MCAT content library / Research & data / Confidence intervals, error bars, and sample size
Identify what the bars represent, which units are independent, and which interval answers the comparison.
About 16 minutes, including the worked examples and checks.
From a dataset to a supported claim · Lesson 4 of 6 · Continue in this path
Useful foundations
On this page
Concept: an error bar needs a definition
Two figures show the same means, but one uses broad bars and the other narrow bars. The figures need not disagree. A bar can describe the spread of individual observations, uncertainty in an estimated mean, or something else entirely. Before comparing its length or overlap, read the legend and identify the quantity it represents.1
Standard deviation describes dispersion of observations. The standard error of the mean describes sampling variability of an estimated mean; under the usual independent, finite-variance model its estimate is s/√n. A confidence interval uses a specified procedure to express uncertainty about a population parameter. These are related quantities, but they answer different questions.2
None is an automatic measure of the measurement system’s bias. A precise mean can still be displaced by an uncorrected systematic error. A narrow interval produced from an inappropriate sampling or analysis model can be misleading even when its arithmetic is correct.
Pattern: redraw the same data three ways
Consider two independent groups of 64 experimental units each. In this original illustrative dataset, their means are 40 and 43 units and both sample SDs are 8 units. Assume the observations and design justify the stated large-sample normal approximation. For the approximate 95% intervals in this example, use the supplied multiplier 1.96; other methods and small-sample settings may require different multipliers.
For either group, SE = 8/√64 = 1 unit. An SD bar therefore extends 8 units from the mean, an SE bar extends 1 unit, and the stated approximate 95% interval extends 1.96 units. The mean stays in the same place in every panel.
| Quantity shown | Group A (mean 40) | Group B (mean 43) |
|---|---|---|
| Mean ±1 SD | 32 to 48 | 35 to 51 |
| Mean ±1 SE | 39 to 41 | 42 to 44 |
| Approximate 95% CI | 38.04 to 41.96 | 41.04 to 44.96 |
The SD bars overlap widely, while the SE bars do not overlap. The approximate 95% intervals overlap slightly. Choosing a different bar has not changed the observations, the estimated 3-unit difference, or the research design. It has changed the information encoded by the bar.
Rule: interpret the interval for the parameter being compared
For a conventional frequentist 95% confidence procedure, the long-run coverage refers to intervals produced by repeating the specified sampling and analysis process under its assumptions. About 95% of those intervals would contain the fixed parameter. A particular computed interval either contains that parameter or does not. It is not a statement that 95% of individual observations fall within the interval.2
If the question concerns a difference between group means, examine an interval for that difference. Overlap of two separate group intervals is not a universal significance test. The uncertainty of a difference depends on the groups’ variability, sample sizes, and whether the observations are independent or paired.3
For independent group means in this example, SEdifference = √(SEA² + SEB²) = √2 ≈ 1.414 units. The stated approximate 95% interval for B − A is 3 ±1.96√2, or about 0.23 to 5.77 units. It excludes zero even though the separate group intervals overlap. This concrete calculation is why a visual overlap shortcut can give the wrong conclusion.
The interval concerns an estimated mean difference under the supplied model. It does not establish a cause, make the difference important for every use, or show that every individual in B exceeds every individual in A. Those claims require different evidence or a different parameter.
Application: change the sample size without changing the spread
Suppose another sample has n = 16 and s = 8. Its estimated SE is 8/4 = 2. A sample of 64 with the same SD has SE = 1. Quadrupling the number of independent observations halves this standard error; simply doubling n reduces it by a factor of √2, not by a factor of 2.
This is a comparison holding the observed SD fixed. Real new samples can have different SDs. Also, a small-sample t interval uses a multiplier that depends on degrees of freedom, so its width is not determined by √n alone. If a problem supplies a multiplier, use that value rather than silently replacing it with 1.96 or 2.
Increasing independent sample size improves the potential precision of the mean estimate. It does not force the underlying population to become less variable. If individual response differences are the scientific question, the SD and the distribution remain relevant even when a large sample produces a narrow interval for the mean.
Count the independent units
A researcher takes samples from 12 independently selected participants and measures each sample three times. Those 36 readings do not represent 36 independently selected participants. Averaging the technical replicates can improve each participant’s measurement, but a population inference about people must respect the 12-participant sampling structure.
