Distribution

In statistics, a distribution is a mathematical function or empirical description that shows all possible values of a variable and how frequently they occur. Distributions characterize the shape, center, and spread of data and are foundational to statistical inference, as the choice of statistical test depends on the distribution of the data. Common theoretical distributions include the normal (bell shaped), binomial, Poisson, chi square, t, and F distributions, each defined by specific parameters and properties that make them appropriate for different types of data and research questions.

The distribution of a variable determines which statistical methods are appropriate for analyzing it. Parametric tests such as t tests and ANOVA assume that the data within each group (more precisely, the model residuals) follow a normal distribution, or that sample sizes are large enough for the central limit theorem to apply. When data are strongly non normal, these tests may produce misleading results, and nonparametric alternatives or data transformations may be needed. Understanding distributions also helps researchers detect data quality issues: unexpected distributional features such as bimodality, heavy tails, or floor and ceiling effects may indicate measurement problems, data entry errors, or the presence of distinct subgroups. In meta analysis, the assumed distribution of true effect sizes across studies underpins the choice between fixed effect and random effects models.

A distribution can be described empirically (by examining the actual data) or theoretically (by specifying a mathematical function). Empirical distributions are visualized using histograms, density plots, box plots, or quantile quantile (QQ) plots. Theoretical distributions are defined by parameters: the normal distribution by its mean and standard deviation, the binomial distribution by the number of trials and the probability of success, and the Poisson distribution by its rate parameter.

The normal distribution is the most important distribution in inferential statistics because many biological and behavioral measurements approximately follow a bell shaped curve, and because the central limit theorem shows that the sampling distribution of the mean becomes approximately normal as the sample size grows, whatever the shape of the underlying data (provided the observations are independent and their variance is finite). The t distribution, which has heavier tails than the normal distribution, is used for inference about means when the population standard deviation is unknown and has to be estimated from the sample; its tails are heaviest for small samples, and it approaches the normal distribution as the sample size grows. The chi square distribution arises in tests of independence and goodness of fit for categorical data. The F distribution is used in ANOVA and in testing the overall significance of regression models.

Skewness describes the asymmetry of a distribution. A positively skewed distribution (right skew) has a long tail extending toward higher values, common in data like income, hospital length of stay, and reaction times. A negatively skewed distribution (left skew) has a long tail toward lower values. Kurtosis describes the heaviness of the tails relative to a normal distribution: leptokurtic distributions have heavier tails than the normal and platykurtic distributions lighter ones. Despite a common description, kurtosis says little about whether the peak is sharp or flat. These properties influence the reliability of parametric tests and guide decisions about data transformation or test selection.

Understanding distributions is useful at nearly every stage of quantitative research. During data exploration, examining distributions helps identify outliers, coding errors, and distributional features that inform analytical choices. During analysis, the distributional assumptions of statistical tests must be checked and met (or the analysis must account for violations). During reporting, descriptions of data distributions help readers evaluate the appropriateness of the analyses and the generalizability of the results. Knowledge of theoretical distributions is essential for understanding the logic of hypothesis testing, confidence interval construction, and Bayesian inference.

A frequent error is assuming normality without checking it, applying parametric tests to data that are clearly non normal. Another common mistake is judging normality from a histogram alone; researchers sometimes report that data are "approximately normal" from a histogram's rough shape when a QQ plot, or a formal test such as Shapiro Wilk or Kolmogorov Smirnov, reveals clear departures, for example in the tails. Confusing the distribution of individual observations with the distribution of sample means leads to inappropriate reasoning about the applicability of the central limit theorem. Using the normal distribution for data that are bounded (such as proportions, which are bounded between 0 and 1) or count data (which are discrete and non negative) can produce invalid results; binomial or Poisson distributions are often more appropriate. Finally, over reliance on formal normality tests with large samples is also problematic, as these tests become overly sensitive and reject normality for trivial departures that have no practical impact on the analysis.

A researcher analyzing hospital length of stay data for 4,000 surgical patients finds that the distribution is strongly positively skewed, with a median of 3 days, a mean of 5.2 days, and a long tail extending to 45 days. A histogram confirms the skew, and a Shapiro Wilk test decisively rejects normality (p < 0.001). The researcher reports median and interquartile range rather than mean and standard deviation, uses the Mann Whitney U test for group comparisons, and presents a log transformed version of the variable for regression analysis, where approximately normal residuals are assumed. With 4,000 patients, a t test comparing mean stays would still be valid by the central limit theorem, but the long tail pulls the mean well above the typical stay, which is why the median is reported.