To review our complete methodology, please view our Technical Appendix.
The Philanthropy Outlook produces forecasts for the annual growth rates and levels of individual/household, foundation, estate, and corporate giving and giving to education, health, and public society benefit for the years 2015 through 2018.101 The forecast for total giving is produced as the sum of the four donor components. Collectively, 25 different variables, plus lagged values for many of these variables, were incorporated into the final models for giving by donor and recipient subsectors.
In the initial stages of methodological development, all possible combinations of variables were compared, resulting in more than 100,000 regressions for individual/household giving alone. Fewer regressions were needed for the three remaining components. For each component, the best model was selected by first considering its explanatory power through 2015. Those models with the best explanatory power were then re-estimated through 2003. One-year-ahead forecasts were constructed through 2015 for these models, and the best model was selected as the one with the lowest root-mean squared error.102 We relied on historical data from Giving USA: The Annual Report on Philanthropy and available IRS data. See Figure 1 in the Technical Appendix for a comparison of actual versus predicted growth rates for total giving for the years 2015 to 2018 and also the section titled, “Variable Definitions and Sources,” for a list of the candidate variables. We know that sometimes an event can have a delayed effect on giving. For that reason, weconsidered previous-year and contemporaneous values of the explanatory variables as well as previous-year values of the dependent variables (i.e., historical giving values).
For the individual/household and corporate giving models, it is not practical to test all of the variables at the same time. Instead, we adopted a three-step approach. In the first step, only the current values of the candidate variables were included in the regression. The best model within this set is referred to as the “base model.” The selection procedure was implemented over all possible combinations of the lagged variables added to the base model. The best model following this step is the “revised model.” In the third step, the selection procedure was run over all possible combinations of variables in the revised model. The result is the “final model.” The estate and foundation models were estimated in a single step, because the number of candidate variables was small enough that the previous and current values of the variables could be evaluated in one program.
The models for estimating giving to the recipient subsectors were developed using a modified version of the aforementioned individual/household and corporate giving models. In general, giving to the recipient subsectors is difficult to predict, as each of the subsectors experiences unique conditions that affect giving. Moreover, because there are several subsectors that receive gifts from the four major donor types, the subsectors experience more variance in their giving on a year-to-year basis than do the sources of giving. To adjust for these factors, additional steps were added to the original three-step approach. When using the first-step “base model” approach, we tested all combinations of a set of subsector-specific variables. These additional variables were derived from the different types of personal consumer expenditures, which allowed us to evaluate variables more specific to each particular subsector. The variables were then tested with the lag of all personal giving variables, and that resultant list was then tested with the lag of all subsector-specific variables. This “revised model” was then tested against all possible permutations of itself, which results in the “final model” for each subsector.
Tables 2 and 3 in the Technical Appendix describes the models for each source of giving and for giving to the recipient subsectors. Note that for each source of giving, with the exception of giving by estates, the adjusted R2s (coefficients of determination) are high. Moreover, the signs of the coefficients are generally consistent with the economic theory that giving responds positively to increases in the ability to give and general economic conditions. See Table 4 in the Technical Appendix to reference the ratio of root-mean-squared error to the standard deviation for each giving prediction.
The forecasts of the different components were processed using the final version of each model. The forecasts covered 2015 to 2018.103 Implementing the forecasts entailed auxiliary models for the explanatory variables (i.e., independent variables). These auxiliary models are described in the Technical Appendix.