Understand how to interpret histograms in quality and assess distribution, dispersion, skewness, outliers, and multimodality in Engineering processes.
Check it out!
A histogram is a frequency-distribution graph used to show how the values of a numerical variable are distributed. Instead of looking only at the mean, minimum, and maximum, it makes it possible to see concentration, dispersion, skewness, tails, outliers, and the possible presence of more than one pattern in the same sample. In Engineering, this helps understand variability in measurements, tests, process times, document quality, manufacturing, inspections, and operational performance.
A histogram does not preserve the time sequence of the data. For that reason, it answers the question “how are the values distributed?” very well, but by itself does not answer “when did the process change?”. This distinction is fundamental to avoid using the graph out of context: distribution and stability are related but different problems.
What Is a Histogram?
A histogram divides the data range into classes — also called bins — and counts how many observations fall into each class. The horizontal axis represents intervals of the variable; the vertical axis represents frequency, relative frequency, or density, depending on the chosen construction.
Unlike a table, the histogram makes the shape of the distribution visible. This makes it possible to assess quickly whether the data are concentrated around a value, spread over a wide range, shifted to one side, divided into two groups, or affected by unusual observations.
The tool is part of the classical set of Quality Tools in Engineering and becomes more valuable when used together with structured data collection, stratification, time analysis, and cause investigation.
Histogram vs. Bar Chart
The two graphs may look similar, but they represent different things.
In a histogram, the bars represent intervals of a continuous or quantitative numerical variable. The order of the intervals is natural and should not be rearranged by preference. In general, the bars are contiguous because the intervals are contiguous as well.
In a bar chart, each bar represents a category: discipline, supplier, failure type, system, area, or another nominal class. Categories can often be reordered without changing the statistical meaning.
| Characteristic | Histogram | Bar chart |
| Main data type | Numerical | Categorical |
| Horizontal axis | Value intervals | Categories |
| Natural order | Yes | Not always |
| Contiguous bars | Usually yes | Usually no |
| Shows distribution shape | Yes | No |
| Useful for Pareto | Not directly | Yes, when categories are ordered |
This distinction avoids a common error: calling any graph with bars a histogram.
What Can a Histogram Reveal?
NIST highlights five characteristics that are especially useful when reading distributions: center, spread, skewness, outliers, and multiple modes. In Engineering practice, each can point to different issues.
Center of the distribution
Shows where values are concentrated. It may be near the mean, median, or another typical point.
If a measurement process should produce values close to a nominal value, the position of the histogram helps reveal overall shifts.
Dispersion
Shows how widely values are spread. A very wide distribution may indicate excessive variability even when the mean is correct.
Skewness
A long tail on one side may indicate physical constraints, process behavior, mixed populations, or the occurrence of less frequent and more extreme events.
Outliers
Isolated values may be measurement errors, incorrect records, special conditions, or legitimate results. The histogram signals a need for investigation but does not determine the cause.
Multimodality
Two or more peaks may indicate that the sample combines different processes: shifts, suppliers, equipment, teams, disciplines, methods, or operating regimes.
When the distribution reveals tails, multiple peaks, or excessive dispersion, the objective is not merely to explain the graph: it is to identify which process sources are mixed together and what evidence must be collected to distinguish common variation from a structural problem.
How to Choose the Number of Histogram Bins
The number and width of bins strongly influence the appearance of the graph. Too few bins hide structure; too many can turn noise into apparent complexity.
There is no single universal number. Statistical rules exist to estimate bin width or number of bins, but judgment should consider sample size, measurement resolution, and the purpose of the analysis.
Some practical principles are:
- do not use bins so wide that they eliminate relevant characteristics;
- do not use bins so narrow that almost every bar contains only a few points;
- maintain consistent widths unless a specific methodology requires otherwise;
- ensure bin boundaries are clear and do not create ambiguity;
- test visual robustness: if small changes in binning completely change the story, the sample may be insufficient or the interpretation fragile.
