EDUCATION SUBJECT ATLAS · STATISTICS
What Is Statistics?
Statistics is the discipline of learning from data while remaining explicit about variation and uncertainty. It asks how observations should be collected, summarised, compared and modelled; how far conclusions can travel beyond the observed sample; how much uncertainty remains; and what decisions are justified by the evidence.
Statistics is not merely the calculation of averages. Its deeper problem is inference: we observe only part of reality, the observations vary, measurements contain error, samples may be imperfect and patterns can arise by chance. Statistical reasoning provides tools for separating signal from noise without pretending uncertainty has disappeared.
Statistics turns variation from an inconvenience into information and uncertainty from a weakness into something that can be measured.
Data are observations, not reality itself
A dataset is a structured record of observations. Each row, measurement or response is produced through a process: someone defined a variable, chose a population, selected a sample, designed an instrument and recorded an outcome. Statistical analysis begins by understanding that process.
If the data-generating process is poor, sophisticated mathematics cannot fully rescue the result. A precise estimate from a biased sample can be confidently wrong.
Populations and samples
A population is the full set of units about which we want to learn. A sample is the subset actually observed. Statistical inference uses the sample to make claims about the population.
The quality of that inference depends on how the sample was obtained. A large sample chosen badly can be less informative than a smaller sample drawn through a sound design.
Variables
Variables are measured characteristics that can take different values. They may be numerical, categorical, ordinal, binary, continuous, discrete, time-indexed or spatial.
Variable type matters because it affects which summaries and models are meaningful. Calculating an ordinary mean of arbitrary category labels does not create useful information.
Measurement
Statistics depends on measurement. A variable must represent the construct we care about with enough reliability and validity to support the intended conclusion.
Measurement error can be random or systematic. Random error adds noise. Systematic error can bias results in a consistent direction. The second can be especially dangerous because larger samples do not automatically remove it.
Descriptive statistics
Descriptive statistics summarise observed data. Measures of central tendency describe typical values; measures of spread describe variation; tables and graphs reveal structure.
- Mean: arithmetic average.
- Median: middle value when observations are ordered.
- Mode: most frequent value or category.
- Range: distance between minimum and maximum.
- Variance and standard deviation: measures of spread around the mean.
- Interquartile range: spread of the middle half of the data.
No summary is automatically best. The median can represent a skewed income distribution more robustly than the mean, while the mean remains useful for many additive processes.
Distributions
A distribution describes how values are spread across possible outcomes. Shape matters: symmetry, skew, multiple modes, heavy tails and outliers can all change interpretation.
Two datasets can have the same mean and standard deviation while having very different shapes. Visualisation therefore complements numerical summaries.
The normal distribution
The normal distribution is a symmetric bell-shaped model that appears in many statistical methods and approximations. It is mathematically convenient and often useful, but real data should not be forced into normality by assumption alone.
Some phenomena naturally produce skewed, bounded, count or heavy-tailed distributions. Model choice should follow the data-generating process.
Outliers
An outlier is an observation far from the bulk of the data. It may be an error, a rare but real event or evidence that the assumed model is incomplete.
Outliers should be investigated, not automatically deleted. Removing inconvenient values can erase the very phenomenon that matters most.
Probability
Probability provides the mathematical language for uncertainty. It can describe long-run frequency, degree of belief under a model or a formal distribution of possible outcomes.
Probability allows statisticians to reason about what patterns are expected under uncertainty before observing the actual result.
Random variables
A random variable maps uncertain outcomes to values. Its probability distribution specifies how likely different values are under the model.
Expected value summarises the long-run centre of the distribution, while variance measures spread. Neither predicts exactly what one future observation will be.
Conditional probability
Conditional probability asks how probability changes when new information is known. Medical testing, credit risk, weather forecasting and machine learning all depend heavily on conditional reasoning.
A common mistake is to confuse the probability of evidence given a condition with the probability of the condition given the evidence. These are not generally equal.
Bayes’ theorem
Bayes’ theorem connects prior probability, likelihood and posterior probability. It formalises how beliefs should update when new evidence arrives under a probability model.
Bayesian statistics uses this logic to combine prior information with observed data, producing a posterior distribution that represents updated uncertainty.
Sampling
Sampling design determines how observations enter a study. Simple random sampling gives each unit a defined selection chance. Stratified sampling deliberately samples within important subgroups. Cluster sampling selects groups of units. Complex surveys can combine these designs.
Sampling is part of inference, not merely logistics. Selection probabilities and design effects can influence estimates and standard errors.
Sampling bias
Sampling bias occurs when the observed sample differs systematically from the target population in ways relevant to the outcome. Convenience samples and voluntary response surveys are especially vulnerable.
More observations reduce random error but do not automatically remove systematic selection bias.
