A t-test asks whether a mean difference is too large to blame on sampling noise — one sample against a benchmark, the same units measured before and after, or two independent groups against each other. It answers with a t value, degrees of freedom and a p-value that report in APA style as t(28) = 2.15, p = .04, with an effect size d beside them when the course wants magnitude. Three group means or more call for ANOVA instead, and category outcomes call for chi-square. Work through the three questions below, then paste your output numbers into the builder and copy the line.
An honesty note before the helper: this page routes decisions and formats numbers, it does not calculate them. The t, df and p come from your statistics software; the builder makes the output say what you mean in APA order, and the effect size it shows is an estimate computed from t, not your software’s exact d. Every example number on the page is invented for teaching and labeled illustrative. Course conventions beat templates — if your sheet wants Welch’s test by default or effect sizes first, the sheet wins.
The decision helper: three questions, one verdict
Test choice for means breaks down into three facts, and every stats course asks them in the same order. How many samples supply the means — one against a benchmark, the same units measured twice, or two different groups? How many cases feed the test? And what do you actually know about the spread in the population? Answer the three below and the verdict re-renders on each change: a t-test flavor when the design calls for one, or the z-test and non-parametric alternatives when they fit honestly. Your choices save to this browser only, about a third of a second after you stop changing them; Reset erases them. Nothing is uploaded anywhere.
T-Test decision helper
100% private — runs in your browserDifferent people per group = independent; before-and-after on the same units = paired.
The classic rule of thumb — below it, the test leans harder on the shape of the data.
A given population sigma almost never happens outside textbooks; estimated is the honest default.
Use an Independent Samples T-Test
Two different groups of people, one outcome each, thirty-plus per group, spreads that are estimated but comparable — the classic independent-samples t-test with pooled variance. Large groups forgive moderate bends in the shape.
Report it as: t(n1 + n2 − 2) = ..., p = ... — the builder below assembles it.
The helper reads the three answers, not the data — it cannot see your spreadsheet. Borderline cases (ranked scores, spreads that visibly disagree with tiny groups, paired designs with missing slots) belong in a conversation with your stats course; the verdict is a router, not a statistician.
When to use a t-test: the three questions, answered
The helper compresses the reasoning; here is the reasoning itself. Each question removes tests that cannot fit, and by the third question only the honest options survive.
Question one: the count of means picks the family
One mean has nothing to pair with, so the comparison needs a fixed reference — a pass mark, an industry figure, a value the methods chapter named in advance. That is the one-sample t-test. The same units measured twice — before and after, two raters, two conditions — are a paired design, and the paired t-test runs on each unit’s difference score, which is why it usually detects smaller changes with the same number of people. Two different groups supply the classic setup: an independent-samples t-test on one outcome, one group each. Three group means or more stop pairing cleanly — that is where ANOVA takes over, because stacking pairwise t-tests quietly multiplies your false-alarm rate. The sibling guide for that step is the ANOVA page; this one owns the one-and-two-mean world.
Question two: the thirty mark is about shape, not legality
Below roughly thirty cases, the t-test leans harder on its one big ask — that the scores run roughly normal, bell-shaped, without heavy tails. Thirty or more, and averages of messy scores start behaving themselves well enough that moderate bends stop moving the verdict much; that is the central limit theorem doing quiet work, and the reason stats courses quote the number. The mark is a rule of thumb, not a law: forty heavily skewed salaries still lean on shape, and a paired design with ten tidy difference scores can be fine. Treat the question as how much you owe the reader a shape check, not as a switch between two different tests.
Question three: the spread decides z, t or Welch
The population standard deviation is essentially never known outside tables and textbooks — and when it genuinely is, the z-test is the tighter label, which is why the helper points there instead of pretending otherwise. When the spread is estimated from the sample, as it almost always is, the t-test is the honest choice, because t was built exactly for a noisy sigma. In two-group work one question remains: can the two spreads be treated as comparable? The pooled t-test merges them into one estimate; when they visibly disagree — especially with unequal group sizes — Welch’s unequal-variances variant adjusts the degrees of freedom instead of trusting the merge. Software prints Welch’s output with different df; nothing else about your reporting changes.
