PM R 11 (2019) 309–312
www.pmrjournal.org
Statistically Speaking
Histograms: A Useful Data Analysis Visualization
Regina L. Nuzzo, PhD
Introduction
Visualizing data with histograms is an excellent first
step in any analysis of quantitative data, but many
researchers fail to take advantage of this exploratory
data analysis tool. This article gives an overview of histo-
grams and uses examples to illustrate important data fea-
tures that they can help reveal.
The Nature of Histograms
Histograms were one of the earliest types of data visu-
alizations, with references to their use dating back to the
19th century.1 The goal of these graphs is to visualize the
shape (distribution) of data for a single quantitative var-
iable such as systolic blood pressure, age, or birthweight.
(Notice that histograms are not bar charts. Bar charts are
properly used only for displaying counts of categorical
variables. Histograms and boxplots display quantitative
data.)2
Derived from the Latin root words for “drawn fences,”
histograms typically consist of a number of adjacent,
equal-width vertical columns, drawn so that there is no
space between the columns. The columns correspond to
“bins” that together span the range of the data. The data
are divided among these bins, with the height of each bar
corresponding to the number of data points falling into
each bin. The taller the bar, the more data points fall into
the range of that bin.
Skewness
Figure 1 shows the length of hospital stays for a random
sample of 1000 patients enrolled in Phase I of SUPPORT
(Study to Understand Prognoses Preferences Outcomes
and Risks of Treatment).3 (The dataset is available at
http://biostat.mc.vanderbilt.edu/wiki/Main/DataSets)
© 2019 American Academy of Physical Medicine and Rehabilitation
https://dx.doi.org/10.1002/pmrj.12145
The hospital stays in this study ranged from 3 days to
241 days. The histogram shows 25 bins extending from
0 to 250 days, so that each bin covers a period of 10 days.
The data in Figure 1 can be seen to exhibit a strong
right or positive skew. This happens when the majority
of the data have low values, but a few values are
extremely high relative to the rest (ie, the data have a
long right tail). Here, about 75% of the 1000 study
patients had short hospital stays with a duration of less
than 20 days (almost 475 patients stayed up to 10 days,
and nearly 300 stayed between 10 and 20 days). After
20 days, the probability of a patient staying even longer
drops dramatically over time, which can be seen in the
long, tapering right tail. In fact, only 5 patients stayed
longer than 140 days. Additionally, because hospital stays
must be nonnegative, there is a floor effect (or basement
effect) at zero. This happens when data values tend to
“pile up” near the lowest possible limit of the variable.
The combination of skewness and floor effect suggests
to researchers that they should avoid most traditional sig-
nificance tests for this data (such as t-tests, analyses of
variance, and tests for correlation coefficients). This is
because these tests often rely on the underlying popula-
tion being relatively symmetric. Using these tests when
the data are skewed can lead to making incorrect infer-
ences such as rejecting the null hypothesis when the null
hypothesis is actually true (type I error). Researchers
should also report the median and IQR (interquartile
range) of the sample instead of (or at least in addition to)
the mean and SD. This is because the mean and SD can be
highly sensitive to skewness and outliers, and the median
and IQR are more robust measures of center and spread.
Comparing Subpopulations
Histograms are also useful to compare the distributions
of two subpopulations on the same axes. For example, in
310 Histograms: A Useful Data Analysis Visualization

Histogram of Hospital Stay Durations Overlaid histograms

(each binwidth = 10 days)
group
metabolic disorder
normal
450

400
20

350
Number of patients

300

250
15

Number of participants
200

150

100 10

50

0 20 40 60 80 100 120 140 160 180 200 220 240 260

5
Days until Hospital Discharge

Figure 1. Histogram of length of hospital stays for a random sample of

1000 enrolled in Phase I of SUPPORT (Study to Understand Prognoses
Preferences Outcomes and Risks of Treatment).3 The majority (approxi-
mately 75%) of the patients had stays of less than 20 days, whereas the
remaining quarter had stays ranging from 20 to 241 days. There is a floor
effect and a strong positive skew. 0

