Survey Non-Response
Contents
Non-Response Bias
Addressing Non-Response
Case-wise Deletion
The simplest solution to non-response is to ignore incomplete responses.
This can introduce further bias into the response data, exacerbating systemic undercoverage.
Imputation
Imputations solves non-response by creating a probable response value.
One option for an imputed value is to reflect the collected responses. This would typically be one of the mean, median, or modal response. Alternatively, random response could be selected based on the observed distribution of responses. These approaches artificially lower the standard errors, however.
Another option is to randomly select a valid response from the set of similar respondent. The complete set of respondents would be filtered to just those similar to the non-respondent. One would be randomly selected from that subset, and their response would be 'borrowed'. This also leads to artificially low standard errors.
Multiple imputation by chained equations is an imputation method based on a repetition of a system of Bayesian equations. A primary value would be taken as given, and the system of equations would predict all others. This would be repeated to create a valid state. Analyses would be run multiple times in different states created from different given primary values. Standard errors will reflect uncertainty of the state.