That distinction depends on the design and the intended inference. Measurements within one person, culture, classroom, or experimental run can share influences. The analysis needs to account for that structure rather than treating the number of entries in a spreadsheet as the sample size.1
Similarly, repeated before-and-after observations from the same participants are paired. Their difference may be estimated more or less precisely than a calculation pretending the two sets are independent. Inspect the design and use the stated paired-difference uncertainty; do not insert an independent-groups formula merely because two means appear in a table.
Within the AAMC’s data-reasoning scope4, a useful figure-reading routine has four steps: identify the estimated quantity, identify the bar definition, identify the independent unit and comparison, and then interpret the interval actually relevant to that comparison. It allows the display to answer its intended question without making a broader claim than the data support.
ACTIVE PRACTICE · DATA REASONING
Change Sample Size, Not the Observed Spread
Keep the worked example’s group means at 40 and 43 units and the observed standard deviation at 8 units for both independent groups. Use an approximate normal 95% multiplier of 1.96.
Standard error per group: . Approximate 95% interval A:; interval B:.
For B − A, approximate 95% interval: .
The intervals describe uncertainty in group means and their difference, not the spread of individuals. Overlap between separate intervals is not a substitute for calculating the interval for B − A. Sample size represents independent units, not technical replicates.
Check your reasoning
Choose an answer before opening its feedback. Each option explains the reasoning behind that choice.
Check 1
An estimated mean difference is +2 units, with a stated 95% confidence interval from −1 to +5. The method’s assumptions are satisfied. Which interpretation is best supported?
A. A fixed population difference has a 95% probability of being between −1 and +5.
Review this choice. The confidence level describes the coverage of the procedure under repeated sampling; it is not a probability assigned to the fixed parameter after this interval has been computed.
B. The data rule out any positive population difference.
Review this choice. Values above zero remain inside the interval, including the point estimate of +2. The result is not evidence that every positive effect is excluded.
C. The interval includes zero and also includes both negative and positive differences.
Correct. The reported range contains zero, negative values and positive values. Its breadth leaves uncertainty about the direction; it does not establish exact equality.
D. The interval shows that 95% of individual differences fall between −1 and +5.
Review this choice. This is an interval for a mean difference, not a range designed to contain 95% of individual observations.
The principle to carry forward
Name the parameter and inspect its interval before deciding which differences are excluded.
Check 2
Study A measures 25 independent units and Study B measures 100 independent units. Both report a sample SD of 10 units, and the usual s/√n calculation applies. What is the ratio SEB/SEA?
A. 0.25
Review this choice. The ratio of sample sizes is 1/4, but standard error depends on their square roots.
B. 0.50
Correct. SE_A is 10/5 = 2 and SE_B is 10/10 = 1, so the ratio is 1/2. Four times as many independent units halves SE when SD is held fixed.
C. 1.00
Review this choice. Equal SDs do not imply equal standard errors when the sample sizes differ.
D. 2.00
Review this choice. This reverses the comparison. Study B has the larger sample and the smaller estimated standard error.
The principle to carry forward
Use the square-root dependence on independent sample size; separate it from the spread of individual observations.
Continue learning
From a dataset to a supported claim
Lesson 4 of 6.
- Previous in From a dataset to a supported claim: Accuracy, precision, and sources of measurement error
- Next in From a dataset to a supported claim: Statistical significance and effect size: what the result means
- Return to this path’s overview
See every step in From a dataset to a supported claim
From a dataset to a supported claim
Interpret variability and uncertainty before combining evidence from multiple figures.
- Reading graphs: axes, scales, and the comparison that matters
- Mean, median, and spread: describing a dataset
- Accuracy, precision, and sources of measurement error
- Confidence intervals, error bars, and sample size · You are here
- Statistical significance and effect size: what the result means
- Reading a multi-figure experiment from claim to conclusion
Related lessons
- Mean, median, and spread: describing a dataset
- Accuracy, precision, and sources of measurement error
- Statistical significance and effect size: what the result means
- Reading a multi-figure experiment from claim to conclusion
Browse research & data · MCAT content library
MCAT preparation and roadmap · Aim High, Score Higher
Sources
1. Cumming, Fidler and Vaux: Error bars in experimental biology
2. NIST/SEMATECH: Confidence Limits for the Mean