A histogram is a representation of the data, not the data themselves. Its shape needs to be analyzed together with descriptive statistics and process context.
Sample Size: When Does a Histogram Become Useful?
With very little data, the shape of the distribution is unstable. A histogram of ten observations can be drawn, but it will hardly support strong conclusions about skewness or multimodality.
The required amount depends on variability, the question being asked, and the expected complexity. The more detail of the distribution one intends to interpret, the larger the data set should be.
In Engineering processes, there is often tension between statistical depth and actual data availability. In this situation, the correct approach is to state the limitation and use the graph for exploration rather than as conclusive proof.
How to Build a Histogram Correctly
1. Define the variable
The variable must represent exactly the phenomenon to be analyzed. “Design time” may be too broad. “Time between submission and return of the technical review for Class A documents” is much more specific.
2. Validate unit and measurement method
Mixing minutes and hours, different measurement methods, or changed criteria produces a meaningless distribution.
3. Check the source population
Before aggregating, ask whether the data really belong to the same process.
4. Choose appropriate bins
Define clear and consistent intervals.
5. Count frequencies
Each observation must fall into one and only one bin.
6. Plot and interpret
Observe shape, center, dispersion, tails, gaps, peaks, and extreme values.
7. Stratify when suspicions arise
If bimodality or excessive width appears, separate the data by relevant factors and compare.
Example: Document Review Lead Time
Imagine 120 technical reviews with lead times between 0.5 and 12 days. The overall mean is 4.2 days.
By itself, the mean does not reveal that the histogram has two peaks: one near two days and another near seven days. After stratification, the first group is found to correspond to documents reviewed by a dedicated team, while the second depends on an additional approval.
In this case, the wide distribution does not represent merely a “variable process.” It combines two different workflows.
The analysis can then advance to End-to-End Engineering Processes or Bottlenecks in Engineering Processes, depending on the operating cause.
Example: Manufactured Component Dimension
Consider a nominal dimension of 50 mm with a technical tolerance defined on the drawing.
The histogram shows most measurements between 49.7 and 50.2 mm, but with an elongated tail extending to 50.8 mm. The overall mean may remain close to 50 mm, but the tail suggests behavior that deserves investigation.
Possible questions include:
- is there progressive tool wear?
- is a particular raw-material lot associated with the higher values?
- is there a difference between shifts?
- is the measurement system stable?
- does the tail correspond to a specific piece of equipment?
The histogram guides the questions, but the answers require evidence.
Histogram and Specification: Be Careful with Capability Conclusions
It is common to overlay specification limits on the histogram and visually assess whether the distribution “fits” within the tolerance. This is useful as an initial reading but does not replace a formal capability analysis.
Assessing capability requires consideration of stability, distribution, variability estimates, and appropriate indices. If the process is not stable, an aggregated snapshot of the distribution can create a false sense of predictability.
For that reason, the methodological sequence is usually:
- confirm the measurement system and data quality;
- assess process stability;
- analyze the distribution;
- only then discuss capability relative to the specification.
Histogram vs. Control Chart
The histogram removes time order to emphasize the distribution. The Control Chart preserves time order to identify changes and special signals.
Consider a process that operated for three months with a mean of 20 and then permanently shifted to a mean of 25. A histogram of the entire period may show two peaks. The control chart can indicate precisely when the level changed.
Therefore:
- use a histogram for shape and distribution;
- use a control chart for stability and change over time.
The two views together are more powerful.
Histogram vs. Pareto
The Pareto Chart works mainly with categories ordered by frequency, impact, or another metric. A histogram works with intervals of a numerical variable.
If the question is “which causes of returns are most frequent?”, Pareto is appropriate. If the question is “how is return time distributed?”, a histogram is more suitable.
A mature analysis can use both: Pareto to choose the critical category and a histogram to understand quantitative behavior within that category.
Histogram and Check Sheet
Histogram quality begins with data collection. If occurrences are recorded inconsistently, no graph will correct the problem.
