I'm doing an experiment for which I've distributed a survey. People were asked in this survey to rate the attractiveness of 5 other people. I provided four groups of pictures and 1 of those groups would be randomly assigned to a respondent. They also had to state how likely they were to follow these people (shown in the pictures) on different social media platforms (5 options provided).

Now my problem is that there are a lot of missing values. Since not everyone is active on every social media platform some of these platforms have a lot of missing values. The same goes for the attractiveness score since one respondent only got 1 group of people.

First I thought it would be more convenient to put the attractiveness scores in one column (making it one variable instead of the 20 columns/variables for every picture that were automatically exported from the survey tool). Now I want to make a regression with 'likelihood to be followed on a social media platform' as the dependent variable and the attractiveness score as the independent variable (plus some control variables). It doesn't have to be anything fancy, but I've been told an OLS regression would not work here since every five rows correspond to one observation/respondent. Using the average of the five pictures is not an option since that would mean I lose the actual information in the independent variable.

I had been given the advice to use a repeated measures analysis. However, this requires to have multiple levels or timestamps. I thought maybe I could interpret the different pictures or even platforms as levels but doing so would mean creating even more missing values or splitting up the 'likelihood to be followed' variable meaning that I can't use it as one dependent variable anymore.

I've also tried creating separate datasets for every platform, but this brings the same problems. I've also thought about interpolating the missing values but some platforms have A LOT of missing values, and I'm not sure if that would be the best idea.

So, is there any way to use repeated measures or is there maybe another solution? I've also been looking at multilevel regressions lately. Maybe it can be done with that?

I know this might be a lot to take in, so for clarity, I've also attached a link to the part of my data that I'm having trouble with (of course I got rid of any sensitive information, everything is anonymous).

Any help would be greatly appreciated!

Link: https://mega.nz/#!EG5yUQzS!sWoTGO849iUDbAcrYUxe81gOilHRhxvxkfwesM2dfLQ

  • $\begingroup$ What are you trying to predict or model? $\endgroup$ Commented Oct 3, 2018 at 16:01
  • $\begingroup$ Basically, I want to know if the attractiveness of a person affects their chances to be followed on social media (and how large or small that effect would be, of course). $\endgroup$
    – Rutger
    Commented Oct 4, 2018 at 14:14

1 Answer 1


Repeated measures - Just because your data is set up where you have multiple rows per participant doesn't necessarily mean you have repeated measures. From your description of the data it sounds like the survey software output the measure of attractiveness of five different people on separate rows. This would not be repeated measures, if you had them rate the attractiveness of the same person 5 different times then you would be correct in treating this data as repeated measures.

I believe what you need to do is restructure your data so that each participant has one row with 5 columns of the attractiveness of the 5 people who's picture they were exposed to as well as whatever other variables you measured.

With the missing data, it sounds like you know what social media platforms participants use, therefore if they said they were not on a specific platform you can assume that this is actually not missing data but a NO or a zero or a "Not at all likely" or whatever other value you want to give it but the missing data is not actually missing its just impossible because they don't use that platform.

Missing data on the outcome - this is a similar answer to the perceived missing value on the social media platforms question. If a participant was given one group of five pictures, out of four groups of pictures, then they are not missing data on the other three groups of pictures because this was the design of your study. Therefore, in order to compare between participants you will have to take measurements at the group level of the pictures provided so that each person has an average attractiveness score per group of pictures but not for each individual person's photo within those groups.

If every participant had the chance of seeing every picture but some opted not to or missed an answer, then this would be considered missing data.

Please let me know if this made sense to you!

  • $\begingroup$ Thanks for your answer! This will be really helpful. Most of this makes sense to me, except for the part "in order to... within those groups". My teacher/supervisor advised me not to take the average of a groups attractiveness score. However, I found this to be the only way to compare them, so this is a little confusing for me right now. Thanks for your answer anyway! $\endgroup$
    – Rutger
    Commented Oct 15, 2018 at 15:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.