Causal Machine Learning for Evaluating the Effectiveness of Business Strategies

Abstract

Firms adopt strategies selectively: those that adopt a new pricing model, a digital channel or a sustainability programme differ systematically from those that do not, so a comparison of adopters with non-adopters confounds the strategy’s effect with the characteristics that drove adoption. This paper examines how causal machine learning can recover strategy effects from observational data and, beyond the average, identify which firms benefit most. A simulation harness generates 6,000 firms in which adoption depends on size, prior growth and digital maturity, outcomes depend nonlinearly on the same variables, and the true treatment effect is heterogeneous, averaging 1.69 outcome units. Five estimators are compared: the naive adopter–non-adopter difference, covariate-adjusted regression, double/debiased machine learning (DML), and T- and X-learner meta-algorithms for conditional average treatment effects (CATE). The naive difference overstates the effect almost threefold (4.85), whereas DML (1.74), the X-learner (1.72) and adjusted regression (1.72) recover it closely. The X-learner recovers firm-level effects with a correlation of 0.96 to the truth, and a targeting policy that adopts the strategy only where its estimated effect exceeds its cost achieves 98% of the oracle policy’s value, against 67% for uniform adoption. The paper sets out a workflow for strategy evaluation that moves from identification assumptions through estimation to policy learning, and discusses the organisational conditions under which the required assumptions are credible.