Every strong career story starts messier than it looks from the outside: uncertain options, imperfect experience, and a few decisions made before everything felt ready. Statistics for Data Science Interviews: Key Concepts & Questions helps you make those decisions with more confidence. Most candidates can define p-value, but few can explain when a p-value of 0.05 is misleading. The interviewers at top Indian analytics firms are looking for practitioners, and they want to know when you would use a t-test versus a z-test and what assumptions you are making.
Descriptive Statistics
Measures of Central Tendency
The mean is the sum of values divided by the count, but it is sensitive to outliers. The median is the middle value when sorted and is robust to outliers. The mode is the most frequent value. An interviewer might ask: "You are analyzing customer spend data. The mean is Rs 1,200 but the median is Rs 450. What does this tell you?" The answer is that the distribution is right-skewed, with a few high-spending customers pulling the mean up. For practical purposes like setting a typical customer budget, the median is more representative, while for revenue forecasting, the mean matters.
Measures of Dispersion
The range, which is max minus min, is simple but sensitive to outliers. Variance is the average of squared deviations from the mean. Standard deviation is the square root of variance and is in the same units as the data. The interquartile range, or IQR, is Q3 minus Q1 and is robust to outliers. In interviews, you might be told that the standard deviation of daily orders for Swiggy in Bangalore is 15,000 orders and asked what that tells you and what you would do about it. High variance might indicate weekends versus weekdays, festivals, or operational issues, so you would segment by day type, identify root causes, and potentially adjust staffing and inventory planning.
Skewness and Kurtosis
Skewness measures the asymmetry of a distribution. Positive skew means a tail on the right, as seen in income distribution, while negative skew means a tail on the left, as seen in age at death. Kurtosis measures tail heaviness. High kurtosis indicates more outliers, such as in financial returns, while low kurtosis indicates fewer outliers, such as in height distribution.
Probability Distributions
Discrete Distributions
The binomial distribution represents the number of successes in n independent trials. It is used when asking questions like "What is the probability that 8 out of 10 users click on an ad given a 60% click-through rate?" and follows the formula P(X=k) = C(n,k) x p^k x (1-p)^(n-k). The Poisson distribution represents the number of events in a fixed interval, used for questions like "How many customers arrive at a store per hour?" Its key property is that the mean equals the variance. A common Indian example is asking "What is the probability of 3 or more orders being placed at a Zomato restaurant in a given minute if the average rate is 1.5 orders per minute?" An interviewer might ask when you would use binomial versus Poisson. The answer is that binomial is used when you have a fixed number of trials with success or failure, while Poisson is used when you are counting events in a continuous interval with no fixed upper bound.
Continuous Distributions
The normal distribution is the bell curve found in natural phenomena. Its properties include symmetry and the 68-95-99.7 rule. The standard normal is calculated as Z = (X - m) / s. An interviewer might ask: "The average height of Indian men is 167 cm with standard deviation 7 cm. What percentage of men are taller than 174 cm?" The answer is Z = (174-167)/7 = 1.0, and using the 68-95-99.7 rule, approximately 16% of Indian men are taller than 174 cm. The exponential distribution describes the time between events in a Poisson process and is used for questions like "What is the probability that the next Uber arrives in less than 5 minutes if the average wait time is 8 minutes?" The uniform distribution means all outcomes are equally likely, used for questions like "If the delivery time for Swiggy is uniformly distributed between 20 and 40 minutes, what is the probability the delivery takes more than 35 minutes?" The answer is 5/20, or 25%.
Hypothesis Testing
Null and Alternative Hypothesis
The null hypothesis, denoted H0, states there is no effect or no difference, such as "The new checkout flow does not change conversion rate." The alternative hypothesis, H1, states there is an effect, such as "The new checkout flow changes conversion rate."
Type I and Type II Errors
A Type I error, or false positive, occurs when you reject the null hypothesis when it is actually true. A Type II error, or false negative, occurs when you fail to reject the null hypothesis when it is actually false. An interviewer might ask: "You are testing whether a new ML model reduces fraud. You set a = 0.01. The p-value comes out to 0.03. What do you conclude?" The answer is that since the p-value of 0.03 is greater than alpha of 0.01, you fail to reject the null and cannot conclude the new model reduces fraud, though this is borderline and you might recommend a larger sample size for a more definitive test.
