Non-response bias in gambling surveys

The representativeness of study samples is a perennial concern in survey research. Here we examined how gambling survey responders differed from non-responders. Australian customers of an electronic gaming machine venue (Study 1; n = 4,301) and two online wagering sites (Study 2; n = 24,829) were invited to participate in a survey about their gambling. We accessed account data for all invitees and compared responders’ and non-responders’ characteristics and gambling behaviors. In Study 1, differences between the 183 responders (4.25%) and non-responders were minimal (d = −0.23 to 0.21). In Study 2, the 1,956 responders (7.88%) were older, gambled more frequently and more recently (d = −0.51–0.58), and were more likely to be flagged by the site’s risk detection system (OR = 1.91). Subsample analyses restricted to those who had gambled in the past 30 days (n = 12,799) found smaller differences (d = −0.35–0.31, OR = 1.34), demonstrating how eligibility criteria affect sample representativeness. The ability to predict survey uptake via logistic regression models was poor in Studies 1 (Nagelkerke’s R2 = 4.6%) and 2 (8.5%). While gambling survey samples are broadly representative, responders – particularly in online studies – tend to be more involved in gambling, highlighting the need for caution when generalizing findings.