Is Multi-Collinearity Destroying Your Causal Inferences In Marketing Mix Modelling?

This article discusses the issue of multi-collinearity in machine learning, particularly in marketing mix modeling (MMM), and explores methods to address it. Multi-collinearity arises when independent variables in a regression model are highly correlated, making it difficult to isolate their individual effects. This can lead to biased coefficient estimates, inflated standard errors, and even sign flipping, where the direction of a variable's effect is reversed. In MMM, multi-collinearity is common because marketing budgets are often set based on demand forecasts, leading to correlated spending patterns across channels. Detecting multi-collinearity can be done through correlation matrices, variance inflation factors (VIF), and analyzing standard errors. Addressing multi-collinearity involves techniques like removing or combining correlated variables, applying regularization methods like Ridge or Lasso regression, and using Bayesian priors. Bayesian priors introduce prior beliefs about parameter values, regularizing the model and mitigating the impact of multi-collinearity. Another approach, specific to MMM, is introducing random budget adjustments to decouple spending patterns. The article concludes with a case study demonstrating these techniques, highlighting the use of Bayesian priors and random budget adjustments to improve the accuracy of MMM results in the presence of multi-collinearity.