Logistic regression: translate odds coefficients back to risk

Last updated: 7 Oct 20265 min read

tutorial

A logistic coefficient changes log odds; its exponent is an odds ratio, not an automatic risk ratio or causal effect.

Identify the outcome and comparison

A claims team models whether a submitted claim needs manual verification within seven days. Define the eligible claims, prediction-time fields, label maturity and the contrast of interest. A logistic model keeps fitted probabilities between zero and one by modeling log odds as a linear function of inputs. A coefficient compares cases at fixed included covariates; it does not describe a guaranteed change from intervening on that field. Coefficient interpretation still begins with the design.

Keep odds and probability distinct

If the baseline probability is 0.20, its odds are 0.25. Multiplying those odds by 1.8 gives 0.45, corresponding to probability about 0.31, not 0.36. The probability difference depends on baseline risk, so one odds ratio does not map to one risk difference for every claim. The code performs that conversion explicitly. Model fitting and log loss are related, while this lesson focuses on interpretation and inference.

Audit data and uncertainty

Inspect class counts, repeated customers, case-mix shifts and complete separation. A huge coefficient with a huge standard error can indicate sparse outcome support rather than a dramatic stable effect. An outcome collected only for reviewed claims produces selective labels; a well-fitted model on those rows need not describe all submissions. Rare outcome counts and observation bias deserve explicit reports.

Report useful quantities

Show predicted risks for a few prespecified, supported claim profiles, with uncertainty and calibration on a future cohort. Include baseline prevalence and action threshold so the odds ratio is not left to stand alone. If the goal is an intervention effect, use a causal design. The project compares interpretation of binary risk with a separate count-rate model; calibration tests whether predicted probabilities match observed frequencies.

Implementation

from math import exp, log

def risk_after_odds_multiplier(baseline_risk, odds_multiplier):
    if not 0 < baseline_risk < 1 or odds_multiplier <= 0:
        raise ValueError("invalid risk or odds multiplier")
    baseline_odds = baseline_risk / (1 - baseline_risk)
    changed_odds = baseline_odds * odds_multiplier
    return changed_odds / (1 + changed_odds)

candidate_risk = risk_after_odds_multiplier(0.20, 1.8)
assert round(candidate_risk, 6) == round(0.45 / 1.45, 6)
assert candidate_risk != 0.20 * 1.8
assert round(exp(log(1.8)), 6) == 1.8

Performance and operating cost

The odds conversion is O(1) time and space. Fitting a logistic model costs repeated passes through the data and depends on sample and feature size. Collecting representative mature labels and auditing sparse groups are more important for a trustworthy estimate than the cost of this arithmetic.

Common Mistakes

Reading an odds ratio as a risk ratio.
Calling an observational coefficient the effect of changing the claim field.
Ignoring sparse-event separation and unstable coefficient uncertainty.
Validating only on reviewed claims while describing all submitted claims.