What's the Difference Between a t-Distribution and a Normal Distribution?
The t-distribution and the normal distribution are both symmetric, bell-shaped probability distributions centered on a mean of zero in their standard forms. The main difference is in the tails: the t-distribution has heavier, fatter tails and a slightly lower peak. The normal distribution is used when the population standard deviation (σ) is known. The t-distribution is used when σ is unknown and must be estimated from a sample (s) — an extra source of uncertainty that its heavier tails account for. A second parameter, degrees of freedom (df), controls exactly how heavy those tails are; as df increases, the t-distribution converges to the normal distribution.
The t-distribution was developed in 1908 by William Sealy Gosset, a chemist working in the quality-control lab at the Guinness brewery in Dublin. Gosset needed a way to draw conclusions from very small batches — sometimes fewer than 10 samples — where the population standard deviation of barley or yeast quality was simply not known. Guinness did not allow staff to publish under their own names, so Gosset wrote up the result in the journal Biometrika under the pseudonym "Student," which is why it is still called Student's t-distribution today.
The normal distribution has a much older history. Abraham de Moivre described its mathematical form in 1733, and Carl Friedrich Gauss and Pierre-Simon Laplace developed it further in the early 1800s — hence its other name, the Gaussian distribution. Unlike the t-distribution, the normal distribution does not depend on sample size at all: a normal curve with μ = 100 and σ = 15 looks identical whether it describes 20 measurements or 2 million.
- Both distributions are symmetric and bell-shaped around their mean.
- The t-distribution has one extra parameter — degrees of freedom (df) — that the normal distribution does not have.
- For a one-sample test, df = n − 1. Lower df means heavier tails and larger critical values.
- By around df = 30, the t-distribution is nearly indistinguishable from the normal distribution.
- The normal distribution's shape never changes; the t-distribution's shape changes for every df.
t-Distribution vs Normal Distribution: Full Comparison
The table below lines up every major property side by side. The practical takeaway is in the last few rows: the t-distribution exists specifically to handle the case where you don't know σ and have to estimate it from your own data.
| Feature | t-Distribution | Normal Distribution |
|---|---|---|
| Shape | Symmetric, bell-shaped | Symmetric, bell-shaped |
| Parameters | Degrees of freedom (df) | Mean (μ) and standard deviation (σ) |
| Tails | Heavier — more probability in the extremes | Lighter |
| Peak | Slightly lower and flatter | Taller, more concentrated near the mean |
| Depends on sample size? | Yes, through df = n − 1 | No — the shape is fixed |
| Used when σ is | Unknown (estimated by s) | Known |
| Typical sample size | Often small (n < 30), but valid for any n | Any n, but the σ-known case is rare outside large established processes |
| As n increases | Converges toward the normal distribution | Shape stays the same |
| Critical value at 95% (two-tailed) | Larger than 1.96, depends on df | Fixed at ±1.96 |
| Named after | William Sealy Gosset ("Student"), 1908 | Carl Friedrich Gauss / Abraham de Moivre |
One row deserves a second look: because the t-distribution's critical value is always at least as large as the corresponding z critical value, any confidence interval or margin of error built with t will be at least as wide as the one built with z for the same data and confidence level. That width difference is the price of not knowing σ — and the worked example in Section 6 shows exactly how big that price is for a small sample.
When to Use a t-Distribution vs a Normal Distribution
The decision comes down to two questions: is the population standard deviation (σ) known, and if not, how large is your sample? The flowchart below follows the same logic used throughout the hypothesis testing section of this site.
📊 Distribution Selection Guide
The table below lays out the same logic from the dataset's point of view rather than the test's. Notice that even with a non-normal population, the answer for large samples is still "normal" — that's the Central Limit Theorem at work, not an assumption that the data itself is bell-shaped.
| Population | Sample Size (n) | Population σ | Correct Distribution |
|---|---|---|---|
| Normal | Small (n < 30) | Known | Normal (Z) |
| Normal | Small (n < 30) | Unknown | Student's t |
| Non-normal | Large (n ≥ 30) | Known | Normal (Z) |
| Non-normal | Large (n ≥ 30) | Unknown | Student's t (≈ Z for large df) |
t-Test vs z-Test: How the Distributions Translate to Hypothesis Tests
The t-distribution and normal distribution aren't just abstract curves — they're the reference distributions behind two of the most common hypothesis tests. The formulas below look almost identical; the only difference is which standard deviation goes in the denominator.
x̄ = sample mean
σ = known population SD
Reference: Normal distribution
s = sample SD
df = n − 1
Reference: t-distribution
| Feature | t-Test | z-Test |
|---|---|---|
| Population SD (σ) known? | No — estimated by s | Yes |
| Reference distribution | Student's t (df = n − 1) | Standard normal (Z) |
| Small-sample behavior | Accounts for extra uncertainty in s | Can understate uncertainty if σ is only assumed |
| Critical value at α = 0.05, two-tailed | Varies by df; e.g. ±2.228 at df = 10 | Fixed at ±1.960 |
| Where it converges | → z-test as n grows | n/a |
For fully worked z-test and t-test examples with step-by-step calculations, see the hypothesis testing examples guide. For the z-test specifically, including how to read the standard normal table, see the z-score guide.
