Monday, 11 January 2010

Controlling for culture

Economists like to control for various confounds in empirical work.

Here's a potential new one for the right hand side of your favourite regression: zip-code level data on Netflix rentals. Half Sigma points to a New York Times piece on neighbourhood patterns in movie rentals. For the top 100 movies rented overall in 2009, you can check zipcode level differences in movie rankings for a few different cities. So Half Sigma argues that you should never move to a neighbourhood where Paul Blart: Mall Cop was popular. Says Half Sigma:
Mad Men makes the list of top 50 rentals in very few zip codes, but it’s heavily rented in sophisticated urban areas, like Manhattan and especially western Brooklyn. In fact, it’s interesting to note that western Brooklyn, where Mad Men is in the top 10, has stronger preference for Mad Men than Manhattan.

Last Chance Harvey is a movie that’s only enjoyed in wealthy suburbs like Scarsdale and Great Neck. It’s very much a non-prole movie, but there’s also very little appreciation for this movie in places where Mad Men is liked.

The vastly different rental patterns Mad Men vs. Last Chance Harvey demonstrates that there are two very different types of upper middle class.

Without knowing anything about the city of Chicago, for example, I can look at the Netflix map and tell you that zip code 60523, which includes West Chester and Oak Brook, is a wealthy family-oriented suburb because of the high number of rentals of Last Chance Harvey, and that the area along the lake, north of the downtown area, between McCormickville and Evanston, is where the hip intellectual child-free people live because they rent Mad Men, and that Addison is full of white proles because Paul Blart: Mall Cop is heavily rented but the Tyler Perry movies are not...
I wonder whether including rankings on a few roll-call movies of this sort might pick up some cultural effects that are missed in other measures. If you could get the cardinal differences within zip codes for the ordinal rankings, you could run a measure of cultural heterogeneity based on movie-watching patterns. The folks doing social capital work often want to check whether heterogeneity promotes or discourages trust; I think this kind of heterogeneity is more interesting than racial or income heterogeneity.

I wonder whether Netflix would make the raw data available.....

Update: See also Social Science Statistics Blog


  1. Sure, but WHAT cultural effects? I can't very well write "and to control for bad taste in movies, we add MALCAOP variable..." What significant differences in culture can we use movie rental patterns to approximate?

  2. Unobserved aspects of education for starters?

  3. Hey Eric,

    Almost sounds like customer profiling. For the demand profiling work I've done for customers in the past, there are two data sources which nicely discriminate relatively homogeneous household sector groups: Mosiac New Zealand's data ( and the University of Otago's Depreviation Index (

    Both of them are based on Census data and are centered around the notion that like people tend to geographically cluster together. Both data sources discriminate changes in the demand for goods and services nicely.

    Enjoy... ;)

  4. Didn't know about Mosaic...looks neat, but also looks like it's a for pay data source....

  5. Yup, its a pay for data source. I used it at Telecom. Running a statistical sample of one, I dropped in where I have lived before and it spookingly gave me a profile of the socio-demographic make up that was 'about right', and suggested what their consumer preferences might be. Neat stuff, although there's significant collinearity between it and the Dep index.

    Its neat because it brings in the information from other sources about their tastes, like "Group X has high spend on Magazines relative to others" so pitching your advertising spend to magazines has higher likelihood of being seen by them.