Why a Renewal Association Is Not Automatically a Price Effect
A renewal model can estimate that policies receiving larger price increases renew at a lower rate than policies receiving smaller increases. That association is useful for forecasting, but it does not by itself establish that the increase caused the difference. The groups may have differed before the price action: customers selected for larger changes may have different risk characteristics, tenure, coverage, shopping behavior, or access to alternatives.
This distinction matters when using a model to describe likely behavior under a new intervention. A prediction based on historical associations is safest when the new situation resembles the data used to fit the model. When prices or other conditions change in a way that also changes who is exposed, or how customers respond, historical relationships may not transport. These interpretation limits are aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.
The actions must be defined precisely. “A price increase” could mean a particular percentage change in the total premium, a change in one coverage component, or a change relative to a different benchmark. It is also important to specify the outcome: renewal with the same policy, retention with the same insurer after a product change, or retention in a broader group. Without clear definitions, even a carefully calculated difference may answer the wrong question.
Selection and Measured Customer Mix
Price actions are often not assigned at random. A rating or business rule may give larger changes to certain risks, territories, coverage combinations, or tenure groups. Customers who receive different actions can therefore have different expected renewal rates even before the action is considered. This is selection: membership in the price-change group is related to characteristics that also relate to renewal.
Stratifying or standardizing by measured characteristics can show whether a crude comparison is driven by different observed mix. It does not automatically remove differences in unmeasured characteristics. Standardization answers a descriptive question: what would the group rates look like if both groups were summarized using the same chosen weights? A causal interpretation requires additional assumptions.
Worked Example: Crude Versus Standardized Renewal Rates
A hypothetical insurer compares policies with an increased renewal price to policies with no increase. The records are grouped by whether a customer obtained a competing quote. Among increased-price policies, 30 of 60 customers with a quote renew, and 32 of 40 without a quote renew. Among unchanged-price policies, 20 of 40 with a quote renew, and 48 of 60 without a quote renew. Calculate the crude rates and equal-weight standardized rates, then interpret the comparison.
Method. First combine the counts within each price group for the crude rates. Then calculate each quote-status rate and average the two status-specific rates using equal weights for both price groups.
For the increased-price group, renewals total \(30+32=62\) out of \(60+40=100\). Its crude renewal rate is \(62/100=62.0\%\). For the unchanged-price group, renewals total \(20+48=68\) out of \(40+60=100\), giving \(68/100=68.0\%\). The crude difference, increased minus unchanged, is \(62.0\%-68.0\%=-6.0\) percentage points.
Among customers with a quote, the increased-price rate is \(30/60=50.0\%\), and the unchanged-price rate is \(20/40=50.0\%\). Among customers without a quote, the corresponding rates are \(32/40=80.0\%\) and \(48/60=80.0\%\). Thus, using equal weights of \(0.5\) for the two quote-status groups, each standardized rate is \(0.5(50.0\%)+0.5(80.0\%)=65.0\%\).
The observed crude difference is entirely associated with the different quote-status mix in these data: 60% of the increased-price group has a quote, compared with 40% of the unchanged-price group, and the quote group has lower renewal in both price groups. The equal-weight standardized rates match. This is not proof that price has no causal effect: quote status may be measured imperfectly, other differences may remain, and the observed data do not reveal both potential outcomes for each customer.
Common mistake to avoid. A wrong calculation might use \(30+32=62\) but report \(62/100=62.5\%\), or average the two strata as though their denominators were equal without first finding the stratum rates. The correct crude increased-price rate is \(62.0\%\); the correct equal-weight standardized rate is \(65.0\%\). The crude difference is \(-6.0\) percentage points, while the standardized difference is \(0.0\) percentage points.
