Frequency Distribution

A frequency distribution is a complete summary of how the observations in a dataset are spread across the possible values or classes, pairing each distinct value (or interval of values) with the number of times it occurs. Where a single frequency is one count, the distribution is the whole pattern: which values are common, which are rare, where the data cluster, and how they trail off. A list of 500 raw exam scores is hard to read, but the same scores arranged by score band show instantly whether most students sit in the middle, whether a second hump appears, or whether a long tail of low scores drags the picture to one side.

A frequency distribution is the shape of the data, and shape is information that no single summary number carries. Two datasets can share the same mean and standard deviation yet have entirely different distributions: Francis Anscombe's 1973 quartet of four small datasets with nearly identical means, variances, and correlations but wildly different plots is the standard demonstration, and its lesson is to look at the distribution before trusting a summary. The vocabulary used to describe shape is itself the reason distributions are examined: symmetric or skewed (a longer tail to the right or left), unimodal or bimodal (one peak or two), light-tailed or heavy-tailed, and whether any isolated points sit far from the rest.

Building a distribution for continuous data requires choosing class intervals, and that choice is a genuine analytical decision. Classes must not overlap and must cover the whole range; equal widths are the default, since unequal widths require adjusting heights so that area, not height, represents frequency. Too few classes smooth away real structure, while too many leave erratic, near-empty bins that reflect sampling noise. Rules of thumb exist, such as Sturges' rule of 1926, which suggests about one plus log base 2 of n classes, and the Freedman-Diaconis rule, which sets width from the interquartile range and the cube root of n, but none replaces trying several widths and checking that the story is stable. The histogram is the standard picture of a grouped distribution, and the bar chart is its counterpart for categories.

The relationship between a distribution observed in data and the probability distributions of theory is the connecting idea of statistics. The relative-frequency form of a frequency distribution is an empirical estimate of an underlying probability distribution, and the historical discovery that real-world measurements often arrange themselves into a bell-shaped distribution is what drew Adolphe Quetelet, in the 1830s and 1840s, to apply the error curve to human characteristics such as height and chest measurements. Comparing a frequency distribution to a candidate model, such as the normal distribution, is the idea behind Karl Pearson's 1900 chi-square goodness-of-fit test. In practice, distributions drive decisions on whether to use a mean or a median, whether a transformation (such as a logarithm for right-skewed incomes) is warranted before modelling, and whether a feature in a machine-learning pipeline has drifted since training, which is monitored by comparing its present distribution to the training distribution.

The principal pitfalls are over-reading and mislabelling. Small samples produce lumpy distributions whose bumps are noise, so a second peak in 30 observations deserves scepticism. Distributions of different groups plotted with different bin widths or axis ranges cannot be compared by eye. A distribution of individual values is also not the same thing as a sampling distribution of a statistic, which describes how an estimate such as a mean varies across repeated samples and is far narrower. Finally, a distribution says nothing about causes: a bimodal pattern signals that two populations may be mixed, not what separates them.

A frequency distribution is the concept; a frequency table and a chart are two ways of presenting it. The table lists classes with frequency, relative frequency, and cumulative frequency in columns, the histogram or bar chart draws the same information, and a pie chart shows a categorical version as shares of a whole. All of them are views of one object, and which view is clearest depends on whether the variable is categorical or numeric, and on the audience.

Read a frequency distribution as the data's fingerprint: an arrangement of every observation by value that reveals centre, spread, skew, and clusters at once, which summary statistics compress and sometimes conceal, and which is always the first thing worth looking at.