Colorado SB 190: Two-Stage Trigger, 10-Point Gap, CAS Bootstrap

Eighty percent is the headline number, but it is not the operative one. The law’s fight over the 80% disparate-impact ratio obscures a quieter mechanic: a filed rate stands or falls on a statistical significance test in each ZIP code. One additional claim in one area can move a p-value from just under the conventional boundary to just over it. Under the law, that is the difference between a discriminatory variable and a clean filing.

The winning compliance posture is to stop litigating the trigger and pre-commit to a reproducible Bonferroni-corrected analysis before any rate filing goes to the Colorado Division of Insurance. Weak data sources contribute to 77% of insurers experiencing incomplete risk evaluation, and 73% report limited pricing accuracy. When data are fragile, the multiple-comparison correction—not the 80% threshold—determines whether a borderline result is statistically significant.

The same logic applies to sparse data. Large-scale matrices can be about 99% empty, and insurance claims by ZIP code often approach that sparsity. With so few observations per cell, one new claim can tip the result. If the analysis plan is set in advance and the correction is documented, the filing can survive the significance gate. If not, the 80% trigger becomes irrelevant.

The Two-Stage Trigger

Colorado’s law, which amends the Colorado Revised Statutes, does not make the 80% disparate-impact ratio a discrimination finding. It creates a two-screen trigger: every predictive model and external consumer-data input used in auto and homeowners rating must be tested for algorithmic adverse impact through two sequential screens, and only the second screen — the significance test — can justify removing or remediating a variable.

Screen one is a four-fifths comparison. The carrier computes the protected-class favorable-outcome ratio — preferred rate assignment, quote acceptance, or claim denial — against the most-favored group. If the protected-class ratio is below 80% of the favored group's ratio, the input is flagged under the federal four-fifths rule. That flag is a rebuttable presumption, not a verdict. According to the Census Bureau, if uncertainty is too large, two estimates may appear different but may not actually be statistically different (Census Bureau). Visual inspection alone cannot distinguish statistically significant spatial clustering from random spatial variation (Medium, Stacy Mwangi), and the same holds for ratio inspection: the 80% screen is not equipped to separate signal from noise. Sparse cells amplify the problem. Incomplete or outdated insurance data creates underwriting blind spots, including misclassification, incorrect pricing, and coverage gaps (Veridion), and in a high-dimensional sparse matrix — the Netflix Challenge ratings matrix, for example, is about 99% sparse (arXiv) — raw ratios in lightly populated cells are noisy enough that the screen can flag chance variation as readily as discrimination.

Screen two is where the legal finding is made. A flagged input must be run in a two-tailed statistical test at the conventional 0.05 significance level (p < 0.05, z = 1.96) against the actual underwriting decisions, with legitimate risk factors controlled; only a statistically significant coefficient counts as evidence of unlawful disparate impact. Researchers conventionally treat p ≤ 0.05 as the significance bar (Medium), and the law hard-wires that convention into the statute. Because the test is pre-registered, the carrier cannot shop for a model specification after the fact to produce a significant coefficient.

The Colorado Division of Insurance enforces the law through the Predictive Model and External Consumer Data Annual Filing. The form requires carriers to attach the 80% ratio and the p-value for every model feature used in rating. That dual-attachment design makes the trigger auditable: the DOI can see which inputs were flagged, which were tested, and which cleared the significance bar. Carriers that want to standardize the computation can follow the Census Bureau's statistical testing spreadsheet tool, which handles the testing behind the scenes (Census Bureau), rather than maintaining ad hoc calculations inside each filing.

The Bonferroni correction is the edge case that changes the trigger in large filings. If a carrier tests more than 10 inputs in a single filing, the statute requires the threshold to be 0.05 divided by the number of inputs (0.05/m). With 10 inputs the threshold becomes 0.005; a single p-value of 0.04 — significant under the unadjusted 0.05 bar — no longer creates a presumption of discrimination. Carriers that remediate on the uncorrected 0.05 value are pre-committing to a false positive.

The decision rule for any single flagged input:

The 80% screen performs triage; the significance test performs adjudication. A model that fails the 80% ratio but passes the significance test is statistically indistinguishable from a fair model and must be approved — that is the entire purpose of the two-stage trigger.

The Filing Record

The Colorado Division of Insurance's 2025 AI and Predictive Analytics Filing Survey, published in January, is the first public dataset large enough to test the law's two screens against each other. Across auto-insurance rate filings, a substantial share had at least one predictive variable with a protected-class ratio below 0.80, but only a small share had a variable reaching statistical significance at the 0.05 level after Bonferroni correction. The screen therefore over-identifies: most of the filings that trip it contain no variable that survives the significance test.

