Estimating Components of Mean Squared Error to Evaluate the Benefits of Mixing Data Collection Modes

Abstract

Mixed mode data collection designs are increasingly being adopted with the hope that they may reduce selection errors in single mode survey designs. Yet possible reductions in selection errors achieved by mixing modes may be offset by a potential increase in total survey error due to extra measurement error being introduced by the additional mode(s). Few studies have investigated this empirically, however. In the present study, we compute the Mean Squared Error (MSE) for a range of estimates using data from a mode comparison experiment. We compare two mixed mode designs (a sequential web plus mail survey, and a combined concurrent and sequential CATI plus mail survey) with a single mode mail survey. The availability of auxiliary data on the sampling frame allows us to estimate several components of MSE (sampling variance, non-coverage, nonresponse and measurement bias) for a number of sociodemographic and target variables. Overall, MSEs are lowest for the single mode survey, and highest for the CATI plus mail design, though this pattern is not consistent across all estimates. Mixing modes generally reduces total bias, but the relative contribution to total survey error from different sources varies by design and by variable type.

Effects of Question Characteristics on Item Nonresponse in Telephone and Web Survey Modes

“…Since respondents self-select into different modes to some extent (Voorpostel et al 2020) these coefficients reflect both, self-selection, and measurement effects. Because it is difficult to disentangle the two effects (Roberts and Vandenplas 2017), we only briefly describe the coefficients without interpreting them in the discussion. In addition, our main focuses are effects from question characteristics and mode, so we treat respondent characteristics as controls rather than as substantively interesting covariates.…”

A Validation of R-Indicators as a Measure of the Risk of Bias using Data from a Nonresponse Follow-Up Survey

“…This limitation may be particularly relevant where subjective variables are concerned, however, and may not be entirely due to a lack of, or only weak correlations with the auxiliary variables. In particular, subjective variables may additionally be affected by substantial measurement biases (Roberts and Vandenplas 2017), which could account for some of the results observed in the comparison between the main survey and the NRFU. Nevertheless, the results suggest the need for some caution when interpreting the magnitude of the R-indicator -as well as that of related bias indicators.…”

Effects of Changing Modes on Item Nonresponse in Panel Surveys

“…The main independent research variable is the transition to web versus repeating a telephone interview. The regression models described below include socio-demographic variables associated with survey participation and panel attrition (Roberts and Vandenplas 2017). The goal is to (partially) control for differences in the sample composition between the two groups.…”