An optimal contact model for maximizing online panel response rates

We develop and test an optimization model for maximizing response rates for online marketing research survey panels. The model consists of (1) a decision tree predictive model that classifies panelists into “states” and forecasts the response rate for panelists in each state and (2) a linear program that specifies how many panelists should be solicited from each state to maximize response rate. The model is forward looking in that it optimizes over a finite horizon during which S studies are to be fielded. It takes into account the desired number of responses for each study, the likely migration pattern of panelists between states as they are invited and respond or do not respond, as well as demographic requirements. The model is implemented using a rolling horizon whereby the optimal solution for S successive studies is derived and implemented for the first study. Then, as results are observed, an optimal solution is derived for the next S studies, and the solution is implemented for the first of these studies, etc. The procedure is field tested and shown to increase response rates significantly compared to the heuristic currently being used by panel management. Further analysis suggests that the improvement was due to the predictive model and that a “greedy algorithm” would have done equally well in the field test. However, further Monte Carlo simulations suggest circumstances under which the model would outperform the greedy algorithm.