Explain correlation vs causation and heatmap limitations in EDA

What's really being asked

This question checks three things at a senior level. First, do you understand that correlation is a mathematical summary of linear association while causation is a structural claim about interventions? Second, can you use an EDA visualization appropriately without overstating what it reveals? Third, do you know the specific failure modes, like confounders and non-linearities, that make correlation matrices dangerous for causal inference?

The full answer

A strong response moves in four stages. First, define correlation as a measure of linear co-movement, typically Pearson, bounded between minus one and one, and note that it is symmetric and unitless. Define causation as a directional claim that an intervention changes another variable. Supporting it requires a defensible identification strategy, such as randomization or an observational design with credible assumptions. Second, describe the heatmap use case during EDA: scanning for multicollinearity before regression, spotting missing value patterns if correlations drop after imputation, and prioritizing which bivariate relationships deserve deeper modeling. Third, list the heatmap limitations explicitly: a standard correlation heatmap captures the selected correlation measure, often linear correlation; it can hide confounding variables, be sensitive to outliers and range restriction, and invite narrative fallacies because color intensity feels like importance. Fourth, name remedies such as partial correlation, domain knowledge, randomized experiments, or causal graphs to move from association to inference.

The mistakes people make

The biggest red flag is saying correlation never helps with causation or that a high correlation coefficient is meaningless. Both extremes are wrong. Another error is claiming heatmaps show causal strength or that removing one correlated feature is always safe. Some candidates also forget that correlation matrices assume numeric variables and miss that categorical or ordinal data need different treatments.

What usually comes next

An interviewer might ask how you would detect non-linear dependence, perhaps mentioning mutual information or scatter plots. They might ask how to handle a pair of features with correlation zero point nine five, or how you would design an A/B test to establish causation for a specific product metric. Be ready to sketch a causal diagram or explain Simpson's paradox.

A concrete example

Imagine an e-commerce dataset where daily ice cream sales and sunscreen sales show a correlation of zero point eight five. A heatmap lights up bright red. The right EDA reaction is to check for a confounder like outdoor temperature or season, not to conclude that discounting sunscreen will boost ice cream revenue. You would then look at partial correlations controlling for temperature, or run a natural experiment, before making any causal business recommendation.