Statistics for Data Science: What You Actually Need to Know

The statistics a working data scientist actually uses — descriptive measures, probability and Bayes, distributions, sampling, hypothesis testing, confidence intervals, correlation and regression — with worked examples, what to skip, and how interviewers test it.

  SAP Training   Oct 14, 2026   89  Add to Reading List
Quick Answer: You need eight areas — descriptive statistics, probability and Bayes' theorem, common distributions, sampling and the central limit theorem, hypothesis testing, confidence intervals, correlation versus causation, and regression. You do not need a statistics degree or proofs. What employers test is whether you can choose the right method, interpret the result correctly and spot when a number is misleading.

Statistics is the part of data science beginners most often try to skip and most often regret skipping. Python can be learned from tutorials and machine learning libraries hide their maths, so it is tempting to treat statistics as optional. Then the interview asks what a p-value means, or why a model's accuracy is meaningless on imbalanced data, and the gap shows immediately.

The good news is that the statistics a working data scientist uses is a bounded set. This guide covers exactly that set, in the order it builds, with one worked example per idea and a clear note on what you can safely leave for later.

Why Statistics Still Matters When Libraries Do the Maths

A library will happily fit a model to any data you give it and print a confident number. It will not tell you that your sample was biased, that your test was underpowered, that two variables move together only because a third drives both, or that a 2% improvement is noise. Those judgements are the job. Statistics is the toolkit for making them, and it is why companies pay data scientists more than people who can only run code.

1. Descriptive Statistics

Start by summarising data: mean, median and mode for the centre; range, variance and standard deviation for spread; percentiles and the interquartile range for position; and skewness for shape.

The most useful lesson here is when the mean misleads. Salaries, order values, session lengths and house prices are right-skewed — a few very large values pull the mean upward. If ten employees earn ₹5 lakh and one earns ₹1 crore, the mean is over ₹13 lakh, which describes nobody. The median, ₹5 lakh, does. Reporting the median with the interquartile range for skewed data is a habit interviewers notice.

2. Probability and Bayes' Theorem

You need the basics: probability of events, independence, and above all conditional probability — the probability of A given that B has happened. Bayes' theorem is how you reverse a conditional probability, and it explains results that surprise almost everyone.

Worked example. A fraud model flags 95% of fraudulent transactions and wrongly flags 5% of genuine ones. Fraud is 1% of all transactions. If a transaction is flagged, how likely is it to be fraud?

Out of 10,000 transactions, 100 are fraud and 95 of those are flagged. Of the 9,900 genuine ones, 5% — 495 — are also flagged. So 95 of 590 flagged transactions are fraud: about 16%, not 95%.

This is the base-rate effect, and it is why accuracy alone is a poor measure for rare events, and why precision and recall exist. It is also one of the most common interview puzzles in disguise.

3. Distributions You Will Actually Meet

The normal distribution describes many measurement errors and averages; roughly 68% of values lie within one standard deviation of the mean, 95% within two and 99.7% within three. The binomial counts successes in a fixed number of yes-or-no trials — conversions out of visitors. The Poisson counts events in a fixed interval — support tickets per hour. And many business quantities, such as revenue per customer, are long-tailed or roughly log-normal, which is why taking logarithms before modelling them is common.

You do not need the formulas memorised. You need to recognise which shape your data has, because the choice of test, model and summary statistic depends on it.

4. Sampling and the Central Limit Theorem

Almost all analysis works on a sample and draws conclusions about a population, so how the sample was collected matters more than any later calculation. A survey answered only by satisfied customers, or a model trained only on users who stayed, is biased before a single number is computed.

The central limit theorem says that the average of a reasonably large random sample is approximately normally distributed, even when the underlying data is not. That is why so many tests work on non-normal data. The spread of that sample average is the standard error, which shrinks with the square root of the sample size — quadrupling your sample halves your uncertainty, a useful fact when someone asks how much more data you need.

5. Hypothesis Testing

This is the most tested area in interviews and the most misunderstood. You state a null hypothesis — usually "no difference" or "no effect" — and ask how surprising your data would be if it were true.

The p-value is the probability of seeing a result at least as extreme as yours if the null hypothesis were true. It is not the probability that the null hypothesis is true, and it is not the probability that your result is due to chance. Getting this definition exactly right is worth practising aloud.

Know the two error types: a Type I error is a false positive, declaring an effect that is not there; a Type II error is a false negative, missing one that is. The significance level, often 0.05, caps the first. Power — the chance of detecting a real effect — guards against the second and depends on sample size and effect size.

Know which test fits which question: a t-test compares means of two groups; a chi-square test checks association between categories; ANOVA compares means across more than two groups; and non-parametric alternatives such as Mann-Whitney help when assumptions fail.

Worked example — an A/B test. A new checkout page converts 5.4% of 10,000 visitors against 5.0% for the old page on another 10,000. Is it better? A two-proportion test gives a p-value of roughly 0.2. That is not evidence of a real improvement; a difference this size appears by chance fairly often at this sample size. The right conclusion is "inconclusive — we need more traffic or a larger effect," not "the new page lost."

