Why Telematics Measures Need Careful Review
Telematics can record mileage, time of travel, acceleration, braking, and other features of vehicle use. These measures may add information about risk, but a recorded value is not automatically a valid rating variable. A low count of harsh-braking events, for example, could reflect smoother driving, fewer miles driven, fewer trips in which hard braking was needed, or less complete device observation. Those explanations have different implications for a rating plan.
A sound review asks what the measure represents, whose behavior could be observed, and how the resulting rating effect is distributed. This is especially important when participation, device operation, or driving opportunity differs across policyholders. A statistical association with losses is relevant evidence, but it does not by itself establish that the measure is valid, fairly observed, or appropriate to use.
Match the Measure to Its Denominator
A count needs context. Ten braking events during 500 miles do not represent the same event frequency as ten events during 5,000 miles. When a measure counts events, a natural descriptive rate is the event count divided by the amount of relevant observed exposure. For driving events, this might be events per 1,000 observed miles. The denominator must match the question: miles may suit a driving-event frequency, while trips or driving hours may be more appropriate for some other measures.
The rate is not automatically a fair or risk-appropriate rating variable. It describes observed event frequency per unit of exposure. Its interpretation depends on whether the exposure is measured consistently and whether the observed driving is representative of the driving the variable is meant to describe.
A short derivation clarifies why this normalization matters. Suppose, as a model, that events occur with a constant rate \(\lambda\) per mile for a particular driving pattern, and that events are completely recorded over \(M\) observed miles. If \(N\) is the event count, then \(E[N\mid M]=\lambda M\). Dividing by \(M\) gives \(E[N/M\mid M]=\lambda\), so the count per mile estimates the per-mile event rate without simply rewarding or penalizing a driver for having more observed miles. This justification requires a stable event rate over the observed miles and complete, comparable event capture. If driving conditions change the event rate, or recording is incomplete in a selective way, the ratio alone does not remove the resulting bias.
Worked Example: Mileage Changes the Comparison
A hypothetical insurer reviews two groups over one year. Group A has 1,000 insured vehicles, 20 claims, and 10,000,000 miles driven. Group B has 800 insured vehicles, 12 claims, and 4,000,000 miles driven. Compare claim frequency per insured vehicle-year and per million miles. State what the comparison does and does not show.
Method. Calculate each frequency with its denominator stated explicitly. This avoids treating policy count and driving exposure as interchangeable.
For Group A, claim frequency per insured vehicle-year is \(20/1{,}000=0.020\) claims per vehicle-year. Its mileage is \(10{,}000{,}000/1{,}000=10{,}000\) miles per vehicle-year. Claim frequency per million miles is \(20/(10{,}000{,}000/1{,}000{,}000)=2.000\).
For Group B, claim frequency per insured vehicle-year is \(12/800=0.015\) claims per vehicle-year. Its mileage is \(4{,}000{,}000/800=5{,}000\) miles per vehicle-year. Claim frequency per million miles is \(12/(4{,}000{,}000/1{,}000{,}000)=3.000\).
The ranking reverses: Group A has the higher claim frequency per vehicle-year, while Group B has the higher frequency per million miles. This is consistent with the groups having different average mileage, but it does not establish that mileage caused the difference or that a per-mile rate is the right basis for a final premium. Severity, mix of driving, claim reporting, and other relevant differences would need review.
Check. The total claims are \(20+12=32\). Total miles are 14,000,000, so the combined frequency is \(32/14=2.286\) claims per million miles, between the group-specific rates of 2 and 3. The common mistake is to interpret 20 versus 12 claims without accounting for the groups’ different sizes or miles driven.
Unequal Observation and Opportunity
Telematics data may be missing because a device was not installed, lost connection, was not used, or recorded only some trips. Even when two drivers have the same true behavior, one can have fewer recorded events if fewer miles or trips were observed. Conversely, two drivers with the same event rate can have different raw counts because their observed mileage differs.
A common adjustment is to calculate an event rate using observed exposure rather than the raw count. This is informative only if observed exposure is measured consistently and the observed activity is sufficiently representative. For instance, if a device tends to record daytime trips but miss nighttime trips, dividing events by recorded miles does not correct the missing nighttime behavior. A simple completeness percentage also does not solve the problem if the missing portion differs systematically by driver or driving condition.
