Practical utility of a new propensity score weighting technique for estimates based on landline-only cases in a dual frame landline/mobile telephone US national alcohol survey

Background The 2010 NAS was a dual-frame landline and cell phone CATI alcohol survey, using RDD sampling. We examined whether propensity score (PS) weighting methods could be used to re-weight landline cases to provide better estimates than census-based weights in the landline-only sample. Methods PS weighting methods were used to reweight the landline sample cases taking account of the ‘missing' mobile sample. In the dual-frame sample, a probability for each land-line individual was estimated using sampling weighted logistic regression, with key predictors of being in the cell sample including demographics and drinking onset age. This probability was transformed to normalized PS weights for the landline sample, constructed by adding the new PS weights to the dual-frame weights. Test variables were 7 key alcohol related estimates available on both landline and cell questionnaires: current drinking, drink-size adjusted, unadjusted, and graduated frequencies volumes, frequency of 5+ drinks/occasion, DSM-IV alcohol dependence and consequences. Results Landline census-weighted estimates for these outcomes differed by on average -4.9% from the ‘silver standard' dual-frame estimates. PS-re-weighted landline estimates differed by -2.2% overall, with all but consequences showing smaller differences from the silver standard. The best performing PS-weighted estimate was GF volume, differing by only 0.2% versus 4.6% for the Census-weighted estimate. Conclusion Propensity score weighting appears a practical way to weight landline-only samples by up-weight cases most like the missing cell phone cases. For important variables in a telephone survey omitted on a shorter, mobile survey, this will likely improve the estimates compared to standard census-based weighting.