We partnered with an opera house that provides a social youth program for children from disadvantaged rural areas offering access to culture and music. The project is financed through donations and the recipients of the donation ask were individuals from the database of opera customers. The opera started engaging in this type of fundraising just two years earlier and had run a total of two fundraising drives prior to this one. Thus, we have a (small) set of past donors we can draw on and previous non-donors who can be partitioned into a set of regular customers and a set of new customers. For the regular customers we know a number of individual characteristics including the number and value of tickets purchased that serve as proxies for income and affinity with the opera house, as well as (self-indicated) gender, family, academic status, and the place of living. For the set of new customers the personal information was not available ex ante.
Unlike the personalization studied in Edwards and List (2014), we did not want to make the connection between the personal characteristics and the threshold obvious. Therefore, we offered a fixed matching of €10 for donations exceeding a specific threshold that was referred to as “large donation” and was not flagged as personalized. In total, we sent 10,004 letters to the subset of opera goers who purchased at least one opera ticket in the last season and, based on their past purchasing behavior, were expected to donate the largest amounts, including all past donors. The recipients consisted of three groups: those who had donated at least once in the two last fundraising campaigns (769 past donors), customers who had attended the opera house in the last three seasons and who had received a fundraising call in the last two calls but did not donate (3,859 regular customers), and new customers (5,376) for whom it was the first fundraising call from the opera house.
The letter informed the recipients that a generous lead donor had been found who would top up an individual donation with €10 if this donation met a minimum threshold (called “large donation”) or exceeded it.
The literature has documented sizable persistence in donative choices. Charitable giving in one year is the best predictor for giving next year (Meier 2007; Landry et al. 2010) and the amounts chosen are usually very close to previous amounts (Adena and Huck 2019). The data from previous campaigns of the opera house reveals that a subset of past donors gave twice in the previous years (a retention rate of 36.5% in the second call). There is a high correlation between the gift levels of repeat donors (0.778) with a paired t-test p-value of 0.482. Consequently, we assume that past behavior is a good proxy for the optimal donation in the absence of a match and we use the (maximum) past donation for the 769 past donors in our sample as such proxy.
For established customers (past non-donors) we predict optimal donations by extrapolating from the estimated giving equation of past donors. More specifically, guided by a lasso selection procedure, we use information on ticket purchasing behavior (from 2016: ticket revenue, ticket revenue (log), average price, dummy equal to one if any tickets bought in a particular year; from 2015: number of tickets, ticket revenue, ticket revenue (log), average price) and individual characteristics (dummies for living in Dresden, living in Germany, for subscription holders, female, couple, an academic and a professorial title). The raw predicted donation is, of course, almost never a round number, and on average, somewhat smaller than the average donation of past donors. In order to address this issue, we ordered individuals according to their predicted donation and then assigned them to the same rank of the actual distribution of past donations. We shall simply refer to the resulting amount as the predicted donation.
We assigned the following thresholds: For one third of past donors and regular customers the threshold was set equal to either the maximum past donation (for donors) or to the predicted donation (for non-donors). For another third of these recipients the above thresholds were lifted up to the next “category” of previously observed donations. With few exceptions this resulted for past donations below €40 in threshold increases of €5, for donations up to €95 in increases of €10, for donations up to €120 in increases of €20, and for higher donations in increases of €50. For the remaining past donors, established customers and all new customers, the threshold values were drawn at random from the distribution of past donations (for the first two groups excluding own past or predicted donations).