P-value
The p-value is the probability of observing data as extreme as what you saw, assuming the null hypothesis is true. A common misconception is that a p-value of 0.03 means there is a 3% chance the null is true, which is wrong. The p-value assumes the null is true and tells you how surprising your data would be if there were no effect. Many interviewers love this question: "What is wrong with this interpretation: p = 0.03, so there is a 97% chance our A/B test result is real?"
Choosing the Right Test
When comparing the mean of one group to a known value, use a one-sample t-test. For comparing means of two independent groups, use a two-sample t-test. For comparing means before and after, use a paired t-test. For comparing means of three or more groups, use ANOVA. For comparing proportions, use a z-test for proportions. For testing association between categorical variables, use a chi-square test. For measuring correlation between continuous variables, use Pearson correlation. An interviewer might ask: "We have 3 pricing strategies and want to know if revenue differs between them. Revenue data is not normally distributed. What do you do?" The answer is that if assumptions for ANOVA, such as normality and homogeneity of variance, are violated, use the non-parametric Kruskal-Wallis test instead, which tests whether samples come from the same distribution without assuming normality.
Statistical Significance vs Practical Significance
An interviewer might ask: "You run an A/B test with 10 million users. The conversion rate goes from 5.00% to 5.01%. The p-value is 0.001. Should you launch the feature?" The answer is that the result is statistically significant but may not be practically significant. A 0.01% improvement means one extra conversion per 10,000 users, so you need to weigh the engineering effort, maintenance cost, and UX impact against this tiny gain. For a low-effort change, launch, but for a major redesign, it is probably not worth it.
Regression Analysis
Linear Regression
The linear regression equation is Y = b0 + b1X1 + b2X2 + ... + bnXn + e. The key assumptions, known by the acronym LINE, are linearity, meaning the relationship between X and Y is linear; independence, meaning observations are independent; normality, meaning residuals are normally distributed; and equal variance or homoscedasticity, meaning residuals have constant variance. An interviewer might ask: "R-squared is 0.85. What does this mean?" The answer is that 85% of the variance in the dependent variable is explained by the independent variables, though R-squared always increases with more variables, so you should always check adjusted R-squared and consider whether there is overfitting. A common Indian example is predicting customer lifetime value based on order frequency, average order value, tenure, and city tier.
Logistic Regression
Logistic regression is used for binary outcomes such as will convert, will churn, or is fraud. Its equation is log(p/(1-p)) = b0 + b1X1 + ... + bnXn. A one-unit increase in X1 multiplies the odds by e^(b1). An interviewer might ask: "Explain odds vs probability." If the probability of churn is 0.2, the odds are 0.2/0.8 = 0.25, meaning for every one churning customer, four stay. Logistic regression models log-odds linearly, which allows the output to range from negative infinity to positive infinity while keeping the probability between 0 and 1.
Overfitting and Regularization
Bias refers to error from wrong assumptions, while variance refers to error from sensitivity to training data. The bias-variance tradeoff means high bias leads to underfitting with a model that is too simple, while high variance leads to overfitting with a model that is too complex. Regularization techniques include Lasso or L1, which shrinks some coefficients to zero for feature selection; Ridge or L2, which shrinks coefficients but keeps all features; and Elastic Net, which combines L1 and L2.
Bayesian Thinking
Bayes' Theorem states that P(A|B) = P(B|A) x P(A) / P(B). A classic interview question goes: "A disease affects 1% of the population. A test is 99% accurate. You test positive. What is the probability you have the disease?" The prevalence P(Disease) is 0.01, P(No Disease) is 0.99, the sensitivity P(Positive | Disease) is 0.99, and the false positive rate P(Positive | No Disease) is 0.01. Applying Bayes, P(Disease | Positive) = (0.99 x 0.01) / (0.99 x 0.01 + 0.01 x 0.99) = 0.0099 / 0.0198 = 0.5. So you only have a 50% chance of actually having the disease despite testing positive, a counterintuitive result that is classic Bayes and a common interview question.
