example, is your measure a vocabulary test where many missing items could be the participant choosing to not even guess? If so, you might treat those like incorrect responses. Alternatively, if your measure is a self-perception of vocabulary knowledge scale, you might try to extracting a score from a factor analysis with mean imputation on the missing items. Or, if there is one item particularly often left blank, you might consider if there is something odd about the item. Is it worth excluding for all participants? Ultimately, you'll probably want the strongest rationale as possible for your decisions so you can explain those in your results. Hope this helps, Semra. ~ Kevin

Before horse power computers were available, missing values were replaced by the means over groups by pre- or post-test. Nowadays, one can replace the missing values (if there are no more that, say 10%) by multiple imputation. See attachments.

Mohialdeen Alotumi

Thank you for your response. I have not to check them yet. I am working on cleaning and getting ready the data. I will check the sources. Thanks again.

Kevin Grobman

Thank you for your response. It looks like the reason is participant choosing to not even g​uess. I did explain my proposal that I will count it as an incorrect response. But while I am doing the analysis my advisor said there should be a different procedure to handle. I searched so many sources but could not find anything that really explains. If you have any references on this could you please share them with me? Thank you!

Jos Feys

Thank you for your respond. I will checked the sources you suggested and those looks very beneficial. Do you have any reference on how to decide the percentage of missing value? Usually they explain the process of how to replace the mising data but not tell extaclly how we can decide to do imputation. Thank you again.

Another procedure could be to replace missing data with chance (e.g., 1 for correct, 0 for incorrect, & .25 for missing if 4 choices). You might also be able to empirically test if missing data reflects a lack of knowledge. Are missing values disproportionately among harder items? Are the children with more missing values those who were more likely to answer incorrectly on the items they did answer? If so, that would provide some additional evidence non-responses reflect a lack of knowledge. Unfortunately I don't know of any citations formalizing this kind of thinking. For me, it's more of the implicit knowledge from brownbags and conferences. Hopefully somebody else has citations to share. I'll be curious to read them. Best wishes with your project, Semra. ~ Kevin