Statistics for Data Scientists 2026 — What You Actually Need

Most of the statistics a data scientist uses every week fits into five areas: summarizing data, modeling uncertainty with distributions, testing whether a difference is real, fitting regressions, and reasoning about cause and effect. Graduate curricula cover far more, and some of it matters in specialized roles. But the daily work of a product or analytics team rests on a small core, and the failures that hurt teams come from misapplying that core, not from missing exotic theory.

This guide walks through the five areas at the depth a hiring manager might probe and a code reviewer would expect. It includes worked numbers you can check by hand, the mistakes that produce confident but wrong conclusions, and an honest list of topics you can defer.

The statistical knowledge that actually matters

The U.S. Bureau of Labor Statistics describes the job in plain terms: data scientists identify useful data, clean and analyze it, build and test models, visualize findings, and recommend business decisions. The same page lists a bachelor's degree in mathematics, statistics, computer science, or a related field as the typical entry requirement, with some employers preferring a master's or doctorate. It reports a median pay of $120,230 per year (May 2025) and projects 35% employment growth from 2025 to 2035, according to the BLS Occupational Outlook Handbook. Every item in that duty list leans on statistics.

The table below maps the five areas to the questions they answer. The interview and on-the-job columns are editorial judgments, not survey results.

Category	Question it answers	Interview emphasis	Day-to-day use
Descriptive statistics	What does this data look like?	Light	Constant
Probability distributions	How much variation should I expect?	Heavy	Frequent
Hypothesis testing	Is this difference bigger than noise?	Heavy	Constant
Regression	How does an outcome change with its drivers?	Heavy	Constant
Causal inference	Did this change cause that outcome?	Moderate	Growing

Everything beyond these, including Markov chain Monte Carlo, Bayesian nonparametrics, and advanced experimental design, pays off in particular domains but is not table stakes. For the career context around the role, the data engineering vs data science comparison shows where statistical work sits relative to pipeline work.

Descriptive statistics: look before you model

Descriptive statistics get the least attention and prevent the most damage. If you fit a model before you have looked at the distribution of each variable, you are driving with the windshield painted over.

Center: mean, median, and mode

The mean is the arithmetic average. The median is the middle value when the data are sorted. The mode is the most frequent value. The three agree for symmetric data and diverge for skewed data.

A small example shows how far they can drift. Five people earn $40,000, $45,000, $50,000, $55,000, and $500,000. The mean is $138,000, which describes nobody in the group. The median is $50,000, which describes most of them. One outlier moved the mean by almost a factor of three while leaving the median untouched. The article on why outliers can change averages goes deeper into this effect.

Spread: variance, standard deviation, and the interquartile range

Spread tells you how much to trust the center. Variance is the average squared distance from the mean. Standard deviation is its square root, which returns the measure to the original units. The interquartile range (IQR) is the distance between the 25th and 75th percentiles, and it ignores the extremes, which makes it the spread companion to the median.

A useful mental picture: the mean is where a seesaw balances, and the standard deviation is how far the riders sit from the pivot. Two seesaws can balance at the same point with riders huddled near the middle or perched at the ends.

Shape: skewness, kurtosis, and boxplots

Skewness measures asymmetry. Right-skewed data, such as incomes, response times, and order values, have a long tail of large values. Kurtosis measures how heavy the tails are relative to a normal distribution. You do not need to compute either by hand. You need to notice them, because many classical tests and intervals assume approximate normality, a point the NIST/SEMATECH Engineering Statistics Handbook makes when it notes that significant skewness and kurtosis indicate data are not normal.

Boxplots, histograms, and quantile tables are the fastest ways to see shape. A boxplot compresses a distribution into five numbers: the minimum, first quartile, median, third quartile, and maximum, with outliers drawn as points.

