PricewaterhouseCoopers along with The Interactive Advertising Bureau (IAB) publishes an annual report on internet advertising revenue. In 2001, they expanded their categorization of ad types to reflect the growth in alternative formats of online advertising. Between 2000 and 2001, annual advertising revenues dropped by $1 billion with banner advertising dropping by 12%–from 48% to 36%. Instead, keyword searches, a new category started with Google Adwords on October 23, 2000, picked up the slack with 4% of ad revenue in 2001. Classifieds also doubled between 2000 and 2001–growing from 7% to 16% of ad revenue.
This analysis tries to estimate how much the shift in the mix of advertising formats impacted revenues in 2001 with a regression discontinuity strategy. First, I plot total revenues by year in Figure 1 below, to demonstrate that advertising revenue grew exponentially between 1996 and 2015.
I then take the natural log of annual revenues to transform Figure 1 into a log-linear model, which allows for the application of a linear regression as the log-linear model is a linear combination of its parameters. The log transformed model looks like Figure 2, below. The red line indicates year 2001 and the figure clearly shows there is a discontinuity in slopes before and after 2001.
Figure 3 adds a linear trend line to the graph. The trend lines plot the fitted values of a regression of log revenues on year. There is a clear discontinuity at year 2001 where log revenues drop by 1.26 log points. The idea is that, absent the change in the mix of ad formats, log revenues would be 10.01 log points. However, with the change in formats, log revenues equal 8.76 log points. The interpretation, in the original metric of revenues, is that revenues drop by 72%. [(e^(-1.26)-1)*100=-72%]
The next step is to address the curvature in the plot of log revenues by year. Figure 4 shows the plot with a quadratic trend line. The quadratic line plots the fitted values of a regression of log revenues on year and year-squared. There is still a clear discontinuity at year 2001 where log revenues drop by .74 log points. Again, the idea is that, absent the change in the mix of ad formats, log revenues would be 9.41 log points. However, with the change in the formats, log revenues equal 8.67 log points. The interpretation is that revenues drop by 52%. [(e^(-.74)-1)*100=-.52%]
Finally, I implement the full regression, which estimates the jump in revenues at the 2001 cutoff point. The regression includes a treatment variable, D, and controls for year and year-squared. The regression coefficient on D estimates the jump in 2001 due to the change in the mix of ad formats. The result is that the change in the mix of ad formats decreased revenues by 56% in 2001. [(e^(-.83)-1)*100=-.56%]
To contextualize the decrease by 56%, note that prior to 2001, ad revenues were doubling every year. Therefore, absent the change in mix of ad revenues in 2001, ad revenues should have been on a trajectory to double in 2001. This analysis looks at the difference between what should have happened absent the change in mix of ad formats and what did happen with the change. The result in the change of formats was a decrease in revenues by 56%.
Ad fraud clearly changes with year. One study said that click fraud is currently growing at 50% a year. To separate the trend in advertising fraud from the jump in ad revenue in 2001, this analysis controls for changes in ad fraud by controlling for year.
This analysis shows that the change in the mix of ad formats in 2001 led to a temporary drop in ad revenues. The shift away from banner advertising in favor of keyword searches with the advent of Google Adwords and also classified ads were largely responsible for this drop. The keyword search format eventually grew to 34% of ad revenues in 2015 with over $20 billion attributable to search, but the initial introduction in 2001 led to a drop in ad revenues.