If you've never run one before, "t-test" sounds like something that belongs in a statistics course rather than on a shop floor. In practice it's one of the most directly useful tools in quality engineering, and the idea behind it is simpler than the name suggests: a t-test tells you whether a difference you observed is big enough to trust, or small enough that it's probably just the noise of the sample you happened to take. ## The problem it solves Say your line is supposed to fill containers to 250 ml. You measure 20 containers and the average comes out to 249.3 ml. Is the machine actually running under target, or would any random sample of 20 come out somewhere around there just by chance, even if the true average is exactly 250? That's the question a t-test answers. Every sample you take will have some average slightly off from the true process mean — that's unavoidable. The t-test quantifies how much "slightly off" you should expect from chance alone, and compares your actual gap against that expectation. ## Where the name comes from The test is built around a statistic called t, calculated from your data: t = (observed difference) / (standard error of that difference) The standard error measures how much your sample mean would be expected to bounce around from one sample to the next, based on how much variation is in your data and how many observations you took. Dividing the observed difference by that expected bounce gives you a number that says, in effect, "how many units of ordinary sampling noise does this gap represent?" A t of 0.5 means your observed difference is smaller than the noise you'd expect anyway — nothing to conclude. A t of 3.5 means the difference is much larger than ordinary sampling noise would produce — worth taking seriously. The exact cutoff between "nothing to conclude" and "worth taking seriously" comes from the t distribution, a bell-shaped curve slightly wider than the normal distribution to account for the fact that you're estimating variation from the same limited sample. ## The three flavors Every real-world situation a t-test can help with falls into one of three shapes, and figuring out which one you have is really the only judgment call involved — the arithmetic follows automatically once you know which test applies. **One-sample t-test** — you have one set of measurements and you're comparing it against a fixed number: a specification, a target, a historical average, a supplier's claim. The filling-machine example above is this case. **Two-sample t-test** — you have two separate, unrelated groups and you want to know if their averages differ: two suppliers, two machines, two shifts, two material lots. Nothing links a specific row in one group to a specific row in the other. **Paired t-test** — you have two measurements of the *same* units: the same parts measured before and after a process change, the same operators timed before and after training, the same gauges compared on the same parts. If you can point to a row and say "these two numbers belong to the same thing," the data is paired. Getting this choice right matters more than anything else in running the test correctly. Treating paired data as though it were two independent groups is the single most common mistake, and it doesn't just produce a slightly worse answer — it can turn a real, easily-detectable effect into "no significant difference," because it throws away the information that comes from comparing each unit to itself. ## A full worked example Twenty containers filled by a line with a 250 ml target. The sample gives a mean of 249.3 ml and a standard deviation of 1.2 ml. **Step 1 — state what you're testing.** This is a one-sample test, since you're comparing one set of measurements against a fixed target. The hypotheses: H₀: μ = 250 (the true process mean equals target) H₁: μ ≠ 250 (it doesn't) **Step 2 — calculate the standard error.** SE = s / √n = 1.2 / √20 = 0.268 **Step 3 — calculate t.** t = (249.3 − 250) / 0.268 = −0.7 / 0.268 = −2.61 **Step 4 — compare against the t distribution.** With n = 20, degrees of freedom = 19. Against a two-tailed test at α = 0.05, the critical value is 2.093. Since |−2.61| exceeds 2.093, the result is significant — p works out to about 0.017. **Step 5 — read the confidence interval, not just the verdict.** The 95% interval comes out to roughly (248.74, 249.86) — the true process mean is plausibly somewhere in that range, comfortably below the 250 target. That's the sentence worth putting in a report: not "p = 0.017," but "the line is running about 0.7 ml under target, and that's a real, detectable shortfall, not sampling noise." ## What the test doesn't tell you A significant t-test tells you a difference is *detectable*, not that it's *important*. With enough data, even a trivial, meaningless difference becomes statistically significant — the test is comparing your gap against sampling noise, and more data shrinks that noise, making smaller and smaller real differences visible. This is why a proper t-test result always comes with more than a p-value: the confidence interval (how big is the difference, plausibly?) and the effect size, usually Cohen's d (is that difference large relative to the process's own variation?). Read all three together, in that order of priority — interval and effect size first, p-value last — and you'll rarely be misled by a result that's "significant" but doesn't actually matter. ## What it assumes A t-test works best when a handful of conditions hold, roughly in order of importance: - **The observations are independent of each other** — this comes from how the data was collected, not from anything visible in the numbers themselves. - **The process is stable** — a t-test assumes there's a single, consistent mean to estimate. A process that's drifting doesn't have one, and testing it produces a number that describes when you happened to sample rather than what the process actually does. - **The data is roughly normal, or the sample is large enough that it doesn't matter much** — this assumption is the one most people worry about and, in practice, the one that matters least at moderate sample sizes. By around 30 observations, the test tolerates real-world non-normality quite well. None of these need to be perfect. The test is forgiving of small departures, especially as sample size grows — but stability and independence are worth checking deliberately, since nothing in the arithmetic will flag a violation of either one. ## Getting started The good news is that none of the arithmetic above needs to be done by hand in practice — it's shown here so the mechanics are visible, not because you should be computing standard errors on paper. The [hypothesis testing calculator on SigmaDesk](https://sigmadesk.app/hypothesis-test-calculator/) runs all three t-tests — one-sample, two-sample, and paired — on the same spreadsheet, so you can switch between them without re-entering your data, and every result comes back with the p-value, confidence interval, effect size, and a set of plain-language assumption checks together, rather than a bare significance verdict. The part worth spending your own attention on is upstream of any calculator: is the process stable, is the data actually independent, and — critically — is this a one-sample, two-sample, or paired situation. Get those three questions right and the test itself takes seconds. If you want to try it on your own numbers, the tools at [SigmaDesk](https://sigmadesk.app/) are free to use and need no account to start.