From an Actuarial Question to a Regression Model
A regression model starts with a question about how an average outcome changes with an explanatory quantity. An insurer might ask whether the average annual servicing cost changes with the size of a policy portfolio, or whether average repair severity varies with vehicle age. To write a useful model, first decide which quantity is the response and which is the predictor. Then specify how the mean response is expected to depend on the predictor, and assess which assumptions make sense for the observations being modeled.
The response is the variable whose conditional distribution or mean is being described. The predictor is the variable on which that description is conditioned. These roles follow the investigation’s question; they are not interchangeable just because both quantities are measured. A simple linear regression model describes a conditional mean that is a straight-line function of one predictor.
The units belong in the model description. If \(Y_i\) is annual cost in dollars and \(X_i\) is portfolio size in hundreds of policies, then \(\beta_1\) is measured in dollars per hundred policies. Changing the units changes the numerical coefficient and its interpretation. Also check whether \(X_i=0\) is meaningful in the setting: the intercept is mathematically part of the line even if zero is outside the observed range or has no practical interpretation.
Why the Error Form Gives a Conditional Mean
The error representation is useful because it separates the systematic mean pattern from deviations around that pattern. For the conditional mean to equal the stated line, the error must have conditional mean zero. This condition says that at each predictor value, positive and negative deviations balance on average.
Thus, if \(E[\varepsilon_i\mid X_i=x]=0\) for predictor values in the modeled range, then \(E[Y_i\mid X_i=x]=\beta_0+\beta_1x\). This is the justification for interpreting the line as a conditional mean. Without zero conditional mean, a line written in the error form does not necessarily describe the actual conditional mean.
Other assumptions concern the errors around that mean. A classical model may additionally assume constant conditional variance, independent errors across observations, and normally distributed errors conditional on the predictors. These are distinct claims: a straight-line conditional mean does not by itself guarantee constant spread, independence, or normality. Normal errors are especially relevant when using procedures that require a normal conditional response; they are not needed merely to define a linear conditional mean.
Worked Examples: Specifying the Model in Context
Worked Example: Portfolio Size and Annual Servicing Cost
A hypothetical insurer compares separate annual portfolio units. For unit \(i\), let \(X_i\) be the number of policies, measured in hundreds, and let \(Y_i\) be annual servicing cost in dollars. The proposed mean cost is \(12{,}000+4{,}500X_i\). Identify the response and predictor, calculate the mean for a unit with 600 policies, and describe assumptions that would be needed for a classical error model.
Method. Define the model in the units supplied, substitute the unit’s predictor value, and separate the mean-model claim from additional error assumptions. This avoids treating the coefficient as a cost per individual policy when it is stated per hundred policies.
The response is annual servicing cost \(Y_i\), and the predictor is portfolio size \(X_i\). Since 600 policies is 6 hundreds, \(x=6\). The conditional mean model is
At \(x=6\), the predicted conditional mean is
The intercept represents a mean cost of $12,000 at zero policies, though a zero-sized unit may not be a useful operational case. The slope means that the modeled mean cost rises by $4,500 for each additional hundred policies. It does not say that every hundred policies add exactly $4,500 to every unit’s actual cost.
If the portfolio units are treated as independent observations, a classical model would further require zero conditional mean errors, a common conditional error variance, and independence across units. Those claims need to be assessed from how units were formed. For instance, units sharing staff or systems may have dependent costs. Normal conditional errors might be a working approximation, but should not be presumed simply because the mean is linear. Check: the substituted predictor is 6 hundreds, and the computed mean is $39,000.
Worked Example: Vehicle Age and Repair Severity
In a hypothetical study of completed collision claims, \(X_i\) is the damaged vehicle’s age in years and \(Y_i\) is the paid repair amount in dollars. A proposed conditional mean is \(2{,}400-150x\). For a four-year-old vehicle with an observed repair amount of $1,950, calculate the conditional mean and the error relative to that mean. Then state what the model does and does not establish.
Method. Substitute the stated age into the conditional mean, then calculate observed response minus conditional mean. This avoids confusing a residual for one claim with the model’s mean response.
The response is paid repair amount; the predictor is vehicle age in years. For \(x=4\),
The observed amount is $1,950, so its deviation from the stated conditional mean is
Under the error representation, this observation has error \(\varepsilon_i=\$150\). The model describes a mean that falls by $150 for each additional year of vehicle age, within the range where the model is being used. It does not establish that increasing a vehicle’s age causes repair costs to fall; vehicle type, claim circumstances, and repair choices could also matter.
Because paid amounts are nonnegative and can be strongly right-skewed, a normal conditional response may be a poor description, particularly if the model predicts appreciable probability of negative amounts. A linear conditional mean could still be a useful approximation over a limited age range, but the range and error assumptions require scrutiny. Check: the fitted mean is $1,800, and the observed amount is $150 above it.
Worked Example: Claims Handled and Monthly Processing Time
A hypothetical service center records, for each branch-month, the number of claims handled \(X_i\) and total processing time \(Y_i\) in staff-hours. Suppose the proposed mean is \(18+6.2x\) staff-hours, where \(x\) is the number of claims. In a month with 10 claims, observed processing time is 91 staff-hours. Compute the mean and deviation, and assess whether the classical assumptions are automatically credible.