Why summaries are not enough: Anscombe's quartet

In 1973, the statistician Francis Anscombe built four small datasets of 11 points each. According to Wikipedia's summary of the quartet, all four share the same mean of x (9), the same mean of y (7.50), nearly the same variance of y (about 4.125), a correlation of 0.816, and the same fitted line, y = 3.00 + 0.500x. Yet one set is a clean linear trend, one is a smooth curve, one is a line with a single outlier, and one is a vertical stack with a single leverage point.

The lesson is blunt: identical summary numbers can hide entirely different structures. Plot the data. The piece on how data visualization can mislead covers the opposite problem, where a chart distorts good data.

Four datasets with identical means, variances, correlation, and regression line can look nothing alike when plotted. Summary statistics describe; they do not diagnose.

Probability distributions: the working vocabulary

A probability distribution describes how likely each possible value of a random quantity is. You rarely derive distributions from scratch. You recognize which one fits the process generating your data, because the choice determines which test or model is appropriate.

Distribution	Typical process	Data-science example
Normal (Gaussian)	Sums of many small independent effects	Measurement error, standardized test scores
Binomial	Count of successes in n yes/no trials	Conversions out of visitors
Poisson	Count of events in a fixed interval	Support tickets per hour, defects per batch
Exponential	Waiting time between Poisson events	Time until the next request or failure
Uniform	Every value in a range equally likely	Random assignment, a null baseline
Log-normal	Product of many small multiplicative effects	Revenue per user, file sizes, latency

The normal distribution and the 68-95-99.7 rule

For a normal distribution, about 68% of values fall within one standard deviation of the mean, about 95% within two, and about 99.7% within three. The more precise figure for the 95% band is 1.96 standard deviations, which is why 1.96 appears in nearly every confidence interval formula you will meet.

Binomial and Poisson: when one approximates the other

Conversions follow a binomial distribution: each visitor either converts or does not. When the number of trials n is large and the success probability p is small, the binomial is well approximated by a Poisson distribution with mean n times p.

You can check this yourself. Take 1,000 trials with a 0.2% success probability. The mean count is 2. The exact binomial probability of zero successes is 0.998 raised to the 1,000th power, which is about 0.1351. The Poisson probability of zero events with mean 2 is e to the power of negative 2, about 0.1353. The two agree to three decimal places, which is why Poisson models are standard for rare events.

Why revenue is usually modeled as log-normal

Revenue per user is bounded below by zero, has a long right tail, and is driven by multiplicative effects: a big customer is big because of size times frequency times price. Taking the logarithm turns multiplication into addition and pulls in the tail, so the logged values look roughly normal. That is the practical reason analysts log-transform revenue, latency, and similar quantities before using methods that assume symmetric errors.

The Central Limit Theorem

The Central Limit Theorem (CLT) says that the distribution of sample means approaches a normal distribution as the sample size grows, provided the underlying data have finite variance. It holds even when the individual observations are far from normal.

The CLT is why a t-test on average order value works even though order values are skewed. It is also the reason heavy-tailed data can break the shortcut: with extreme skew or infinite variance, "large enough" can mean far larger than the folk rule of 30 observations. The standard error of the mean, which is the standard deviation divided by the square root of n, shows the price of precision: quadrupling the sample only halves the uncertainty. The explainer on why sample size changes confidence builds on this relationship.

Hypothesis testing: the workhorse

Hypothesis testing is the framework behind A/B tests, before-and-after comparisons, and feature impact studies. It asks whether an observed difference is larger than you would expect from random variation alone. The experimentation side has its own treatment in the A/B testing statistics guide.

The logic in four steps

1. State a null hypothesis, usually "there is no effect," and an alternative hypothesis, "there is an effect."
2. Choose a significance level (alpha), commonly 0.05, which is the false-positive rate you are willing to accept.
3. Compute a test statistic and its p-value: the probability, assuming the null is true, of seeing a result at least as extreme as yours.
4. If the p-value falls below alpha, reject the null. Otherwise, fail to reject it.

Notice what is missing: you never "accept" the null or "prove" the alternative. A court analogy helps. A verdict of "not guilty" means the evidence did not clear the bar, not that the defendant was shown to be innocent.

