Peer Effects in the Diffusion of Solar Photovoltaic Panels
The authors build a model for predicting how people who buy solar panels cause other people to buy solar panels. They build a model by using zip code data and solar panel adoption data. They augment their data set with features about those zip codes, including income, install base, gender, and commuters. They then borrow from other well known models, including Bass (1969), and then run their own model against the data. The authors find "strong evidence for causal peer effects, indicating that an extra installation in a zip code increases the probability of adoption in the zip code by 0.78 percentage points when evaluated at the average number of owner-occupied homes in a zip code.” (p. 910).
The managerial implication is that "targeting marketing efforts in areas that already have some installations is a promising strategy” (p. 911), and that demographic and behavioral targeting can enhance this effectiveness.
The appendix contains details how to estimate potential buyings at street level from zip code data, and a brief summary of the Bass Diffusion model.
Bass (1969) used aggregate market data to estimate innovators and imitators. He looked at aggregate adoption data for consumer products, plotted it over time, and noticed that it followed an S-curve. Bass observed two groups of people, innovators, who adopt the product early, and imitators, who adopt later on. One group causes the other to adopt a product. His model agreed with nature, and this is how the words 'innovator' and 'imitator' became adopted in text books.
Advances in digital measurement have enabled the likes of Iyengar and Godes to cheaply measure the effect of peer influence on product adoption from the transactional record. To replicate their approach, you need to have some information about how people are related to each other, how they regard each other, and the order in which they adopted, used, and increased their usage of a product. The methodology is available and very executable. The data isn't always readily available.
Bollinger and Gillingham go one level of abstraction up, to the zip code level, to observe the same fundamental phenomenon. Solar panels are unique in that visible ones (and the uglier the better!) are more desirable than the transparent variety. People want other people to know that they have solar panels. This approach to a model, by way of a social-geographic diffusion, is appropriate for solar panels, and, the finding of incremental adoption probability as adoption increases is compelling. Data about places, and the people in those places, is a fair bit more available. General adoption data for a given product is also, comparatively, an easier query to execute against a CRM database.
Their approach is quite a bit more generalizable beyond solar panels. There are a whole range of products that become incrementally more desirable the more other people adopt it. Clothing, cellphones, and automobiles are three such social-physical goods. Solar panels are stationary and tend to be visible from the street level. People are mobile, and their natural habitats can span hundreds of kilometres. As such, geographical segmentation would need to be larger if aggregate data were to be used.
If a marketer has a foothold in a given zip code, they will want to increase their efforts for a period of time dictated by the diffusion curve, as opposed to the arbitrary length of the campaign, before moving onto the next set of zip codes. The rapid turn-on/turn-off capability of digital media is particularly attractive for this type of optimization strategy.
This modelling approach, combining facts that are known and readily accessible about locations, and is known about how people react to each other, can be extended into forecasts and decision models.
I recommend members of the DAA read this paper.