A check sheet can be structured to record events, measurements, categories, and minimum context. The collected data can then feed histograms, Pareto charts, control charts, or other analyses.
The principle is simple: sophisticated analysis of poor data produces a sophisticatedly wrong conclusion.
How to Interpret Common Histogram Shapes
Approximately symmetric and unimodal distribution
The data concentrate around one main peak and spread similarly to both sides. This does not prove normality, but it may be compatible with a relatively homogeneous process.
Right skew
Values are concentrated at the lower end with a tail toward higher values. This is common for waiting times, lead times, and durations because there is a lower bound near zero and exceptional events can extend the tail.
Left skew
This may occur when there is a physical upper limit or a process concentrated near a ceiling.
Bimodality
Two peaks suggest mixed populations or two process regimes. Examples include two pieces of equipment, two suppliers, or two methods.
Very wide distribution
This may indicate high variability, mixed sources, or an inadequate measurement system.
Narrow peak with extreme values
This may represent a normally consistent process with isolated special events — or a recording error. Investigation must determine which.
Stratification: the Main Response to Confusing Histograms
When the distribution shows multiple peaks, unexpected tails, or high dispersion, the next step is usually to stratify.
In Engineering, stratification criteria may include:
- discipline;
- supplier;
- equipment;
- shift;
- team;
- project phase;
- document class;
- location or work front;
- software version;
- operating condition;
- material type;
- criticality.
The objective is to discover whether the observed variation comes from a single process or from a combination of different processes.
If the histogram changes radically when data are separated by supplier, discipline, equipment, or phase, management must treat those differences as part of the process — not hide them in a global average.
Histograms in Inspection and QA/QC
In inspections, a histogram is useful when the results are quantitative. Examples include:
- coating thickness;
- insulation resistance;
- measured torque;
- geometric dimension;
- response time;
- temperature;
- vibration;
- pressure;
- voltage drop;
- values obtained in repetitive tests.
It helps visualize dispersion and pattern but does not replace acceptance criteria. An item may fall within the predominant distribution and still fail to meet the technical requirement.
The article Quality Inspection addresses how criteria, sampling, records, and acceptance should be structured within the broader QA/QC context.
Histograms in Engineering Processes
Applications in a project office include:
- distribution of review lead time;
- distribution of comments per document;
- number of review cycles;
- time to close an RFI;
- duration of supplier analysis;
- mobilization time;
- rework cost per occurrence;
- Engineering hours by deliverable class.
These analyses may reveal that the “process average” is not very representative because there is strong skewness or more than one population.
Histograms in DMAIC
In DMAIC, the histogram is especially useful in Measure and Analyze.
In Measure, it helps characterize the baseline and dispersion. In Analyze, it allows comparison of stratified distributions, observation of shifts, and verification of whether specific factors are associated with different patterns.
After improvement, “before vs. after” histograms can show changes in center and dispersion. However, sustainability verification should also consider the time dimension — for example, with a control chart.
Common Errors in Construction and Interpretation
Choosing bins to produce the desired shape
Manipulating intervals until the graph “looks normal” or “looks bimodal” is a form of analytical bias.
Ignoring time
Mixing periods before and after a structural change can generate a misleading distribution.
Aggregating different processes
Mixing suppliers, equipment, or complexity classes can create artificial multimodality.
Interpreting normality only visually
An appearance similar to a normal curve is not statistical proof of normality.
Inferring causality from the shape
The histogram shows a pattern, not a cause.
Using excessively rounded data
Low resolution can create artificial bars or hide the true dispersion.
Confusing frequency with importance
A frequent event may have low impact, while a rare event may be critical. Priority should consider risk and consequence when necessary.
How to Compare Histograms Before and After an Improvement
Comparing distributions requires ensuring that the data sets are comparable.
Check:
- the same variable definition;
- the same unit;
- equivalent measurement methods;
- similar populations;
- appropriate periods;
- reasonable sample size;
- consistent bins or another appropriate visualization.