Sampling distributions
A sampling distribution describes how a statistic would vary across repeated samples from the same process. This idea underlies standard errors, confidence intervals and many significance tests.
We usually observe one sample. Statistical theory imagines the repeated-sampling behaviour to quantify how unstable our estimate might be.
The central limit theorem
Under broad conditions, the distribution of sample means becomes approximately normal as sample size grows, even when individual observations are not normally distributed.
This theorem supports many approximations, but its practical quality depends on sample size, tail behaviour and dependence.
Estimation
Estimation uses sample data to infer unknown quantities such as a population mean, proportion or regression coefficient. A point estimate gives one best value under a method; an interval estimate expresses uncertainty around it.
Good estimation considers bias, variance, efficiency and robustness rather than reporting a number without its uncertainty.
Standard error
The standard error measures the estimated variability of a statistic across repeated samples. It should not be confused with the standard deviation of individual observations.
As sample size grows, standard errors often shrink, reflecting greater precision when other conditions remain suitable.
Confidence intervals
A confidence interval is constructed by a procedure designed to capture the true parameter at a stated long-run rate under specified assumptions. It provides a range of values compatible with the data and method.
A narrow interval is precise, but precision does not guarantee accuracy if bias, measurement error or model failure is present.
Hypothesis testing
Hypothesis testing evaluates how compatible observed data are with a specified null model. A test statistic summarises evidence; a p-value measures how extreme the observed statistic would be under that null model.
A p-value is not the probability that the null hypothesis is true. It also does not measure effect size or practical importance.
Statistical significance and practical significance
With enough data, tiny effects can become statistically detectable. A statistically significant result may still be too small to matter in practice.
Responsible interpretation therefore reports effect size, uncertainty and domain consequence rather than a significance threshold alone.
Type I and Type II errors
A Type I error rejects a true null hypothesis under the chosen decision framework. A Type II error fails to reject a false null. Changing thresholds trades one risk against the other.
Study design should consider which error is more costly in the actual application.
Statistical power
Power is the probability of detecting an effect of a specified size under the model when that effect exists. Power depends on sample size, noise, effect size and decision threshold.
Underpowered studies can produce unstable estimates and exaggerated effects among findings that happen to cross significance thresholds.
Correlation
Correlation measures association between variables. Strong correlation can be useful for prediction, but it does not by itself identify causal direction.
Two variables may move together because one causes the other, because causation runs in both directions, because a third factor influences both or because the pattern is coincidental.
Regression
Regression models relationships between an outcome and one or more predictors. Linear regression estimates average changes under specified assumptions; other regression families handle binary, count, survival and non-linear outcomes.
Regression coefficients must be interpreted in the context of model specification. Adding or removing variables can change meaning when predictors are related.
Prediction and explanation
A model can predict well without revealing causal mechanisms, and a causal explanation can be useful even when individual predictions remain noisy. Statistics distinguishes predictive accuracy from explanatory inference.
The correct evaluation depends on the goal. Forecasting tomorrow’s demand and estimating the effect of a policy are different tasks.
Confounding
Confounding occurs when a third variable influences both an exposure and an outcome, creating or distorting an observed association.
Statistical adjustment can help when confounders are measured appropriately, but no model can adjust for important factors that were never observed or conceptualised.
Randomised experiments
Random assignment helps create comparable groups by making treatment assignment independent of many potential confounders on average. This strengthens causal inference.
Experiments still require attention to compliance, attrition, measurement, treatment implementation and generalisability.
Observational studies
Observational studies examine exposures that were not randomly assigned. They are essential where experiments are impossible or unethical, but causal claims require stronger assumptions and designs.
Matching, weighting, regression adjustment, natural experiments and quasi-experimental methods can reduce bias when their assumptions are credible.
Causal inference
Causal inference asks what would have happened to the same unit under an alternative condition. Because we cannot observe both realities simultaneously, causal analysis uses design and assumptions to estimate the missing counterfactual.
This is why causal claims are not generated automatically by larger datasets. Design logic matters.
Missing data
Missing data can arise randomly, conditionally on observed variables or through mechanisms related to the missing values themselves. Different mechanisms require different treatments.
Deleting incomplete records can introduce bias if missingness is informative. Good analysis investigates why values are missing.
Multiple comparisons
When many hypotheses are tested, some apparently significant results can occur by chance. Multiple-comparison methods control error rates across families of tests.
Researchers should define primary questions before exploring hundreds of possible associations.
Model assumptions
Every statistical model simplifies reality. Common assumptions concern independence, distributional form, linearity, constant variance or missingness mechanisms.
Assumptions should be diagnosed rather than hidden. Sensitivity analysis asks whether conclusions survive reasonable alternative specifications.
