Correlation and Causation in Predictive Reports

Prediction Is Not an Explanation of Cause

A predictive model estimates outcomes from patterns in observed data. If a predictor is useful, that establishes that it contains information about the outcome in the data and setting studied; it does not, by itself, establish that changing the predictor would change the outcome. That distinction matters when a report moves from a statement about model performance to a recommendation about what an organization should do.

For example, a model might find that customers who receive more service calls tend to have longer claim durations. It may use call count to predict duration well. But the calls might be a response to a claim already taking longer, rather than the reason it takes longer. A report that says “more calls increase claim duration” makes a causal claim the predictive association alone cannot justify.

Definition: An association is a pattern of joint variation between a predictor and an outcome in observed data. A causal effect describes how the outcome would differ under a specified intervention that changed the predictor, compared with an alternative intervention, for an appropriate target population. An association can be useful for prediction without identifying a causal effect.

In observational data, predictor values are not necessarily assigned independently of other characteristics. Those characteristics may also be related to the outcome. A model can therefore capture a mixture of relationships: a direct relationship, differences in the populations being compared, reverse direction, or other shared influences. Good predictive performance does not separate these explanations.

Why an Aggregate Association Can Mislead

A third variable can make two quantities appear associated even when their within-group relationships are weak or point in a different direction. This is one reason to be careful with broad claims based on an aggregate comparison. The total covariance identity makes the issue precise.

Result: For random variables \(X\) and \(Y\) with finite second moments, and a grouping variable \(Z\), \(\operatorname{Cov}(X,Y)=E[\operatorname{Cov}(X,Y\mid Z)]+\operatorname{Cov}(E[X\mid Z],E[Y\mid Z])\). The first term summarizes within-group covariance; the second reflects how group averages of \(X\) and \(Y\) vary together.

To see why this holds, write \(X-E[X]\) as \((X-E[X\mid Z])+(E[X\mid Z]-E[X])\), and write the corresponding decomposition for \(Y-E[Y]\). Multiply the two expressions and take expectations. Each cross-term has expectation zero: conditional on \(Z\), the conditional residual, such as \(X-E[X\mid Z]\), has mean zero. The remaining terms are the expected within-group covariance and the covariance of the conditional means. Thus the overall association can reflect differences between groups, even if the within-group pattern does not match it.

This identity does not tell us which explanation is true in a particular study. It does explain why an aggregate comparison alone is not enough to justify a causal statement. Adjusting for measured variables may change an association, but adjustment does not guarantee that every relevant difference has been measured or handled correctly.

Worked Example: An Aggregate Relationship Reverses Within Groups

A hypothetical insurer reviews a voluntary coaching program and subsequent missed-payment outcomes. Customers are classified as higher risk or lower risk using information recorded before the program. In the higher-risk group, 18 of 30 coached customers and 8 of 10 uncoached customers miss a payment. In the lower-risk group, 1 of 10 coached customers and 9 of 90 uncoached customers miss a payment. Calculate the outcome rates in each group and overall. Then assess whether the aggregate comparison supports the claim “coaching increases missed payments.”

Method. Calculate each rate as the number with a missed payment divided by the number in that comparison group. For the aggregate rates, add the relevant counts before dividing.

Higher-risk group. The coached rate is \(18/30=0.600\), or 60.0%. The uncoached rate is \(8/10=0.800\), or 80.0%. The coached rate is 20 percentage points lower.

Lower-risk group. The coached rate is \(1/10=0.100\), or 10.0%. The uncoached rate is \(9/90=0.100\), also 10.0%. The rates are equal in this group.

Aggregate coached group. There are \(30+10=40\) coached customers and \(18+1=19\) missed payments. The rate is \(19/40=0.475\), or 47.5%.

Aggregate uncoached group. There are \(10+90=100\) uncoached customers and \(8+9=17\) missed payments. The rate is \(17/100=0.170\), or 17.0%. The aggregate rate is therefore \(47.5\%-17.0\%=30.5\) percentage points higher for coached customers.

