Kristen Blenk, The Ohio State University
Using Propensity Scores to Control Coverage Bias in Telephone Surveys
Telephone surveys are a convenient method of data collection. However, bias may be introduced into population estimates by the exclusion of non-phone households. It is estimated that six percent of US households are without phone service at any time. The bias introduced can be significant since non-phone households may differ from telephone households in ways that are not handled by poststratification. Many households, called "transients" move in and out of the telephone population, and these transients may be representative of the non-phone population in general. We develop a weighting adjustment for transients in an effort to reduce the coverage bias while controlling variance due to weighting. We use logistic regression to model each household's propensity for transience, using data from a survey of distressed and non-distressed regions of KY, OH, and WV. Weight adjustments are based on the propensity scores. Estimates of the reduction in bias and the error of estimates are computed using our weight adjustments and several alternative weight adjustments. The error in adjusted estimates is compared to that of the standard estimate to assess the effectiveness of the adjustment.
LinChiat Chang, Department of Psychology, The Ohio State University
Improving Election Forecasting
This project evaluates techniques intended to improve the accuracy of election forecasting polls. Potential sources of error in election forecasting include respondents who do not end up voting on election day, failure of some respondents to report their candidate preferences, failure to interview people who will eventually vote, and candidate name order effects. We evaluated solutions for each of these problems using data on candidate races and referenda from Buckeye State Polls conducted in Ohio between 1997 and 1999. Filtering out 50% of respondents was optimal for forecasting candidate races, but filtering out 80-90% of respondents was optimal for forecasting referenda. Random allocation of undecided respondents to candidates or referendum positions improved forecasting more than exclusion of these undecided respondents. Sample weighting did not improve forecasting of candidate races but did improve referendum forecasting. Candidates were forecast to receive more votes when their names were listed last, which suggests that name order effects on ballots should be taken into account when deciding whether to rotate candidate name order in forecasting surveys.