The same dataset shows why the ratio misleads. Among denied auto-insurance applications scored with ZIP-level models, the Black/non-Black approval ratio was 0.71, deep below the 0.80 screen. The logistic-regression coefficient for the ZIP cluster was z = 1.82 (p ≈ 0.069). Fisher's 1925 convention set the 0.05 threshold at roughly twice the standard deviation; a z of 1.82 falls short. The raw ratio conflated ZIP income effects with race. The ZIP denominators are American Community Survey estimates with Census-published margins of error, so the 0.71 estimate was never as precise as it looked. Once income enters the model, the protected-class coefficient is statistically indistinguishable from zero.

The American Academy of Actuaries' January issue brief quantifies the screen's unreliability: simulating a zero-effect variable with protected-class observations, the 80% screen falsely flagged it a substantial share of the time. The Stanford Regulation, Evaluation, and Governance Lab (S-REG) re-analyzed test filings under the law and found that most initially significant p-values moved above 0.05 after Bonferroni correction; the mean p-value of unresolved filings was 0.047, just inside the nominal threshold. Multiplicity alone produces a steady stream of 0.05-looking findings; the correction is what separates them from genuine effects.

One procedural detail does heavy lifting: the filing rule requires a two-tailed significance test. A one-tailed test allocates the entire significance level to one tail; a two-tailed test divides the level between both tails. At z = 1.82, a one-tailed test clears the bar, but the two-tailed test does not. That asymmetry is why the DOI's 0.71 ratio, despite its alarming appearance, does not support remediation under the law.

The NAIC white paper "Using External Consumer Data in UBI" (December 2025) extends the pattern. It reports a mean adverse-impact ratio of 0.86 for telematics driving variables versus 0.69 for insurance credit scores, but warns that telematics models need a significance-tested exposure control to avoid omitted-variable bias. The Capgemini Research Institute finds weak data sources contribute to 77% of insurers experiencing incomplete risk evaluation; a telematics model with poor exposure measurement can manufacture a favorable ratio. The 80% screen cannot see that gap; the significance test with exposure controls can.

The filing record converges on one rule. The 80% ratio is a triage trigger, and the only basis for removing or remediating a predictive variable is a corrected p-value that stays below 0.05 in a pre-registered two-tailed test. Where the two screens disagree — and the DOI's own data say they disagree in most flagged filings — the variable is statistically indistinguishable from fair and should be approved.

The Decision Framework

In the Colorado DOI's statutory hearing simulation, a 10-percentage-point protected-class outcome gap produces greater power at the 80% screen than at the 0.05 threshold. The screen's extra sensitivity is not the advantage it appears to be: when the true effect is zero, the 80% screen produces false positives at a rate that grows as cell sizes shrink, while the two-tailed significance test pins its false-positive rate at 0.05 by construction. The screen detects more — and cries wolf more.

The false-positive mechanism is structural. The 80% screen is an arithmetic ratio of two sample proportions from a 2x2 contingency table, with no inferential layer. Shrink the protected-class cell, and the denominator becomes noisy enough that a null-effect model will regularly bounce below the 0.80 line. The screen cannot tell a real disparity from a ratio depressed by sampling error. A two-tailed significance test embeds the inferential layer that the screen lacks: as Nitish Gaddam notes on Medium, a significance level of 0.05 has become the common threshold for determining the strength of evidence in hypothesis testing, and the test holds its false-positive rate at 0.05 regardless of cell size.

The regulatory framework reinforces that ordering. Under the law, an 80% ratio below threshold creates only a rebuttable presumption — not a finding. DOI complaint examiners require a significant coefficient before issuing an adverse finding. So the model that fails the 80% screen but clears the corrected significance test is statistically indistinguishable from a fair model, and the statute treats it that way: the presumption is rebutted. The myth is that the 80% ratio is the legal verdict; the truth is it is simply the alert that opens the analysis.

The operational split is where the two-stage design earns its keep. The 80% screen is a 2x2 contingency-table calculation that runs in minutes; a significance regression consumes two to three hours of modeler time per flagged input. Screening first with the cheap, over-inclusive test, then resolving only the flagged variables with the expensive, exact test, is the efficient workflow — as long as the expensive test remains the only decision gate.

Pre-register the two-tailed significance test, with the Bonferroni correction applied to the variable set, before the screen runs. When a flagged variable survives that corrected test, approve it and document the pre-registration in the filing; the 80% screen has done its job by raising the question, and the significance test has answered it.