What a p-value is, and what it is not

A p-value is a statement about the data under an assumed model. It is not the probability that the null hypothesis is true, and it is not the probability that your result happened "by chance" in some vague sense. The American Statistical Association's 2016 statement, written by Ronald Wasserstein and Nicole Lazar in The American Statistician, says p-values can indicate how incompatible the data are with a specified statistical model, as summarized on the Wikipedia page on the misuse of p-values. That same page notes the 0.05 threshold is a convention, not a law of nature.

The distinction has practical weight. A p-value of 0.03 does not mean there is a 97% chance the effect is real. It means that if there were no effect, data this extreme or more would turn up about 3% of the time. Whether the effect is real also depends on prior plausibility, study design, and how many comparisons were made. See how p values are misunderstood for the common failure patterns.

Type I error, Type II error, and power

Two kinds of mistakes are possible, and they trade off against each other.

	Null is actually true	Null is actually false
Reject the null	Type I error (false positive), probability alpha	Correct, probability = power (1 - beta)
Fail to reject	Correct, probability 1 - alpha	Type II error (false negative), probability beta

Alpha is set by you. Beta depends on the sample size, the true effect size, the noise in the data, and alpha. Power is 1 minus beta: the chance of detecting a real effect of a given size. A common planning target is 80% power.

Common tests and when each applies

Situation	Typical test	Key assumption
Compare a mean to a benchmark	One-sample t-test	Approximately normal means
Compare means of two independent groups	Two-sample (Welch) t-test	Independent observations; Welch allows unequal variances
Compare before and after on the same units	Paired t-test	Differences approximately normal
Compare two proportions	z-test for proportions	Large enough counts in each cell
Association between two categorical variables	Chi-square test	Expected counts not too small
Two groups, skewed or ordinal data	Mann-Whitney U	Independent samples; compares distributions
Paired data, skewed or ordinal	Wilcoxon signed-rank	Paired observations
Compare means of three or more groups	ANOVA (or Kruskal-Wallis)	Independent groups; roughly equal spread for ANOVA

Two practical habits prevent most errors. First, use Welch's version of the t-test by default, because it does not require equal variances and costs almost nothing when variances are equal. Second, report an effect size and a confidence interval alongside the p-value, because "significant" says nothing about whether the difference matters.

A worked A/B test, with numbers you can verify

Suppose a checkout redesign is tested on 10,000 users per arm. The control converts at 10.0% and the variant at 10.6%. Is that a win?

• Observed difference: 0.6 percentage points.
• Pooled conversion rate: 10.3%. The standard error of the difference is the square root of 0.103 x 0.897 x (2 / 10,000), which is about 0.0043.
• The z statistic is 0.006 / 0.0043, about 1.40. The two-sided p-value is roughly 0.16.
• The 95% confidence interval for the difference runs from about -0.24 to +1.44 percentage points.

The result is not significant at alpha = 0.05. More informative still, the interval includes zero and also includes a 1.4-point lift that would be commercially important. The honest summary is "inconclusive," not "no effect."

Now plan the test properly. To detect a 10.0% to 10.6% lift with 80% power at alpha = 0.05, the standard normal-approximation formula gives about 40,000 users per arm: (1.96 + 0.84)² x (0.1 x 0.9 + 0.106 x 0.894) / 0.006² is roughly 40,300. The example experiment had one quarter of that. Underpowered tests are one of the most common sources of ambiguous results, and they also inflate the apparent size of any effect that does reach significance.

The explainer on what a confidence interval really means clarifies why an interval is a statement about a procedure and not about a single computed range.

Misuse that sinks analyses

• Multiple comparisons without correction. The family-wise error rate across m independent tests at level alpha is 1 - (1 - alpha)^m. With 20 tests at 0.05, the chance of at least one false positive is 1 - 0.95^20, about 64%. Wikipedia's article on the multiple comparisons problem gives the same formula and notes that with 100 tests the probability exceeds 99%.
• Peeking. Checking the p-value daily and stopping when it dips below 0.05 inflates the false-positive rate well above 5%. Fix the sample size in advance, or use a sequential method designed for repeated looks.
• Reporting the winner only. If you test five metrics and report the one that worked, you ran five tests.
• Confusing statistical and practical significance. With enough data, a trivial difference becomes "significant."

