Z Score vs T Score
Short answer: use a z score when the population standard deviation σ is known (or the sample is very large); use a t score when σ is unknown and estimated from a small sample. The t distribution is the z distribution's cautious sibling — same bell shape, heavier tails — and it exists because estimating σ adds uncertainty that the z curve refuses to admit.
The decision table
| Your situation | Use | Why |
|---|---|---|
| Population σ is known | z | The spread is a fact, not an estimate — no extra uncertainty to model |
| σ unknown, large sample (n ≥ 30, roughly) | z (t nearly identical) | With many observations s estimates σ so well that t and z critical values agree to a few thousandths |
| σ unknown, small sample (n < 30) | t | s might badly miss σ; the t distribution's heavier tails price that risk honestly |
| Comparing a single raw value against a known distribution | z score | That is a standardisation, not an inference — no sampling involved at all |
| Data clearly not normal, small n | Neither, naively | Both lean on normality (t a little more forgivingly); consider a transformation or a non-parametric method |
What "heavier tails" actually does to your answer
The t distribution is a family, not a single curve: one member for each degrees of freedom (df = n − 1 for a one-sample test). Small df means very heavy tails and large critical values; as df grows, the t curve relaxes onto the z curve and the two become indistinguishable. The 97.5th percentile — the two-tailed 5% fence — shows the whole story in one column: df = 5 → 2.571; df = 9 → 2.262; df = 15 → 2.131; df = 24 → 2.064; df = 30 → 2.042; df = 60 → 2.000; df = ∞ → 1.960, the z value. Every step down that list is the price of uncertainty shrinking as data accumulates.
The same data, worked both ways
One sample, two verdicts
A sample of n = 16 has mean 54.1 against a hypothesised population mean of 50, with sample standard deviation s = 8. The standard error is 8 / √16 = 2.0, and the test statistic is (54.1 − 50) / 2.0 = 2.05 — the same number whichever label you give it.
Judged as z: the two-tailed fence at α = 0.05 is 1.96. 2.05 clears it (p ≈ 0.0404): reject the null hypothesis.
Judged as t (df = 15): the fence is 2.131. 2.05 does not clear it (p ≈ 0.058): fail to reject.
Same data, same statistic, opposite conclusions — and the t verdict is the honest one, because σ was never known; s = 8 is an estimate from just 16 observations, and had the sample been luckier or unluckier, s might have been 6 or 10. The z method spends certainty it does not have. This is not a contrived edge case: statistics near the boundary are common in real research, which is precisely where the z/t choice bites.
Why the confusion persists
Three reasons. First, the formulas look identical — (estimate − hypothesised value) / standard error — so the statistic's arithmetic never warns you which distribution should judge it. Second, older teaching introduced z-tests with σ "known" in contexts (standardised testing, manufacturing specs) where it genuinely was known, and the habit outlived the context. Third, software defaults differ by procedure, and output labels ("z" in one package, "t" in another) get copied into reports without the reasoning. The protection is a single question, asked every time: was σ known before the data arrived, or did the data have to tell me? Known → z. Told by the data, small n → t. Told by the data, large n → either, and say which you used.
A naming trap: T scores are also a measurement scale
In educational and psychological testing, a T score (capital T) is a standard score with mean 50 and standard deviation 10 — a re-dressing of z via T = 50 + 10z, used so reports avoid negatives and decimals. It has nothing to do with Student's t distribution beyond the letter. A T score of 70 is z = +2.00; a T score of 35 is z = −1.50. If a document's "t score" comes with percentiles and a mean of 50, it is this scale; if it comes with degrees of freedom and a p-value, it is the test statistic. Context always disambiguates — but only if you know both creatures exist.
Frequently asked questions
When exactly is a sample "large enough" for z?
The common teaching rule is n ≥ 30, but it is a rule of thumb about s behaving, not a law. At n = 30 the t fence (df = 29, ≈2.045) still sits noticeably above 1.96; by n = 60 they differ by about 0.04, and by n = 120 by about 0.02. If your statistic lands in that gap, use t and say so.
Is the t statistic calculated differently from z?
The formula is the same shape — difference divided by standard error — with s in place of σ. What differs is the reference distribution used to judge the result: heavier-tailed t with its degrees of freedom, versus the standard normal.
What are degrees of freedom?
For a one-sample t, df = n − 1: the number of observations free to vary once the sample mean is fixed. Degrees of freedom select which member of the t family judges your statistic — fewer observations, heavier tails, higher fences.
Can I convert a z score into a t score?
For the measurement scale, yes: T = 50 + 10z. For test statistics, there is no conversion — the statistic is a single number, and "z versus t" is a choice of the distribution you compare it against, made from how σ was obtained.
Which is more common in published research?
t, by a wide margin — because σ is almost never truly known in research settings. Genuine z-tests survive where σ is a published constant of the instrument or process, and in large-sample approximations where the distinction has faded away.
For the z side of the story, work from the z score table and the critical values; for how percentiles compare with z, see z score vs percentile.