Is this different from proportional reduction in error?

The phrases refer to the same general comparison principle; particular named statistics differ in their error rules.

Can lambda be zero when variables are associated?

Yes. Different conditional probabilities may still produce the same modal category in every group.

Does a 50% reduction mean 50% accuracy?

No. It means half the baseline errors were removed. Accuracy also depends on the baseline error rate and total number of cases.

AI Glossary / Statistics and probability / Intermediate

Proportionate reduction of error

Definition

Proportionate reduction of error is a way of interpreting predictive association by comparing errors made without particular information with errors made when that information is used. The reduction is divided by the original error. This is the same general principle often called proportional reduction in error, or PRE. Different statistics implement it using different prediction rules and error definitions, so the exact measure must always be stated.

What it means

The idea asks whether knowing one variable helps predict another, relative to a defined baseline. In categorical analysis, one example is Goodman and Kruskal's asymmetric lambda: compare predicting the overall most common response category with predicting the most common category separately within each predictor group.

This interpretation is directional. Predicting a customer's request type from a service channel is a different task from predicting the service channel from the request type. The associated reductions need not be equal. A measure can also show no improvement for a specific decision rule even when the variables are not statistically independent.

How it works

First define the response to predict and what counts as an error. Establish a baseline rule that does not use the proposed predictor. Then establish the corresponding rule that uses it. On the same data, calculate baseline error E0 and informed-rule error E1.

The proportionate reduction is (E0 − E1)/E0 when E0 is positive. For modal-category prediction, E0 equals the total count minus the largest response-category total. E1 is the sum, across predictor groups, of each group total minus its largest response-category count. Their ratio gives the directional lambda interpretation.

For other measures, replace the error definition only with clear justification. Squared-error regression and classification mistakes are not the same loss. An in-sample association measure is descriptive; claims about future predictive improvement require appropriate validation. If the baseline already makes no errors, the ordinary ratio is undefined rather than an invitation to divide by zero.

Three practical examples

A service analyst examines whether channel information reduces mistakes when predicting the most likely request category.
A researcher compares directional categorical association measures and reports which variable is being predicted.
An AI evaluation team explains a model's improvement relative to a baseline while distinguishing mistake counts from cost-weighted errors.

Think of it this way

Think of guessing the contents of labeled boxes. Without a label, you always guess the most common item overall. With the label, you can choose the most common item within that box type. The measure asks how many of your original wrong guesses this extra information removes under that specific guessing rule.

Worked example

In an original two-by-two example, channel A has 30 billing requests and 10 technical requests. Channel B has 10 billing and 50 technical requests. There are 100 requests overall: 40 billing and 60 technical.

Without channel information, predict technical for every request. This makes 40 errors. With channel information, predict billing for A and technical for B. That makes 10 errors in A and 10 in B, or 20 total.

The proportionate reduction is (40 − 20)/40 = 0.50. Channel information removes half the mistakes made by the baseline modal-category rule in this table.

The classification accuracy rises from 60% to 80%, an increase of 20 percentage points. It does not rise by 50 percentage points. The 50% figure uses baseline mistakes as its denominator. These are descriptive calculations for a constructed table, not evidence that channel changes cause request types or that the same improvement will appear in deployment.

Illustrative case study

Fictional case study: Cedar Routing evaluates whether a simple intake label helps route requests. Its analyst reports a strong proportionate error reduction using category modes, but managers assume this means the label explains the same proportion of every aspect of the data.

The analyst shows the baseline guesses, informed guesses, and error counts, then evaluates the routing rule on later requests. Costly technical misroutes receive a separate review. This fictional case illustrates how an association statistic can be made concrete without turning it into a universal measure of model quality.

Common misconceptions and mistakes

Zero modal-prediction improvement does not necessarily mean independence. If every predictor group has the same most common response category, the modal guess may never change even when the category probabilities differ substantially.

Do not assume the measure is symmetric. Switching predictor and response can change the result and its business meaning. A proportionate reduction is also different from an absolute accuracy change. Finally, all errors need not have equal costs. A rule that minimizes the number of mistakes may still make the most consequential mistakes more often, so the chosen loss must match the intended interpretation.

Business use

This framework makes association useful by connecting it to an explicit prediction task. Businesses can see what a baseline does, what information changes, and how much error remains after using that information.

For AI-related reporting, name the exact statistic and explain its baseline. Keep descriptive association separate from out-of-sample prediction and causal claims. Include a confusion table or other error breakdown when costs differ. If the aim is to evaluate probabilistic predictions, use a suitable probability-scoring measure as well; modal-category error discards information about how confident the predictions were.

Check your understanding

A baseline makes 30 errors and an informed rule makes 24. What is the proportionate reduction, and how many mistakes remain?

Answer: The reduction is (30 − 24)/30 = 0.20, or 20%. Twenty-four mistakes remain; the percentage describes improvement relative to the baseline rather than the remaining error rate among all cases.

Related terms

Proportional reduction in error (PRE)
Confusion matrix
Loss function
Proportional reduction in error
Goodman and Kruskal's lambda
categorical association
modal prediction
baseline error
directional association

Learn next

Build small contingency tables and compute baseline and group-specific modal errors. Compare both prediction directions and construct an example with association but no modal improvement. Then study other PRE measures and evaluate predictive rules on independent data.

Sources

Howell's glossary states the general residual-error reduction principle: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html

SAS's primary documentation describes asymmetric lambda's predictive interpretation; indexed passages were consulted, while direct page retrieval timed out: https://support.sas.com/documentation/cdl/en/procstat/67528/HTML/default/procstat_freq_details22.htm

The contingency table, calculations, and fictional case are original. The title is preserved as a wording variant, with categorical modal prediction used to illustrate the broader principle.

Previous: Proportionality Next: Proposal distribution