The Usual Susceptibles

Uplift Modelling: Why, How, When?

When designing any intervention to affect change in people's behaviour, a key question is always whom to target. In a world of limited resources, we should always aim to target those most likely to be positively affected and avoid those who may respond negatively.

Consider the example of developing a new pharmaceutical drug. We would like to identify -  based on the details of the ailment being treated, characteristics of the patient, and the known side-effects of the drug - who will benefit most strongly if the drug is prescribed.

In business, whether attempting to market new products or reduce churn, it is essential to aim our campaign at those we stand to gain most from. The most elegantly-designed campaign will fail if it only reaches individuals who have already made up their mind.

Predictive analytics plays an important role in systematically establishing the features that determine the propensity of individuals to respond to a campaign, and based on these features, the exact individuals to target. Propensity modeling typically involves collecting data that is a useful predictor of the characteristic we are interested in influencing. For example, the likelihood of subscribing to a service, defaulting on a loan, or spending a certain amount of money, We 'train' machine learning algorithms to predict propensity based on these features and then select the highest ranked individuals based on the assumption that it will be most advantageous to target those with the highest propensity.

This workflow, while commonly used and sometimes met with moderate success, makes some flawed assumptions about where our campaign will make the most difference. Assuming that those with the highest propensity to buy, subscribe or default are the same individuals we ought to target, often leads to wasted resources. It can even reduce success compared to doing nothing. It is much more useful to break a campaign population into four groups namely: the Susceptibles, Sleeping Dogs, Sure Things and Lost Causes.  

Sure Things are the people who will respond in the manner we desire regardless of whether we intervene or not. Lost Causes on the other hand, will not change their behavior, no matter what we do. Resources spent targeting either of these groups are simply resources wasted.

Sleeping Dogs can be inadvertently nudged to behave in the opposite direction we were hoping. For example, customers who might have had no incentive to cancel a subscription contract but are motivated to do so after a phone call or a text message. Or people who would normally not vote for Party X, but wake up on election day and do just that. Why?  Because they received an irritating phone call from Party Y.

The final group, the Susceptibles, are the only individuals from whom a campaign stands to benefit. These are people who currently do not respond as we would like, but who could be convinced to change because they are contacted. Predictive analytics can help us identify these individuals in the population and the analytical method design that does this is called Uplift Modelling or Persuasion Modelling.

One of the best known examples of this technique was Barack Obama's 2012 presidential campaign that used uplift modelling to optimise the placement of calls, door knocks, flyers, television and radio ads. The data scientists working on the Obama for America 2012 campaign, actively directed volunteers and marketing spend towards the most susceptible constituents and did so with great effect.

The key to identifying the Susceptibles using uplift modeling, is the ability to predict the behavior of targeted individuals before and after our campaign. No amount of analytical effort can accomplish this without first testing a variety of campaigns on your customer base - including doing nothing. With this information on hand, it is possible to predict the propensity of the behavior we are interested in for any hypothetical or real customer and for any of the campiness we have tested.

The importance of experimentation

Perhaps the lesson here is the importance of experimentation in business. If you don’t try a range of different campaigns and experiments, you have no way of knowing who you should be actively targeting and who you're just wasting money on. Uplift modeling is only useful if you are farsighted enough to experiment with the ways in which you interact with your customers.