The aggregate association points in the opposite direction from the within-group comparisons: the coached rate is lower among higher-risk customers and equal among lower-risk customers, but the coached group has a much higher overall rate. The coached customers are disproportionately in the higher-risk group: \(30/40=75\%\), compared with \(10/100=10\%\) of uncoached customers. The aggregate difference does not establish that coaching increases missed payments.

Report wording. A defensible statement is: “In this dataset, coached customers had a higher overall missed-payment rate; the comparison differed by recorded risk group, so the aggregate association should not be interpreted as evidence that coaching increases missed payments.” The common mistake is to report the 30.5-percentage-point aggregate difference as a causal effect. The group-specific rates show why that conclusion is unsupported.

Interpreting Predictors Without Turning Them into Causes

A fitted model describes an association under its specified form and included variables. For instance, in a logistic regression, the model for a binary outcome \(Y\) with predictor values \(x_1,\ldots,x_p\) is

\log\left(\frac{P(Y=1\mid x_1,\ldots,x_p)}{1-P(Y=1\mid x_1,\ldots,x_p)}\right) =\beta_0+\beta_1x_1+\cdots+\beta_px_p.

Here the model assumes that the conditional log-odds are represented by the specified linear predictor. For a one-unit increase in \(x_j\), holding the other included predictors fixed, the modeled odds are multiplied by \(\exp(\beta_j)\). This is a conditional model association. The equation does not say that an intervention changing \(x_j\) would multiply an individual's odds by that amount.

The distinction remains important even when a coefficient is statistically clear, the model predicts well, or the report adjusts for several relevant variables. A causal interpretation would require a causal question and a design and assumptions that support that interpretation. Observational prediction alone does not establish those conditions.

Worked Example: Rewriting a Logistic-Model Finding

A hypothetical model predicts whether an application receives a high-cost review. The predictor “document submitted late” is an indicator \(X\), equal to 1 for late submissions and 0 otherwise. Holding the other included predictors fixed, the fitted coefficient is \(\beta_X=\ln(1.5)\). For an application with \(X=0\), the model predicts a 20% probability of high-cost review. Calculate the modeled probability for \(X=1\), and revise the draft statement: “Late submission causes a 7.27-percentage-point increase in high-cost review.”

Method. Use the logistic-model odds relationship. The fitted coefficient implies an odds ratio of \(\exp(\beta_X)\), conditional on the other predictors in the model.

For \(X=0\), the modeled probability is 0.20. The corresponding odds are \(0.20/(1-0.20)=0.20/0.80=0.25\).

The odds ratio is \(\exp(\ln(1.5))=1.5\). Therefore, for \(X=1\), the modeled odds are \(0.25\times1.5=0.375\). Convert these odds back to a probability:

\frac{0.375}{1+0.375}=\frac{0.375}{1.375}=\frac{3}{11}\approx 0.2727.

The modeled probability is about 27.27%, a difference of \(27.27\%-20.00\%=7.27\) percentage points. The arithmetic checks: the starting odds of 0.25, multiplied by 1.5, give odds of 0.375 and a probability below 1.

Report wording. “Holding the other included predictors fixed, late submission was associated with 1.5 times the modeled odds of high-cost review. At the stated reference profile, the modeled probability was about 27.27% for a late submission and 20.00% otherwise.” This describes the fitted model; it does not establish that making a submission late causes the outcome. The draft’s probability difference is numerically consistent with this profile, but its causal verb is not justified by the model alone. The common mistake is treating an adjusted coefficient as an intervention effect.

Timing and Reverse Causation

A causal interpretation also requires attention to when variables are measured. If an outcome process has already begun before a predictor is recorded, the outcome may influence the predictor. A model can still use that predictor to estimate the outcome, but a report should not assume the predictor came first in the causal chain.

Worked Example: Contact Counts and Claim Duration

A hypothetical claims team compares two groups of closed claims. For 40 claims closed within 10 days, the recorded total is 48 customer contacts and the average duration is 8 days. For 20 claims open for more than 30 days, the recorded total is 120 contacts and the average duration is 40 days. Calculate average contacts per claim in each group and describe what the comparison can and cannot support.

Method. Divide total contacts by the number of claims in each group. Then compare the group averages, while noting that the recorded contacts accumulated during the claims process.

