Project: compare claims centers under a shared severity mix

Last updated: 7 Oct 20265 min read

project

Reconcile severity coding, absent cells and changing workload mix before ranking claims centers by escalation rate.

Set the benchmark population

An operations lead wants to rank two claims centers by avoidable escalation rate. One center receives more complex claims. Freeze the eligible claim frame, escalation definition, observation window and pre-assignment severity classes, then choose a common reference distribution reflecting the future portfolio. Preserve each center’s actual workload mix for capacity planning. A standardized rate answers how rates compare under the reference mix, not how many escalations each center will handle tomorrow. The standardization lesson defines the calculation.

Catch a classification change

Midway through the quarter, one center changed its severity rule; claims previously marked complex became routine. The first export mixes old and new labels, so the stratum rates are not comparable. Reconstruct severity under one stable pre-outcome rule where possible and report the share that cannot be harmonized. One center has no observed claims in a small high-severity class. Do not insert a zero rate; combine classes only with a predeclared rationale or hold the comparison. The missingness ledger records classification gaps.

Compare crude and common-mix estimates

Show actual counts and crude rates, then compute stratum rates and a direct standardized rate using the same weights for both centers. If ordering reverses, explain the composition arithmetic with denominators. Carry uncertainty from sparse rates and from any estimated target weights. Explore whether routing or staffing created additional unmeasured differences before suggesting a process change. The reversal lesson explains why two valid summaries can answer different questions.

Make the release decision auditable

The packet contains source-query version, case definitions, severity-rule history, observed support by center and stratum, reference weights, rate intervals and both crude and standardized results. The gate below blocks a ranking when an eligible reference stratum lacks support or the classification change remains unreconciled. A passing gate starts review; it cannot turn an observational center comparison into a randomized effect. The target frame and small-cell uncertainty remain visible.

Implementation

def claims_center_gate(audit, limits):
    if audit["severity_rule_unreconciled"]:
        return "hold:severity-contract"
    if audit["unsupported_target_strata"]:
        return "hold:target-support"
    if audit["smallest_stratum_count"] < limits["minimum_cell_count"]:
        return "review:sparse-cells"
    if not audit["reference_mix_predeclared"]:
        return "hold:reference-mix"
    return "review:scoped-center-comparison"

limits = {"minimum_cell_count": 23}
audit = {"severity_rule_unreconciled": 1,
         "unsupported_target_strata": 0, "smallest_stratum_count": 31,
         "reference_mix_predeclared": True}
assert claims_center_gate(audit, limits) == "hold:severity-contract"
assert claims_center_gate({**audit, "severity_rule_unreconciled": 0},
                          limits) == "review:scoped-center-comparison"

Performance and operating cost

The gate is O(1), while assembling a stable case and severity ledger is O(n) expected time across claims. Standardization across s strata is O(s). A fast crude rank can be actively misleading when centers serve different case mixes.

Common Mistakes

Assigning zero to a severity class with no observations.
Mixing old and new severity definitions in the same stratum.
Publishing only adjusted rates while hiding actual queue burden.
Describing an observational adjusted contrast as a causal center effect.