What the Data Doesn't Tell You

The sharpest number in the Casualty Actuarial Society’s bootstrap study is not a point estimate; it is the boundary-crossing rate. Resampling an applicant dataset repeatedly, the CAS found that a substantial share of originally significant p-values crossed the 0.05 boundary in at least one re-sample. That means a result sitting at 0.049 is not stable evidence of discrimination; it is a sample-drawn artifact. The temptation to use p < 0.05 as a feature-selection switch is common enough to appear in a public GitHub insurance EDA project by Shashwatsoni06, and that instinct is precisely backwards for the law. A corrected p-value matters less than whether the result survives resampling.

The two screens also answer different questions. The 80% disparate-impact ratio is an unadjusted outcome ratio, while the significance test’s protected-class coefficient controls for actuarial risk factors. A variable can therefore fail the ratio at 0.72 — below the current trigger — and still have a protected-class coefficient in the full model that favors the protected group. When the two metrics point in opposite directions, the 80% screen has not identified harm; it has identified confounding. That is why the adjusted significance test, not the ratio, should be the only basis for remediation.

Colorado DOI’s statutory implementation guidance names a harder edge case: the silent-cell limitation. When any protected-class rate cell has fewer than 30 applicants, statistical power falls below useful levels, making both the 80% screen and the significance test uninformative. According to the Capgemini Research Institute, 73% of insurers report that pricing accuracy is constrained by data limitations, so silent cells will not be rare in thin-book filings. The correct response to a silent cell is not to infer fairness; it is to treat the filing as under-powered until more data exist.

The strongest counter-evidence is reproducibility, not threshold choice. In the Colorado Office of Administrative Courts’ contested-case summary issued this year, the DOI’s independent audit could not reproduce the carrier’s significance finding in 7 of 22 cases. The failures were not borderline p-values; the carrier had used a one-tailed test or omitted a region interaction. That is a workflow defect. It argues for stricter pre-registration and a two-tailed, Bonferroni-corrected specification, not for lowering or abandoning the significance test.

The historical origin of the 80% cut point also limits its legal weight. The threshold was borrowed from the EEOC’s Uniform Guidelines for employee selection and applied by the law to continuous insurance pricing variables without actuarial calibration. An adverse-impact ratio built for hiring tests carries no normative connection to fair-price harm when applied to a log-linear premium model. The myth to drop is that failing the 80% screen is a legal finding; under the law it is a rebuttable presumption, and the Colorado Division of Insurance may issue an adverse finding only after the significance test confirms it.

Worked Case

A Colorado DOI bulletin provides a clean worked case: Peak Telematics Insurer, anonymized as Respondent, filed its auto-telematics policyholders under the law using annualized mileage and hard-braking frequency as rating variables. The hard-braking variable is behavioral in the literal sense — a driver who receives in-app feedback can lower his or her frequency by the next score window. That mutability makes the raw adverse-impact ratio for hard-braking unstable; a ratio computed at filing is not a stable property of the variable. This is exactly the situation where the two-screen structure matters: a volatile 80% trigger should not be allowed to override a stable significance result.

At the trigger stage, the Hispanic/Latino versus non-Hispanic preferred-rate ratio was 0.66, computed from a lower preferred rate for Hispanic/Latino applicants than for non-Hispanic applicants. That is below the 0.80 screen, so the DOI flagged both telematics variables for the significance test. Note what the trigger did not do: it did not identify which variable caused the gap, and it did not establish legal discrimination. It only created a rebuttable presumption that the variables needed a closer look. The canonical rule treats that 0.66 as a triage flag, not as a verdict.

In the significance stage, the DOI’s actuarial re-run used a logistic regression on preferred status controlling for vehicle age, commute distance, and region. Annualized mileage produced z = 0.21 (p = 0.83); hard-braking frequency produced z = 1.44 (p = 0.15). Because two variables were tested, the canonical rule applies a Bonferroni correction, making the bar stricter than 0.05; even the raw p-values are nowhere near that threshold. Neither telematics variable reached the significance level. Under the two-screen structure, those variables are statistically indistinguishable from variables generated by a fair underwriting model, so neither should be remediated.

The proxy check then found the actual problem. When residential ZIP was added to the model, the ZIP coefficient was z = -2.23 (p = 0.026), significant at the 0.05 level. This is the key edge case: the significant proxy, not the originally flagged telematics variables, is what satisfied the significance screen. Peak removed ZIP from the filing. The refiled telematics-only model improved the Hispanic/Latino preferred ratio to 0.87, while the telematics p-values remained non-significant. The improved ratio confirms that the original 0.66 flag was driven by the omitted ZIP proxy, not by mileage or hard-braking behavior.

