PHAS0083 Chap.9 Interval Estimation and Coverage
Interval Estimation and Coverage
Interval Estimation and Coverage replaces a single estimate with a set that comes with a frequency guarantee. An interval estimator is a pair of statistics, so the interval is random while the parameter is fixed, and its coverage probability is the chance, over repeated samples, that it contains the true value.
Coverage can depend on the parameter, so the guarantee is the confidence coefficient, the smallest coverage over all parameter values; a fixed-length interval can see its coverage fall towards zero as the scale of the problem grows, while a scale interval for a uniform maximum keeps a constant coverage. With discrete data coverage moves in steps and sits above the nominal value between them.
The notes construct intervals almost exclusively by inverting tests: keep every tested value that a level-alpha test would not reject, which gives a set with confidence at least one minus alpha, read from a confidence belt. The alternative hypothesis fixes the shape, so a one-sided alternative towards larger values produces a lower limit.
When thresholds lack a formula, each tested value on a grid is calibrated by simulation, and grid spacing limits endpoint precision.
What this chapter covers
- 01
Interval estimators and estimates
- 02
Coverage probability
- 03
Confidence coefficient
- 04
One-sided limits
- 05
Inverting tests and the confidence belt
- 06
Numerical inversion on a grid
- 07
Coverage with discrete data
Worked example · free
Set an upper limit after observing zero events
- 1A test of lambda against smaller values rejects lambda when observing a count this small would have probability at most 0.05. This allocation is practice weighting, not a UCL marking scheme.
- 1With zero observed, that probability is P(N = 0 given lambda) = exp(-lambda).
- 1Values with exp(-lambda) at most 0.05 are rejected, so the boundary is lambda = -ln 0.05 = 3.00.
- 1Keep all values not rejected: the 95% upper limit is lambda at most 3.00 events.
Key terms
- Confidence coefficient
- The smallest coverage probability of an interval estimator over all parameter values.
- Confidence belt
- The union of the acceptance regions of tests at every tested parameter value, read at the observed statistic.
- Upper limit
- A one-sided confidence set extending from minus infinity or zero up to a computed bound.
Interval Estimation and Coverage FAQ
Why can a fixed-length interval have zero confidence coefficient?
Its coverage can depend on the parameter. For a uniform sample on zero to theta, an interval of the form maximum plus a fixed shift covers less and less often as theta grows, so the smallest coverage over all theta is zero. Scaling the interval with the maximum instead keeps the coverage constant.
Does a 95% interval contain the parameter with probability 0.95?
Not once it has been computed. The parameter is fixed and the interval is either right or wrong. The 95% describes the procedure: over many repeated experiments, intervals built the same way contain the true value at least 95% of the time.
How do I turn a hypothesis test into a confidence interval?
Test every candidate value of the parameter at level alpha with the observed data and keep the values the test does not reject. Theorem-level reasoning guarantees the kept set has confidence at least one minus alpha, and the confidence belt shows it graphically.
Why do intervals for counts over-cover?
With discrete data the achievable test sizes jump, so tests at nominal level alpha usually have smaller actual size. Inverting them gives intervals whose coverage stays at or above one minus alpha and moves in steps as the parameter changes.
Exam move
Start by simulating thirty intervals for a normal mean with known variance and counting how many miss, so coverage becomes a frequency you have seen rather than a definition. Then compute coverage analytically for two constructions on a uniform maximum, a scale interval and a shift interval, and explain why only one has a positive confidence coefficient.
Practise inverting tests in three settings: a two-sided normal-mean test, a one-sided test that gives a lower limit, and a Poisson count with zero events that gives an upper limit. For each, say which alternative produced which shape. Rehearse the correct sentence for interpreting a computed interval until it is automatic.
Finally write the grid algorithm for numerical inversion in words and identify when one threshold can be reused for every tested value.
Working through Interval Estimation and Coverage in PHAS0083? Sia is AskSia’s AI Statistics tutor — ask any PHAS0083 Interval Estimation and Coverage question and get a clear, step-by-step explanation grounded in how PHAS0083 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.