Overfitting
Overfitting occurs when a model learns accidental features of the training data that do not generalise. Flexible models are especially vulnerable when data are limited.
Validation, regularisation, cross-validation and held-out test data help estimate out-of-sample performance.
Bayesian and frequentist approaches
Frequentist and Bayesian statistics represent uncertainty differently. Frequentist methods evaluate procedures through repeated-sampling properties. Bayesian methods represent uncertainty about parameters with probability distributions updated by data.
Both traditions offer powerful tools. The appropriate approach depends on the question, assumptions, available prior information and decision context.
Resampling
Resampling methods such as the bootstrap approximate sampling variability by repeatedly recomputing statistics from resampled data. Permutation methods assess how unusual a pattern is under rearrangements consistent with a null hypothesis.
These approaches reduce dependence on some closed-form formulas but still rely on assumptions about what the observed data represent.
Time series
Time-series data violate simple independence because observations close in time can be related. Trend, seasonality, autocorrelation and structural breaks must be considered.
Forecasting requires honest backtesting because models that fit historical data can fail when the underlying process changes.
Spatial statistics
Spatial data also contain dependence: nearby locations can resemble one another. Geographic scale, boundaries and spatial autocorrelation affect inference.
Maps can make patterns visually persuasive before statistical evidence is strong, so spatial analysis requires careful modelling.
Robust statistics
Robust methods are designed to perform reasonably when assumptions are imperfect or when outliers exert excessive influence. Medians, trimmed means and robust regression are examples.
Robustness is not a substitute for understanding anomalies, but it prevents a few observations from controlling every conclusion.
Reproducibility
A statistical analysis should be reproducible: another qualified person should be able to trace data, transformations, code, model choices and outputs.
Reproducibility supports error detection and turns analysis into an inspectable evidence pipeline rather than a one-off calculation.
Statistical ethics
Statistics can mislead through selective reporting, hidden exclusions, inappropriate graphs, p-hacking, overconfident causal language or omission of uncertainty.
Ethical statistics makes the analytical degrees of freedom visible and avoids presenting fragile results as settled fact.
A systems model of statistics
- ENTITY: units, observations, variables, samples and populations.
- STATE: measured values, distributions, model parameters and uncertainty.
- OCCURRENCE: sampling, measurement, treatment, missingness and updating.
- RELATIONSHIP: association, dependence, causal link, hierarchy and clustering.
- INTENT: research question, estimand, prediction goal and decision threshold.
- OBSERVATION: recorded measurements and data-generating processes.
- ARTIFACT: datasets, code, models, tables, graphs and reports.
- CLAIM: estimates, intervals, predictions and causal conclusions.
- VOID: unmeasured confounders, missing values, model uncertainty and unobserved counterfactuals.
structured systems analysis statistics keeps the claim attached to the process that produced it. Data, sampling, measurement, model and uncertainty are rotated before the conclusion is released.
How to think statistically
- Define the question and target population.
- Identify how data were generated.
- Inspect distributions before modelling.
- Separate description from inference.
- Quantify uncertainty.
- Distinguish association from causation.
- Check assumptions and alternative models.
- Report effect size as well as significance.
- Test sensitivity and reproducibility.
- State what the data cannot answer.
Common misconceptions
- “A larger sample fixes bias.” Larger samples mainly reduce random error.
- “A p-value is the probability the hypothesis is true.” It is not.
- “Correlation proves causation.” Design and assumptions are needed.
- “A precise estimate must be accurate.” Precision can coexist with systematic bias.
- “One average describes everyone.” Distribution and heterogeneity matter.
Mini case: an average test score rises
Suppose a school’s average score rises by five points. Statistics asks whether the cohort changed, how variable scores are, whether the same test was used, whether missing students differ systematically and whether the change exceeds ordinary year-to-year fluctuation.
The five-point increase is an observation. Statistical reasoning determines what can responsibly be inferred from it.
Mini case: a treatment appears effective
If people choosing a treatment improve more than those who do not, the difference may reflect treatment benefit, baseline health, motivation, access or other confounding factors. A randomised design or credible quasi-experiment is needed for stronger causal inference.
Statistics converts “the groups differ” into the harder question “what caused the difference?”
Statistics across the learning journey
Young learners can begin with tables, charts, averages and variation. Secondary learners can study probability, sampling, correlation and inference. Advanced study adds regression, experimental design, Bayesian methods, causal inference, time series, multivariate analysis and statistical computing.
The progression is from describing data to reasoning about what unseen reality the data can support.
Why statistics belongs inside education
Statistics teaches intellectual humility with precision. It shows how much evidence supports a claim, how uncertainty should be represented and why method matters as much as the final number.
In a world of dashboards, polls, experiments, rankings and machine learning, statistical literacy is one of the core defenses against confident misunderstanding.