| Design you actually have | Test the design calls for | Why |
|---|---|---|
| One sample vs a benchmark, sigma estimated | One-sample t-test | One mean, one reference, df = n − 1 |
| Same units before and after | Paired t-test | Difference scores; each unit is its own control |
| Two different groups, one outcome | Independent-samples t-test | Two means, pooled or Welch variance |
| Two groups, sigma genuinely known | Z-test | Textbook slot; say where the sigma came from |
| Ranked or heavily skewed small samples | Wilcoxon / Mann-Whitney | Same questions without the shape promise |
| Three or more group means | One-way ANOVA | One omnibus F; pairwise t-tests inflate error |
What the t-statistic actually compares
The t value is a ratio, and the whole test lives in that comparison. The top is the signal: the difference you observed, measured against its reference — a benchmark, a zero-change in paired scores, or the other group’s mean. The bottom is the noise: how far means that size wobble from sample to sample, built from the spread of the scores and shrunk by how many cases you have. If the treatment does nothing, the observed difference sits inside ordinary wobble and t lands near zero. The further the difference climbs above the noise, the larger t grows in either direction — and the software attaches a p-value to that size.
| Part of t | What it carries | What moves it |
|---|---|---|
| Numerator (the difference) | Observed gap: mean vs benchmark, mean change vs 0, mean gap between groups | The effect itself |
| Denominator (the noise) | Standard error — spread divided by sample size | More cases shrink it; wider spread grows it |
| df, e.g. 28 | Cases minus design costs: n − 1, n1 + n2 − 2, or pairs − 1 | — |
Two habits make t legible in your own paper. First, the sign is direction, not trouble: a negative t says the first-listed mean sat below the second, and the sign survives into the report unchanged. Second, t without its degrees of freedom is half a sentence — t(28) tells the reader how many cases stood behind the number, which is what lets a marker check the arithmetic of your design. The same 28 means twenty-nine cases in a one-sample test or thirty split across two groups; the df encode the design, and the prose you wrap around the line supplies the rest.
Reading p: significant, and how big
The p-value answers one question: if the treatment truly did nothing, how often would sampling noise alone produce a t this far from zero? Small p, rare coincidence — the classic cutoff is .05, meaning the coincidence runs under five in a hundred. A significant t-test says the difference is real enough to trust at the bar you set; it says nothing about how big the difference is, and that gap is where results sections wobble. A tiny p on a large sample can ride a difference too small to matter in practice, and a clean, useful effect can miss the cutoff in a small one.
So the habit is mechanical: report the t line for the evidence, then answer size separately — Cohen’s d beside the line, or the raw means where the reader can see them. d near 0.2 reads as a small effect, 0.5 medium, 0.8 large, by the usual classroom bands — labels, not laws. And when p lands above .05, the honest sentence is “no statistically significant difference was found” — not “the groups are identical”. Non-significant says the evidence fell short of the bar you set; the means themselves can still be worth a sentence of description.
The APA reporting template: assemble the line live
Every t-test paragraph hangs on one formatted line, and the style sheet holds the rules firmly: degrees of freedom in parentheses directly after the statistic, t and p italicized, two decimals, no leading zero on p, and p below .001 written as p < .001 because an exact value of .000 cannot happen. Fill in the three numbers from your software output and the line below assembles itself, with the format checks running as you type — a whole-number df, a t that keeps its sign, a p between 0 and 1, and an effect-size estimate appended because a results line that carries d answers the size question in the same breath. The estimate names its formula so nobody mistakes it for software output; the group-means box is optional, feeds the copied paragraph, and is never saved. Copy the result, italicize t, p and d in your document, and the results paragraph has its spine.
APA reporting template
Format checks run live — numbers stay on your deviceA negative t is normal — it keeps its sign on the line.
One sample: n − 1. Two groups: n1 + n2 − 2. Paired: pairs − 1.
Below .001 reports as p < .001 — never p = .000.
One entry per group or occasion, each starting with M =. Feeds the copied paragraph; nothing here saves.
Fill t, df and p to assemble the results sentence.
Degrees of freedom — one whole numberwaiting
A count, not a measurement: n − 1, n1 + n2 − 2, or pairs − 1. No decimals, no negatives.
t — any sign, decimals finewaiting
Unlike F, a negative t is direction, not error — the sign stays on the line.
p — between 0 and 1, zero droppedwaiting
Probabilities cannot exceed 1, so APA strips the leading zero: .04, not 0.04.
Effect size d — estimated from twaiting
Computed from t and df: 2t/√df for equal-sized groups, t/√n for one sample or paired differences. Your software’s exact d wins when it prints one.
