Arbitrary ad hoc rules such as "10% missing data is OK but not more" are unnecessarily restrictive and not helpful. I recommend that you take a look at some modern, user-friendly/applied introductions to missing data theory and handling such as:

Enders, C. K. (2022). Applied missing data analysis. 2nd edition. New York: Guilford Publications.

Schafer, J. L., & Graham, J. W. (2002). Missing data: our view of the state of the art. Psychological Methods, 7(2), 147-177.

Multiple imputation (MI) may be an option for you. MI works for both continuous and categorical data and only requires missing data to be "missing at random (MAR)", which is a weaker assumption than "missing completely at random (MCAR)" data.

Thank you very much! I'll look into the recommended sources.

I was just genuinely confused because in my simplistic way of thinking it made a difference if I have "scattered missing values" throughout my whole data set (for whatever reason) or if most of my missing values are due to dropouts and the remaining items automatically marked as "missing" not considering that, the person dropped out.

If SPSS does an analysis it usually only uses the data where all variables that are needed are available. Therefore in my later analysis, the dropouts are automatically not considered in the calculations because (for whatever reason) those people did not finish the questionnaire.

Therefore, considering those "missing values" within the missing data analysis makes no sense since I already report the dropout rate in my sample description. That's the main point where I fail to understand why I should consider people (nearly 50% of the initial sample) who considered at one point (I know it was due to fatigue in most cases) to just close the online questionnaire and therefore all remaining unanswered items were marked as missing.

It depends on why people "dropped out," i.e., the missing data mechanism. If they dropped out simply due to fatigue and fatigue is not relevant to the outcome variables considered in your study, then this would be considered MCAR data (which is benign and can be handled well using MI or maximum likelihood estimation).

Here is another recent resource that you might want to check out:

Enders, C. K. (2023). Missing data: An update on the state of the art. Psychological Methods. Advance online publication. https://doi.org/10.1037/met0000563

Thank you so much for the time, patience, and very fast answers as well as resources!

Should the calculation of missing values be used across all variables of the dataset or are they only applied to those variables that are relevant for my hypothesis testing? I have found different sources and handling in articles regarding this and depending on how I calculate it, I get very different results in the "missing data analysis" of SPSS and the resulting percentages.

They only gave us a very short introduction on "missing values" in the lectures but I would like to have the correctly performed calculations and also a solid LDQ-detection and missing value detection in my work, so I try to acquire this knowledge myself now as good as I can to raise the quality of my work.

Christian Geiser, thanks for alerting me to the Enders (2023) article. Another resource I like is this Annual Review chapter by John Graham:

Article Missing Data Analysis: Making It Work in the Real World

https://www.personal.psu.edu/jxb14/M554/articles/Graham2009.pdf

Cheers,

Bruce

I would not impute those missing values if possible. In order to impute missing values, they need to be "missing at random". From what you describe, these are definitely not "missing at random" thus should not be imputed, no matter how clever the imputation method. Remember, when you impute, you are literally making up data. It's good to remember that.

Randall Chadwick Thank you for the answer! I felt a bit insecure to impute data from people who just dropped out and decided not to finish the whole questionnaire. I can totally understand the imputation solutions for data if it's missing in between the questionnaire (which is not possible in my case because they couldn't skip a question so if there was no dropout and ending of the questionnaire, there are no missing values).

The problem I'm facing is, that my professor wants me to report the amount of missing data (unfortunately he sticks to the rules of thumb with a max. of 10 - 15% of all data missing). But that's impossible because nearly half of the initial sample dropped out at one point or another. And that messes up the whole missing data analysis because those are all included in the percentages. And above those 10 - 15% he expects some sort of reporting on how I dealt with the missing data in my data set. (He never asked for anything like that in the past he just told it to me specifically because he knows that I'm a huge fan of data analysis so he probably assumed that I'm more committed to looking into it on my own).

Now I was wondering if I would just report the percentages but add, that a huge amount is due to the dropout of the people and therefore the greater amount of missing values. But I have no clue if that's a smart way to address the issue.

It almost feels like altering my data and not really presenting what actually can be seen in the data set and that's a higher amount of missings because of the dropouts. I know, that in my questionnaire people were forced to do a fair amount of reading before making a judgment (it was a factorial survey) and those who completed left the input, that it was just too much reading. But of course, there is a vast amount of reasons why people could have dropped out (time, distraction, disinterest, fatigue,...) and I can never be sure about the reason.

I'm just nervous about the "reporting part" because if I report the actual percentages as shown in the missing data analysis I am pretty sure that my professor might ask why I wasn't using the MI technique to reduce the number of missing values. Can I just argue (maybe in my discussion) that the amount of missing data seems to be high but that it needs to be considered that a lot of people dropped out and I decided against the MI technique to alter the data in a way that probably wouldn't be accurate? Or is that non-scientific? I will have an oral exam about my thesis and I would argue it the same way if I was asked that question in person. But in the written report I'm really struggling with how (and at what point) I should address that issue. I tried to look at some other thesis projects that my professor was taking care of but those students neither performed a missing data analysis nor anything concerning LDQ detection. They just left it out but I think those things are important information (especially IF there would be missing values in between and not due to dropout and IF there would only be a few so that replacing them might be worth considering)

I'm sorry for asking so many questions but (for me) it's not an easy topic since I never actually was lectured in that aspect (apart from being told the different forms of missing values) and I just want to hand in a solid work that shows, that I actually want to research (maybe make a publication one day if I make it that far) not based on what I was told during University lectures but following the standards of scientific research.

Melissa Maderthaler, you mentioned SPSS as a topic, so I assume that is the software you use. Notice that the Multiple Imputation command has two options, one being to just analyze the patterns of missing data. See this web page for an example:

http://bayes.acs.unt.edu:8083/BayesContent/class/Jon/SPSS_SC/Module6/SPSS_M6_2.htm

The output should include everything your professor wants to see.

One would normally only impute the variables that are actually used in the statistical analysis, but include all available auxiliary variables in the imputation phase (auxiliary variables are variables that are correlated with dropout/a missing data indicator and variables that are correlated with the outcome variables to be imputed).

I don't see why you should not impute. If dropout is unrelated to the variables of interest in your study, imputing would most certainly be better than using listwise deletion of cases. Listwise deletion not only leads to a loss of a lot of cases (and therefore reduced statistical power) but is also more likely to result in biased results because it rests on the more restrictive assumption of "missing completely at random" (MCAR) data, whereas multiple imputation only requires the weaker "missing at random" (MAR) data assumption. See the references that I listed above for more detailed explanations.

Thank you so much to everyone! I think I finally got it!

I have not yet had time to read that Enders (2023) article in its entirety. But after reading his explanation of the MCAR, MAR, and MNAR missingness mechanisms, I cobbled together this slide, which some readers may find helpful.

As Bruce Weaver points out, the key is to understand the level of missingness. That is more an issue of careful reasoning, rather than doing tests. If you come to the conclusion that MNAR is a likely, you are screwed. For MAR and MCAR, most modern models (such as linear models, multi-level models) are unaffected.