Does monotone mean strictly increasing?

Not necessarily. Nondecreasing ratios can include flat portions.

Does the likelihood ratio give the chance that the larger parameter is correct?

No. It gives relative support from the observed data under the compared parameter values.

Why are one-sided tests emphasized?

The likelihood-ratio ordering aligns naturally with evidence in one parameter direction; two-sided questions require separate analysis.

AI Glossary / Statistics and probability / Advanced

Monotone likelihood ratio

Definition

A monotone likelihood ratio is a property of a statistical family in which the likelihood ratio comparing a larger parameter value with a smaller one changes in only one direction as a specified statistic increases. Using the nondecreasing convention, larger values of the statistic provide increasingly strong relative evidence for the larger parameter. The property concerns ratios of probabilities or densities across parameter values, not whether either density is itself increasing. It is important in statistical decision theory and supports simple one-sided testing rules under appropriate conditions.

What it means

A likelihood ratio compares how well two parameter values explain the same observation. Monotonicity means that ordering observations by a statistic also orders their relative support for those parameter values. For a family with densities or probability masses f_theta, the nondecreasing condition requires f_theta2(x)/f_theta1(x) to be nondecreasing in T(x) whenever theta2 exceeds theta1, with support issues handled appropriately. The condition must hold across the relevant parameter pairs. Some presentations reverse the ratio or use a decreasing statistic, so the stated direction must always be checked.

How it works

Identify the model family, parameter ordering and statistic. Form the likelihood ratio for two arbitrary ordered parameter values and simplify it. Determine whether its dependence on the statistic is consistently monotone over the relevant support. In suitable one-parameter settings, this structure allows a rejection region based on a threshold in the statistic. The Karlin-Rubin result establishes uniformly most powerful one-sided tests under its assumptions. Select the threshold to control the null rejection probability. Discrete statistics may require randomization at a boundary to achieve an exact requested size, rather than only a conservative test.

Three practical examples

Binary outcomes: compare success probabilities using a count of successes from a fixed number of trials.
Count modeling: investigate one-sided evidence for a higher rate in an appropriate count distribution.
Measurement analysis: construct a one-sided test of a location parameter when the model has the required likelihood-ratio ordering. The conclusion depends on the family and assumptions, not just on an observed upward trend in data.

Think of it this way

Think of a scale that orders evidence rather than physical weight. Moving upward on the scale consistently favors the larger parameter relative to the smaller one. The scale does not directly report the probability that a hypothesis is true, and its calibration depends on the statistical model being used.

Worked example

Let K count successes in four independent Bernoulli trials. Compare p0 = 0.25 with p1 = 0.5. The binomial coefficients cancel in the likelihood ratio, leaving L1/L0 = (0.5/0.25)^K times (0.5/0.75)^(4-K) = (16/81) times 3^K. For K = 0, 1, 2, 3 and 4, the ratios are approximately 0.1975, 0.5926, 1.7778, 5.3333 and 16. Each extra success multiplies the ratio by three. A rule rejecting for K at least 3 has null probability 13/256, approximately 0.05078, at p = 0.25. Its rejection probability at p = 0.5 is 5/16 = 0.3125. This original calculation demonstrates the ordering and a particular threshold's operating characteristics. The unrandomized rule slightly exceeds a 5% size target at the boundary and must not be described as an exact 5% test.

Illustrative case study

In a fictional case, Spruce Quality studies whether a process success rate exceeds a specified benchmark. Analysts first check whether the independent-trial model is reasonable rather than assuming that successive items are unrelated. They use the success count to describe the direction of evidence and calculate exact threshold probabilities for the planned sample size. When the available threshold does not match the desired error rate exactly, the report explains the discrete limitation instead of rounding it away. The team also reports the limited power of a very small sample.

Common misconceptions and mistakes

A monotone likelihood ratio is not the same as a monotone likelihood function in the parameter. A density can rise and fall while its ratio to another density is monotone in the statistic. The property does not imply that every possible testing problem has a uniformly most powerful solution, especially for two-sided alternatives. A likelihood ratio is not a posterior probability without a prior and the rest of the Bayesian calculation. Checking one parameter pair numerically does not prove the property for an entire family. Dependence or model misspecification can invalidate operating characteristics derived under independence.

Business use

Businesses benefit from this concept when a defensible statistical model permits a transparent threshold rule for a one-sided question. Document the null hypothesis, parameter range, statistic and error-rate calculation. Explain whether a threshold is exact, conservative or randomized, and examine power at practically important alternatives. Do not equate mathematical optimality within a restricted model with operational superiority under all conditions. Model checks, measurement quality and the consequences of both error types remain necessary when translating a statistical test into a business decision.

Check your understanding

If reversing the likelihood ratio turns an increasing function into a decreasing one, has the underlying evidence changed?

Answer: No. The direction changes because the numerator and denominator are reversed, so the convention must be stated.

Related terms

Sufficient statistic
Uniformly most powerful test
Likelihood ratio
Karlin-Rubin theorem
one-sided hypothesis
test size
statistical power

Learn next

Review likelihood ratios and the Neyman-Pearson lemma, then derive the ordering for a binomial count. Calculate threshold probabilities explicitly before applying an asymptotic approximation.

Sources

Karlin and Rubin, Distributions Possessing a Monotone Likelihood Ratio: https://www.tandfonline.com/doi/abs/10.1080/01621459.1956.10501355 Fadoua Balabdaoui, On Hypothesis Testing, Karlin-Rubin theorem: https://stat.ethz.ch/~fadouab/OnHypothesisTesting.pdf

Previous: Monolithic Machine Learning Pipeline Next: Monotone spline