The Digital Media Consortium undertook to evaluate different methods of attribution. These methods included: a household level regression, a household level matched panel, a household level exposure/lift model known as “cognitive analytics” and store level market mix modeling. The Consortium uses a combination of simulation and statistical analyses to identify the conditions under which each method performs best. The analysis covered eleven CPG brands in nine categories, $30 BN of sales analyzed, four thousand ad campaigns and 3.6 BN digital display impressions.
The optimum results varied by the quality of the data used. So, for very clean data, the household regression technique performed the best. This technique was a HH-Level logit model predicting brand purchase based on previous HH base sales, digital ads, TV ads, price, in-store promos and seasonality.
However, for typical CPG data, the matched panel performed the best for digital advertising. This approach used HH-level propensity-adjusted matched panel analysis. It is based on James Heckman’s (U Chicago) selection bias methods, for which he won a Nobel prize. Accounts for HH’s historical brand purchasing, demographics, and media exposure are included in this model. However, for linear television, traditional market mix model appeared to work better.