A measurement is more than a number
Suppose you measure the length of a table and report 120 centimetres. That number is useful, but it does not say whether you measured to the nearest centimetre or used a tool precise enough to distinguish millimetres. Measurement uncertainty describes the range of values reasonably compatible with a measurement, given the method and its limitations. It is not an admission that the measurement was careless; it is part of reporting what the measurement can tell us.
In a graph, error bars are lines extending from plotted points or columns. They make some kind of spread or uncertainty visible. The crucial phrase is “some kind”: error bars do not have one universal meaning. Before drawing conclusions, look for a caption or methods description that says what they represent and how they were calculated.
Three common meanings of error bars
A graph might show the standard deviation of a set of observations. This describes how much individual observations tend to vary around their average. If several people measure their own walking speeds, standard-deviation bars could convey how different those speeds are. Adding more people may help describe the group better, but it does not necessarily make the people’s speeds less varied.
Another graph might show the standard error of an estimated average. This concerns the precision of the average rather than the spread of individual observations. It generally becomes smaller as more independent observations are collected, even if those observations remain just as varied. Confusing standard error with standard deviation can make a dataset look more uniform than it is.
A third possibility is a confidence interval, often labelled with a confidence level such as 95%. It is calculated by a specified procedure intended to capture the underlying value at that rate over repeated uses of the procedure. It is not a guarantee that the plotted interval contains the underlying value, nor does it mean that 95% of individual observations fall inside it. These distinctions are why an unlabelled error bar is difficult to interpret.
What overlap can and cannot tell you
When two groups appear on the same graph, it is tempting to treat overlapping error bars as proof that there is no difference, or non-overlapping bars as proof of an important discovery. Neither rule works in every situation. The interpretation depends on what the bars represent, how many observations were made, and whether the measurements are independent or paired. A formal comparison may estimate the uncertainty of the difference directly, rather than compare two separate sets of bars by eye.
Even a difference that is clear relative to measurement uncertainty may be small in practical terms. Imagine two thermometers that consistently give slightly different average readings. Whether the difference matters depends on the question: checking room comfort and monitoring a temperature-sensitive experiment demand different levels of precision. Read the axis scale and the size of the effect as well as the error bars.
Uncertainty does not capture every possible error
Repeated measurements can differ because of small, unpredictable influences: where a ruler is placed, how a sensor responds, or ordinary changes in the thing being measured. This is often called random variation. Taking repeated, independent measurements and reporting their spread can help reveal it. But repetition alone cannot fix a ruler whose markings are misplaced or a sensor that consistently reads high. Those are examples of systematic error, which shifts results in a particular direction.
Error bars calculated from repeated readings may look reassuringly narrow even when a shared bias remains. Researchers can address possible biases by checking calibration, comparing methods, measuring a known reference, or changing parts of a procedure. No single bar on a chart necessarily includes all these concerns. A useful question is: which sources of uncertainty were counted, and which were assessed separately?
A practical checklist for reading a graph
Start by identifying what each plotted point represents: one measurement, an average, or a model estimate. Then find the definition of the error bars. Check the sample size, the units, and whether the horizontal or vertical axis begins at zero; a narrow-looking bar can take on a different visual meaning when the scale changes. If measurements were taken from the same subjects or objects before and after a change, look for information about those paired differences.
Finally, ask whether the graph shows variation among observations, uncertainty in an estimate, or both. Consider plausible systematic errors and whether the size of any difference matters for the question being asked. Error bars are most helpful when treated as a guide to better questions, not as a stamp declaring a result certain or uncertain.
