Adjustment of Selection Bias in Marketing Campaign
A standard practice to evaluate the effect of a marketing campaign is to divide the targeted prospects into two testing groups, a control group without the marketing intervention and the other with the intervention, and then to compare the difference of results between two groups given observed characteristics. The sole purpose is to see if the intervention is the causal effect of results that we are interested in, e.g. response or conversion rate. From the statistical perspective, it is desirable to randomly assign targeted individuals into one of the testing groups such that the background information of individuals in different groups are comparable or balancing and the assignment of individuals to groups is independent of outcomes of the intervention. This practice is also called Randomization. However, in many marketing campaigns that we observed, the randomization mentioned above is prohibitively difficult due to various constraints in the real world. An example is that the assignment of an individual to the specific group might be somehow determined by his background information, which might be related to his response to the campaign. As a result, this characteristic heterogeneity of targeted individuals between different testing groups will give a biased estimation for the campaign effect, which is also called selection bias. A typical observation of such bias is that the campaign effect looks more optimistic than it is supposed to be.
The selection bias is a common issue in many observational studies in social science. While different methods have been purposed to adjust or correct this bias, we’d like to demonstrate two modeling techniques to correct the selection bias in the marketing campaign, namely propensity score method and Heckman selection method.
Introduced by Rosenbaum and Rubin (1983), propensity score can be defined as the conditional probability of an individual receiving a specific exposure (treatment or campaign) given a certain observed background information. The idea of propensity score is to partial out the observed characteristic heterogeneity of individuals between testing groups so as to make the assignment of groups conditionally independent of the campaign outcome. While there are numerous ways, such as matching or weighting, to implement propensity score, we are particularly interested in the one proposed by Cela (2003) due to its simplicity and flexibility. The implementation of Cela’s method is straight-forward and considered a two-step approach.
First of all, we build a logistic regression using the assignment of groups as the response variable and then estimate the probability of an individual assigned to the group with marketing intervention given a set of background variables. The propensity score model can be formulated as:
Prob(d = 1 | x) = p = EXP(T * x) / [1 + EXP(T * x)]
where d = 1 for an individual assigned to the intervention group, p is the propensity score, and x is the covariate matrix of background information. In this step, while we propose using a logistic regression, other models designed for binary outcome, such as probit model or classification tree, might also work well.
Secondly, we build another logistic regression to estimate the casual effect of marketing campaign such that
Prob(y = 1 | z, d, p) = EXP(A * z + B * d) / [1 + EXP(A * z + B * d + C * p)]
where y = 1 for the individual with a positive response and z is the covariate matrix of control factors and background information. In this model, the propensity score acts as a control factor such that the outcome y is independent of the group assignment d after partialling out the observed characteristic heterogeneity incorporated in the propensity score p. In the formulation of our second model, the parameter C is the estimated causal effect of marketing intervention that we are interested. While we are using logistic regression for the binary outcome in our second model, any model within the framework of Generalized Linear Models (GLM) should be applicable for the appropriate response variable.
While the propensity score method is able to adjust the selection bias due to observed characteristics, it is under the criticism that it fails to address the bias arised from the unobservable. Originated by Heckman (1976), Heckman selection method is able to control the bias due to the unobservable, if the assumption about the underlying distribution is valid. In the framework of Heckman model, it is assumed that there are 2 simultaneous processes existed in the model, one called Selection Equation and the other Outcome Equation, of which the error terms follow a Bivariate Normal Distribution with the nonzero correlation. While two equations can be modeled simultaneously, we prefer a 2-stage estimation method due to its simplicity and its advantage in algorithm convergence.
Similar to the propensity score method, we fit a probit model to estimate the probability of an individual assigned to the group with marketing intervention.
Prob(d = 1 | x) = G(T * x)
where G(.) is the cumulative density functin of the Normal Distribution. Please note that the use of probit model is determined by the assumption of Heckman model for the Normal Distribution. Based upon the result from this probit model, we calculate Inverse Mills Ratio (IMR) for each individual, which is considered a measure of selection bias and can be formulated as
IMR = g(T * x) / G(T * x)
where g(.) is the probability density function and x is the covariate matrix in Selection Equation.
We build a second probit model by including IMR as one of the predictors, which can be formulated as
Prob(y = 1 | z, d, IMR) = G(A * z + B * d + C * IMR)
where y = 1 for the individual with a positive response and z is the covariate matrix of control factors and background information. While the significance of parameter C indicates the existence of selection bias, the lack of such significance doesn’t necessarily imply that there is no bias. The consistency of estimates in Heckman model strongly replies on the distributional assumption, which is difficult to justify in the real-world data.