GivingPulse 2024 Year in Review

Methodology Note

Predicting Value of Unsolicited Respondents

Test Accuracy Test Recall Test Precision Test F1
0.67 0.29 0.77 0.43

We use a decision tree classifier to identify GivingPulse respondents who were more likely to respond generously to solicitation among the unsolicited subset of the sample. A decision tree is a flowchart-like model used to split respondents up into increasingly homogenous groups with respect to an observed outcome. We trained a decision tree on the subset of solicited respondents in GivingPulse and used the model to predict solicitation response rates among the unsolicited subset of respondents.

The questions we used for the predictive model in this report are the following:

  • “I feel everyone has the responsibility to give and help those in need” – strongly disagree to strongly agree; 4-point scale

  • “I generally trust nonprofits and the services they provide” – strongly disagree to strongly agree; 4-point scale

  • Full-time employed - Yes or no

Once potential responders were identified, we used donation value data from the US Bureau of Labor Statistics Consumer Expenditure Survey (CE) to generate ranges of donation values for each GivingPulse respondent. The CE provides estimates of average donations to charitable organizations, religious institutions, educational institutions, and political organizations in the past quarter broken down by their demographic characteristics. To calculate a lower bound and upper bound estimate for individuals in the GivingPulse dataset, we use the following procedure:

  1. We extracted from CE summary tables the mean donation value in each of the categories above for the years 2014 to 2024 broken down by the following demographics: age group, income group region of residence, retiree status, full-time employment status, type of area of residence (urban or rural), and ethnicity. To obtain a singular value, we used the most recent available estimate for each subgroup and donation category (sometimes CE subgroup estimates are suppressed due to high relative standard errors).

  2. Within each demographic subgroup, the mean values for each donation subcategory were summed to calculate the average total donation across all categories for each subgroup.

  3. Each GivingPulse respondent was then assigned a set of average donation values based on their demographic traits. For example, if a respondent was full-time employed, between 18 and 30 years old, and living in an urban area, their set of average donation values would contain one CE subgroup estimate value for each of these demographic cells.

  4. The lower bound and upper bound donation values for each respondent are calculated as the minimum and maximum average donation values from the set of demographic subgroup averages.

  5. We then compute the average lower bound and upper bound donation value among respondents who were identified by the decision tree as predicted responders ($1010 and $2530 respectively, rounded to the nearest 10).

The final value was calculated as:

(Predicted response rate) * (% of the population unsolicited within the past 7 days) * (Adult population in the US) * (Lower or upper average donation value among predicted responders)
= 0.1 * 0.71 * 258,300,000 * [1010, 2530] = [~19 billion, ~46 billion]

Civic Intent Score

Our measure of what we call “civic intent” consolidates the prosocial generosity behaviors we track with depolarizing beliefs and actions, trust in people and/or nonprofits, and one's good intentions to participate in and strengthen one's community. By combining these features in a single index, we can compare a single metric against all other beliefs, attitudes, worldviews, awareness, and feelings of belonging in our survey. 

Each item is weighted equally and adds one point to the overall score, or a 0-1 scaled value as noted. Scores are then normalized to a 0-100 point scale for ease of comparison.

Attitudes

  • Attitude: everyone has a responsibility to give and help those in need (0-1)

  • Attitude: it is important to me personally to help those in need (0-1)

  • Attitude: giving is somewhat or very important to me (0-1)

  • Attitude: intend to give as much or more in the future (0-1)

Trust

  • Trusting of people in general (0-1)

  • Trusting of nonprofits (0-1)

Depolarizing actions

  • I help others, even people whose beliefs, politics, or lifestyle I don’t agree with (0-1)

  • I strive to help those most in need, even if that means helping those from my community less (0-1)

  • I am helping to make my community a better, more civil place (0-1)

  • I have performed a random act of kindness for someone before (0-1)

Generosity behaviors

  • Gave in any form (1 point, or -1 point if did not give in any form)

  • Giving in multiple forms to multiple recipient types (0-1)

  • Recently started an effort to help others or an organization in one's community (0-1)

  • Choose any form of giving as one's most significant form of giving in the past year – and not "none of these" (0-1)

  • Recently gave in some form (0-1 based on recency)

  • Recently helped a local community person or organization (0-1)

We converted 4-point Likert scale questions into a 0-1 range, using: strongly agree = 1; agree=0.75, disagree = 0.25, strongly disagree = 0).

In some cases we use six questions from the above list that approximate civic intent (r=0.93 correlation with the overall civic intent scoring rubric above):

  • Attitude: It is moderately to very important for me to help those in need

  • Depolarization: I am helping to make my community a better, more civil place

  • Depolarization: I strive to help those most in need, even if that means helping those from my community less

  • Giving: Yes, I performed some act of generosity in the last 7 days

  • Giving: Yes, I recently gave in some form, weighted for recency

  • Trust: In general, most people can be trusted

ACP Clusters

The American Communities Project (ACP) is an initiative out of the Michigan State University’s School of Journalism that uses demographic, socioeconomic, political, and consumer survey data to group US counties into 15 distinct community types based on shared characteristics unrelated to their geographic proximity to each other. Full descriptions of each community type as well as a map of counties pertaining to each type can be found on the ACP website.

GivingPulse has been asking respondents to identify their county of residence since Q2 2024. Respondents can then be linked to their county’s corresponding ACP community type, which allows us to compute summary statistics of GivingPulse data points within each community type. Summary statistics for county types lacking a minimum representative sample (e.g. less than 30 people in a given period) are not shown.

Pew Political Typology

The Pew Political Typology is a model of the left-right political spectrum in the US developed by Pew Research Center based on a nationally representative survey of around 10,000 respondents every 3 years. The survey asks respondents for their opinions on a set of 27 social, economic, and political issues (including party affiliation and voting history) and uses responses to group similar types of people together via cluster analysis. The most recent study, conducted in 2021, identifies nine segments of the typology ranging from Faith and Flag Conservatives (far right) to Progressive Left (far left) with a low political engagement group called Stressed Sideliners in the center.

For the GivingPulse questionnaire, we reduced Pew’s original 27 question survey to a set of eight questions that most significantly distinguished the typologies from each other. Response options for each question are yes, no, or unsure, which are mapped to the numeric values 1, 0, and 0.5 respectively. We then compute the sum of squared errors (SSE) between each GivingPulse respondents’ set of answers and the responses to the corresponding questions on the original survey within each segment of the Pew typology. We assign Pew types to GivingPulse respondents by choosing the type that minimizes the SSE difference between how they answered questions and the percent agree/disagree/not-sure for the closest Pew type.