Data Analyst Interview Questions & Answers (2026)

Explain how you would handle missing values in a dataset.

First, I would assess the pattern of missingness—whether it is MCAR, MAR, or MNAR—using visualizations and statistical tests. For MCAR, simple imputation like mean or median works; for MAR, I’d consider model‑based imputation such as K‑Nearest Neighbors or regression. If the missingness is systematic (MNAR), I might create an indicator variable or consult domain experts before deciding to drop or model the missingness. The key is to justify the method based on data distribution and business impact.

What is the difference between INNER JOIN and LEFT JOIN in SQL?

INNER JOIN returns rows only when there is a matching key in both tables, effectively intersecting the datasets. LEFT JOIN returns all rows from the left table and matches from the right; when no match exists, NULLs fill the right‑side columns. Interviewers look for awareness of result set size, handling of NULLs, and performance implications—INNER JOIN can be faster because it eliminates unmatched rows early, while LEFT JOIN may require additional processing.

How do you determine if a variable is normally distributed?

I start with visual checks—histograms, Q‑Q plots, and boxplots—to spot deviations from symmetry. Then I apply statistical tests like Shapiro‑Wilk or Kolmogorov‑Smirnov, noting sample size sensitivity. If the p‑value exceeds the chosen alpha (commonly 0.05), I fail to reject normality. I also examine skewness and kurtosis values; absolute skewness <0.5 and kurtosis close to 3 support a normal assumption. Explaining both visual and test‑based evidence shows thoroughness.

Describe a time you turned a complex dataset into a clear business insight.

At my last role, I received raw clickstream logs with millions of rows and dozens of columns. I aggregated sessions by user, calculated churn probability using logistic regression, and visualized the top three drivers in a Tableau dashboard. The insight revealed that users who accessed the mobile app more than three times per week were 20% less likely to churn, prompting the product team to prioritize mobile feature enhancements. I highlighted the data pipeline, model choice, and actionable recommendation.

What is a p‑value and how do you interpret it?

A p‑value quantifies the probability of observing data as extreme as the sample, assuming the null hypothesis is true. A small p‑value (typically <0.05) suggests the observed effect is unlikely under the null, leading to rejection of the null hypothesis. However, it does not measure effect size or practical significance. Interviewers expect you to discuss Type I/II errors, the importance of context, and why a statistically significant result may still be irrelevant to business decisions.

How would you explain the concept of confidence intervals to a non‑technical stakeholder?

I would say a confidence interval provides a range where we expect the true metric—like average revenue per user—to fall, based on our sample data, with a certain level of confidence (e.g., 95%). It’s like saying, ‘We are 95% sure the real average lies between $12 and $15.’ This conveys uncertainty without overwhelming technical detail, and I’d illustrate with a simple graphic to reinforce the idea of variability around the estimate.

Write a SQL query to find the top 5 products by total sales in the last quarter.

SELECT product_id, SUM(sales_amount) AS total_sales FROM sales WHERE sale_date BETWEEN DATEADD(quarter, -1, CURRENT_DATE) AND CURRENT_DATE GROUP BY product_id ORDER BY total_sales DESC LIMIT 5; This query filters sales to the last quarter, aggregates revenue per product, orders descending, and limits the result to the five highest sellers. Interviewers look for correct date functions, aggregation, and efficient limiting.

SELECT product_id, SUM(sales_amount) AS total_sales FROM sales WHERE sale_date BETWEEN DATEADD(quarter, -1, CURRENT_DATE) AND CURRENT_DATE GROUP BY product_id ORDER BY total_sales DESC LIMIT 5;

What is multicollinearity and how does it affect regression models?

Multicollinearity occurs when two or more independent variables are highly correlated, inflating variance of coefficient estimates. This leads to unstable coefficients, wide confidence intervals, and difficulty interpreting individual predictor impact. I detect it using Variance Inflation Factor (VIF) thresholds (>5 or >10) and correlation matrices. To mitigate, I may drop redundant features, combine them via PCA, or use regularization (Ridge) that shrinks coefficients, thereby improving model robustness.

Explain the difference between a left‑skewed and right‑skewed distribution and how you would transform each.

A left‑skewed distribution has a longer tail on the left side (negative skew), often requiring a power transformation like square or cube to pull the tail rightward. A right‑skewed distribution (positive skew) has a longer right tail; applying a log, square‑root, or Box‑Cox transformation compresses high values, making the distribution more symmetric. I would choose the transformation based on visual inspection and test normality after transformation to ensure suitability for parametric models.

How do you assess the performance of a classification model beyond accuracy?

