Welcome. This applet is designed to allow you to explore the robustness of the results of Shapiro, Hitsch & Tuchman (2020) to various modeling choices. Updated results will be loaded in as they become available.
Please cite any use of the figures from this applet both to the URL (https://advertising-effects.chicagobooth.edu/) and the paper:Shapiro, Bradley and Hitsch, Günter J. and Tuchman, Anna, Generalizable and Robust TV Advertising Effects (August 7, 2020). Available at SSRN: https://ssrn.com/abstract=3273476 or http://dx.doi.org/10.2139/ssrn.3273476
This tab shows our results for ROI (return of investments). All calculations are based on estimated advertising effect from border-strategy with carry-over parameter set to 0.9.
Average ROI of Weekly Advertising: Set one week's advertising equal to zero and leave all other weeks unchanged. Compute ROI of going from zero ads in that week to the observed level of ads. Average across weeks where we observe positive advertising.
Average ROI of All Observed Advertising: Compute ROI of going from all weeks with zero advertising to all weeks having the observed level of advertising.
We analyze how much larger the ad effect need to be in order for the ROI to be equal to zero.
The numbers in the histogram represent by how much you would have to multiply the estimated ad effect in order for the ROI to be equal to zero. Only brands with positive estimated ad effect are included. Observations in blue indicate that the break-even ad effect is within the confidence interval of the estimated ad effect.
This tab is specific only to specifications where we estimate the best delta.
This tab shows the residual variation in advertising stock, net of market/store, month, and week-of-year FEs and price, promotions, and feature/display controls for different main specifications and assumed delta. The coefficient of variation is computed as the ratio of the residual variation to the average DMA-level weekly GRPs for that brand. The residual SD is computed as the standard deviation of the residuals.
This tab shows our main results estimated using different functional forms. We included polynomials up to degree 10, polynomials of the log(1+A) up to degree 10 and a cubic b-spline with 9 knots, using cross-validated LASSO to select the functional form. We compare the results to our main log(1+A) model. The comparison is shown in a scatter plot, where the predicted quantities are on the vertical axis and ad stock is on the horizontal axis.
To compare the results at similar level, We adjusted the mean value of predicted quantities to be the same for some brands.