In a previous post I offered top predictive variables for Annual Giving. Today let’s talk about **Planned Giving**, which I think is a lot more challenging.

Why more challenging? Because compared with Annual Giving, there’s just not as much data to work with. Predictive modeling requires a certain amount of historical data in which to sift for characteristics associated with the desired behaviour we’re trying to model. If you already have many donors of a certain type, you’re better able to identify the characteristics that flag similar individuals in the non-donor pool. No past donor data, no model.

For Planned Giving, your existing pool of expectancies is probably small, both in absolute terms and as a proportion of your total alumni population. The haystack is huge and the needles are few! Experts will offer you advice on the traits that characterize good prospects for Planned Giving (“many small gifts over time” is a popular notion), but when you approach the problem from a predictive modeling point of view, your job is to study many traits (i.e. variables) in your own data – NOT rely on assumptions or rules of thumb.

The approach is not to reject rules of thumb, but to **test them**. In other words, try to convert the assumptions (eg. “many small gifts over time”) into **variables**.

Fortunately, for this type of model we are free to use variables based on giving, which is something we could NOT do with the Annual Giving model. Why? Because our predicted value in the Annual Giving model was ‘Giving’ – we can’t use Giving to predict Giving. But for our Planned Giving model, our predicted value is a binary Yes/No value (1 for Yes, 0 for No) – Yes meaning the individual is a Planned Giving expectancy, No for everyone else. That’s not the same as ‘Giving’.

(Sorry – that’s probably clear as mud. In another post I will discuss how to create the Predicted Value, a.k.a. dependent variable, which is absolutely key in defining your model. For now I’m talking strictly about predictor variables, a.k.a. independent variables.)

Here are two examples of variables based on giving patterns.

According to one expert opinion in the field of Planned Giving, an institution’s best prospects are those with the **greatest number of years of giving**, regardless of the size of gift. There’s some database query work involved here, but if you can count up the number of years that an individual has given a gift, for all the years you have data, you’ve got a good variable to work with. My institution has a little over 20 years of donor data, so values for this variable range from 0 to 20.

A slightly more involved variable is **‘Frequency’**. A formula for calculating “frequency of giving” is below. Again, there’s some work involved: You’ll need to query the database for the minimum and maximum fiscal years of everyone’s giving records, and then you’ll need to create some derived variables in your stats or modeling software to create the variable.

**[NOTE: Please see correction to this formula below!]**

In other words, ‘Frequency’ is equal to the number of gifts made by the donor, divided by the number of years the individual has been a donor. Note that if a donor’s first and last year of giving is the same, the denominator will be zero, which is a no-no; in your derived variable, you’ll recode those zeroes to ‘1’.

Donor frequencies in our university database range from a high of 5 gifts per year, down to as few as one-tenth of a gift per year. (Actually, I count pledges rather than gifts, to avoid distortion caused by monthly automatic payments such as payroll deduction.) All non-donors are given a ‘Frequency’ of zero.

‘Frequency’ might be superior to ‘Years of Giving’, as it levels the playing field somewhat between older and younger donors, but let’s not be too hasty to reject a variable. In **Part 2**, I will show you how both of these new variables can be **explored and tested**.

**CORRECTION**:

Peter Erhardt, Data & Research Specialist with Development & Alumni Relations at the Rochester Institute of Technology in Rochester NY, helpfully pointed out a problem with the formula I’ve got above. The denominator should read “Last year of giving minus First year of giving **+ 1**“.

The way I had it (missing the “plus one”), the calculation of the number of years an individual has been a donor will always be off by one year. For example, if a donor gave once per year in 2005, 2006, 2007, 2008, 2009 (5 gifts, 5 years, so once per year) then the formula would give 5 / (2009-2005) = 5/4 = 1.25 gifts per year, which is inaccurate. The + 1 in the denominator will turn this into 5 / (2009 â€“ 2005 + 1) = 5/(4+1) = 5/5 = 1, which is the correct one gift per year.

Not only does this correct the “number of years” problem, but it removes the need to recode a “zero” denominators to 1’s. **Thanks, Peter!**

[…] — kevinmacdonell @ 1:13 pm It’s time to explore the two variables we created in Part 1. The first was ‘Years of Giving’, and the second was ‘Frequency of Giving’. […]

Pingback by The 15 top predictors for Planned Giving – Part 2 « CoolData blog — 8 January 2010 @ 1:14 pm

good advice for computing frequency, thanks – I’ll try it out in my next planned giving predictive model.

Comment by Diane — 29 March 2010 @ 11:13 am

Diane – Please note the correction above.

Comment by kevinmacdonell — 29 June 2010 @ 8:23 am