Description
Summarising how representative a respondent set is from the variation in estimated response propensities, and using partial indicators to locate which groups drive it.
Response rates say little about bias. R-indicators measure representativeness directly: using auxiliary variables known for respondents and nonrespondents alike, response propensities are estimated, and R = 1 - 2 x SD of the propensities. R of 1 means every unit was equally likely to respond on those variables. The coefficient of variation of propensities gives an upper bound on nonresponse bias for estimates related to the auxiliaries. Partial R-indicators, unconditional and conditional, at variable and category level, show which groups contribute most to the lack of representativeness. These feed adaptive fieldwork (which groups to chase) and the choice of weighting variables. The indicator is only as good as its auxiliary variables, and a low R does not prove bias in any given estimate.
Attributes
- Discipline
- Survey & Psychometrics
- Loss type
- Unmigrated
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Days
Steps
- Assemble auxiliary variables available for both respondents and nonrespondents (from the frame, account data or logs).
- Estimate each sampled unit's response propensity with a logistic model.
- Compute R = 1 - 2 x standard deviation of propensities, and the coefficient of variation as a bias bound.
- Compute unconditional and conditional partial R-indicators by variable and category.
- Use the partial indicators to target fieldwork at under-represented groups and to choose weighting variables; track R across waves.
Sources
- The SAGE Handbook of Survey Methodology (Wolf et al., 2016)
Also known as
- R-indicator
- Representativeness indicator
- Partial R-indicators
- Coefficient of variation of response propensities
Categories
- Analytics & Behaviour
- Survey & Psychometrics