Data Fusion: Assessing the Feasibility of Combining Alcohol Data Collected via Probability and Non-probability Samples

We assess the feasibility of using data fusion to combine data from a telephone interview collected from a dual-frame random digit dialing (RDD) sample, with data collected from an opt-in web panel.  While RDD telephone samples are based on probability sampling, they are challenged by decreasing response rates and increasing costs.  Data collected via opt-in web-panel raises concerns regarding population representativeness, yet is less expensive than a telephone sample. 

We hypothesize that a hybrid telephone/web panel approach will offer probability-based prevalence estimates for drinking behaviors (e.g., drank wine in past month), while collecting detailed data (e.g., alcohol expenditures, volume and brands consumed) via web panel.

Our research has three specific aims:

• Identify linking variables to fuse the data and assess conditional independence
• Compare the results of estimates from the telephone and web samples
• Data fusion as a method for a) imputing responses for mid-survey terminates, and b) producing accurate population-based estimates

In the first aim, we use the telephone survey to evaluate the linking variables that provide conditional independence between data from the two sources, a critical assumption for data fusion. In the second aim we compare responses from web and telephone data, controlling for the linking variables.  These comparisons will increase our understanding of whether web panel responses will produce estimates comparable to telephone given that there could be mode effects and/or differences due to population representativeness. In the third aim, we assess the degree to which web data can be integrated with telephone.