Why an Observed Frequency Shift Needs Diagnosis
A change in claims per exposure is a signal to investigate, not by itself proof that the insured risk changed. A new reporting channel may capture claims that were previously missed; a coding revision may bring a claim type into the count; a claims system conversion may create duplicates. Each can increase the recorded count without increasing the number of claim events that actually occurred.
This tutorial focuses on distinguishing those operational effects from a genuine change in incidence. The earlier material on frequency trends and portfolio mix provides the context for comparing counts over time: exposure and mix matter, but even a carefully standardized comparison can mislead if the underlying data process changed.
A useful diagnosis asks whether the numerator means the same thing in both periods. Check how claims entered the system, when they were counted, which events qualified, how records were deduplicated, and whether claim status or data extraction rules changed. Then compare against independent evidence, such as audit samples, system logs, stable operational measures, or a parallel count constructed under consistent rules.
A Simple Model of Reporting Completeness
One way to formalize a reporting change is to suppose that eligible claims are recorded independently with probability \(q_t\) in period \(t\). Let \(\lambda_t\) be the underlying incidence rate per exposure. Assume the number of eligible events is Poisson with mean \(E_t\lambda_t\), each event has the same recording probability \(q_t\), and there are no false positives or duplicate records. Under these assumptions, the expected observed count is \(E_t\lambda_tq_t\), so the expected observed frequency is \(\lambda_tq_t\).
The key implication is that observed frequency combines risk and recording. If \(q_t\) changes, observed frequency can change even when \(\lambda_t\) does not. When \(q_t\) is known or estimated reliably, a simple adjusted rate is \(N_t^{\text{obs}}/(E_tq_t)\). This adjustment is only as credible as its assumptions and the estimate of \(q_t\).
This expectation argument justifies the rate relationship. Under the stated Poisson and independent-recording assumptions, independent thinning also means \(N_t^{\text{obs}}\) is Poisson with mean \(E_t\lambda_tq_t\). The simple adjustment does not automatically apply when missed claims are concentrated in particular types, recording probabilities vary materially by segment, duplicates occur, or the denominator does not match the counted events.
Operational Changes to Investigate
The investigation should identify what changed in the production of the data, then determine how that change could affect the numerator or its relationship to exposure. A useful review includes:
- Reporting access and timing: Did a new portal, call center, or notification rule change the probability or speed of reporting?
- Claim definition and coding: Did covered event types, claim-status filters, cause codes, or inclusion rules change?
- System conversion and record handling: Did migration create duplicates, split one event into multiple records, or omit records?
- Operational practice: Did claim intake, screening, or classification change, even if the contract and underlying events did not?
- Measurement consistency: Do the numerator, period of assignment, and exposure denominator refer to compatible populations and time windows?
These checks do not replace examination of genuine risk changes. A rate shift that remains after consistent definitions and credible operational adjustments may be evidence of a real change, but it should still be evaluated alongside exposure quality and portfolio characteristics. Conversely, an operational explanation is not established merely because a process changed near the same time; there should be evidence connecting the process change to the counts.
Worked Example: A New Reporting Channel
A hypothetical insurer observes 600 recorded claims on 12,000 exposure units in one year and 660 on the same exposure in the next. A review estimates that 80% of eligible events were recorded in the first year and 88% in the second. Assume the independent-recording model above applies, with no duplicates or false positives. Determine the observed change and compare adjusted rates.
Method. Calculate each observed frequency, then divide by the applicable recording probability to estimate the incidence rate under the model.
First-year observed frequency: \(600/12{,}000=0.0500\) claims per exposure.
Second-year observed frequency: \(660/12{,}000=0.0550\) claims per exposure. The observed rate ratio is \(0.0550/0.0500=1.10\), an apparent increase of 10%.
First-year adjusted rate: \(0.0500/0.80=0.0625\). Equivalently, the estimated eligible count is \(600/0.80=750\), and \(750/12{,}000=0.0625\).
Second-year adjusted rate: \(0.0550/0.88=0.0625\). Equivalently, \(660/0.88=750\), and \(750/12{,}000=0.0625\).
The adjusted rates are equal, so these data are consistent with unchanged incidence and improved recording. The calculation is checked by recovering 750 estimated eligible events in each year. Common mistake avoided: treating the raw 10% rise as a risk increase without considering the change in recording completeness.
Worked Example: A Revised Claim-Inclusion Rule
A hypothetical program recorded 750 claims on 15,000 exposure units in its first year. In the next year, it recorded 990 claims on 18,000 exposure units. The next year's data include 90 claims from an event category that was excluded under the first year's counting rule. Assume those 90 records are valid, are the only definition change, and can be removed to create a comparable count. Compare the rates under the original common definition.
First-year rate: \(750/15{,}000=0.0500\) claims per exposure.
Unadjusted next-year rate: \(990/18{,}000=0.0550\). Relative to the first-year rate, the apparent change is \(0.0550/0.0500-1=0.10\), or a 10% increase.
Comparable next-year count: \(990-90=900\), after excluding the newly included category.