For the claims closed within 10 days, average contacts per claim are \(48/40=1.2\). For the claims open more than 30 days, average contacts per claim are \(120/20=6.0\). The difference is \(6.0-1.2=4.8\) contacts per claim; the longer-duration group has \(6.0/1.2=5\) times the average.

A suitable report statement is: “In these closed claims, claims open more than 30 days had an average of 6.0 recorded contacts, compared with 1.2 for claims closed within 10 days.” It would be unsupported to conclude that reducing contacts would shorten claim duration. Longer-running claims may create more occasions for customers to make contact, and other claim characteristics may influence both duration and contact volume. The common mistake is to treat an observed difference as evidence for the effect of an intervention, without establishing that the proposed cause precedes the outcome or ruling out other explanations.

A Practical Standard for Report Language

A careful report distinguishes the data-supported statement from the causal interpretation a reader might infer. Name the population and setting, identify whether the finding is a model association or a direct comparison, and qualify the conclusion to match the evidence. Avoid causal verbs such as “causes,” “leads to,” “reduces,” and “prevents” unless the analysis was designed to support a causal claim.

State what was observed.
Name the predictor, outcome, population, and relevant period. Distinguish an observed group difference from an adjusted model association.

Check the direction and timing.
Ask whether the predictor was recorded before the outcome and whether the outcome could have influenced the predictor.

Consider alternative explanations.
Ask whether groups differ in characteristics related to both the predictor and outcome, and whether those differences were measured and addressed.

Choose wording that matches the evidence.
Use “was associated with,” “had a higher observed rate,” or “was a predictor of” for predictive evidence. State that the analysis does not establish causation when that limitation matters to the decision.

A causal claim may be supportable in some studies, but it needs more than a predictive model’s fit or a nonzero coefficient. The study must define the intervention being compared and justify why the comparison identifies its effect for the target population. Randomized assignment can support that goal when implemented appropriately; observational analyses require their own defensible design and assumptions. When those requirements are not established, keep the report at the level of association.

This discipline is useful in actuarial reporting because a predictive finding may inform triage or forecasting without showing that changing a predictor will improve outcomes. Separating prediction from intervention avoids overstating what the evidence establishes and helps decision-makers see what further evidence a proposed action would require. This reporting distinction is aligned with the SOA Exam PA (Predictive Analytics) syllabus.

Key takeaway: A predictor can be informative without being a cause. Report the observed or modeled association precisely, check for group differences and timing concerns, and do not describe an association as an intervention effect unless the study design and assumptions support that conclusion.

Check Your Understanding

For each situation, identify the defensible interpretation or wording. The final answer gives a concise check.

A hypothetical model finds that customers with a recorded address change have 1.8 times the modeled odds of a missed payment, conditional on included predictors. Rewrite “An address change causes missed payments.” (Answer: “A recorded address change was associated with 1.8 times the modeled odds of a missed payment, conditional on the included predictors; this finding alone does not establish causation.”)
In a voluntary training program, 12 of 30 trained workers and 8 of 40 untrained workers make a processing error. Calculate both rates and state whether the difference alone proves training is harmful. (Answer: Trained: \(12/30=40\%\); untrained: \(8/40=20\%\); the 20-percentage-point observed difference does not by itself prove a harmful causal effect.)
A model uses the number of service calls made during a policyholder’s claim to predict whether the claim will remain open after 30 days. Is a strong association evidence that calls cause claims to remain open? Give one alternative explanation. (Answer: No; claims that are already complex or slow to resolve may generate more calls and also remain open longer.)
A report says a new reminder policy reduced late payments because late-payment rates were lower after rollout. Name one issue that must be addressed before attributing the change to the policy. (Answer: A concurrent change, such as a seasonal or population shift, could explain some or all of the difference; the before-and-after comparison alone does not isolate the policy effect.)
Within each of two customer groups, a predictor-outcome relationship is weak, but the groups have different average predictor values and outcome rates. What should an analyst check before reporting a single overall association? (Answer: Check group-specific comparisons and how group composition contributes to the aggregate association; group differences may distort or reverse the overall pattern.)

What You'll Learn