How to Learn Statistics
Statistics is one idea repeated in many costumes: a sample is not the population, and everything else is bookkeeping about how far off you might be. It feels hard because it is usually taught backwards — as a catalog of named tests reached through a flowchart — so students end up able to run a procedure and unable to say what the output means. Budget around 265 hours for real fluency, from core vocabulary through inference, Bayesian and frequentist reasoning, regression, and experimental design. Along the way you will meet the p-value, the single most misunderstood object in the subject, and discover that the replication crisis in published science is largely downstream of that one misunderstanding.
Updated
Why Learn Statistics?
It is the only discipline that tells you how much to believe something
Every other field produces claims; statistics is where you decide how much weight a claim can carry. That makes it the load-bearing skill under medicine, policy, machine learning, and business decisions alike — and the reason a statistically literate person can evaluate an argument in a field they do not otherwise know.
Your intuition about uncertainty is reliably wrong, and this fixes it
Humans see patterns in noise, underestimate how much small samples wobble, and confuse a rare event with an impossible one. Statistics is not primarily a set of formulas — it is a systematic recalibration of instincts that are wrong in predictable, well-documented directions.
It is the layer AI tools cannot supply for you
Software will happily run any test on any data and return a confident number, because software cannot tell whether the assumptions hold or whether the design supports the conclusion. Knowing what a method assumes, and what a violated assumption does to the answer, is exactly the judgment that automation has made more valuable rather than less.
It lets you read primary research instead of headlines about it
Once confidence intervals, effect sizes, and study design are readable to you, a paper's methods section becomes the most interesting part of it. That single ability changes how you consume health, education, and economic claims permanently, and it takes a fraction of the full curriculum to acquire.
Your Learning Path
Core vocabulary and the one idea underneath it
~25 hoursPopulation versus sample, parameter versus estimate, variance, and above all sampling variability — the fact that a statistic computed from a sample would have come out differently with a different sample. Every technique later in the curriculum is a way of accounting for that wobble, and students who internalize it early stop experiencing statistics as disconnected recipes.
Probability, properly, before any inference
~45 hoursRandom variables, conditional probability, expectation and variance, the common distributions and what generating process each one describes, plus the law of large numbers and the central limit theorem. Inference is applied probability, so weakness here shows up later as an inability to say why a formula has that particular shape.
The inference pipeline end to end
~50 hoursEstimation, standard errors, confidence intervals, hypothesis testing, and what a p-value is and is not — stated as a full conditional sentence every single time until it is automatic. Learn resampling and permutation here too: simulating the null hypothesis yourself makes testing concrete in a way no table of critical values manages.
Bayesian reasoning as a genuine second lens
~30 hoursPriors, likelihood, posteriors, and updating — plus the base-rate problems that expose why an accurate test can still produce mostly false positives on a rare condition. Bayesian thinking answers the question people wrongly believe a p-value answers, which makes learning it the fastest cure for misinterpreting one.
Frequentist theory — why these estimators and not others
~35 hoursMaximum likelihood, bias and variance of estimators, sufficiency, and statistical power. This is the pillar that converts a procedure user into someone who can tell whether a study was capable of detecting the effect it went looking for, which is the question most reported research quietly fails.
Regression as the workhorse of applied statistics
~40 hoursLinear and logistic regression, interpreting coefficients honestly, diagnostics, multicollinearity, and confounding. Most real statistical work is a regression with careful thought about what belongs in it, and the difficult part is never the fitting — it is deciding which variables you are controlling for and why.
Stochastic processes — when observations are not independent
~20 hoursMarkov chains, random walks, queues, and time series, where each observation depends on the last. Independence is the assumption most quietly violated in practice, and this is where you learn what dependence does to your effective sample size and to every interval you computed under the assumption it was absent.
Experimental design and the replication crisis
~20 hoursRandomization, blocking, power calculation, multiple comparisons, preregistration, and the specific analytic freedoms that produced a body of published findings that later replication projects could not reproduce. Study the crisis as a design and incentives problem — it is the most useful case study in the subject, because it shows exactly how correct arithmetic yields false conclusions.
Common Mistakes to Avoid
Learning which test to use instead of what a test does
For every procedure you meet, write three lines: what quantity it estimates, what it assumes, and what the answer looks like when that assumption fails. Then write one permutation test by hand — shuffle the labels a few thousand times and count how often you see your observed difference. Doing that once replaces the whole decision flowchart with actual understanding.
Reading a p-value as the probability that the result is real
Say it back as a full conditional every time: 'if there were genuinely no effect, data at least this extreme would appear this often.' It is a statement about data given a hypothesis, never about a hypothesis given the data. Lead your reporting with the effect size and its interval and let the p-value be a footnote, because that ordering makes the error nearly impossible to commit.
Analyzing until something turns significant
Write the analysis plan before you see the data — the outcome, the model, the exclusions, and the stopping rule — and then count every comparison you actually made, including the ones you abandoned. Report that count. Twenty honest looks at noise produce a significant result roughly once, and the difference between exploration and evidence is entirely whether you disclosed the looking.
Treating statistical significance as practical importance
Always report the estimate in real units with a confidence interval, then ask whether the smallest value in that interval would change any decision. With a large enough sample every trivial difference becomes significant, so significance tells you the effect is probably not exactly zero and nothing whatsoever about whether it matters.
Assuming independent observations because nothing warned you otherwise
Before any interval or test, ask what could make two observations related: repeated measurements on the same person, students within a classroom, customers within a store, or anything ordered in time. Clustered data has a far smaller effective sample size than its row count suggests, so model the structure with clustered standard errors or a hierarchical model rather than hoping it away.
Structured Roadmaps
Follow a guided learning path on Mochivia:
Frequently Asked Questions
Is statistics hard to learn?
How long does it take to learn statistics?
What does a p-value actually mean?
Do I need calculus to learn statistics?
Should I learn Bayesian or frequentist statistics?
Is statistics still worth learning when software runs every test for you?
Learn Statistics in short lessons
Interactive lessons, a few minutes each, in the Mochivia app for iPhone.
See the roadmap →