Means — optional, M = per entrywaiting
Leave empty to copy the line alone. Each entry starts with M =.
The copied text is plain — italicize t, p and d after pasting, and add your means sentence if the box has entries. Your three helper choices and these three numbers save to one browser key about a third of a second after typing; the means box never saves.
The t-line, part by part
Four parts, four rules. The table is the whole style guide for the line — everything else in the results paragraph is prose around it.
| Part | Example | Rule |
|---|---|---|
| Statistic + df | t(28) | t italic, df as a whole number in parentheses right after it, no space before the parenthesis |
| Equals and value | = 2.15 | Two decimals; a negative value keeps its sign; a space both sides of the equals |
| Probability | p = .04 | p italic, two decimals, leading zero dropped; exact values only in (0, .001) |
| Tiny probabilities | p < .001 | Below .001 the style switches to an inequality — never p = .000 |
| Effect size | d = 0.70 | d italic, two decimals; the honest number is your software’s exact d — an estimate says so |
| Descriptives | (M = 11.2, SD = 3.0) | Means and standard deviations ride beside the line, in a table or a sentence |
One placement rule belongs beside the table: the t line lands after the claim it supports, not before it. The sentence commits to what the test found, the line certifies it — “scores improved after the workshop, t(23) = 3.41, p = .002” — and the marker reads cause and evidence in one pass.
A worked results paragraph (illustrative numbers)
Here is the whole shape in one paragraph. The study, the numbers and the finding are invented for teaching — a real paper swaps in its own output and keeps the skeleton.
Twenty-four employees completed a focus rating scale one week before and one week after a two-week deep-work workshop. A paired-samples t-test found a statistically significant increase in focus scores, t(23) = 3.41, p = .002, with a within-persons effect size of d = 0.70. Mean focus rose from 11.2 (SD = 3.0) before the workshop to 13.1 (SD = 3.2) after it. The gain was consistent across the team: twenty of the twenty-four employees rated higher afterwards, and no rating fell by more than one point.
Count what the paragraph does: names the design and the test in one clause, hangs the t line on the finding it supports, quotes the effect size with its label, puts both means where the reader can compare them, and closes with the consistency check a marker looks for in small samples. Four sentences, and the results section has its spine. Everything after it — why the workshop moved the number — belongs to the discussion, not the results.
Assumptions worth one line in your paper
The t-test leans on three assumptions, and most coursework satisfies them without drama. The observations are independent — one score per person, no borrowing across the table — unless the design says otherwise, which is what the paired flavor exists for. The scores run roughly normal in each group, or in the difference scores for a paired design; with thirty-plus cases this is forgiving enough that a one-line note covers it, while small samples owe the reader a shape check or a non-parametric fallback. And in two-group work the pooled version assumes comparable spreads — software tests this with one click, and when the spreads disagree, Welch’s variant adjusts the degrees of freedom instead of pretending. Name which variant you ran in the methods; the t line itself does not say.
Five t-test mistakes that show up in margins
The same five problems account for most of the red ink on results sections:
- The pairing ignored. An independent t-test run on before-and-after scores from the same people. The pairing is the design’s strongest asset; treating two readings from one person as independent throws the control away and usually buries a real change.
- The bare t. “t = 2.15, significant” — without the df the reader cannot check the design: how many cases, which flavor. The parentheses are half the statistic; put them back.
- p = .000. An exact zero probability cannot happen; the output rounded. Write p < .001 and the line is both correct and more professional than the raw printout.
- The size skipped. A significant t on two hundred cases may ride a difference too small to matter. Without d — or the raw means — the paragraph answers “is it real?” and stays silent on “is it worth anything?”.
- The z detour. Reporting a z-test because a spreadsheet function was nearest to hand. When the spread came from the sample, the honest label is t — and markers notice the letter that does not match the design.
Notice the pattern: none of these is a calculation error. They are reporting habits, which is why they survive proofreading — the writer knows what the test meant and forgets the reader does not.
Test picked, output open, deadline loud?
A t-test takes one afternoon; the chapter wrapped around it often takes the week. If the deadline is nearer than the analysis schedule, a quoted, deadline-bound writer can turn your output and materials into the referenced sections the course expects, while the numbers and the argument stay yours. The form below is the way to ask.