Visualizing the Curves: Tails and Degrees of Freedom
The Geometry of the Tails
Both curves approach the horizontal axis without ever quite touching it — mathematicians call this "asymptotic." But for any given distance from the center, the t-distribution's curve sits higher above the axis than the normal curve does. That extra height in the tails is exactly the extra probability the t-distribution assigns to unusually large or small values, which is what makes its critical values larger than the corresponding z critical values.
How Degrees of Freedom Change the Shape
At df = 1, the t-distribution is extremely heavy-tailed — so heavy that its variance is technically undefined. At df = 5, the tails are noticeably fatter than the normal curve's, and the 95% critical value is still well above 2. By df = 30, the curve has narrowed enough that the 95% critical value (2.042) is barely distinguishable from the normal distribution's 1.960. The four numbers below show that progression at a 95% confidence level.
(95% confidence)
(95% confidence)
(95% confidence)
(95% confidence)
Worked Example: Same Data, Two Confidence Intervals
The example below uses one small dataset and builds a 95% confidence interval two different ways — once treating the standard deviation as known (the normal/Z approach), and once treating it as estimated from the sample (the correct t approach). The gap between the two results is the cost of the extra uncertainty discussed in Section 2.
Problem: A clinic measures the drop in systolic blood pressure for 10 patients eight weeks after starting a new medication. The sample mean drop is x̄ = 8.4 mmHg with sample standard deviation s = 3.2 mmHg. Build a 95% confidence interval for the true mean drop.
SE = s / √n
n = 10, df = 9
Standard error: SE = s/√n = 3.2/√10 = 3.2/3.162 = 1.012
If σ were known and equal to 3.2 (a hypothetical, since in practice it almost never is for a new measurement): use z* = 1.960 → CI = 8.4 ± 1.960 × 1.012 = 8.4 ± 1.98 → (6.42, 10.38), width = 3.97 mmHg
Since σ is actually unknown and s is only an estimate from 10 patients: df = n − 1 = 9, t* = 2.262 → CI = 8.4 ± 2.262 × 1.012 = 8.4 ± 2.29 → (6.11, 10.69), width = 4.58 mmHg
Compare: the t-based interval is about 0.61 mmHg wider on each side — roughly 15% wider overall — purely because of the uncertainty in estimating σ from a sample of 10.
✅ Conclusion: With n = 10 and σ unknown, the correct interval is (6.11, 10.69) mmHg, built with the t-distribution at df = 9. Reporting the z-based interval (6.42, 10.38) here would understate the true uncertainty by about 15% — a meaningful difference for a clinical decision.
Real-World Applications
The choice between t and z is not just a classroom exercise — it shapes how confidence intervals and test results are reported across several fields.
Clinical Trials
Early-phase drug trials often involve small cohorts (10–30 patients) where σ is unknown. The one-sample t-test and two-sample t-test are the default tools for comparing treatment effects.
Manufacturing QC
A production line that has run for years often has a well-established process σ from historical data. With σ known, quality engineers use the normal distribution and z-scores to flag out-of-spec batches.
A/B Testing
Conversion-rate tests with large traffic typically rely on the normal approximation through the Central Limit Theorem, while tests on continuous metrics (time on page, revenue per user) with smaller samples use a two-sample t-test.
Finance & Risk
Daily stock returns tend to have more extreme days than a normal distribution predicts. Some risk models replace the normal distribution with a t-distribution specifically because its heavier tails better match real market behavior.
Machine Learning
Comparing a model's accuracy across a handful of cross-validation folds is a classic small-n, σ-unknown situation — handled with a paired t-test.
Large Surveys
A national survey with thousands of respondents has df well into the hundreds, so its t critical values and z critical values are effectively identical — either distribution gives the same answer in practice.