Comparable next-year rate: \(900/18{,}000=0.0500\). The difference from the first-year rate is \(0.0500/0.0500-1=0\%\). The additional category accounts for \(90/18{,}000=0.0050\) claims per exposure, exactly the gap between the unadjusted and first-year rates.
The common-definition comparison shows no rate change. The arithmetic checks because \(0.0500+0.0050=0.0550\). Common mistake avoided: comparing counts that use different event definitions as if they measured the same outcome.
Worked Example: Faster Reporting at a Fixed Evaluation Age
A hypothetical organization compares claims recorded within 12 months of the relevant event. It has 10,000 exposure units in each of two periods. The earlier period has 720 claims recorded by 12 months; the later period has 810. Independent audits estimate that 80% and 90%, respectively, of eligible events have been recorded by that 12-month point. Assume these completeness estimates apply to the counted events and there are no duplicates or false positives. Diagnose the raw comparison and calculate the adjusted rates.
Earlier observed rate: \(720/10{,}000=0.0720\). Its estimated incidence rate is \(0.0720/0.80=0.0900\).
Later observed rate: \(810/10{,}000=0.0810\). The raw rate ratio is \(0.0810/0.0720=1.125\), an apparent increase of 12.5%.
Later adjusted rate: \(0.0810/0.90=0.0900\). The audit-based adjustment gives the same incidence rate in both periods. As a count check, \(720/0.80=900\) and \(810/0.90=900\) estimated eligible events.
The evidence is consistent with faster or more complete reporting by the fixed evaluation point, not a change in incidence. This is a diagnostic comparison, not a general method for estimating claim liabilities: it depends on reliable completeness estimates and the stated assumptions. Common mistake avoided: treating claims observed at the same evaluation age as equally complete when reporting speeds differ.
Reading the Evidence Carefully
A sound diagnosis separates three questions: did the recorded count change, did the process that turns events into records change, and does the change remain under a consistent measurement basis? The calculations above answer the first question and illustrate possible adjustments for the second. The third requires evidence, not just arithmetic.
For example, a completeness estimate from an audit is more useful when the audit uses the same eligibility definition across periods and can detect both missed events and duplicates. A coding bridge is more persuasive when the newly added and removed categories can be identified directly. If an adjustment relies on an unsupported assumption that the operational effect is uniform across all risks, the adjusted figure may be less reliable than the raw comparison.
When a process change cannot be quantified credibly, report the break in comparability rather than presenting a precise adjusted trend. Where possible, construct a parallel series using a stable definition, show results both before and after a documented adjustment, and state the limitations. This makes clear which conclusions follow from the data and which depend on assumptions.
Common Mistakes
- Calling every increase a risk increase. First check whether the claim definition, reporting channel, or data capture changed.
- Adjusting only the denominator. If exposure is consistent but the numerator’s capture probability changed, adjusting exposure does not correct the recording effect.
- Applying an adjustment without matching its scope. A completeness estimate for one claim type or reporting channel may not apply to all claims.
- Ignoring duplicates and false positives. The simple completeness adjustment handles missed records under its assumptions; it does not correct duplicate or ineligible records.
- Claiming causation from timing alone. A process change and a frequency change occurring together do not prove that the process caused the change.
Check Your Understanding
For each problem, distinguish the raw comparison from the operationally comparable result.
- A hypothetical portfolio has 8,000 exposure units in each year, with 360 claims recorded in the first year and 396 in the second. Estimated recording completeness is 75% and 82.5%, respectively. Assuming the independent-recording model applies, what are the adjusted rates and the implied incidence change? (Answer: Both adjusted rates are 0.0600 claims per exposure; implied incidence change is 0%.)
- A hypothetical program records 950 claims on 19,000 exposures in one year. The next year it records 1,050 claims on 20,000 exposures, including 70 claims in a category newly added to the count. If those 70 claims are the only definition change, what is the comparable rate change? (Answer: First-year rate 0.0500; comparable next-year rate \(980/20{,}000=0.0490\); decrease of 2.00%.)
- Two hypothetical periods each have 9,000 exposure units. Counts recorded by the same evaluation age are 540 and 600, with estimated completeness of 75% and 80%. What are the raw and adjusted rate comparisons? (Answer: Raw rates are 0.0600 and 0.0667, an increase of about 11.11%; adjusted rates are both 0.0800, so the adjusted change is 0%.)
- A hypothetical system records 540 claims in one period and 648 in the next, with 12,000 exposure units in each. An audit identifies 48 duplicate records in the second count and no duplicates in the first. What are the raw and deduplicated rate changes? (Answer: Raw rates are 0.0450 and 0.0540, a 20% increase; the deduplicated second rate is \(600/12{,}000=0.0500\), an 11.11% increase over the first.)
- A hypothetical portfolio has 25,000 exposure units in each of two periods. It records 1,250 and 1,375 claims, with estimated completeness of 90% and 99%. Under the independent-recording assumptions, what do the adjusted rates indicate? (Answer: Both adjusted rates are \(0.05556\) claims per exposure, approximately; the adjusted incidence change is 0%.)