Method. Treat branch-month processing time as the response, substitute the claim count as the predictor, and compare the observed total with the conditional mean. Then distinguish a plausible mean specification from assumptions about the errors.
The response is total processing time \(Y_i\), and the predictor is claims handled \(X_i\), measured in claims. At \(x=10\),
The observed deviation from the conditional mean is
The slope is 6.2 additional mean staff-hours per additional claim, under this model. The intercept of 18 staff-hours may represent setup or fixed monthly work, but only if that interpretation is plausible for a month with no claims.
The number of claims can vary naturally across branch-months, so it is a random predictor in this data setting. The conditional-mean statement remains meaningful: it describes average processing time among observations with a given claim count. However, total-time variation may grow with the number of claims, which would call constant conditional variance into question. Branches may also share staff, creating dependence across observations. A normal-error assumption is not automatic for positive total times. Check: 80 plus the observed 11-hour deviation gives 91 staff-hours.
Matching Assumptions to the Data Setting
Before using a regression model, state what one observation represents and how those observations were collected. The same equation can have very different credibility for independently sampled policies, repeated months from the same branch, or aggregated portfolio segments. A useful specification names the response, predictor, units, and population or observation unit before listing its assumptions.
Use the response for the outcome whose mean is of interest; use the predictor for the explanatory quantity on which that mean is conditioned.
State \(E[Y_i\mid X_i=x]=\beta_0+\beta_1x\), define the measurement units, and check that the proposed straight-line pattern is sensible over the intended range.
Consider zero conditional mean, constant conditional variance, independence across observations, and—when needed for a procedure—normal conditional errors.
For a fixed-predictor design, values of \(X_i\) are set by the study plan, and responses vary conditionally at those values. For a random-predictor design, \(X_i\) varies across sampled observations. In either case, a regression conditional mean concerns the response given the predictor. In a random-predictor setting, zero conditional mean is still a statement about the average error at each predictor value, not merely an overall average error of zero.
Independence must be judged using the observation structure. Repeated observations from one policyholder, branch, or geographic region may share unmodeled influences. In such a setting, treating every row as independent may be unjustified even if the conditional mean line is reasonable. Likewise, constant variance means the conditional variance of the response around its mean is the same at every predictor value; it is not a claim that all observations have the same variance in every possible grouping.
Common Model-Building Traps
- Reversing the roles. If the question is how average repair cost varies with vehicle age, repair cost is the response and age is the predictor. Reversing them changes the conditional question.
- Ignoring units. A slope of $4,500 per hundred policies is not $4,500 per policy. Convert the predictor before substitution.
- Reading the line as an individual outcome. The equation gives a conditional mean. Individual observations can be above or below it.
- Assuming every error condition follows from linearity. A linear conditional mean does not establish constant variance, independence, or normality.
- Claiming causation from a regression association. A conditional-mean relationship alone does not rule out confounding or justify a causal interpretation.
- Interpreting an unsupported intercept. The intercept is the model’s mean at predictor value zero, but that value may lie outside the data range or lack a practical meaning.
Check Your Understanding
For each setting, identify the response and predictor, perform the requested calculation, and consider which assumptions need attention.
- A hypothetical insurer models annual inspection expense \(Y_i\), in dollars, by the number of inspections \(X_i\): \(E[Y_i\mid X_i=x]=700+48x\). Identify the response and predictor and calculate the conditional mean for 15 inspections. (Answer: response is annual inspection expense; predictor is inspection count; mean is $1,420.)
- A study models average monthly claim handling time \(Y_i\), in hours, by monthly claim count \(X_i\), with \(E[Y_i\mid X_i=x]=25+3.5x\). A month with 20 claims has 101 hours of handling time. Find the conditional mean and the observed deviation. (Answer: mean \(25+3.5(20)=95\) hours; deviation \(101-95=6\) hours.)
- A hypothetical portfolio model uses annual benefit amount \(Y_i\), in dollars, as response and policy duration \(X_i\), in years, as predictor. Its proposed mean is \(5{,}000+250x\). State the slope’s units and find the conditional mean at 8 years. (Answer: $250 per year of duration; \(5{,}000+250(8)=\$7{,}000\).)
- A researcher records several monthly costs for each of a small number of branches and proposes a linear conditional mean in monthly policy volume. Give two reasons why the classical independent, constant-variance error assumptions may fail even if the mean line is plausible. (Answer: repeated months from the same branch may have dependent errors; cost variability may change with policy volume.)
- A hypothetical model for repair cost uses \(E[Y_i\mid X_i=x]=3{,}100-120x\), where \(x\) is vehicle age in years. For a six-year-old vehicle with observed cost $2,500, calculate the model mean and observed deviation, and explain whether the negative slope alone proves a causal effect. (Answer: mean \(3{,}100-120(6)=\$2{,}380\); deviation \(2{,}500-2{,}380=\$120\); no, the association alone does not establish causation.)