To learn statistics for data science, study descriptive statistics first, then probability and distributions, sampling and confidence intervals, hypothesis testing, correlation and regression, and finally experiment design such as A/B tests. Learn each idea with a small dataset in code or a spreadsheet, not only with formulas. Understanding why a method works and when it misleads matters more than memorising every test.
How much maths do you need?
Less than most people fear. For analyst roles you need comfortable algebra, reading a graph and a feel for proportions. For machine learning roles, add some linear algebra and calculus later. Start with the statistics that appear in everyday decisions.
The order to learn it in
| Step | Topic | You should be able to |
|---|---|---|
| 1 | Descriptive statistics: mean, median, mode, variance, standard deviation, quartiles | Summarise a column and explain which average to use |
| 2 | Visualising data: histograms, box plots, scatter plots | Spot skew, outliers and clusters |
| 3 | Probability: events, conditional probability, Bayes' rule | Work out a simple conditional probability |
| 4 | Distributions: normal, binomial, Poisson; the central limit theorem | Explain why sample means are roughly normal |
| 5 | Sampling and confidence intervals | Say what a 95% interval does and does not mean |
| 6 | Hypothesis testing: null and alternative, p-values, t-tests, chi-square, errors | Pick a test and interpret the result |
| 7 | Correlation and linear regression | Read coefficients, R-squared and residuals |
| 8 | Experiment design and A/B testing; power and sample size | Plan a test before collecting data |
Ideas worth taking time over
Mean against median
Incomes, house prices and response times are skewed, so the mean is pulled up by a few large values. Report the median or both, and look at the distribution.
What a p-value means
A p-value is the probability of seeing data at least as extreme as yours if the null hypothesis were true. It is not the probability that the null is true, and it says nothing about how large or important an effect is. Report effect sizes and intervals alongside it.
Confidence intervals
A 95% confidence interval comes from a procedure that, over many repeated samples, captures the true value about 95% of the time. It does not mean there is a 95% chance the true value lies in this particular interval. This is one of the most commonly misstated ideas and a favourite interview question.
Correlation is not causation
Two variables moving together may share a cause, be linked by chance, or have the causal direction reversed. Randomised experiments are the standard way to establish cause. Observational data needs care.
Multiple comparisons
If you test 20 hypotheses at the 5% level, you expect about one false positive by chance alone. Adjust or pre-register your analysis.
How to practise
- Simulate. Use Python or R to draw random samples and watch the central limit theorem happen. Seeing it beats reading it.
- Work with a real dataset, such as survey data, sports results or open government statistics, and ask a question that needs a test.
- Explain each result in a sentence a manager would follow.
- Do problems by hand a few times (a small t-test, a probability tree) so the software is not a black box.
- Use free resources: Khan Academy for fundamentals, OpenIntro Statistics (a free textbook), and StatQuest videos for intuition.
Make the concepts stick
Statistics is full of definitions and conditions that sound alike: Type I and Type II error, standard deviation and standard error, population and sample. These are ideal for retrieval practice. Write the definition from memory, then check. Type to Study can build a statistics course from a topic or from a textbook you upload, with typed recall, flashcards on a spaced schedule and exam-style questions with model answers, so you can test whether you can explain a p-value without looking. It teaches the concepts; for running tests you will still use Python, R or a spreadsheet.
A weekly rhythm
- Two sessions to learn a new topic, one reading and one with data.
- One session of mixed problems from earlier topics, to keep them fresh.
- A short daily review of flashcards, ten minutes at most.
- At the end of each month, a mini-analysis that uses several topics together.
Common traps
- Memorising which test to use without understanding its assumptions.
- Treating p < 0.05 as a switch between true and false.
- Ignoring sample size and data quality.
- Learning only from textbooks with clean data. Real data is messy.
- Skipping probability. Many later ideas depend on it.
Statistics rewards patience. Return to the same ideas several times at greater depth, and they will click.