What this page cannot do
The honest limits, because a page about statistical reporting should survive an audit:
- It does not calculate. No sums, no tests, no p-values: the t, df and p come from your statistics software, and the builder only formats them. A wrong number in, a confidently wrong line out.
- It cannot read your assignment sheet. Some courses want Welch’s everywhere, some want exact p to three decimals, some want effect sizes before p. Where the sheet disagrees with the template, the sheet wins.
- The helper is a router, not a statistician. Twelve routes cover the standard designs — and the four “consider the alternative” verdicts are honest stops, not diagnoses. Messy realities (ranked scores, missing pairs, nested data) need a stats course, not three dropdowns.
- The effect size is an estimate. The d the builder appends is computed from t and df with the equal-groups or difference-scores conversion. When your software prints its own d, its number wins; when group sizes are very unequal, the estimate degrades — swap in the software value.
- It stores almost nothing. The three helper choices and three APA numbers save to one browser key; the means box, quiz and poll never persist. Real results belong in your notes and your data file.
Where the t-test sits in the research family
The t-test is one station on a line the rest of the shelf runs. Upstream, the research question generator turns a topic into the question, and the hypothesis statement generator states in advance what a yes would look like — the prediction the t either supports or fails to support. The multi-group sibling is the ANOVA guide: from three group means up, its helper routes the design to one-way, repeated measures or factorial ANOVA. The design that fixes how many cases you have belongs to the methods chapter, where the lab report results guide builds the table this line usually sits under. Downstream, the research paper abstract compresses what the test found into two sentences, and the study guides hub keeps the whole research shelf together. The free writing tools cover the mechanical neighbors a finished paper also owes — word counts, citations and readability. One page, one job: this one stops at the t line.
Quiz: five questions on choosing and reporting
Five questions on the calls t-test work makes: which flavor fits, what the sign means, the APA line, the honest non-significant sentence, and the unequal-spreads case. Every wrong answer reveals the right one.
1. The same eighteen employees rate a process before and after a training week. Which test fits?
2. Your output reads t = −2.44, df = 27. What does the minus sign carry?
3. Which line reports the test in APA style?
4. Twelve sites per condition, p = .18. The honest sentence for the results section is:
5. Independent groups of n = 9 and n = 14, and the two spreads visibly disagree. The safer call:
Frequently asked questions
When is a t-test the wrong choice?
From three group means up, the comparison belongs to ANOVA — the t-test pairs two means at a time, and stacking several of them inflates the false-alarm rate. Categorical outcomes (passed or failed, chose option A) have no mean to test, so the chi-square family reads the counts instead. Ranked scores or heavily skewed small samples break the t-test's shape assumption, which is where the Wilcoxon and Mann-Whitney tests carry the same questions without it. And when the population standard deviation is genuinely known, the z-test is the tighter, legal label.
What do the t value and its df actually say?
t is a signal-to-noise ratio: the difference you observed, divided by the noise sampling puts around differences that size. t(28) says twenty-nine cases fed the test in a one-sample design, or thirty split across two groups — one sample spends df on n − 1, two independent groups on n1 + n2 − 2, a paired design on pairs − 1. A bigger t means the difference stands further above the noise, and the software attaches the p-value to that size. The df matter because the same t carries different evidence at ten cases than at a hundred.
How do I write the t-test results line in APA style?
The line pairs the statistic with its degrees of freedom and a probability: t(28) = 2.15, p = .04. Italicize t and p, keep a space on both sides of the equals sign and the comma, report both to two decimals, and drop the leading zero from p because a probability cannot exceed 1 — p = .04, never p = 0.04. A negative t is fine and keeps its sign; probabilities below .001 report as p < .001, never p = .000. Add Cohen's d beside the line when the course asks for effect size, two decimals as well.
My t came out negative — is that a problem?
No — the sign carries direction, not error. It says the first-listed mean sat below the second, and which one that is depends only on the order your software listed the groups. Report the value with its sign exactly as printed, because flipping it would misstate the direction of the effect. If the sign surprises you, check which group the software called group one before rewriting any prose around the number.
Does this page store my numbers and test choices?
Your three helper choices and the three numbers behind the line — t, df and p — save to one local storage key, csw-ttt-v1, about a third of a second after you stop typing, so a reload restores the assembled line. Reset erases the key. The means box, quiz picks and poll votes never save, because pasted statistics belong in your notes, not in a browser. Nothing leaves the device; there is no account and no server copy.
Did the page settle your test flavor and the t line?