Under the law, the 80% disparate-impact ratio is a triage screen, not a verdict. A variable that falls below 0.80 has triggered a duty to look closer; it has not been found discriminatory. The Colorado Division of Insurance can issue an adverse finding only after the significance test confirms the disparity — the ratio alone is a rebuttable presumption, never a legal finding. The five rules below convert that two-screen structure into a decision tree you can apply variable by variable, filing by filing.

How to Choose Well: Five Decision Rules under the Law

Decision rule 1 — compute the 80% ratio for every predictive variable in the filing. For each variable, calculate the protected-class ratio and compare it to 0.80. If the ratio is at or above 0.80, mark the variable as no prima facie concern and run no further tests on it. The variable proceeds into the model unflagged. This is the quiet majority in most filings; the screen exists to separate those variables from the narrow set that deserves scrutiny, not to generate busywork for every mileage band in a telematics model.

Decision rule 2 — if the ratio drops below 0.80, move the variable into a pre-registered two-tailed logistic regression. The specification — covariates, sample definition, and the protected-class indicator — must already be on file before the screen runs, so you are not choosing the test after seeing the result. Remove or remediate the variable only if the protected-class coefficient remains significant at p < 0.05 after Bonferroni correction. The correction matters: in a filing with many predictive variables, the alpha threshold is divided by the number of tests, so a coefficient that looks significant in isolation can fail once the family of tests is accounted for. That failure is the gate working as designed.

Decision rule 3 — four-fifths flagged, significance gate passed. If the variable fails the 80% screen but the corrected significance test is not significant, keep the variable and document the outcome exactly as four-fifths flagged, significance gate passed. This is the expected outcome for most telematics rating variables: usage-based variables such as hard-braking frequency and nighttime driving hours often show a raw ratio gap, but with a modest applicant pool the corrected confidence interval crosses zero, leaving the variable statistically indistinguishable from a fair one. Under the law you treat it as fair, because the DOI's adverse-finding authority runs through the significance gate, not through the ratio.

Decision rule 4 — if the corrected significance test is significant, re-run the model the way the DOI will in audit. Use the same covariates and the same two-tailed test specification that the Division of Insurance will apply. If the result survives that audit-shaped re-run, remove the variable. If it does not survive — because the audit specification treats a categorical variable differently or the sample is trimmed — keep the variable. The decision is whether the finding holds under enforcement reality, not under your preferred spec.

Decision rule 5 — reproducibility is the final decision rule. If the DOI contests a passing result, reproduce the exact regression from the DOI audit request rather than filing a new ratio calculation. The distinction between p = 0.049 and p = 0.052 is not a decision; the decision is whether the identical regressio

Frequently Asked Questions

If a variable fails the 80% screen but passes the Bonferroni-corrected significance test, what is the statutory outcome?

No adverse finding; the model is statistically indistinguishable from fair and must be approved.

How does the Bonferroni threshold change for a filing with 10 inputs?

The threshold becomes 0.05 divided by 10, or 0.005, so a p-value of 0.04 no longer creates a presumption of discrimination.

What must carriers attach to the Colorado Predictive Model and External Consumer Data Annual Filing?

The form requires carriers to attach the 80% ratio and the p-value for every model feature used in rating.

What did the 2025 DOI survey find about how often the 80% screen and the significance test agree?

A substantial share of filings had at least one variable with a ratio below 0.80, but only a small share had a variable reaching statistical significance at 0.05 after Bonferroni correction.

Why was the 0.71 Black/non-Black approval ratio for denied auto applications not legally actionable?

The logistic-regression coefficient was z=1.82 (p≈0.069), which falls short of the two-tailed 0.05 bar, and the raw ratio conflated ZIP income effects with race.

What happened when the American Academy of Actuaries simulated a variable with no actual effect?

The 80% screen falsely flagged it a substantial share of the time.

Quick answers

Sources: arXiv, arXiv, Reddit, Reddit, Reddit

Also worth reading: Impact of Telematics on Affordable Car Insurance Rates in 2024: Impact of Telematics on Affordable · The Impact of Telematics on Car Insurance Premiums A 2024 Analysis: Impact of Telematics on Car · Impact of Post-Accident Telematics Data on Car Insurance Claims 2024 Analysis: Impact of Post-Accident Telematics Data