Estimating price elasticity of risk-taking in an auto-insurance market

The experiment focuses on a set of small firms (“borrowers”) that have an auto insurance with our partner firm (“company”). The typical insured firm is a closely-held small business with one owner-operated vehicle or a small fleet, and has vehicle financing as well as insurance from our partner firm. Most insurers already have GPS devices installed in their vehicles due to credit contract covenants. The insurer decided to run the experiment on 1,798 borrowers (with 1,978 loan accounts).

The randomization aims to obtain four equally split samples across the control and three treatment conditions (described below). The unit of randomization is at the borrower level. The sample consists of all loans that are classified as operating on a long-distance route based on a proprietary algorithm. All loan accounts (i.e. loaned vehicles) held by a borrower within the sample (i.e., long-distance route) receive the same offer.

The experiment randomly assigns 1,798 borrowers (with 1,978 vehicles) monthly rewards based on share of distance covered by them in that month below the speed of 100 km/h out of the total distance covered with speed limit of 80 km/h or more in that month (“safe driving statistic”). The rewards will be in the form of a cash-rebate and will be distributed at the start of the month, and will be based on the safe driving statistic of the previous month. As we detail below, the safe driving statistic captures the extent to which the driver avoids engaging with high speed driving. Rewards following the following schedule:
(i) Reward Amount = Monthly Premium*Maximum Discount if safe driving statistic is >= 0.9 and <= 1
(ii) Rewards Amount = Monthly Premium *0.75*Maximum Discount if safe driving statistic is >= 0.8 and < 0.9
(iii) Rewards Amount = Monthly Premium*0.5*Maximum Discount if safe driving statistic is >= 0.7 and < 0.8
(iv) Rewards Amount = Monthly Premium*0.25*Maximum Discount if safe driving statistic is >= 0.6 and < 0.7
(v) Rewards Amount = 0 if safe driving statistic < 0.6
The monthly premium considered for reward amount calculation for each account is from February 2026.

The firms will be randomized into one of the four groups that vary the maximum discount:
(1) Control group: no discount (i.e. maximum discount = 0%)
(2) Treatment 1: maximum discount = 10%;
(3) Treatment 2: maximum discount = 30%;
(4) Treatment 3: maximum discount = 50%

We stratify by one variable: an indicator for whether a borrower has more than one loan account in our sample of loan accounts.

The lender will manage the communication with the customers. The experiment is structured in a way that customers are automatically enrolled in the new program, but they are allowed to opt out of this option at no cost. Customers will first receive the communication about their enrollment into the offer via an SMS. The message will contain a link to a PDF describing the safe driving reporting (all arms) and incentive offer (treatment arms only). If the customer decides to reject the offer, then she needs to opt-out within five business days. A full-time hire will be available to answer any questions that borrowers might have about the offer, which they can ask by calling a number mentioned in the SMS and the PDF document.

Customers will also be sent two updates per month with information on their driving. The first update will be sent mid-month that tells them their safe driving statistic since the start of the month. The second update will be sent at the end of the month providing them with information about their driving, safe driving statistic, and incentive earned (treatment only) over the month. The experiment is expected to run for six months.

2026-04-01

2026-11-30

(1) Speeding: We will measure the share of distance driven above 100 km/h of the total distance driven by operators on roads and freeways with a speed limit of 80 km/h or more. Both measures come from GPS devices installed in the vehicles. This variable is also what is used to identify the eligibility to the discount (i.e., the safe driving statistic mentioned above).

(1) Insurance claims: We will also study how speeding affects insurance claims filed with our partner insurer.

(2) Loan performance: First, we plan to examine outcomes that capture the repayment activity by borrowers. This includes (i) how much borrowers pay month by month; (ii) total arrears; (iii) whether the borrower is current on the loan; (iv) scaled arrears (=arrears/ required monthly payment); (v) a summary index of above measures.

(3) Effort Outcomes: Second, we plan to examine the outcomes of firm’s effort measure during the intervention period. For instance, using the lender’s proprietary data on borrowers’ driving, we can the distance driven, the time on job, and the number of days worked.

(4) Repayment and borrowings from other debt sources: We will analyze spillovers on other forms of borrowings reported in the credit bureau data. This includes any adverse flag, amount of debt overdue, late payments, new account openings, the share of borrowings overdue, total borrowings, and defaults on those debts.

Each borrower is randomly assigned to one of four experiment arms. At inception, each borrower, including the control arm receives a message with a link to a PDF that informs them of the potential incentive they can receive. Borrowers in the control group also receive a detailed PDF, but because they receive no monetary incentives, we show that they can receive a top safe driver status every month based on the safe driving statistic (T also gets this status report; i.e., the only difference between T and C is the monetary incentive). This allows us to provide the control group the same type of information and signaling. The four arms are the following:

(1) Control group: no discount (i.e. maximum discount = 0%)
(2) Treatment 1: maximum discount = 10%;
(3) Treatment 2: maximum discount = 30%;
(4) Treatment 3: maximum discount = 50%

Sample: Our sample consists of all borrowers who are (1) driving on a long-distance route (as determined by our proprietary algorithm based on past driving behavior); (2) have a loan account with the partner lender; (3) have insurance on that collateralized vehicle with the partner lender/insurer; (4) are not deeply delinquent on their loan with the lender.

Not available

Randomization done in office by a computer.

The unit of randomization is at the borrower level. All loan accounts (i.e. loaned vehicles) held by a borrower and driving long-distance route receives the same offer.

Yes

1,798 borrowers

1,978 loan accounts

495 loan accounts; 448 borrowers (control – 0% maximum discount)
492 loan accounts; 450 borrowers (10% maximum discount)
495 loan accounts; 450 borrowers (30% maximum discount)
496 loan accounts; 450 borrowers (50% maximum discount)