Causal Models for Business Decision Making

Managerial decision makers need reliable methods for achieving business objectives, whether they seek to maximize revenue, profit, or some other measure of business performance. However, businesses as well as their external environments are quite complex and difficult to analyze, and in practice it is hard to predict the results of a given managerial action. For example, if a retailer spends a million dollars redesigning the floor layout of their stores, will the retailer at least break even in terms of increased sales? Or would it be wiser to invest in improving the quality of service from the salespeople? For guidance in making these investment decisions, the business community currently relies at best on unreliable correlational models, or on gut instinct. There has been a lack of sufficiently powerful scientific methods for building models that can reliably predict the results of a business strategy, and thus enable the decision maker to maximize return on investment.

However recent developments in the emerging field of causal modeling offer substantial advances that can change the practice of managerial decision making. Unlike the traditional `predictive’ models commonly used in marketing, causal modeling techniques are used to reliably discover, represent, and reason about cause-and-effect relationships instead of correlations. The resulting quantitative models support estimation of the true expected returns from a given business intervention, enable computation of the optimal set of interventions, and support what-if analyses to handle uncertain situations. Exploiting the full power of the new approach requires integration of techniques from many disciplines, not just statistics. This paper presents such a unified framework for modeling decision processes and making the best investment decisions.

Our general framework was developed and tested in the process of building a specific causal model of retail sales, which we also present in this paper. The model captured numerous factors that affect a consumer’s choice of retailer from whom they purchased a large electronics product. Not only did the model answer the above question about the effect of improving store layout---and many other questions relating to product, price, promotion, etc.---but it also showed how the retailer could double their sales via just three factors among the thousands studied.

The framework presented in the paper is used as the basis for introducing several state-of-the-art methods and tools for gaining customer insight. It shows how to rigorously use ethnographic techniques in support of variable discovery and causal modeling. Itpresents methods for fusing qualitative and quantitative research methods with scientific rigor, avoiding the weaknesses of each. It describes new statistical algorithms and tools for causal discovery. The paper also makes significant theoretical contributions by introducing the notion of quasi-deterministic relationships which are a fundamental type of choice behavior.