Best Practive for Customer/Universe Comparison

[daniel_mmi]

I've been working on a process that features the following steps:

- Spatial match a customer list against all block groups to assign the relevant group for each

- Use the same customer list to choose appropriate larger geos, like DMAs, CBSAs, 'Places', etc.

- Append behavioral/preference data from Experian and Simmons to both sets

- Compare the values for each vs. the general population to get an Index value. For instance, if a certain factor, like 'Luxury Merchandise Buyer' appears in 0.1% of the general population and 0.2% of the customer population, the index would be 200(%), because the factor appears twice as often.

This part has been pretty straightforward, but when I start switching up my geographic units, my results sometimes change dramatically. If I do the most basic comparison of DMA (a large area) vs block groups, I see what seems to be a profile that lines up with my expectations, and as I shrink the comparison area down to CBSA and then census tract, the profile stays similar, though the index values shrink. This makes sense, as there will likely be more similarity within smaller units.

Is there a best practice in terms of what kind of comparison scale is most useful/likely to provide usable results? I've also experimented with using Custom Geographies to locate the representative customer locations at an even finer grain than block groups, but that seems to be the point where my 'Vegetarian' and 'Will pay more for organic' groups turn into 'Drink to get drunk' and 'Fast food fits my schedule and budget' groups, and I realize I'm really just fumbling around in the dark.

Thanks in advance for any insight/suggestions/witty comments.

[Philip]

Are you familiar with the Modifiable Areal Unit Problem?

http://gispopsci.org/maup/

This would be the starting point, and different researchers have provided different ways of determining spatial units.

To determine the best geographical unit, what is the question you're trying to answer or problem you're trying to solve?

[daniel_mmi]

That's a pretty great start, thanks.

I had already stumbled across the MAUP, though I wouldn't have known what to call it. I guess it's a testament to Alteryx that we have *so* many ways of organizing spatial information, even if we can also set ourselves up to run afoul of that particular problem.

[Philip · Accepted solution]

It's a really common problem. You might use customer data and a convex hull to create catchment areas. Then you could intersect the smallest demographic unit information you have with the catchment to create customer zones.

[daniel_mmi]

That's a good suggestion. I had to do something like that yesterday for someone who needed me to describe the area of a set of ZIP codes but could only take information in the form of lat/long and radii. I filled the hull with a grid and manually adjusted the centroid size until everything was covered. Reminded me of an old county fair game where you're trying to cover a star with three circles to win a prize.

It's such a nebulous concept, whether or not it's better to say what differentiates customers from whe market as a whole, from their locality, or even from their neighborhood. I take some consolation in the process because the general set of factors stays similar as you zoom in, though the index values get closer to 100. Along the way, though, there are a couple levels where the results go wacky, almost invert themselves.Seems like a practical example of those MAUP illustrations.

[Philip]

As another approach, you could find clusters using SatScan (http://www.satscan.org/) inside a custom R tool using rsatscan. I was just using it last week and it's really handy.