1964 cancer researcher Frits de Waard and colleagues pub- 30 35 40 45 50 55 60 65 70 75 80 85

lished a paper that investigated the bimodality in age-
specific breast cancer incidence that had been observed
since before World War II.4 Figure 2 shows a histogram with
data simulated to recreate their published histograms.
The researchers stratified their sample of 240 patients
with mammary carcinoma into two groups: those who
exhibited obesity, hypertension, and/or decreased glu-
cose tolerance and those who exhibited none of these
symptoms. The former group had its peak around
65-69 years of age, whereas the latter peaked at around
45-49 years. The dark shaded area shows the overlap in
the distributions of the two groups. After the age of
75 breast cancer was seen only in women with one of the
indications. These histograms combined with other data
led the researchers to propose that breast cancers develop
along two different pathways, ones that researchers now
realize correspond to estrogen-receptor-positive cancers
and estrogen-receptor-negative cancers.
Bin Width: The Tuning Parameter for Histograms
The most important decision a researcher faces when
developing a histogram is the width of the bins.
Age of breast cancer incidence
Figure 2. Overlaid histograms of age-specific breast cancer incidence,
reconstructed from published histograms in de Waard et al.4 The group
of patients with indications of metabolic disorder ranged from 30 to
85 years of age, whereas the group without indications ranged from
30 to only 74 years. The dark shaded area shows the overlap between
the two groups. Using histograms to compare these subgroups help lead
the researchers to propose the existence of two distinct mechanisms for
breast cancer.
(Alternatively, the number of bins can be chosen: the
greater the number of bins, the smaller the bin width.)
Bin widths control the “resolution” of the histogram. If
the bins are too wide, the histogram becomes very “soft
focus,” without a clear shape and with many interesting
data features obscured. On the other hand, using too-
narrow bins will result in a histogram with an overly
choppy result; this tends to accentuate random artifacts
in the data sample and makes it difficult to discern the
true distribution of the underlying population data.
See Figure 3, for example, which shows the effect of
different bin widths on the interpretation of the data.
The variable plotted here is average systolic blood pres-
sure (SBP) of each patient early in the SUPPORT study.
 R. L. Nuzzo / PM R 11 (2019) 309–312 311

(A) Undersmoothed Histogram (B) Oversmoothed Histogram

(70 bins) (7 bins)
400

40

300
Number of participants

Number of participants
30

200

20

100
10

0 0

0 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 120 140 160 180
Mean Systolic BP Mean Systolic BP

(C) Balanced Histogram

(20 bins)

150
Number of participants

100

50

0 20 40 60 80 100 120 140 160 180

Mean Systolic BP
Figure 3. Histograms (of mean systolic blood pressure from participants enrolled in Phase I of the SUPPORT study) showing the effect of bin width on
interpretability.2 The bins in panel A are too narrow, resulting in a plot that accentuates random artifacts in the data. The bins in panel B are too wide,
obscuring the bimodality in the data. Panel C follows the Rice rule and displays a balanced plot that reveals an almost-smooth bimodal shape with
peaks at about 70 mm Hg and 110 mm Hg.
312 Histograms: A Useful Data Analysis Visualization