Two standard fixes exist for multiple testing. The Bonferroni correction divides alpha by the number of tests m, which controls the family-wise error rate but, as the Wikipedia article notes, can be overly conservative. The Benjamini-Hochberg procedure, introduced in 1995, controls the false discovery rate, the expected proportion of false positives among the results you call significant. It is often the better choice when you screen many metrics and can tolerate a few false leads.

Regression: where statistics meets machine learning

Linear and logistic regression are both core statistical tools and baseline machine learning algorithms. Understanding them well is the bridge between the statistician's and the engineer's view of a model.

Linear regression

Linear regression fits a line (or plane) that minimizes the sum of squared residuals, where a residual is the gap between an observed and predicted value. scikit-learn's documentation describes LinearRegression as minimizing exactly this residual sum of squares, as detailed in its linear models guide.

What the coefficients mean: each one is the expected change in the outcome for a one-unit change in that predictor, holding the other predictors constant. The last clause is the part people skip. If square footage and number of bedrooms move together, the bedroom coefficient answers "what does an extra bedroom add at the same square footage?" which may be a strange question.

R-squared is the share of variance in the outcome the model explains. It is not a grade. A model can have a high R-squared and badly patterned residuals, and adding any predictor, even random noise, never lowers in-sample R-squared. Use held-out error or adjusted measures to compare models.

The assumptions, and what breaks when they fail

Assumption	If violated
Linearity of the relationship	Biased predictions; systematic residual patterns
Independent errors	Standard errors too small; false confidence (common in time series and clustered data)
Constant error variance (homoscedasticity)	Coefficients still unbiased, but standard errors and intervals are wrong
Roughly normal residuals	Matters mainly for small-sample intervals and tests
No severe multicollinearity	Coefficients become unstable and hard to interpret

On the last point, the scikit-learn documentation explains that when predictors are nearly linearly dependent, the design matrix becomes nearly singular and the least-squares estimate becomes highly sensitive to random errors in the target, producing large variance. Practical remedies are dropping or combining redundant predictors or using regularization.

Regularization in one paragraph

Ridge regression adds a penalty on the sum of squared coefficients, shrinking them toward zero and making estimates more stable under collinearity. Lasso adds a penalty on the sum of absolute coefficients and can set some exactly to zero, so it doubles as feature selection. Per the same documentation, Lasso tends to pick one of several correlated features, and Elastic-Net is suggested when you want correlated features handled together. Regularization trades a little bias for a large reduction in variance, which usually improves out-of-sample accuracy.

Logistic regression

Logistic regression models a binary outcome such as churn or no churn. It predicts a probability by passing a linear combination of predictors through the logistic (sigmoid) function. The linear part is on the log-odds scale, where odds are p / (1 - p).

To interpret a coefficient, exponentiate it. A coefficient of 0.4 gives exp(0.4), about 1.49, so a one-unit increase in that predictor multiplies the odds by roughly 1.49, a 49% increase in odds. Odds are not probabilities: a 49% increase in odds is not a 49% increase in probability. That confusion is one of the most common interview traps.

A practical note from the scikit-learn documentation: its LogisticRegression applies regularization by default, which it says is common in machine learning but not in statistics. If you want classical, unpenalized estimates for inference, you must turn the penalty off. Otherwise your coefficients will be shrunk, and comparisons against a statistics package will not match.

Classification metrics and imbalanced classes

Metric	Definition	Use when
Precision	True positives / predicted positives	False alarms are costly
Recall	True positives / actual positives	Missed cases are costly
F1	Harmonic mean of precision and recall	You need one balance number
ROC AUC	Probability a random positive outranks a random negative	Comparing ranking quality across thresholds

With imbalanced classes, accuracy misleads. If 1% of transactions are fraudulent, a model that predicts "not fraud" every time is 99% accurate and useless. Precision, recall, and the precision-recall curve describe such problems honestly.