Potential Outcomes and the Selection Component
A potential-outcome formulation makes the limitation precise. Let \(T=1\) denote receiving a specified price action and \(T=0\) denote the comparison action. Let \(Y\) be the observed binary renewal outcome. The consistency condition says that the observed outcome equals the potential outcome under the action actually received: \(Y=Y(1)\) when \(T=1\), and \(Y=Y(0)\) when \(T=0\). For the group receiving the action, the average treatment effect on the treated is \(E[Y(1)-Y(0)\mid T=1]\).
Under consistency, the observed difference in renewal rates can be split into the effect for the treated group and a selection component. The derivation is an add-and-subtract step:
The first bracket is the average causal effect among those who received the action. The second is the difference in their average outcomes under the comparison action: a group-composition or selection component. It would be zero if the treated and comparison groups would, on average, have had the same renewal rate under the comparison action. That condition is not guaranteed by observing similar records or by fitting a predictive model.
For a causal estimate from observational data, analysts commonly need conditional exchangeability: after conditioning on measured covariates, treatment assignment is independent of the relevant potential outcomes. They also need consistency, a well-defined intervention, and positivity: each covariate pattern being analyzed must have a nonzero chance of receiving each action being compared. These conditions are substantive, not merely technical. If an important driver of both price assignment and renewal is unobserved, adjustment for recorded variables alone may not establish exchangeability.
Worked Example: Decomposing an Observed Difference
In a hypothetical observational study, 72 of 100 policies receiving a specified increase renew, compared with 80 of 100 policies receiving no increase. Suppose, for illustration, an external design or a credible comparison establishes that the increased-price group would have had 76 renewals per 100 under no increase. Calculate the observed difference, the effect among increased-price policies, and the selection component.
Method. Express each rate as a proportion, then apply the decomposition. The assumed 76% is the increased-price group’s counterfactual comparison rate; it is not directly observed in an ordinary one-action-per-policy dataset.
The observed increased-price rate is \(72/100=72\%\), and the observed no-increase rate is \(80/100=80\%\). Therefore the observed difference is \(72\%-80\%=-8\) percentage points.
For the increased-price group, the assumed comparison-action rate is \(76\%\). Its effect on the treated is \(72\%-76\%=-4\) percentage points. The selection component is \(76\%-80\%=-4\) percentage points: even under no increase, the increased-price group is assumed to have had a lower renewal rate than the comparison group.
The check is \(-4+(-4)=-8\) percentage points, which matches the observed difference. This calculation illustrates why an observed \(-8\)-point association need not equal an \(-8\)-point causal effect. In real observational data, the treated group’s counterfactual rate is not directly available, so its value must be identified by a credible design or supported by assumptions and analysis.
Unobserved Alternatives and Customer Responses
Customers respond to more than the price shown in a renewal record. They may compare offers from other insurers, change coverage or deductibles, switch to another product, contact an agent, or lapse. A competing offer may be unobserved even when the policy file records no quote. Prior shopping, perceived service, financial circumstances, and willingness to tolerate a price change may also be missing or poorly measured. If such factors affect both which price action a customer receives and whether the customer renews, the observed association can be confounded.
The alternative set matters as well. A customer’s decision may depend on prices and terms available elsewhere, not only the insurer’s own renewal premium. If those alternatives change over time, a relationship estimated in one period may not describe behavior in another. Even a well-calibrated prediction for the historical environment is not automatically a causal forecast under a different competitive environment or a different intervention.
The outcome definition can conceal meaningful responses. A binary same-policy renewal indicator treats a customer who renews with reduced coverage the same as one who renews unchanged, and may count a switch to another product as a nonrenewal. The right measure depends on the question, but one measure should not be interpreted as another. Retention alone also does not establish the change in premium, coverage, expected loss, or contribution.
Worked Example: Defining the Retention Outcome
A hypothetical renewal cohort contains 120 policies offered a price increase. Of these, 72 renew the same policy at the same coverage, 18 renew after reducing coverage, 12 move to another product at the same insurer, and 18 leave the insurer. Calculate same-policy, same-coverage renewal and retention with the insurer under a definition that counts any of the first three outcomes.