Worked Example: Incomplete Recording Reverses Raw Counts
A hypothetical telematics pilot records harsh-braking events for two driver groups. Group A drove 300,000 total miles, of which 240,000 were observed; 480 events were recorded during observed miles. Group B drove 120,000 total miles, of which 72,000 were observed; 216 events were recorded. Compare raw counts and observed event rates. Then estimate full-exposure event counts under an explicit assumption.
Method. First calculate observed rates per 1,000 observed miles. For a separate full-exposure estimate, assume observed miles are representative of total miles within each group and events are captured on observed miles. Without that assumption, extrapolation is not justified.
Group A’s observed-mile fraction is \(240{,}000/300{,}000=0.80\), or 80%. Its event rate is \(480/(240{,}000/1{,}000)=2.000\) events per 1,000 observed miles.
Group B’s observed-mile fraction is \(72{,}000/120{,}000=0.60\), or 60%. Its event rate is \(216/(72{,}000/1{,}000)=3.000\) events per 1,000 observed miles.
The raw counts, 480 and 216, are larger for Group A, but the normalized observed rate is higher for Group B. Under the stated representativeness assumption, Group A’s estimated full-exposure count is \(480/0.80=600\), and Group B’s is \(216/0.60=360\). Dividing by total miles gives \(600/(300{,}000/1{,}000)=2.000\) and \(360/(120{,}000/1{,}000)=3.000\) events per 1,000 total miles.
Check. The extrapolated counts imply the same rates as the observed-mile calculations because the observed fraction was applied to both exposure and events. The calculation does not prove that missing miles are representative. The common mistake is either to compare raw counts as if exposure were equal or to treat an extrapolation as valid without stating and testing its missing-data assumption.
Other variables need an equally careful definition. “Night driving” could mean the share of recorded miles within a specified time window, the count of nighttime trips, or simply whether any nighttime driving occurred. Each has a different denominator and can treat a driver with one short trip differently from a driver with extensive nighttime mileage. Braking measures should distinguish a count from a rate, specify the event definition and threshold, and be checked for consistency across devices and vehicle types.
Disparate Effects, Participation, and Privacy
A telematics program can affect groups differently even when its formula does not explicitly use a protected characteristic. Differences can arise through access to compatible devices, ability or willingness to enroll, patterns of travel, incomplete observation, or the way a threshold awards a discount. Review should examine both the distribution of the measure and the resulting rate impact. A disparity is a reason to investigate; it does not, by itself, identify a cause or determine whether a rating distinction is justified.
The analysis should distinguish at least three questions: Is the behavior measured accurately? Does the measure relate to the intended risk outcome on a suitable basis? And do access, observation, or rating rules create material differences in effects across groups? Evidence supporting one answer does not settle the others. A validated relationship to claims, for example, does not show that every driver had an equal opportunity to generate the data used in the rating.
Worked Example: Discount Access and Average Premium
A hypothetical program applies an 8% discount to a $1,200 annual base premium for policyholders who meet a telematics threshold. In Group X, 180 of 500 policyholders qualify; in Group Y, 60 of 500 qualify. For a limited calibration check, suppose expected annual losses are $900 for qualifying X policyholders, $1,000 for nonqualifying X policyholders, $920 for qualifying Y policyholders, and $1,020 for nonqualifying Y policyholders. Calculate qualification rates, average premiums, and average expected losses. Interpret the results cautiously.
Method. Apply the discount only to qualifying policies, then calculate group averages with all 500 policies in each denominator. This prevents confusing the premium among qualifiers with the average effect on the entire group.
The qualification rate is \(180/500=36\%\) for Group X and \(60/500=12\%\) for Group Y. The discount amount is \(0.08(1{,}200)=\$96\), so a qualifying policy pays \(\$1{,}200-\$96=\$1{,}104\).
