The paper examines the role of interviewers in sample selection and its impact on data quality, especially the response rates. They use the European Social Survey - a face-to-face survey of behavior and opinions in many EU countries. Selection methods vary from using countries' registries of individuals (no interviewer involvement) to household and address registries, where the sample may have more than one house and the interviewer selects a housing unit and a person to interview) to random walks, where the interviewer selects every k-th unit and conducts interviews.
The total survey error framework associates the descrease in data quality with the following sources of errors:
- undercoverage (some persons have no chance to be selected)
- nonresponse (not all selected persons participated)
- sampling error
- measurement error (the response given does not match the true value)
This study focused on undercoverage and nonreponse as selection bias and used an external (another larger survey in the EU) and internal (difference from 50% female sample) measures of this bias.
The results suggest that when interviewers are not involved in sample selection, response rates are unrelated to selection bias. However, when interviewers are involved in sample selection, the response rates are higher, but they're associated with more selection bias. The paper concludes:
The most important issue for researchers who rely on survey data is how we can prevent manipulation of selection by interviewers. We recommend using sampling methods that minimize interviewer selection, as far as possible. Improved training and supervision of interviewers could also reduce interference in the selection process. If interviewers did not feel pressured to achieve high response rates, they might allow the selection process to be fully random and selection bias would be smaller.