Logistic regression remains the right baseline before reaching for anything fancier, because it is fast, interpretable, and hard to beat by a wide margin on tabular data with modest feature engineering. For the tree-based and neural alternatives, the machine learning algorithms guide lays out the options.

Missing data and its effect on regression

Real tables have holes. Dropping rows with missing values silently assumes the data are missing completely at random, which is rarely true. If high-income customers skip the income question, a complete-case regression underrepresents them. The article on how to handle missing data covers imputation strategies and when each is defensible.

Causal inference: the growing edge

Most prediction work asks "what will happen?" Causal work asks "what would happen if we changed something?" The two questions need different methods, and mixing them up is a leading source of bad business decisions.

Correlation, confounding, and Simpson's paradox

A confounder is a variable that influences both the treatment and the outcome, creating a link that is not causal. The classic illustration is ice cream sales and drownings rising together because both follow temperature. The article on how correlation differs from causation expands on the idea.

Simpson's paradox shows how confounding can reverse a conclusion. In a kidney stone study described on Wikipedia, Treatment A succeeded in 93% of small-stone cases (81 of 87) and 73% of large-stone cases (192 of 263), beating Treatment B in both groups (87% and 69%). Yet combined, A succeeded in 78% of cases (273 of 350) and B in 83% (289 of 350). Treatment B had been given mostly to the easier small-stone cases, which lifted its overall rate.

The 1973 University of California, Berkeley, graduate admissions data is the other famous case. Overall, men were admitted at 44% and women at 35% (8,442 male and 4,321 female applicants). Broken down by department, the gap largely disappeared, because women applied more often to highly competitive departments. The aggregate number mixed together groups with very different base rates.

Why randomization works

If you assign treatment by coin flip, then on average the treated and control groups look the same on every characteristic, measured or not. Confounders cannot line up with the treatment, because the coin does not know anything about the users. This is why an A/B test supports a causal claim that a dashboard comparison cannot. The article on why randomized trials are powerful develops the argument.

When you cannot randomize

Method	Idea	Typical use
Difference-in-differences	Compare the before-after change in a treated group to the change in an untreated group	A policy or feature launched in some regions but not others
Regression discontinuity	Compare units just above and just below a cutoff	An eligibility threshold, such as a score or an age
Instrumental variables	Use something that shifts treatment but affects the outcome only through treatment	Encouragement designs, lotteries
Propensity score matching	Match treated and untreated units with similar probabilities of treatment	Observational data with rich covariates

Each method buys identification with an assumption. Difference-in-differences needs the groups to have followed parallel trends absent the intervention. Regression discontinuity needs units not to manipulate which side of the cutoff they land on. Propensity matching can only adjust for confounders you measured. Your job is to state the assumption and test it where possible. For worked examples, see the causal inference guide.

What you can safely defer

Some topics carry prestige in textbooks but rarely drive decisions early in a career:

• Formal proofs of theorems. Valuable for research, seldom needed to ship an analysis.
• Most of Bayesian statistics. It is essential in some fields, such as clinical trials and hierarchical modeling, but frequentist methods handle most product analytics. Learn it once the foundations above are solid.
• Survival analysis. Worth learning when you model time-to-event outcomes such as churn timing or equipment failure.
• Bootstrap and resampling. Useful and quick to pick up when a closed-form interval is unavailable.
• Advanced forecasting. Basic decomposition and baselines cover many needs. When you need more, the explainer on how time series forecasting works is the next stop.

This list is opinionated, and teams differ. Calibrate to the roles you are targeting.

