A histogram is a visual representation of the distribution of quantitative data. The term was first introduced by Karl Pearson.[1] To construct a histogram, the first step is to "bin" (or "bucket") the range of values— divide the entire range of values into a series of intervals—and then count how many values fall into each interval. The bins are usually specified as consecutive, non-overlapping intervals of a variable. The bins (intervals) are adjacent and are typically (but not required to be) of equal size.[2]
Histograms give a rough sense of the density of the underlying distribution of the data, and often for density estimation: estimating the probability density function of the underlying variable. The total area of a histogram used for probability density is always normalized to 1. If the length of the intervals on the x-axis are all 1, then a histogram is identical to a relative frequency plot.
Histograms are sometimes confused with bar charts. In a histogram, each bin is for a different range of values, so altogether the histogram illustrates the distribution of values. But in a bar chart, each bar is for a different category of observations (e.g., each bar might be for a different population), so altogether the bar chart can be used to compare different categories. Some authors recommend that bar charts always have gaps between the bars to clarify that they are not histograms.[3] [4]
This is the data for the histogram to the right, using 500 items:
| Bin/Interval | Count/Frequency |
|---|---|
| −3.5 to −2.51 | 9 |
| −2.5 to −1.51 | 32 |
| −1.5 to −0.51 | 109 |
| −0.5 to 0.49 | 180 |
| 0.5 to 1.49 | 132 |
| 1.5 to 2.49 | 34 |
| 2.5 to 3.49 | 4 |
The words used to describe the patterns in a histogram are: "symmetric", "skewed left" or "right", "unimodal", "bimodal" or "multimodal".
It is a good idea to plot the data using several different bin widths to learn more about it. Here is an example on tips given in a restaurant.
The U.S. Census Bureau found that there were 124 million people who work outside of their homes.[5] Using their data on the time occupied by travel to work, the table below shows the absolute number of people who responded with travel times "at least 30 but less than 35 minutes" is higher than the numbers for the categories above and below it. This is likely due to people rounding their reported journey time. The problem of reporting values as somewhat arbitrarily rounded numbers is a common phenomenon when collecting data from people.
| Interval | Width | Quantity | Quantity/width | |
|---|---|---|---|---|
| 0 | 5 | 4180 | 836 | |
| 5 | 5 | 13687 | 2737 | |
| 10 | 5 | 18618 | 3723 | |
| 15 | 5 | 19634 | 3926 | |
| 20 | 5 | 17981 | 3596 | |
| 25 | 5 | 7190 | 1438 | |
| 30 | 5 | 16369 | 3273 | |
| 35 | 5 | 3212 | 642 | |
| 40 | 5 | 4122 | 824 | |
| 45 | 15 | 9200 | 613 | |
| 60 | 30 | 6461 | 215 | |
| 90 | 60 | 3435 | 57 |
This histogram shows the number of cases per unit interval as the height of each block, so that the area of each block is equal to the number of people in the survey who fall into its category. The area under the curve represents the total number of cases (124 million). This type of histogram shows absolute numbers, with Q in thousands.
| Interval | Width | Quantity (Q) | Q/total/width | |
|---|---|---|---|---|
| 0 | 5 | 4180 | 0.0067 | |
| 5 | 5 | 13687 | 0.0221 | |
| 10 | 5 | 18618 | 0.0300 | |
| 15 | 5 | 19634 | 0.0316 | |
| 20 | 5 | 17981 | 0.0290 | |
| 25 | 5 | 7190 | 0.0116 | |
| 30 | 5 | 16369 | 0.0264 | |
| 35 | 5 | 3212 | 0.0052 | |
| 40 | 5 | 4122 | 0.0066 | |
| 45 | 15 | 9200 | 0.0049 | |
| 60 | 30 | 6461 | 0.0017 | |
| 90 | 60 | 3435 | 0.0005 |
This histogram differs from the first only in the vertical scale. The area of each block is the fraction of the total that each category represents, and the total area of all the bars is equal to 1 (the fraction meaning "all"). The curve displayed is a simple density estimate. This version shows proportions, and is also known as a unit area histogram.
In other words, a histogram represents a frequency distribution by means of rectangles whose widths represent class intervals and whose areas are proportional to the corresponding frequencies: the height of each is the average frequency density for the interval. The intervals are placed together in order to show that the data represented by the histogram, while exclusive, is also contiguous. (E.g., in a histogram it is possible to have two connecting intervals of 10.5–20.5 and 20.5–33.5, but not two connecting intervals of 10.5–20.5 and 22.5–32.5. Empty intervals are represented as empty and not skipped.)[6]
The data used to construct a histogram are generated via a function mi that counts the number of observations that fall into each of the disjoint categories (known as bins). Thus, if we let n be the total number of observations and k be the total number of bins, the histogram data mi meet the following conditions:
n=
| k{m | |
| \sum | |
| i}. |
An alternative to kernel density estimation is the average shifted histogram,[7] which is fast to compute and gives a smooth curve estimate of the density without using kernels.