I examine precision, recall, and F1‑score to balance false positives and false negatives, especially when classes are imbalanced. The ROC‑AUC provides a threshold‑independent measure of separability. I also look at confusion matrices to understand error types, and use calibration curves to verify probability estimates. For business impact, I might compute cost‑sensitive metrics, such as expected monetary loss, aligning model evaluation with the organization’s risk tolerance.

Describe a time you used A/B testing to influence a product decision.

We wanted to test a new checkout flow. I defined the primary metric (conversion rate), calculated required sample size using a power analysis (80% power, 5% significance), and randomly assigned users to control and variant groups. After 2 weeks, the variant showed a 3.2% lift with a p‑value of 0.02. I presented the statistical findings, confidence intervals, and potential revenue impact, leading the product team to roll out the new flow globally.

What is the difference between a data warehouse and a data lake?

A data warehouse stores structured, curated data optimized for fast SQL queries and reporting, often using schemas like star or snowflake. A data lake holds raw, unstructured, or semi‑structured data in its native format, enabling flexible ingestion and future processing. Warehouses prioritize consistency and performance; lakes prioritize scalability and cost‑effectiveness. Understanding both helps candidates discuss architecture choices and integration strategies for analytics pipelines.

What is the purpose of dimensionality reduction and when would you use PCA?

Dimensionality reduction simplifies data, reduces noise, and speeds up computation while preserving variance. Principal Component Analysis (PCA) projects data onto orthogonal axes that capture maximal variance. I use PCA when features are highly correlated, the dataset is large, and I need to visualize high‑dimensional data or improve model training time. I always check explained variance ratio to retain enough components (e.g., 90% variance) and verify that the transformed features still support the business objective.

How would you design a data pipeline for near‑real‑time analytics?

I’d start with a streaming ingestion layer (Kafka or Kinesis) to capture events. A processing engine like Flink or Spark Structured Streaming would clean, enrich, and aggregate data in micro‑batches. Results would be written to a low‑latency store (Redis or ClickHouse) for dashboarding, while raw data lands in a data lake (S3) for archival and batch analytics. I’d implement schema evolution, monitoring, and idempotent writes to ensure reliability and scalability.

Explain the bias‑variance trade‑off in machine learning models.

Bias reflects error from erroneous assumptions (under‑fitting), while variance reflects sensitivity to training data fluctuations (over‑fitting). A high‑bias model may miss patterns, leading to systematic errors; a high‑variance model captures noise, hurting generalization. The trade‑off is managed by selecting model complexity, using regularization, or employing ensemble methods. I illustrate this by plotting training vs. validation error as model complexity changes, showing the sweet spot where total error is minimized.

What is a time‑series decomposition and why is it useful?

Time‑series decomposition splits a series into trend, seasonal, and residual components, typically using additive or multiplicative models. Trend captures long‑term direction, seasonality reflects periodic patterns, and residuals contain irregularities and noise. Decomposition helps diagnose underlying structures, improve forecasting accuracy (by modeling components separately), and communicate insights clearly to stakeholders. I often use STL (Seasonal‑Trend decomposition using Loess) for flexible, robust decomposition.

How do you handle outliers when building a predictive model?

First, I identify outliers using statistical methods (z‑score >3, IQR) and domain knowledge. I then decide whether they represent data errors or genuine extreme cases. For errors, I may cap or impute them; for legitimate extremes, I might keep them but use robust algorithms (e.g., tree‑based models) that are less sensitive. I always document the impact on model performance and explain the rationale to stakeholders, showing that the decision aligns with business risk tolerance.

Write a Python snippet to calculate the monthly churn rate from a user activity dataframe.

Assuming a DataFrame df with columns user_id, activity_date, and a churn flag, I would first create a month column, then group by month to compute churn rate as churned users divided by total active users. This concise code demonstrates pandas proficiency and clear aggregation logic, which interviewers expect for data‑wrangling tasks.

df['month'] = df['activity_date'].dt.to_period('M')
monthly = df.groupby('month').agg(total_users=('user_id', 'nunique'), churned=('churn_flag', 'sum'))
monthly['churn_rate'] = monthly['churned'] / monthly['total_users']

How would you evaluate the ROI of an analytics project?

I’d quantify benefits (e.g., revenue uplift, cost savings) and costs (personnel, tooling, data acquisition). Using a time‑horizon, I’d calculate net present value (NPV) and internal rate of return (IRR) to capture cash‑flow timing. I also include intangible benefits like improved decision speed. Presenting a clear business case with sensitivity analysis shows that I can translate analytical work into measurable financial impact.