Critical Values: z* vs t* Side-by-Side
The table below shows how much larger the t critical value is than the z critical value at the same confidence level, and how that gap shrinks as df increases.
| Confidence Level | z* (df = ∞) | t* (df = 4) | t* (df = 10) | t* (df = 29) |
|---|---|---|---|---|
| 90% | 1.645 | 2.132 | 1.812 | 1.699 |
| 95% | 1.960 | 2.776 | 2.228 | 2.045 |
| 99% | 2.576 | 4.604 | 3.169 | 2.756 |
Reading across any row, the t critical value gets smaller as df increases, moving steadily toward the fixed z value on the left. The table below tracks that convergence specifically at the 95% confidence level.
| Degrees of Freedom (df) | t* at 95% Confidence | Difference from z* (1.960) |
|---|---|---|
| 1 | 12.706 | +10.746 |
| 5 | 2.571 | +0.611 |
| 10 | 2.228 | +0.268 |
| 20 | 2.086 | +0.126 |
| 30 | 2.042 | +0.082 |
| 60 | 2.000 | +0.040 |
| 120 | 1.980 | +0.020 |
| ∞ (z) | 1.960 | 0.000 |
For the complete table covering every degree of freedom from 1 to 120 across multiple confidence levels, see the full t-distribution table. For the standard normal distribution's area-under-the-curve values, see the z-table.
Distribution & Critical Value Selector
Enter your sample size, whether the population standard deviation is known, and your desired confidence level. The calculator recommends the correct distribution, calculates degrees of freedom where relevant, and returns the matching critical value.
🔬 t vs Normal Distribution Selector
Common Mistakes When Choosing Between t and Normal
| What People Do | Why It's Wrong | What's Correct |
|---|---|---|
| "My sample is n ≥ 30, so I'll use z" | Sample size alone doesn't determine the distribution — whether σ is known does | If σ is unknown, t with large df is still technically correct (and barely different from z) |
| Treating t as only for "tiny" samples (n < 5) | t applies whenever σ is estimated, at any sample size | Large-df t-tests are standard in regression and large studies, not just small ones |
| Looking up z* = 1.96 but plugging in the sample SD (s) | Mixing a z critical value with an s-based standard error understates uncertainty | Pair s with t*(df); pair a known σ with z* |
| Assuming a higher df always means "better" data | df reflects sample size, not data quality — a large biased sample is still biased | df only affects the shape of the reference curve, not whether assumptions hold |
| "The t-distribution means I don't need normal data" | Both t-tests and z-tests assume the underlying population is roughly normal, or that n is large enough for the CLT | Check the normality of your data separately from which reference distribution you use for the test statistic |
Frequently Asked Questions
Both are symmetric, bell-shaped distributions centered at zero in their standard form, but the t-distribution has heavier tails and a slightly lower peak. The normal distribution has one fixed shape, while the t-distribution has a different shape for every value of degrees of freedom (df), and df = n − 1 for a one-sample test.
Use the t-distribution whenever the population standard deviation (σ) is unknown and must be estimated from the sample standard deviation (s) — which is the case for almost all real-world data. Use the normal distribution when σ is genuinely known, such as a process with an established long-run standard deviation.
The heavier tails account for the extra uncertainty that comes from estimating σ with the sample standard deviation s. Because s itself varies from sample to sample — especially with small n — extreme values of the test statistic occur somewhat more often than the normal distribution alone would predict, and the t-distribution's shape reflects that.
Yes. As degrees of freedom increase, the t-distribution's tails get thinner and its peak gets taller. By df ≈ 30 its critical values are within about 0.08 of the normal distribution's, and as df → ∞ the t-distribution converges exactly to the standard normal distribution.
Degrees of freedom (df) equal the sample size minus one for a one-sample test (df = n − 1). They determine the exact shape of the t-curve: lower df produces heavier tails and larger critical values, while higher df produces a curve closer to the normal distribution. See the degrees of freedom guide for how df is calculated for other tests.
A t-test assumes the underlying population data is approximately normally distributed (or that the sample is large enough for the Central Limit Theorem to apply). However, the test statistic itself follows the t-distribution, not the normal distribution, because it's built using the sample standard deviation s rather than the population σ.
Use a z-test when the population standard deviation σ is known from prior data, such as a long-running manufacturing process. In practice, a z-test is also commonly used as an approximation for proportions and for very large samples, where the t and z critical values are nearly identical anyway.
The t-distribution produces a wider confidence interval than the normal distribution at the same confidence level and standard error, because t* is always at least as large as z*. The worked example in this guide shows a 95% interval that is about 15% wider under the t-distribution than under the normal distribution for a sample of 10.
Sources and References
This guide draws on the following primary and secondary sources for the history, formulas, and critical values discussed above.
- Student (William Sealy Gosset) (1908) — "The Probable Error of a Mean." Biometrika, 6(1), 1–25. The original paper introducing the t-distribution.
- Encyclopædia Britannica — "Student's t-test." britannica.com
- Penn State STAT 415 — Introduction to Mathematical Statistics. Penn State Eberly College of Science. online.stat.psu.edu
- OpenStax Introductory Statistics — Confidence Intervals and the t-Distribution. Rice University. openstax.org
- NIST/SEMATECH Engineering Statistics Handbook — National Institute of Standards and Technology. itl.nist.gov