Before joining HelloWallet, I was a microtargeter; I’d co-founded a small microtargeting company that helped non-profits find potential volunteers for their causes. We used machine learning techniques on huge datasets to identify people who might become active. Our success was judged with randomized experiments showing whether the people we identified really did care about the cause and would volunteer (generally, they did).
One of the things we thought a lot about was what types of information you need, to discover if someone actually wants to, and is ready to, try something new. The best algorithm in the world doesn’t help if you don’t have good data to use it on. That’s especially true when you’re working with a rare and time-consuming behavior like volunteering, instead of voting.
So what data seems to give the most bang for the buck?
To make things concrete, I’ll talk about a client of ours that’s an advocacy organization; I’ll call them “ActMore” so they don’t sue me. ActMore is an environment-related NGO, and wanted to encourage its members to become more involved in their community of like-minded folks. They had a large number of people who had given money to the organization, and had said they wanted to do more, but hadn’t yet really gotten involved. First, ActMore needed the basics – age, location, income, gender, etc. They already had most of that information, and we supplemented it with third party sources to provide hundreds of data fields.
We wanted to understand how strong the members’ interest would be in a specific event the organization was putting together – a rally on Earth Day. We were looking specifically for divisions within the member base: groups of people who would respond differently to appeals to join the rally. Each group would get its own personalized appeal that made sense given their background and level of experience. Potential volunteers would interact with outreach materials and an associated website (and, as I remember, there was a phone-calling component as well).
Microtargeting entails throwing algorithms at large amounts of data. But, with ActMore, and with other companies and organizations I’ve worked with, there are certain types of data rise to the top. You can focus on them when you don’t havemounds of data, and get most of the way there. Here are questions that I’ve found are most relevant to ask about a population, from the perspective of product-mediated behavior change:
- Prior experience with the action: do the people have experience taking similar actions? (For ActMore – have they been to other, similar rallies? In the voter microtargeting world, one question is: have they voted before?) It’s much easier to increase an action than to start a new one. Existing habits around the action are especially important. Prior experience is almost always the strongest predictor of future political action.
- Prior experience with similar communication channels and products: if the product used to reach people employs an email and a website, do the people have regular access to a computer (and know how to use them)?
- Relationship with the company or organization: or, to put it more bluntly, do they trust you? You’ll have a harder time making your case for the people who don’t trust you versus those that already know you and love you . Again, this is especially true for rare behaviors like volunteering; I think it’s less so with voting.
- Existing motivation: why would the people want to take action in the first place? In other words, what can the product build upon, and so it doesn’t need to do all of the work itself?
- Physical, psychological, or economic impediments to action: this isn’t as common, but sometimes arises. Are there groups of people for whom the action is especially difficult? People that are homebound, or don’t have the money to travel to the rally, etc. (We faced this with ActMore, in fact).
These five things provide a broad picture of whether people care about your action, and whether they are likely to act. In a data poor environment (like most product-mediated behavior change), you can focus on them. To gather it, you can use the standard tools of market research and product development – collecting existing quantitative data on user demographics, using field surveys, running qualitative research with one-on-one interviews, etc. Just make sure you add these five questions: their actual prior experience with action and with the communication channel, existing motivations, existing relationships with the company, and problems vis. a vis. the action.
Since we were doing microtargeting, the end result of our process was a set of machine-learning models of the propensity of ActMore’s members to respond to different outreach campaigns, which we then field tested before rolling out the product for real. We used quantitative data from the organization and from third-party providers. But in a less quantitative-data heavy environment this approach can help you with the same core goal: who is ready and able to act, but just needs to be asked?
— This post is an extract from a much larger piece I’m writing about designing products for behavior change. It’s a step-by-step, “how to do it” manual. If you’d like to take a look at it, drop me a line, or just enter your email address here.
 While we’ve all heard humorous stories that what liquor people drink predicts their voting preferences, that stuff is all on the margins – it’s helpful after you have handled the really major factors.