How statistics fits into a project

Experienced analysts tend to follow a similar order of operations:

1. Load and inspect. Check types, ranges, duplicates, and missing values. In pandas, df.describe() is the first call. The pandas data analysis guide covers the tooling.
2. Visualize each variable. Histograms and boxplots catch unit errors, impossible values, and surprising shapes.
3. Define the question precisely. "Does X cause Y?" and "Can we predict Y from X?" lead to different methods.
4. Choose a method that matches the data type and question. Use the test table above as a starting point.
5. Check assumptions. Plot residuals, inspect variances, look for dependence.
6. Report honestly. Give effect sizes with intervals, the number of comparisons you made, and the limits of the design.

Data often has to be pulled before any of this begins. The SQL for data science guide covers the querying skills that feed these steps, and the Python vs R comparison helps with the choice of tooling.

The last step separates careful analysts from careless ones. Overselling a result, or quietly ignoring a null finding, damages trust faster than any technical error.

Where to focus your learning

An order that tends to work: first descriptive statistics and plotting, then distributions and the CLT, then hypothesis testing with a real A/B dataset, then regression, then causal reasoning. Do each with data you can inspect, not just formulas. Standard textbook references include Casella and Berger's Statistical Inference for rigor and Bruce, Bruce, and Gedeck's Practical Statistics for Data Scientists for an applied perspective. Interview preparation follows predictable patterns; the SQL interview questions guide covers the query half of the screen.

The takeaway

The core of statistical practice in data science is smaller than most course catalogs suggest, but it has to be understood, not memorized. A p-value answers a narrow question about data under a model. A regression coefficient is conditional on everything else in the model. A causal claim needs a design that deals with confounding. Most of the costly mistakes, from peeking at A/B tests to trusting an aggregate that hides a Simpson's reversal, come from skipping one of those caveats. Learn the five areas well, plot everything, and state your assumptions out loud.

Frequently Asked Questions

Q: Do I need a PhD in statistics to work as a data scientist? A: No. The BLS lists a bachelor's degree in mathematics, statistics, computer science, or a related field as the typical entry-level requirement, and says some employers require or prefer a master's or doctoral degree. Solid grounding in the five areas above, plus SQL and Python, covers most analytical roles.

Q: Should I learn R or Python first? A: Python is the more common choice for general data science work because of its machine learning and engineering ecosystem, while R remains strong for statistics-heavy research. Our Python vs R comparison explains when each wins.

Q: How much linear algebra do I need? A: Enough to follow regression and dimensionality reduction documentation: matrix multiplication, the idea of a design matrix, eigenvalues and eigenvectors at an intuitive level, and basic vector operations. You do not need to derive every result by hand.

Q: Is Bayesian statistics worth learning? A: Eventually, yes, especially for hierarchical models, small samples, or when prior knowledge matters. Early on, frequentist testing and regression cover most product analytics, so build those foundations first.

Q: How large should an A/B test be? A: It depends on the baseline rate, the smallest effect worth detecting, the power you want, and alpha. As the worked example above shows, detecting a lift from 10.0% to 10.6% at 80% power and alpha = 0.05 needs roughly 40,000 users per arm. Smaller effects need much more data.

Q: What is the most common statistics interview question? A: Explaining a p-value to a non-technical stakeholder is a frequent one. A solid answer: "If there were truly no effect, results at least this extreme would appear about X% of the time. It does not tell us the probability the effect is real."

Q: Do machine learning models replace statistics? A: No. Machine learning extends statistical thinking about uncertainty, sampling, and generalization. Understanding those ideas is what lets you tell whether a model's output means anything.

Sources

• U.S. Bureau of Labor Statistics, Occupational Outlook Handbook: Data Scientists
• scikit-learn documentation, Linear Models
• Wikipedia, Anscombe's quartet, Multiple comparisons problem, Simpson's paradox, Misuse of p-values
• Wasserstein, R. L., and Lazar, N. A. (2016). "The ASA Statement on p-Values: Context, Process, and Purpose." The American Statistician.
• NIST/SEMATECH e-Handbook of Statistical Methods, itl.nist.gov/div898/handbook

---

Read next:

• A/B Testing for Data Scientists: Statistical Fundamentals
• How P Values Are Misunderstood
• SQL Interview Questions for Data Scientists 2026