Amazon·Data Scientist·Technical Phone Screen·Intermediate

Intermediate

Jul 2026

Summary

Amazon technical screen, probably for a data science or ML-adjacent role. Just the one question but it's the kind that sounds straightforward until you're actually in it trying to remember all five assumptions on the spot.

Questions Asked (1)

What are the assumptions underlying linear regression, and what are the consequences if those assumptions are violated?

Technical Trade-offsProduct Analytics & Metrics

Author's notes

I could rattle off linearity and independence fine, then kind of fumbled through homoscedasticity.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by listing the core assumptions of linear regression (linearity, independence, homoscedasticity, normality of errors, no multicollinearity, and no endogeneity). Then, for each assumption, explain the specific consequences of violation on coefficient estimates, standard errors, and predictions, and briefly mention detection and remedies. Finally, tie it back to practical implications for model reliability and decision-making.

Pro tip: Emphasize that some violations (e.g., non-normality) matter mainly for inference in small samples, while others (e.g., endogeneity) bias coefficients even in large samples—showing you understand the nuance between statistical and practical significance.

1. List the assumptions

Clearly state the key assumptions: linearity, independence of errors, homoscedasticity, normality of errors, no multicollinearity, and no endogeneity (or exogeneity of regressors).

2. Explain consequences of violation

For each assumption, describe the impact: e.g., non-linearity leads to biased predictions; heteroscedasticity or autocorrelation inflates standard errors; multicollinearity inflates coefficient variance; endogeneity biases coefficients.

3. Discuss detection and remedies

Mention common diagnostic tools (residual plots, VIF, Durbin-Watson, Breusch-Pagan) and potential fixes (transformations, robust standard errors, regularization, instrumental variables).

4. Relate to business impact

Connect the violations to real-world consequences: unreliable feature importance, poor generalization, incorrect confidence intervals, and misguided business decisions.

Key Points to Mention

Linearity: relationship between predictors and outcome is linear; violation leads to biased predictions.
Independence: errors are uncorrelated; violation (autocorrelation) underestimates standard errors, inflating Type I error.
Homoscedasticity: constant error variance; violation (heteroscedasticity) makes OLS inefficient and standard errors biased.
Normality of errors: needed for valid hypothesis tests in small samples; with large samples, CLT mitigates impact.
No multicollinearity: predictors are not highly correlated; violation inflates coefficient variance, making estimates unstable.
No endogeneity: predictors are uncorrelated with error term; violation biases coefficients, even asymptotically.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.