8.7 — Prediction Interval (PI) vs Tolerance Interval (TI)

Offline ✓ Student-friendly
Theme:

PI bounds the next value. TI bounds a chosen fraction of the population. Simple calculators + visuals below.

10-second primer

PIPrediction Interval

  • What: Bounds the next observation from the same process.
  • Use when: “Where will one more part/reading land?”
  • Formula: x̄ ± t*·s·√(1+1/n) (two-sided)
  • Note: Use t* (small n). For very large n, z is ok.

TITolerance Interval

  • What: Bounds a chosen percent of the population.
  • Use when: Make claims like “≥90% of all units lie in [L,U]”.
  • Formula: x̄ ± k·s where k comes from a TI table/algorithm for your γ, n, and TI confidence.
Assumptions: roughly normal data, independent observations, and s is a reasonable estimate of σ. Check a QQ plot or rely on process physics/CLT. Misusing normality ⇒ misleading intervals.

Inputs

α (for PI) = 0.050
Ready.

PIPrediction Interval — for the next observation

We compute x̄ ± t*·s·√(1+1/n). In Excel: t* = T.INV.2T(α, n-1).
Interpretation: If you take one more reading under the same conditions, it falls in [L,U] with ≈ stated confidence (normal model).
Common mistakes: (1) Using a CI when you actually want a PI. (2) Forgetting the √(1+1/n) factor. (3) Using z instead of t for small n.

TITolerance Interval — to cover a chosen % of the population

We use TI = x̄ ± k·s. Enter k from a TI table/software (n, γ, 1−α). For a quick large-n intuition, the button suggests k ≈ z(1+γ)/2 (ignores TI confidence; not exact).
Reading: “With the stated confidence (from your table/software), at least γ% of the whole distribution lies in [L,U].”
Reminder: Exact k needs (γ, n, TI confidence). The quick k here is only a large-n intuition aid.

Which one? (1-liners) & Excel cells

  • CI: bounds a parameter (e.g., μ).
  • PI: bounds a future single value (one random draw).
  • TI: bounds a chosen fraction of the population (coverage).
PI (Excel, two-sided):
t* = T.INV.2T(α, N-1)
half = t* * S * SQRT(1 + 1/N)
PI = XBAR ± half

PI (z-approx large n): use NORM.S.INV(1−α/2) instead of t*

TI (normal, two-sided): TI = XBAR ± k*S (get k from TI table/software for your γ and confidence)