The smallest bars might come from the same data

Imagine two figures on your screen. Both plot an average of 60. In one, the bars reach a dozen units above and below the point. In the other, they extend just four. It looks as though the second experiment was much tidier, until you read the legends and find that both graphs summarize the same observations.

Whenever you see error bars, find the sentence that defines them. It should name the quantity, such as SD, SEM or 95% CI, and say which observations the calculation used. The bar’s length is only interpretable once you have those two details.

Spread, precision and intervals

Standard deviation (SD) describes how spread out the observations are around their mean. It can help you picture whether the values sit close together or tend to stray from the center. Individual observations can give you a clearer view of clusters, gaps or unusual points that a single spread measure cannot display.

Standard error of the mean (SEM) describes the estimated precision of the average. In a simple calculation with independent observations it is the sample SD divided by the square root of the sample size. If that size increases while the observed spread stays similar, SEM gets smaller. You can estimate the average more precisely while still seeing substantial variation between observations.

A confidence interval (CI) adds a stated confidence level and a method for constructing an interval around an estimate. For a frequentist 95% interval, the procedure would cover the underlying parameter in about 95% of repeated samples under the model assumptions. Read the endpoints in the units of the estimate. They describe uncertainty about that estimate, such as a group mean or a difference.

Work through one fictional set of summary numbers

Suppose nine independent observations have a mean of 60 response units and a sample SD of 12. One SD on either side of the mean gives bars from 48 to 72. Dividing 12 by the square root of nine gives a SEM of 4, so mean plus or minus one SEM gives bars from 56 to 64.

For these fictional numbers, a usual two-sided 95% t interval for the mean is about 50.8 to 69.2, assuming the conditions for that interval are appropriate. The center is 60 in all three displays. The endpoints move because the bars describe different quantities.

When you encounter a figure that only says “mean ± error,” write down what is missing. The width alone cannot tell you which formula was used. A very narrow bar offers little help if you are trying to judge how much individual samples varied and you cannot find the underlying observations.

Find the difference the experiment actually estimates

It is tempting to turn overlapping bars into a yes-or-no test. Start instead with the comparison you care about. An interval around the difference can address that question more directly than two intervals around separate means. Overlap depends on what the bars represent and how the estimates are related.

Consider a fictional paired comparison. Four cell preparations give readings of 48, 54, 66 and 72 in one condition. Their matched readings in a second condition are 49, 56, 67 and 74. The preparations differ substantially from each other. The changes within preparations are 1, 2, 1 and 2 units. You can see the pairing immediately if the plot draws a line between the readings from each preparation.

Weissgerber and colleagues show how different relationships can disappear behind the same summary graph. Cumming and Finch explain why the design matters when interpreting interval overlap. For paired data, look for an estimate of the within-pair change and its uncertainty. The fictional readings illustrate the distinction and are too limited for any general conclusion about a peptide.

A ribbon around a fitted curve needs its own caption

In a concentration-response plot, bars on measured points and a shaded band around a fitted line can refer to different quantities. Read what the band surrounds. It might describe uncertainty in the fitted mean response or a range for future observations under a model. The width near one concentration does not tell you the uncertainty of the fitted EC50 by itself.

Keep the reported parameter and its interval together in your notes. If you are comparing midpoints, look for uncertainty on the midpoints or on their contrast. If you are comparing responses at one concentration, look for evidence about that response. Those are separate reading jobs even when the same plot is open in front of you.

Put the result and its limits in the same sentence

Try describing the figure to someone who cannot see it. Give the estimate, its units and the endpoints of the interval. Say whether those endpoints belong to a mean, a difference or a fitted value. For example, in the fictional nine-observation display, you could describe the mean as 60 response units and identify the interval around that mean as 50.8 to 69.2 under the stated assumptions.

Then ask whether that is the estimate the article’s headline is about. A panel about the uncertainty of an average may say little about how similarly individual samples responded. A curve panel that summarizes responses at several concentrations may leave a comparison of fitted parameters to a table. Keep looking until you find the analysis that matches the claim.

Frequently asked questions

Do shorter SEM bars mean the observations varied less?

Look for SD or the individual observations if you want to know the spread. SEM also depends on the number and structure of the observations used to estimate the mean.

Where should I look for uncertainty when the samples are paired?

Look for the paired differences and an interval around the estimated change. A connected plot also shows which two measurements belong to the same preparation.

If two confidence intervals overlap, can the estimated difference still be statistically distinguishable from zero?

Yes. Separate intervals describe the individual estimates. Interpret the interval or test for the difference using the actual study design, especially when observations are paired or otherwise related.

Sources

  1. Basic Guidelines for Reporting Non-Clinical DataDahlin and colleagues, NIH/NCATS Assay Guidance Manual · 2019
  2. Error bars in experimental biologyCumming, Fidler and Vaux, Journal of Cell Biology · 2007 · DOI: 10.1083/jcb.200611141
  3. Inference by eye: confidence intervals and how to read pictures of dataCumming and Finch, American Psychologist · 2005 · DOI: 10.1037/0003-066X.60.2.170
  4. Beyond Bar and Line Graphs: Time for a New Data Presentation ParadigmWeissgerber and colleagues, PLOS Biology · 2015 · DOI: 10.1371/journal.pbio.1002128

Published by EliteGen Research for research education. Studies describe their own materials and experimental conditions; they do not establish the identity, quality, or suitability of an EliteGen product. This article does not provide guidance for personal, human, or veterinary use.