A cumulative histogram is a mapping that counts the cumulative number of observations in all of the bins up to the specified bin. That is, the cumulative histogram Mi of a histogram mj is defined as:
Mi=
| i{m | |
| \sum | |
| j}. |
There is no "best" number of bins, and different bin sizes can reveal different features of the data. Grouping data is at least as old as Graunt's work in the 17th century, but no systematic guidelines were given until Sturges's work in 1926.
Using wider bins where the density of the underlying data points is low reduces noise due to sampling randomness; using narrower bins where the density is high (so the signal drowns the noise) gives greater precision to the density estimation. Thus varying the bin-width within a histogram can be beneficial. Nonetheless, equal-width bins are widely used.
Some theoreticians have attempted to determine an optimal number of bins, but these methods generally make strong assumptions about the shape of the distribution. Depending on the actual data distribution and the goals of the analysis, different bin widths may be appropriate, so experimentation is usually needed to determine an appropriate width. There are, however, various useful guidelines and rules of thumb.[8]
The number of bins k can be assigned directly or can be calculated from a suggested bin width h as:
k=\left\lceil
| maxx-minx | |
| h |
\right\rceil.
The braces indicate the ceiling function.
k=\lceil\sqrt{n}\rceil
which takes the square root of the number of data points in the sample and rounds to the next integer. This rule is suggested by a number of elementary statistics textbooks [9] and widely implemented in many software packages.[10]
Sturges's rule[11] is derived from a binomial distribution and implicitly assumes an approximately normal distribution.
k=\lceillog2n\rceil+1,
Sturges's formula implicitly bases bin sizes on the range of the data, and can perform poorly if, because the number of bins will be small—less than seven—and unlikely to show trends in the data well. On the other extreme, Sturges's formula may overestimate bin width for very large datasets, resulting in oversmoothed histograms.[12] It may also perform poorly if the data are not normally distributed.
When compared to Scott's rule and the Terrell-Scott rule, two other widely accepted formulas for histogram bins, the output of Sturges's formula is closest when .[12]
k=\lceil2\sqrt[3]{n}\rceil
The Rice rule[13] is presented as a simple alternative to Sturges's rule.
Doane's formula[14] is a modification of Sturges's formula which attempts to improve its performance with non-normal data.
k=1+log2(n)+log2\left(1+
| |g1| | ||||
|
\right)
where
g1
| \sigma | |
| g1 |
=\sqrt{
| 6(n-2) | |
| (n+1)(n+3) |
}
Bin width
h
h=
| 3.49\hat\sigma | |
| \sqrt[3]{n |
where
\hat\sigma
k=\sqrt[3]{2n}
The Terrell–Scott rule[12] [18] is not a normal reference rule. It gives the minimum number of bins required for an asymptotically optimal histogram, where optimality is measured by the integrated mean squared error. The bound is derived by finding the 'smoothest' possible density, which turns out to be
| 3 | |
| 4 |
(1-x2)
The Freedman–Diaconis rule gives bin width
h
h=2
| \operatorname{IQR | |
| (x)}{\sqrt[3]{n}}, |
which is based on the interquartile range, denoted by IQR. It replaces 3.5σ of Scott's rule with 2 IQR, which is less sensitive than the standard deviation to outliers in data.
This approach of minimizing integrated mean squared error from Scott's rule can be generalized beyond normal distributions, by using leave-one out cross validation:[20] [21]
\underset{h}{\operatorname{argmin}}\hat{J}(h)=\underset{h}{\operatorname{argmin}}\left(
| 2 | |
| (n-1)h |
-
| n+1 | |
| n2(n-1)h |
\sumk
| 2 | |
| N | |
| k |
\right)
Here,
Nk
The choice is based on minimization of an estimated L2 risk function[22]
\underset{h}{\operatorname{argmin}}
| 2\bar{m | |
| - |
v}{h2}
where
style\bar{m}
stylev
styleh
| style\bar{m}= | 1 |
| k |
| k | |
| \sum | |
| i=1 |
mi
stylev=
| 1 | |
| k |
| k | |
| \sum | |
| i=1 |
(mi-\bar{m})2
Rather than choosing evenly spaced bins, for some applications it is preferable to vary the bin width. This avoids bins with low counts. A common case is to choose equiprobable bins, where the number of samples in each bin is expected to be approximately equal. The bins may be chosen according to some known distribution or may be chosen based on the data so that each bin has
≈ n/k
For equiprobable bins, the following rule for the number of bins is suggested:[23]
k=2n2/5
This choice of bins is motivated by maximizing the power of a Pearson chi-squared test testing whether the bins do contain equal numbers of samples. More specifically, for a given confidence interval
\alpha
k=4\left(
| 2n2 | |
| \Phi-1(\alpha) |
| ||||
| \right) |
Where
\Phi-1
\alpha=0.05
1.88n2/5
3.77n2/5
A good reason why the number of bins should be proportional to
\sqrt[3]{n}
n
n
s
nh/s
\sqrt{s/(nh)}
h/s
h
s/\sqrt[3]{n}
k
\sqrt[3]{n}