Data Analytics Probability Interview Questions and Answers

Probability is where interviewers check whether you can reason about uncertainty without hand-waving. For data analyst roles the bar is not advanced measure theory — it is conditional probability, Bayes' theorem, expected value and a handful of distributions, plus the discipline to define a sample space before computing anything. This page covers the questions that recur across analyst and BI interviews, worked through the way an interviewer wants to see you think.

How to answer probability questions

Say the sample space out loud, write the given quantities, and move one step at a time. Most wrong answers come from skipping straight to a formula. The candidates who score highest narrate their reasoning, which also lets the interviewer give partial credit and nudge them when needed.

Q1. What are the basic rules of probability?

A probability is a number between 0 and 1. For mutually exclusive events, P(A or B) = P(A) + P(B); in general P(A or B) = P(A) + P(B) − P(A and B). For independent events, P(A and B) = P(A) × P(B). The complement rule gives P(not A) = 1 − P(A).

The complement rule is the quiet workhorse. "What is the probability of at least one success in n trials?" is almost always easier as 1 minus the probability of zero successes. Reaching for the complement instead of summing many cases signals fluency.

Interview note: Trap: "P(A or B) = P(A) + P(B), always?" Only when A and B are mutually exclusive. Otherwise you double-count the overlap and must subtract P(A and B).

Q2. What is conditional probability?

Conditional probability P(A given B) is the probability of A occurring given that B has already occurred, defined as P(A and B) / P(B). It updates the sample space to only the outcomes where B is true.

The intuition is a shrinking universe: once you know B happened, you throw away every outcome where it did not, and re-normalize. A concrete framing helps — "given that a customer is on a premium plan, what is the chance they churn?" restricts attention to premium customers only.

Interview note: Follow-up: "does P(A given B) equal P(B given A)?" No, and confusing them is the core Bayesian error tested in Q4.

Q3. Define independence, and how is it different from mutual exclusivity?

Two events are independent if one occurring does not change the probability of the other: P(A given B) = P(A). They are mutually exclusive if they cannot happen together: P(A and B) = 0. These are different, and in fact two events with nonzero probability cannot be both.

This catches many candidates. If A and B are mutually exclusive, then knowing B happened tells you A definitely did not — which is the strongest possible dependence, not independence. Being able to state that they are near-opposites is a clean discriminator.

Interview note: Trap: "mutually exclusive events are independent." False — mutual exclusivity implies dependence when both have nonzero probability.

Q4. Explain Bayes' theorem with the medical-test example.

Bayes' theorem is P(A given B) = P(B given A) × P(A) / P(B). It lets you flip a conditional probability using the base rate. In a disease test, even a 99% accurate test can yield mostly false positives when the disease is rare.

Suppose a disease affects 1 in 1,000 people, and the test is 99% accurate (both sensitivity and specificity). Of 100,000 people, 100 have the disease and about 99 test positive; of the 99,900 healthy, about 1% — roughly 999 — falsely test positive. So a positive result means disease with probability about 99 / (99 + 999) ≈ 9%, not 99%.

p_disease = 0.001
sensitivity = 0.99          # P(positive | disease)
false_positive = 0.01       # P(positive | healthy)

p_pos = sensitivity * p_disease + false_positive * (1 - p_disease)
p_disease_given_pos = sensitivity * p_disease / p_pos
print(round(p_disease_given_pos, 4))   # ~0.0902

The lesson interviewers want stated: with a low base rate, a positive test is far weaker evidence than the accuracy figure suggests. Ignoring the base rate is the "base-rate fallacy".

Interview note: Follow-up: "how do you make the positive result more trustworthy?" Retest, or apply the test to a higher-prevalence subgroup, raising the prior.

Q5. What is expected value, and how do you use it?

Expected value is the probability-weighted average of all possible outcomes: E[X] = sum of (value × probability). It represents the long-run average if the experiment were repeated many times.

Expected value drives business decisions constantly. A promotion that yields ₹500 profit with probability 0.2 and −₹50 with probability 0.8 has expected value 0.2 × 500 + 0.8 × (−50) = ₹60 per customer — positive, so it is worth running at scale even though most individual cases lose money.

Interview note: Trap: "is expected value a value you actually expect to see?" Not necessarily — the expected value of a fair die is 3.5, which is never rolled. It is a long-run average.

Q6. When do you use the binomial, Poisson and normal distributions?

Use the binomial for the count of successes in a fixed number of independent yes/no trials. Use the Poisson for the count of rare events over an interval of time or space with a known average rate. Use the normal for continuous data that clusters symmetrically around a mean.

The recognition skill is what gets tested. "Number of clicks out of 500 ad impressions" is binomial. "Number of support tickets per hour" is Poisson. "Distribution of adult heights" is normal. The Poisson also approximates the binomial when n is large and the success probability is small.

Interview note: Follow-up: "what is the mean of a Poisson?" It equals its rate parameter lambda — and so does its variance, a distinctive property.

Q7. What is the empirical (68-95-99.7) rule?

For a normal distribution, about 68% of values fall within one standard deviation of the mean, 95% within two, and 99.7% within three. It gives a fast mental estimate of how unusual a value is.

Analysts use this for quick outlier screening: a data point more than three standard deviations from the mean is a roughly 0.3% event and worth investigating. Stating this rule confidently signals you can eyeball a distribution without running code.

Interview note: Trap: "does the rule hold for skewed data?" No — it assumes normality. For skewed data use percentiles instead.

Q8. A classic puzzle: two children, one is a boy. What is the probability both are boys?

Given the equally likely combinations BB, BG, GB, GG, the condition "at least one boy" eliminates GG, leaving BB, BG, GB. Only one of the three has two boys, so the probability is 1/3, not 1/2.

This puzzle rewards the exact discipline the section tests: enumerate the sample space, apply the condition to shrink it, then count. The intuitive 1/2 answer ignores that "at least one boy" is different information from "the older child is a boy" (which would give 1/2).

Interview note: Follow-up: "why does specifying the older child change it?" It removes an ordering ambiguity, cutting the space to BB and BG — hence 1/2.

Q9. What is the difference between probability and likelihood?

Probability describes how likely an outcome is given fixed parameters; likelihood describes how well different parameter values explain fixed observed data. Probability runs from model to data, likelihood from data back to model.

Analysts meet likelihood when a model is fit by maximizing it. Being able to say "we hold the data fixed and search for the parameters that make it most probable" shows you understand what "maximum likelihood estimation" actually does under the hood.

Interview note: Trap: "do likelihoods sum to 1?" No — unlike a probability distribution over outcomes, a likelihood function over parameters need not integrate to 1.

What interviewers really test

Every question here rewards the same habit: define the sample space, write what is given, and reason in visible steps. The Bayes and base-rate problems are the most reused because they punish memorization and reward thinking. Pair this page with the statistics question set, which builds directly on expected value and distributions, and with the machine learning basics questions, where probability underpins classification. A structured Data Analytics path plus a timed mock interview will turn these from puzzles into reflexes.