Causal Analysis: The Next Frontier in Analytics?

Sometimes all we need are predictions that are “accurate enough.”  More often than not, though, in marketing research we need to understand the Why and not just the What. For example, why do some consumers prefer a competitor brand to ours, and vice versa?  Why are some consumers heavy (or light) purchasers of our category?  Are our price promotions subsidizing consumers who would have bought our brand anyway?  Are they eroding our brand equity?  Which elements of our marketing mix are working best and why?  To address questions such as these, data mining and predictive analytics is not enough and expertise in causal analysis is required. 

Causal analysis does not necessarily attempt to “prove” cause-and-effect relationships but, instead, assesses plausible reasons for patterns in the data we have observed.  Causal analysis is part of my daily work and a subject I’ve studied for many years.  It’s not an easy one!  Academics are still hard at work on it – especially in psychology, economics and medical fields such as epidemiology – and scholars in different disciplines tend to approach causal analysis from different angles. 

Even laboratory experiments in which subjects are randomly assigned to treatments still turn up false positives and false negatives.  Apparently strong correlations may be spurious or, conversely, strong correlations that are real and causal in nature may be masked by measurement error or variables we haven’t observed or taken into account.  Causal relationships are often mediated or moderated in ways that are very difficult to unravel.  There is also the question of representativeness and generalizability…even in rat studies when we’re not extrapolating our findings to humans. 

In non-experimental research, causal analysis is more challenging than in controlled experiments.  A sophisticated statistical model that includes a large number of control variables may still yield results that are well wide of the mark.  Competing models often suggest different interpretations and courses of action for decision-makers, which is one reason attempts to automate causal modeling are risky.

Nevertheless, causal implications will be drawn – even in qualitative studies – when we have to know the Why driving the What.  This is especially true in a field like marketing research, as I’ve noted.  We frequently do this quite informally, however, which can lead to costly mistakes.  Intelligent speculation is, after all, speculation.

Though it isn’t possible to get into this huge topic in depth in one short article, I have given a sketch of it in Pros and Pitfalls of Observational Research and reproduced a seminal article by renowned statistician Sir Austin Bradford Hill in Causation Matters.  Stuff Happens reviews David Hand’s celebrated book The Improbability Principle.  Meta-analysis and Marketing Research and Propensity Scores:  What they are and what they do are two other brief articles on related topics.

In addition to books, articles and courses on marketing, psychology, sociology and other subjects, below I’ve listed some books on research methods and statistics that have helped me get my head around causal analysis.  Some titles have been abbreviated to conserve space.

• The Improbability Principle (Hand)
• Introduction to Probability (Bertsekas and Tsitsiklis)
• Sampling: Design and Analysis (Lohr)
• Theory Construction and Model-Building (Jaccard and Jacoby)
• Design and Analysis of Experiments (Montgomery)
• Statistics for Experimenters (Box and Hunter)
• Experimental and Quasi-Experimental Designs (Shadish et al.)
• Design of Observational Studies (Rosenbaum)
• Causality: Models, Reasoning and Inference (Pearl)
• Statistical Mediation and Moderation (Jose)
• Regression And Mediation Analysis (Muthén et al.)
• Explanation in Causal Inference (VanderWeele)
• Causality in a Social World (Hong)
• Causal Inference (Imbens and Rubin)
• Counterfactuals and Causal Inference (Morgan and Winship)
• Using Propensity Scores in Quasi-Experimental Designs (Holmes)
• Propensity Score Analysis (Guo and Fraser)
• Multilevel and Longitudinal Modeling (Rabe-Hesketh and Skrondal)
• Longitudinal Analysis (Hoffman)
• Analysis of Panel Data (Hsiao)
• Linear Causal Modeling with Structural Equations (Mulaik)
• Longitudinal Structural Equation Modeling (Little)
• Longitudinal Structural Equation Modeling (Newsom)
• Time Series Analysis (Hamilton)
• Multiple Time-Series Analysis (Lütkepohl)
• Statistics for Spatio-Temporal Data (Cressie and Wikle)
• Principles of Forecasting (Armstrong et al.)
• Discrete Choice Methods with Simulation (Train)
• Applied Choice Analysis (Hensher et al.)
• Best-Worst Scaling: Theory, Methods and Applications (Louviere et al.)
• Introduction to Meta-Analysis (Borenstein et al.)
• Methods of Meta-Analysis (Schmidt and Hunter)
• Modern Epidemiology (Rothman et al.)
• Handbook of Survival Analysis (Klein et al.)
• Exponential Random Graph Models for Social Networks (Lusher et al.)
• Probabilistic Graphical Models (Koller and Friedman)
• Risk Assessment and Decision Analysis (Fenton and Neil)
• An Introduction to Agent-Based Modeling (Wilensky and Rand)
• Artificial Intelligence (Russell and Norvig)
• It’s Not The Size Of The Data – It’s How You Use It (Pauwels)
• Market Response Models (Hanssens et al.)
• Marketing Metrics (Farris et al.)

While space prevents me from offering synopses of these books, your favorite booksellers will probably let you look through their tables of contents and have a peek inside.  Most have been reviewed too, though I personally have found reviews aren’t always that useful. 

Of course, there also are peer-reviewed journals, conferences and seminars in addition to credit courses offered by universities and other educational organizations.  Some other books and resources that cover the “how to” in depth are listed here.

I hope you find this helpful!

Kevin Gray is president of Cannon Gray, a marketing science and analytics consultancy.

Arrange a Conversation