This question was discussed a lot when I attended method courses at university and it is still actual and not easy to answer. Some argue the higher the better. But is a higher response rate (always) needed? What do you think and are there interesting scientific papers you know/recommend that deal with this issue?
I think that more important than the response rate is why people don't respond. That is, if non-response is completely random and not related to the variables of interest, it may not matter as long as you retain sufficient statistical power. In contrast, if non-response is systematic (e.g., if you want to measure the frequency of poverty and poor people choose not to respond), this could be much more problematic in terms of resulting in bias.
Gert Lang I would think think that the general literature on missing data analysis applies here in principle, e.g.,
Enders, C. K. (2022). Applied missing data analysis. 2nd edition. New York: Guilford Press.
There are even "planned non-response designs" (i.e., planned missing data designs), in which participants only respond to subsets of questions "by design" to save resources. Such designs yield "missing completely at random" data and can be quite powerful even though they do not require complete responses from everyone. See, e.g.,
Graham, J. W., Taylor, B. J., Olchowski, A. E., & Cumsille, P. E. (2006). Planned missing data designs in psychological research. Psychological Methods, 11, 323-343.
When using a design-based (probability-of-selection-based) approach, any nonresponse goes contrary to the theoretical foundations upon which your inference to the population is based unless it is (completely) missing at random, which is usually more a matter of degree rather than absolute. As Christian indicated, some missing data are more problematic than others and considering a given percent which may be tolerated is not a feasible way to approach this. Testing is best, which was possible with my work where I could leave out known data and see how well inference handled that, but such testing is generally more problematic, especially for your nonresponse issue. Many surveys are having nonresponse issues, and the methods used to handle that generally require good covariate data which may also not exist. Still, there has been a lot of work done on this. For example, please see the following:
Elliott, M.R., and Valliant, R. (2017). Inference for Nonprobability Samples. Statist. Sci. 32(2): 249-264, May 2017. https://doi.org/10.1214/16-STS598 (Open Access from Project Euclid)
I have a thread which might be of interest:
https://www.researchgate.net/post/Comparison_of_Nonresponse_in_Probability_Sampling_with_Nonprobability_Sampling
You can search for books on missing data. Three people of whom I have heard who have worked in this area are Robert Groves, Don Rubin, and Rod Little, by the way. There should be some additional newer contributors, however, I would think. Christian noted something recent. Cheers.
- Badges
- Science topic
- Similar topics
- Computer Science
- Research Topics