Sampling Techniques
Random sampling methods include simple random sampling, stratified sampling where you divide the population into strata and sample from each, cluster sampling where you divide into clusters and sample clusters, and systematic sampling where you take every nth element. A typical Indian example asks: "You need to survey 10,000 Indian households about their grocery spending. How do you ensure the sample is representative?" The answer is to use stratified sampling by geography, urban versus rural, and income quintile, since simple random sampling would likely underrepresent rural areas and lower-income groups. Sampling bias includes selection bias from a non-random sample, survivorship bias from only observing surviving entities, confirmation bias from seeking evidence that confirms beliefs, and sampling bias in India where online surveys miss rural and offline populations.
A/B Testing
Key concepts in A/B testing include the control versus treatment groups, randomization where users are randomly assigned to groups, sample size calculation to determine how many users are needed, the minimum detectable effect or MDE which is the smallest effect you want to detect, and the test duration which must account for day-of-week effects. A common interview question asks: "Your A/B test shows a 5% lift in conversion with p = 0.04. Should you launch?" Before launching, check whether the sample size was adequate, whether the test was run long enough covering at least one full business cycle, whether there were any novelty effects, whether the lift is consistent across segments, and whether there are any negative impacts on other metrics like revenue or customer satisfaction.
Peeking Problem
An interviewer might ask: "Why should you not check your A/B test results every day and stop when p < 0.05?" Repeated peeking inflates the Type I error rate. If you check daily, the probability of seeing a false positive at some point far exceeds 5%. Use sequential testing methods or set a fixed duration upfront.
Common Interview Questions
Q1: "You have a dataset with 1 million customers. The average order value is Rs 500 with a standard deviation of Rs 200. What is the 95% confidence interval for the mean?"
The confidence interval is x-bar plus or minus 1.96 times sigma divided by the square root of n, which equals 500 plus or minus 1.96 times 200 divided by 1000, giving 500 plus or minus 0.392. The 95% confidence interval is Rs 499.61 to Rs 500.39.
Q2: "Explain p-value to a business stakeholder who does not know statistics."
"Imagine you are flipping a coin and it lands heads 8 times out of 10. Is the coin biased? The p-value answers: 'If the coin were fair, how likely would we be to see such an extreme result?' A low p-value says 'This is surprising if there is no effect, which suggests there might be a real effect.' But it does not tell you how big the effect is, which is what the confidence interval is for."
Q3: "Correlation vs causation, give an Indian example."
"Ice cream sales and drowning incidents both increase in summer. They are correlated, but eating ice cream does not cause drowning. The confounding variable is hot weather because more people swim. Similarly, in India, having more doctors correlates with higher disease rates, not because doctors cause disease, but because cities with more doctors have better diagnosis."
Q4: "What is the Central Limit Theorem and why does it matter?"
"The CLT says that regardless of the population distribution, the distribution of sample means approaches a normal distribution as sample size increases, typically with n greater than 30. This is why we can use normal-distribution-based tests like t-tests and z-tests even when the underlying data is not normal."
Q5: "How would you detect fraud in credit card transactions using statistics?"
Use anomaly detection based on statistical deviation from normal patterns. For each transaction, calculate z-scores for amount, frequency, location, time, and merchant category. Flag transactions where the combined deviation exceeds a threshold. Use logistic regression with historical fraud data. Apply real-time Poisson models for frequency anomalies, so if a card that averages 2 transactions per day suddenly has 15 in an hour, that is statistically suspicious.
Practical Tips for Interviews
Knowing when to use which test is more important than memorizing formulas. Explain assumptions before applying any statistical method, since every method has them. Connect your answers to business impact, for example by saying a p-value of 0.03 means you would expect this result 3% of the time if there were no real effect, which is low enough to recommend launching. Be honest about limitations, such as noting that with a small sample size you might have insufficient power to detect the effect. Finally, know your Indian benchmarks including population, GDP growth, smartphone penetration, and internet user growth.
Conclusion
Statistics is the language of data science. In interviews, focus on conceptual understanding over formula recall, practical application of knowing when to use what, the assumptions and limitations of each method, interpretation of what the numbers mean for the business, and communication of technical concepts to non-technical stakeholders.
Need more data science interview prep? Check out our data analyst interview questions guide and SQL interview prep.
Your Move
- Record three answers using the STAR method: situation, task, action, result.
- Replay each answer and check whether it is specific, concise, and tied to the role you want.
- Rewrite the weakest answer, then practice it once more without reading notes.