Figure 3C shows a histogram using 20 bins, each cover-

Box ing 9 mm Hg (this choice gave the most interpretable bin
width in whole numbers). Here the bimodality is clear,
Method Number of bins Bin width with one peak at around 70 mm Hg and another at
pffiffiffi 110 mm Hg. This shape should be a clear sign to
Square root n ffiffi
Maxp−Min
n
pffiffiffiffiffiffiffiffiffiffiffi researchers that further investigation is warranted to
1000 ≈32 ffiffiffiffiffiffiffi
180
p −0ffi
≈6
1000 investigate if there are distinct subgroups in this sample.
Sturges5 ceil(log2n) + 1 Max −Min
ceilðlog2 nÞ + 1
ceil(log21000) + 1 = 11 180 −0
ceilðlog2 1000Þ + 1 = 16:4
Rice6 2 * n1/3 Max −Min Conclusion
2 * n1=3
2 * 10001/3 = 20 180 −0
=9
2 * 10001=3
Scott7 ðMax −MinÞ * n1=3 3:5 * SD
n1=3
Histograms are useful exploratory data visualizations
3:5 * SD
ð180− 0Þ * 10001=3 3:5 * SD
= 9:6 for spotting outliers, skew, bimodality, and other shape
3:5 * 27:3 ≈19 n1=3
Freedman-Diaconis8 ðMax −MinÞ * n1=3 2 * IQR features in the distribution as well as for comparing sub-
2 * IQR n1=3
2 * 42:25
groups in the data. The presence of strong skewness or
ð180− 0Þ * 10001=3 = 8:45
2 * 42:25 ≈21 10001=3 outliers should lead researchers to investigate the use of
median and IQR as summary statistics and nonparametric
Notes hypothesis tests instead of traditional parametric tests.
ceil(x) refers to the ceiling function, which returns the Bin width/bin number is a tuning parameter that should
smallest integer that is greater than or equal to x. be experimented with to find the right balance to allow
SD refers to the standard deviation. interesting features to emerge from the data. Even if no
IQR refers to the interquartile range. histograms are included in a final publication, they are a
quick and indispensable tool to help researchers catch
potential problems in the data and reveal interesting
Figure 3A shows an undersmoothed histogram of the data, features.
with 90 bins each covering 2 mm Hg. We can see what
appears to be bimodality (ie, 2 modes, or clusters of data)
in the sample, with one peak around 70 mm Hg and References
another around 110 mm Hg. The high resolution makes it
difficult to determine the overall shape and height of each 1. Beniger JR, Robyn DL. Quantitative graphics in statistics: a brief his-
peak, however. In Figure 3B the same data is plotted in an tory. Am Stat. 1978;32(1):1-11.
2. Nuzzo RL. The box plots alternative for visualizing quantitative data.
overly smoothed histogram with only 7 bins, each one cov-
PM R. 2016;8(3):268-272.
ering 30 mm Hg. In this version we have lost interesting 3. Knaus WA, Harrell FE, Lynn J, et al. The SUPPORT prognostic model:
and important data features, such as the bimodality. objective estimates of survival for seriously ill hospitalized adults.
There is no magic formula for determining the perfect Ann Intern Med. 1995;122:191-203.
bin width, but over the years statisticians have developed 4. De Waard F, Baanders-Van Halewijn EA, Huizinga J. The bimodal age
rules of thumb to help guide researchers. See Box for distribution of patients with mammary carcinoma. Evidence for the
existence of 2 types of human breast cancer. Cancer. 1964;17(2):
common formulas for bin width and the corresponding 141-151.
number of bins.5–8 The resulting estimates for the mean 5. Sturges HA. The choice of a class interval. J Am Stat Assoc. 1926;21
SBP example in Figure 3 are also given. In this case, there (153):65-66.
are three approaches that are close in their recommenda- 6. Terrell GR, Scott DW. Oversmoothed nonparametric density esti-
mates. J Am Stat Assoc. 1985;80(389):209-214.
tions: Rice at 20, Scott at 19, and Freedman/Diaconis at
7. Scott DW. On optimal and data-based histograms. Biometrika. 1979;
21 for the number of bins. There was not much visual dif- 66(3):605-610.
ference between histograms with each of these bin num- 8. Freedman D, Diaconis P. On the histogram as a density estimator: L
bers (not shown). 2 theory. Probability theory and related fields. 1981;57(4):453-476.

Disclosure

R.L.N. Department of Science, Technology, and Mathematics, Department of Submitted for publication February 8, 2019; accepted February 11, 2019.
Psychology, HMB S340F Gallaudet University, Washington, DC, NE.
Address correspondence to: R.L.N.; e-mail: regina.nuzzo@gallaudet.edu
Disclosure: nothing to disclose