Method. Use all 120 offered policies as the denominator for both outcomes, but include different outcome categories in the numerators according to each definition.
Same-policy, same-coverage renewal is \(72/120=60.0\%\). Insurer retention under the stated broader definition is \((72+18+12)/120=102/120=85.0\%\). The remaining \(18/120=15.0\%\) leave the insurer; the four outcome counts check because \(72+18+12+18=120\).
The 25-percentage-point difference between the two measures, \(85.0\%-60.0\%\), reflects customers who stayed with the insurer but changed coverage or product. It does not show that the price increase caused either outcome: there is no comparison group or counterfactual here. Nor does 85.0% insurer retention imply that premium or risk exposure was retained unchanged.
Interpreting Predictions With Appropriate Caution
A renewal model can still be valuable when its purpose and interpretation are clear. It may predict renewal under the historical process, identify patterns associated with lapse, or support a forecast when the pricing and customer environment are sufficiently similar to the training data. The key is not to convert predictive accuracy into a causal claim. Calibration, ranking performance, or validation across renewal periods assesses aspects of prediction; none alone demonstrates that changing a price will cause the predicted response.
When interpreting a price association, ask whether the comparison groups were formed by a rule related to expected renewal; whether relevant customer and policy characteristics were measured before the action; whether plausible alternatives and shopping behavior are observed; and whether the action and outcome are defined consistently. If the conclusion depends on unobserved factors, state that limitation rather than treating adjustment as a guarantee.
Specify the price action, comparison action, target population, time period, and renewal outcome.
Compare pre-action characteristics and look for selection, sparse overlap, or groups that are not comparable.
Report observed differences as associations unless a design and defensible assumptions support causal interpretation.
Explain which responses and alternatives are observed, and avoid extending the result to outcomes the data do not measure.
A common exam trap is to interpret a lower renewal rate in the larger-increase group as the causal effect of increasing prices. The correction is to identify the observed comparison, examine selection and measured mix, and state what further assumptions or evidence would be needed for a causal conclusion. Standardization can clarify a measured-mix comparison, but it cannot correct unmeasured selection by arithmetic alone.
Check Your Understanding
For each problem, distinguish the quantity calculated from the causal conclusion, if any, that the data support.
- A hypothetical portfolio has 45 renewals among 75 policies receiving a price increase and 56 renewals among 80 policies with no increase. Calculate both crude rates and their difference in percentage points, increased minus unchanged. (Answer: 60.0% and 70.0%; the difference is \(-10.0\) percentage points, an observed association rather than a causal effect.)
- In a price comparison, quote-status renewal rates are 40% and 70% for the increased-price group, and 40% and 70% for the unchanged-price group. Standardize each group using weights of 0.25 on the first status and 0.75 on the second. (Answer: both standardized rates are \(0.25(40\%)+0.75(70\%)=62.5\%\); this does not rule out unmeasured confounding.)
- A model reports a 6-percentage-point lower renewal rate for policies receiving larger increases. Name one condition under which that difference could be interpreted causally, and one reason the condition might fail. (Answer: for example, conditional exchangeability given measured pre-action covariates would be needed; it may fail if unobserved shopping behavior affects both price assignment and renewal.)
- Among 160 policies offered an increase, 96 renew unchanged, 24 renew with reduced coverage, 16 switch products within the insurer, and 24 leave. Calculate same-coverage renewal and insurer retention counting the first three outcomes. (Answer: \(96/160=60.0\%\) same-coverage renewal; \((96+24+16)/160=85.0\%\) insurer retention.)
- An analyst observes that customers with larger increases have lower renewal and concludes that reducing each increase by one percentage point will raise renewal by a fixed amount. Give two distinct limitations of that conclusion. (Answer: the observed groups may differ because of selection or unmeasured confounding; the association may not transport to a changed price action, competitive environment, or customer response.)