2014–15 MacroMonitor Survey Weighting Procedures

Questionnaire Design

The MacroMonitor target for weighting to the national household population in the United States is economic-household units—a definition that allows households to have more than one economic unit. The US Bureau of the Census defines households as persons sharing a common dwelling unit. The MacroMonitor definition of economic household includes families, individuals living alone, and two or more adults living together in a common dwelling who share basic finances. Thus, adults who live together but are unrelated and unmarried—such as housemates, roomers, a cohabiting couple, resident employees, or adult children or other relatives who might contribute to the housing expenses but otherwise maintain separate finances—count as separate economic households.

The KnowledgePanel sample design began as an equal-probability sample, incorporating several enhancements to improve efficiency. Because any alteration in the selection process is a deviation from a pure equal-probability sample design, statistical weighting adjustments to the data offset known selection deviations. We have incorporated these adjustments in the sample's base weight.

Also, several sources of survey error—such as noncoverage nonresponse because of panel recruitment methods and inevitable panel attrition—are an inherent part of any survey process. We addressed these sources of sampling and nonsampling error by using a panel demographic post-stratification weight.

Last, we constructed a set of study-specific post-stratification weights to adjust for the study's sample design and survey nonresponse. Unless otherwise noted, the 2014–15 MacroMonitor study-specific weights are based on household-level demographic and geographic distributions for the noninstitutionalized civilian population ages 18+ from the Current Population Survey, March 2014 Supplement. Respondents were weighted to the derived benchmark distributions in Table 1.

We calculated comparable distributions using all completed cases from the field data. Because study-sample sizes are typically too small to accommodate a complete cross-tabulation of all the survey benchmark variables, we used an iterative proportional fitting for the poststratification weighting adjustment. This procedure adjusts the sample data to the selected benchmark proportions. The weighted sample data optimally fitted to the marginal distributions through an iterative convergence process.

After this poststratification adjustment, we examined the distribution of the weights to identify and, if necessary, trim outliers at the extreme upper tail of the weight distribution. Extremely large weights, though they help the fit of the total sample to the Census marginals, are statistically very unreliable. And because demographically the underrepresented (and overweighted) persons in the sample are typically the young, the low income, and the poorly educated, large weights also increase the relative importance of questionnaire responses of low validity. For these reasons, we capped the largest weights by winsorizing the distribution: We replaced the top 2.5% of weights by the mean value of the top 2.5% of weights. After winsorizing, the range of weights was 0.0716 to 5.5345. We then scaled the poststratified and trimmed weights to the size of the population of economic households in the United States. The final weights ranged from 2.1679 to 167.6596 (ratio 77.3) and summed to 133,443 (000) economic households.