If Your Goal Is to Show Variation, Use a Scatter Plot
If you want readers to see how much variation exists among your individual data points, do not use error bars at all. Plot the actual data points as a scatter plot. This is more honest and more informative because:
- Readers can see whether the distribution is symmetric, skewed, or bimodal — something summary bars hide.
- With small sample sizes (N < 10), each point carries meaningful information that SD conceals.
- A SD error bar makes it look like you have a large, normally distributed dataset even when you have only 5 values.
If Your Goal Is to Show Precision of the Mean, Use 95% CI
SEM is often misinterpreted. Readers frequently treat SEM bars as if they were some kind of confidence interval, but they are not. The SEM is the standard deviation of the sampling distribution of the mean — it tells you about the precision of your mean estimate, but it answers a question most readers are not actually asking.
A 95% confidence interval is almost always more useful because:
- It has a direct, intuitive interpretation: you are 95% confident the true mean lies within the interval.
- It lets readers visually judge whether two groups differ significantly (non-overlapping 95% CIs suggest P < 0.05, though this is not exact).
- Journals increasingly require CIs over SEMs.
The Problem with SEM Error Bars
SEM bars are smaller than SD bars by a factor of the square root of N. This means:
The more data you collect, the smaller your SEM bars become, even if the biological variability is unchanged. SEM bars will always shrink as N increases, which can make data look more precise than it really is and can be used (intentionally or not) to make graphs look cleaner than the underlying variation warrants.
Recommendation by Goal
| Your goal | Recommended | Avoid |
|---|---|---|
| Show the spread of individual values | Scatter plot (show all points) | SD or SEM |
| Show the precision of the mean estimate | 95% CI error bars | SEM |
| Compare means visually for significance | 95% CI error bars + P value annotation | SD |
| Communicate raw data for a small N | Show all individual data points | Any error bars |
When SD Is Appropriate
SD error bars make the most sense when the variation itself is the scientific message. For example, if you are characterizing how variable a biological measurement is across a population, SD bars appropriately describe that population spread. But even then, showing all individual points alongside the mean and SD is preferable when N is small.