Data Scientist (US)

How do you determine the required sample size for an experiment before running it?

I run a power analysis based on the minimum effect size that would actually matter for the business decision, the expected baseline variance, and the desired statistical power, typically 80%. Skipping this step is how teams end up running underpowered experiments that can't reliably detect real effects, then wrongly concluding a change had no impact.

What's the difference between correlation and causation, and how do you establish the latter with observational data?

Correlation just means two variables move together, which can be driven by a third confounding factor rather than either variable causing the other. With observational data, I look for natural experiments, instrumental variables, or matching techniques to approximate a controlled comparison, since a true randomized experiment usually isn't available and correlation alone can't support a causal claim.

How would you choose between a simpler statistical model and machine learning approach for a forecasting problem?

For data with clear seasonal and trend patterns and limited history, a statistical model like ARIMA or exponential smoothing is often more reliable than machine learning, which typically needs more data to outperform simpler methods. I'd benchmark both against a holdout period rather than assuming the more complex option is automatically better.

How do you handle a situation where your model's predictions need to be explained to justify individual decisions?

I'd use an inherently interpretable model where possible, or apply a post-hoc explanation method like SHAP values on top of a more complex model if the added performance is worth the extra explanation layer. The explanation needs to be accurate to how the model actually makes decisions, not just a plausible-sounding story layered on afterward.