z-test, t-test, chi-squared goodness of fit, F-test: Visual walkthrough

Step 1 — A measurement is a dot on a line

WHY start here. Every test compares "what I saw" to "what was claimed." Both are just positions on the same line, so we must first agree that a number is a place.

PICTURE. The orange dot below sits at $990$ . The magenta tick at $1000$ is the claim — the value the manufacturer swears by. We call the claim $μ_{0}$ (read "mew-nought"): the " $0$ " is a label meaning the number in the null hypothesis, the boring status-quo story from Hypothesis Testing. To be clear about symbols: a single reading is written $x$ ; the average of a whole sample is written $\overset{x}{ˉ}$ ("x-bar"). When you have only one reading, $\overset{x}{ˉ} = x$ — the average of one number is that number. Below, the dot at $990$ is both.

Figure — z-test, t-test, chi-squared goodness of fit, F-test

Step 2 — A raw gap of "10" means nothing yet

WHY. Ten grams off is enormous for a coin's mass but invisible for a person's mass. Size only has meaning relative to natural spread. So we need a ruler whose unit is the wobble.

PICTURE. Two worlds, same gap of $10$ . Top: measurements barely scatter — a gap of $10$ pokes far past the crowd, clearly weird. Bottom: measurements scatter wildly — a gap of $10$ is lost in the noise, totally ordinary. Same gap, opposite verdicts.

Step 3 — The wobble of one measurement: $σ$

WHY this symbol. We need one number that summarises "how far do individual readings stray?" $σ$ is exactly that — measured in the same units as the data (hours, ml).

PICTURE. The magenta bell is centred on $μ$ — the population's true mean, which under the claim equals $μ_{0}$ . The violet arrows span one $σ$ to each side. Roughly $68%$ of single readings land inside that band. $σ$ is literally the half-width of the "usual" zone.

Step 4 — Averaging shrinks the wobble: the standard error

WHY average at all. A single bulb might be a freak. Averaging $n = 64$ bulbs cancels individual flukes: some run long, some short, and they partly wash out. The leftover wobble of the average is the standard error, written $SE$ .

PICTURE. Same claim, three bells. The widest (violet) is one reading: spread $σ$ . Averaging $4$ readings (orange) halves the spread. Averaging $64$ (magenta) shrinks it eightfold — a tall, tight spike. More data → sharper knowledge of the true mean.

Step 5 — Measure the gap in units of wobble: the z-statistic

WHY divide. Division re-expresses the gap in ruler-units. A result of $- 2$ means two full SE-lengths to the left of the claim. This number is now comparable across any problem — hours, grams, votes — because the units cancel. And here is where the Central Limit Theorem earns its keep: it promises that $\overset{x}{ˉ}$ follows a bell shape, so this ratio follows the standard bell.

PICTURE. The standard bell (centre $0$ , one SE per gridline). The orange arrow steps from the claim to our data: exactly $2$ SE-steps left. That count, with its sign, is $Z$ .

Step 6 — Where is "too far"? The rejection tails

WHY $1.96$ . With the standard bell, exactly $2.5%$ of the area sits beyond $+ 1.96$ and $2.5%$ beyond $- 1.96$ — together $5%$ . That $5%$ is our tolerated false-alarm rate $α$ (a Type I error). See p-value for the area interpretation.

PICTURE. The bell with both tails past $\pm 1.96$ shaded (the rejection region). Our $Z = - 2.0$ (orange) lands inside the left shaded tail — so we reject the claim. The bulbs really do run short.

Recall Check yourself

If $Z = - 2.0$ and the cutoff is $\pm 1.96$ , do we reject? ::: Yes — $∣ - 2.0∣ = 2.0 > 1.96$ , so it lands in the tail. What does the shaded area represent? ::: The probability of seeing data this extreme if the claim were true — the two-sided p-value threshold $α$ .

