摘要：While a successful survey requires engaged and attentive respondents, careless survey completion remains a great concern in online market research. In this article, we test metrics of engagement in an online willingness-to-pay (WTP) study for fresh blueberry attributes using a major U.S. panel company and evaluate the impact that poorly behaving respondents have on subsequent data quality. In doing so, we investigate in detail the complex joint relationship between attribute non-attendance (ANA) and measures of respondent engagement in web surveying. Using fixed latent classes, an approach known as the Equality Constrained Latent Class procedure, we export individual probabilistic class assignment of all levels of attribute attendance to cross reference with respondents who fail measures of engagement and fraudulence, and analyze their composition and impact on latent classes, indicating non-attendance of individual and combinations of attributes. We also analyze engagement impacts on the tau variance parameter in the scaled mixed logit model and find strong links to unnecessarily increased heterogeneity when not properly filtering poorly behaving respondents. While WTP estimates between respondents passing and failing engagement metrics are similar with the ECLC model, filtering failing respondents in the scaled mixed logit model reduces overall WTP estimates. Results have implications for both WTP researchers and general online market researchers.
其他摘要：While a successful survey requires engaged and attentive respondents, careless survey completion remains a great concern in online market research. In this article, we test metrics of engagement in an online willingness-to-pay (WTP) study for fresh blueberry attributes using a major U.S. panel company and evaluate the impact that poorly behaving respondents have on subsequent data quality. In doing so, we investigate in detail the complex joint relationship between attribute non-attendance (ANA) and measures of respondent engagement in web surveying. Using fixed latent classes, an approach known as the Equality Constrained Latent Class procedure, we export individual probabilistic class assignment of all levels of attribute attendance to cross reference with respondents who fail measures of engagement and fraudulence, and analyze their composition and impact on latent classes, indicating non-attendance of individual and combinations of attributes. We also analyze engagement impacts on the tau variance parameter in the scaled mixed logit model and find strong links to unnecessarily increased heterogeneity when not properly filtering poorly behaving respondents. While WTP estimates between respondents passing and failing engagement metrics are similar with the ECLC model, filtering failing respondents in the scaled mixed logit model reduces overall WTP estimates. Results have implications for both WTP researchers and general online market researchers.
