When do you use a t critical value instead of z for a confidence interval for a mean?
answer
- what did you have to estimate
- extra uncertainty from an estimated spread
- degrees of freedom are n minus 1
- heavier tails mean a bigger multiplier
- twelve measurements give 2.201, not 1.96
basics
~20 sUse t whenever the population standard deviation is unknown and estimated from the sample, which is almost always. The z shortcut is exact only when that spread is known, and adequate once the sample is large.
solid answer
~40 sThe interval is `xbar ± t × s / sqrt(n)`, where the t critical value has `n - 1` degrees of freedom. The t distribution has heavier tails than the normal, and that extra width is precisely the price of not knowing the true spread and having to estimate it with `s`. For 12 lab measurements the degrees of freedom are 11 and the 95% critical value is 2.201, about 12% larger than 1.96 — a real difference on a small sample. The gap closes fast: at n = 30 the value is 2.045, at n = 100 it is 1.984, so past a few dozen observations the z shortcut costs almost nothing. The trigger is *estimated spread*, not *small sample* — small n is just where the correction is big enough to notice.
go deeper
Be able to write the interval as the sample mean plus or minus t times s over sqrt(n), and to say that t is used because the spread came from the sample.
Explain the mechanics: n - 1 degrees of freedom, heavier tails, and concrete values such as 2.201 at 11 degrees of freedom against 1.96, converging as n grows.
Show you check the assumptions the t interval does not fix — skew at small n, outliers inflating s, and dependent observations that make the effective sample smaller than the row count.
Be ready to argue when a normal-theory interval is the wrong tool entirely for the data your teams handle, and what the organisation should standardise on instead.
## The two intervals For a population mean there are two textbook intervals, and they differ in one symbol. If the population standard deviation `sigma` is genuinely known: ``` xbar ± z × sigma / sqrt(n) ``` If it is unknown and you estimate it with the sample standard deviation `s`: ``` xbar ± t(n-1) × s / sqrt(n) ``` The first case is rare outside classrooms and a few calibrated-instrument settings. In real analysis you estimate the spread from the same data you are summarising, so the second formula is the default. ## Why estimating the spread costs something When `sigma` is known, `(xbar - mu) / (sigma / sqrt(n))` follows a standard normal distribution (exactly, if the data are normal; approximately, by the central limit theorem, if n is reasonably large). Substituting `s` for `sigma` puts a second random quantity into the denominator. Sometimes `s` lands low, and the ratio blows up further than a normal would allow. That extra tail behaviour is exactly what the t distribution encodes: same symmetric bell shape, but fatter tails, so its critical values are always larger than the matching normal ones. That is the sentence that separates a memorised rule from an understood one. Candidates who say *use t for small samples* are describing the symptom; candidates who say *use t because s is estimated, and the correction is only large when n is small* are describing the cause. ## Degrees of freedom The t reference has `n - 1` degrees of freedom for a one-sample interval. The intuition: the sample standard deviation is computed from deviations around `xbar`, and those n deviations must sum to zero, so only `n - 1` of them are free to vary. One degree of freedom was spent estimating the centre. For 12 measurements, that is 11 degrees of freedom. ## The numbers worth carrying Two-sided 95% critical values: - df = 4 (n = 5): 2.776 — about 42% wider than the normal value - df = 11 (n = 12): 2.201 — about 12% wider - df = 29 (n = 30): 2.045 — about 4% wider - df = 59 (n = 60): 2.001 — about 2% wider - df = 99 (n = 100): 1.984 — about 1% wider - df = infinity: 1.96 — the normal value exactly The t distribution *converges* to the normal as the degrees of freedom grow, which is why the last row is the normal value. This table also answers the practical question of when the shortcut is safe: past about n = 60 the difference is inside the noise of everything else in the analysis, and past n = 100 it is negligible. At n = 5 it is not remotely negligible, and quoting 1.96 there produces an interval that is far too narrow. ## What the t interval still assumes Using t fixes the estimated-spread problem. It does not fix everything else. - **Shape.** The exact justification requires the underlying observations to be roughly normal. With a large sample the central limit theorem rescues the mean's sampling distribution even from non-normal data; with a small sample it does not, and a t interval on eight strongly skewed observations is not trustworthy at its stated level. - **Outliers.** Both `xbar` and `s` are sensitive to extreme points. A single wild value inflates `s`, widening the interval, and drags `xbar`, moving it. - **Independence.** The `sqrt(n)` in the denominator assumes n independent observations. Repeated measurements on the same subject, or clustered data, give you fewer effective observations than the raw count, and the interval will be too narrow no matter which critical value you use. ## The practical rule Default to t. It is never wrong when the spread is estimated, it costs nothing to compute, and it converges to the z answer automatically when the sample is large — so the safe habit and the correct habit are the same habit. Reach for z only when you have a genuinely known population standard deviation, or when you are doing arithmetic in your head on a large sample and one percent of width does not change the conclusion. A final framing that lands well in interviews: t versus z is not a choice about the *data*, it is a choice about *what you know*. If you had to estimate the spread, you owe the interval the extra width.
- Why are the degrees of freedom n minus 1 rather than n for a one-sample interval?The sample standard deviation is built from deviations around the sample mean, and those deviations are constrained to sum to zero. One piece of information was consumed estimating the centre, so only n - 1 of the deviations are free to vary. With 12 measurements that leaves 11 degrees of freedom.
- Does using t rescue a small sample of badly skewed data?No. The t correction accounts for having estimated the spread, not for the shape of the underlying distribution. With a small sample and strong skew the interval's true confidence level can be well below the stated one. Large samples are rescued by the central limit theorem; small skewed ones are not.
- At what sample size does using z instead of t stop mattering in practice?Around 30 the two-sided 95% t value is 2.045 against 1.96, roughly 4% wider. By 60 it is 2.001, about 2%, and by 100 it is 1.984, about 1%. Past a few dozen observations the choice rarely changes a conclusion, though defaulting to t costs nothing.
saying these in an interview costs you the question
- Says t is for small samples without mentioning the estimated spread
- Uses 1.96 automatically once n passes 30
- Uses n instead of n minus 1 degrees of freedom
- Believes t corrects for skewness or outliers
- Thinks t critical values can be smaller than the matching z