An improvement may shift the mean without reducing variability, reduce variability without changing the mean, or affect only one tail. Interpretation must go beyond “the graph looks better.”
How to Connect the Histogram to Cause Analysis
The histogram can direct an investigation.
If there is bimodality, ask which factor separates the two populations. If there is a long tail, look for events associated with the extremes. If there is high dispersion, investigate sources of variability.
Tools such as Ishikawa and 5 Whys help structure hypotheses, but validation must return to data and evidence.
Histograms and Executive Indicators
A monthly indicator usually summarizes the process into a single number. The histogram shows what that number hides.
For example, an average SLA of three days may look acceptable. However, the distribution may show that half the cases close in one day while 10% take more than ten days. For capacity and customer-experience decisions, the average alone is insufficient.
Combining percentiles, distribution, and recurring indicators improves performance interpretation and reduces conclusions based on simplified aggregates.
When a Histogram Is Not the Best Tool
It is not the primary tool when:
- the variable is categorical;
- there are very few observations;
- the main question is temporal;
- the focus is the relationship between two variables;
- the objective is to prioritize categories;
- the population is changing so quickly that the aggregated distribution does not represent the process.
In these cases, Pareto, a control chart, scatter plot, boxplot, time series, or another analysis may be better.
How to Turn Histogram Interpretation into a Decision
A management routine can follow four levels:
- describe: how the data are distributed;
- stratify: which sources explain the observed shape;
- investigate: which causal mechanisms are supported by evidence;
- act and verify: what change will be implemented and how its effectiveness will be demonstrated.
The histogram should not end in a presentation. It should generate a better question and guide the next analysis.
A histogram is most useful when it generates a decision: stratify, investigate a tail, review the measurement system, or redesign a step. The analysis should end in verifiable action, not an isolated statistical presentation.
Final Considerations
A histogram is visually simple but demands analytical discipline. Its value lies in revealing structure that the mean and overall standard deviation can hide: skewness, dispersion, multiple populations, and extreme values.
In Engineering, it works best when data collection is reliable, the population is well defined, and interpretation is combined with stratification, control charts, Pareto, cause analysis, and technical criteria. The correct question is not only “what is the shape of the graph?”, but “what process characteristic could explain this shape, and what evidence should we seek next?”.
Technical references
[1] NATIONAL INSTITUTE OF STANDARDS AND TECHNOLOGY (NIST). NIST/SEMATECH e-Handbook of Statistical Methods: Histogram. Available at: https://www.itl.nist.gov/div898/handbook/eda/section3/histogra.htm
[2] AMERICAN SOCIETY FOR QUALITY (ASQ). What is a Histogram? Available at: https://asq.org/quality-resources/histogram
[3] AMERICAN SOCIETY FOR QUALITY (ASQ). Statistical Process Control (SPC). Available at: https://asq.org/quality-resources/statistical-process-control
Frequently asked questions
It is a frequency-distribution graph that divides the values of a numerical variable into intervals and shows how many observations occur in each range.
A histogram represents contiguous intervals of a numerical variable. A bar chart represents categories. Therefore, the order and meaning of the bars are different.
Not by itself. It shows the aggregated distribution but loses time order. A control chart is more appropriate for assessing stability and changes over time.
Not necessarily, but two peaks suggest the sample may contain two different populations or regimes. The next step is to stratify by factors such as supplier, equipment, team, or phase.
No. A histogram can support an initial reading, but capability requires consideration of stability, distribution, variability, and appropriate indices relative to specification limits.
There is no universal number. The choice depends on sample size, data resolution, and the objective. Too many bins show noise; too few hide structure.
Complementary technical materials
Related solutions
- Indicators, Dashboards, and Executive Reports
- Process Management, Workflows, and Technical Approvals
- Requirements, Evidence, and Acceptance Criteria Management
Related services
- Engineering Process Diagnosis and Optimization
- Technical Engineering Consulting
- Ongoing Engineering Consulting Services