Group X’s average premium is \([180(\$1{,}104)+320(\$1{,}200)]/500=(\$198{,}720+\$384{,}000)/500=\$1{,}165.44\). Group Y’s average premium is \([60(\$1{,}104)+440(\$1{,}200)]/500=(\$66{,}240+\$528{,}000)/500=\$1{,}188.48\). The difference is \(\$1{,}188.48-\$1{,}165.44=\$23.04\).
Group X’s average expected loss is \([180(\$900)+320(\$1{,}000)]/500=(\$162{,}000+\$320{,}000)/500=\$964\). Group Y’s is \([60(\$920)+440(\$1{,}020)]/500=(\$55{,}200+\$448{,}800)/500=\$1{,}008\). The qualification and average-premium differences are material facts to investigate, but this small hypothetical comparison does not establish why qualification differs or whether the rating is properly calibrated. Review would need evidence about measurement, exposure, enrollment, and other relevant risk differences.
Check. The group average premiums are between $1,104 and $1,200, as required, and the group with more qualifying policies has the lower average premium. The common mistake is to apply the 8% reduction to every policy or to conclude from a difference in qualification alone that the cause is known.
Privacy is a distinct practical concern. Telematics can reveal patterns of movement and routines, not merely a risk score. A responsible review should ask what data are collected, why each field is needed, how long it is retained, who can access it, how it is secured, and whether it is shared or used for purposes beyond rating. Clear notice, meaningful choices where applicable, data minimization, retention limits, and controls on secondary use help address these concerns. Privacy safeguards do not establish actuarial validity, but actuarial evidence does not remove the need for privacy safeguards.
A Structured Review
The following process keeps measurement, fairness, and implementation questions distinct. It supports a defensible assessment without assuming that a disparity is automatically acceptable or automatically proof of unfair treatment.
Specify the behavior or exposure the variable is intended to represent, the event definition, units, time window, and denominator.
Compare enrollment, device operation, missingness, and observed exposure. Check whether gaps vary by driving conditions or relevant groups.
Evaluate whether the measure relates to the intended insurance outcome on an appropriate exposure basis, and whether results persist across reasonable specifications.
Calculate who qualifies, how much rates change, and whether the effects differ across groups. Investigate plausible mechanisms rather than treating a disparity as its own explanation.
Document data access, retention, security, consumer notice, and limits on use. Monitor the measure and its effects after implementation.
Common Mistakes
- Using event counts without exposure. A larger count can simply reflect more miles or trips. Choose and report a relevant denominator.
- Assuming normalization fixes missing data. Events per observed mile can still be biased if unobserved driving differs systematically from observed driving.
- Equating prediction with validity. Predictive association does not establish that the measurement process is consistent or that the variable represents the intended construct.
- Treating disparity as a complete conclusion. A group difference warrants examination of measurement, opportunity, and rate impact; it does not alone establish cause or justification.
- Ignoring privacy because data are useful. A variable’s potential value does not answer whether the information collected and retained is necessary and appropriately controlled.
Check Your Understanding
Use the stated assumptions in each problem and distinguish a descriptive calculation from a causal conclusion.
- A hypothetical cohort has 15 braking events during 7,500 observed miles. What is the rate per 1,000 observed miles? (Answer: 2.000 events per 1,000 miles.)
- Group R has 90 events in 30,000 observed miles; Group S has 120 events in 60,000 observed miles. Which has the higher observed rate, and by how much per 1,000 miles? (Answer: Group R, by 1.000 event per 1,000 miles: 3.000 versus 2.000.)
- A device observes 70% of a driver’s miles and records 140 events on observed miles. Assuming observed miles are representative and event capture is complete on those miles, estimate the full-mile event count. (Answer: 200 events.)
- A $900 base premium receives a 6% discount. In a group of 250 policyholders, 75 qualify. Find the qualifying premium and the group’s average premium, assuming all others pay the base premium. (Answer: $846 qualifying premium; average premium $883.80.)
- Two groups have the same harsh-braking rate per observed mile, but one group has substantially more unobserved nighttime mileage. Is the equality enough to conclude that the measure is equally valid for both groups? (Answer: No; the unobserved nighttime activity may differ systematically, so comparable observed rates do not establish comparable measurement.)