Two traps worth naming in interviews: stopping a test the moment it looks significant, which inflates false positives, and running many comparisons without correction, which guarantees some will look significant by chance.

6. Confidence Intervals

A confidence interval gives a range of plausible values for a quantity instead of a single number. "Conversion rose by 0.4 points, with a 95% interval from −0.2 to 1.0" tells a manager far more than a p-value: the effect might be negative, might be meaningfully positive, and the data cannot yet tell which. Prefer reporting intervals; decision-makers understand them better, and they show the size of an effect, not just whether it exists.

7. Correlation, Causation and Confounders

Correlation measures how strongly two variables move together; Pearson's coefficient captures linear relationships, and Spearman's captures monotonic ones. Neither implies that one causes the other. Ice cream sales and drowning incidents rise together because summer drives both — a confounder.

In business data confounders are everywhere. Customers who use a feature may retain better because engaged customers use more features, not because the feature causes retention. Recognising this — and knowing that a randomised experiment is the cleanest way to establish cause — is what separates a data scientist from someone who reports correlations. Simpson's paradox, where a trend reverses when data is split into groups, is a favourite interview example of the same idea.

8. Regression: Where Statistics Meets Machine Learning

Linear regression models a numeric outcome as a weighted sum of inputs. You should be able to interpret a coefficient — the expected change in the outcome for a one-unit change in that input, holding the others fixed — read R-squared as the share of variation explained, and check residuals for patterns that signal a poor fit. Know that multicollinearity, strongly correlated inputs, makes individual coefficients unstable even when predictions are fine.

Logistic regression models a yes-or-no outcome as a probability, and its coefficients relate to odds. It remains one of the most used models in industry because it is fast, interpretable and hard to overfit badly.

From here statistics flows directly into machine learning: overfitting is a variance problem, train-test splits are a sampling idea, and cross-validation is resampling. Our Machine Learning course builds on exactly these foundations.

How Interviewers Test This

Statistics questions in Indian data science interviews rarely ask for formulas. They come in three forms. The first is definitions under pressure: explain a p-value, a confidence interval or Type II error in one or two sentences, without the common mistakes. The second is probability puzzles, usually a disguised version of the Bayes example above or a coin, dice or card problem. The third, and most important, is a business scenario: "Sales rose 8% after we launched a campaign — did the campaign work?" or "Design a test for a new pricing page."

For scenarios, a strong answer asks about the baseline and seasonality, names possible confounders, proposes a controlled comparison, states the metric and sample size needed, and says what result would change the decision. Interviewers are checking whether you will protect the business from false conclusions, not whether you can recite a test statistic.

What to Learn Deeply, What to Skim, What to Skip

LevelTopicsWhy
Learn deeplyDescriptive statistics, conditional probability, sampling and standard error, hypothesis testing, confidence intervals, correlation, linear and logistic regressionUsed every week and tested in every interview
Working knowledgeCommon distributions, Bayes' theorem, chi-square and ANOVA, power and sample size, non-parametric tests, bootstrappingNeeded regularly; know when to reach for them
Skip for nowMeasure theory, formal proofs, advanced Bayesian modelling, stochastic processes, deriving estimators by handValuable for research roles, not for getting hired

How to Learn It Without Drowning in Formulas

Learn each idea with code alongside the concept. In Python, scipy.stats and statsmodels run every test above in a line or two, which frees you to focus on choosing and interpreting rather than calculating.

Use simulation to build intuition. Draw a thousand samples from a skewed distribution and plot their means, and the central limit theorem stops being abstract. Run an A/A test — two identical groups — a hundred times and count how often p falls below 0.05, and you will understand false positives permanently.

Then apply it to a real dataset with a real question: did a price change affect sales, does delivery time affect ratings. Write up the conclusion in plain language with an interval, a limitation and a recommendation. That write-up is a portfolio piece and an interview answer at the same time. If you want a guided route, our Data Science course teaches statistics through applied projects rather than theory alone.

People Also Ask / FAQ

Can I become a data scientist without a maths background?

Yes. Class 12 level maths is enough to start. The statistics in this guide is learnable in two to three months of steady practice alongside Python.

Is statistics more important than machine learning?

For analyst and most data scientist roles, yes — more decisions rest on correct analysis and experiments than on complex models. ML builds on statistics, so learn it first.

Do I need to learn linear algebra and calculus too?

Basic matrix operations and the idea of a gradient are enough for most roles. Deeper maths matters for research and deep learning engineering.

Which statistics questions are most common in interviews?

Explaining a p-value, Type I versus Type II errors, mean versus median for skewed data, correlation versus causation, designing an A/B test, and a Bayes-style probability puzzle.

Should I learn Bayesian statistics?

Learn Bayes' theorem and the intuition well. Full Bayesian modelling is a valuable later specialisation, not an entry requirement.

Learn Statistics the Way Data Scientists Use It

Live classes, real datasets and mock interviews — online and in select cities across India Explore the Data Science Course