Step 7 — When the ruler itself is fuzzy: $σ \to s$ , and $t$ is born

WHY a new distribution. In Step 5 the denominator was a rock-solid constant. Now both top and bottom of the ratio jitter. A fuzzy ruler means we occasionally over- or under-measure the gap, producing more extreme values than the Normal predicts → fatter tails. That fatter-tailed curve is Student's t-distribution, tuned by Degrees of Freedom $ν = n - 1$ . As $n \to \infty$ , $s \to σ$ and $t$ melts back into the Normal.

PICTURE. Two curves overlaid: the crisp Normal (magenta) and Student's $t$ with few Degrees of Freedom (violet). Same centre, but $t$ is squashed lower in the middle and heavier in the tails — so its $95%$ cutoff sits further out ( $2.78$ for $ν = 4$ , versus $1.96$ ). Being cautious about a fuzzy ruler pushes the goalposts back.

Step 8 — The degenerate & edge cases (never leave a gap)

Case A — huge $n$ . As $n$ grows, $SE = σ / n \to 0$ : the bell in Step 4 collapses to a spike. Even a tiny real gap eventually produces a giant $Z$ . Interpretation: with enough data, any true difference is detectable — but a detectable difference may still be trivially small in real terms.

Case B — $\overset{x}{ˉ} = μ_{0}$ exactly. The gap is $0$ , so $Z = 0$ : dead centre of the bell. Maximum agreement with the claim; we never reject.

Case C — $σ \to 0$ (no wobble). The ruler shrinks to nothing, so $Z \to \pm \infty$ for any nonzero gap. Interpretation: if measurements truly never wobble, the slightest gap is infinitely surprising. In practice $σ$ is never exactly $0$ .

Case D — $n = 1$ for the t-test (undefined!). The t-formula divides by $s$ , and $s^{2} = \frac{1}{n - 1} \sum (x_{i} - \overset{x}{ˉ})^{2}$ . With $n = 1$ the denominator $n - 1 = 0$ and there are no deviations from the single point — so $s$ is undefined and $t$ cannot be computed. The t-test requires $n > 1$ : you need at least two readings before "how much do they scatter?" even has an answer. (The z-test escapes this because it uses a given $σ$ , not one estimated from the data.)

Case E — from mean to counts ( $χ^{2}$ ) and from counts to spreads ( $F$ ). The same ratio idea extends. For category counts we standardize each cell as $(O_{i} - E_{i}) / E_{i}$ : this treats each count as roughly Poisson (variance $\approx E_{i}$ , so standard deviation $\approx E_{i}$ ). This is an approximation — a large-sample / Poisson-normal limit that only holds when every expected count is reasonably large (rule of thumb $E_{i} ≳ 5$ ); with tiny $E_{i}$ it breaks and you must pool categories. Under that condition, squaring and summing the standardized cells gives the Chi-squared Distribution. For comparing two spreads we take a ratio of two variances, and the race between two chi-squareds is the F-distribution used in ANOVA.

PICTURE. Four mini-panels: (A) SE collapsing with $n$ ; (B) $Z = 0$ at the peak; (C) the ratio blowing up as $σ \to 0$ ; (D) the family tree $Z \to t \to χ^{2} \to F$ .

The one-picture summary

Recall Feynman retelling — say it like you mean it

I measured something and it came out $10$ below what they promised. But " $10$ " alone is a bluff — I don't know if that's a lot. So I ask: how much does the average of my readings usually jiggle around? One reading jiggles by $σ$ , but I averaged $64$ of them, and averaging calms the jiggle by $64 = 8$ , leaving a jiggle of $5$ . Now I re-measure my gap in jiggles: $10 \div 5 = 2$ jiggles below. Two jiggles is far — the standard bell says landing that far by pure luck happens under $5%$ of the time, so I stop believing "just unlucky" and say the bulbs really are worse. If I didn't know the jiggle-size and had to guess it from my tiny sample, my ruler would be shaky too, so I'd be extra cautious and demand almost $2.8$ jiggles before I dare reject — that cautious version is the $t$ -test (and it needs at least two readings, or there's no jiggle to measure). Same trick counts categories (chi-squared) and compares two jiggles (F). Every test is: how far, measured in jiggles.