Back to History Current Version

Predictors of insurance decisions

Last registered on March 15, 2018


Trial Information

General Information

Predictors of insurance decisions
Initial registration date
March 14, 2018

Initial registration date is when the trial was registered.

It corresponds to when the registration was submitted to the Registry to be reviewed for publication.

First published
March 15, 2018, 5:25 PM EDT

First published corresponds to when the trial was first made public on the Registry after being reviewed.


Primary Investigator

University of Georgia

Other Primary Investigator(s)

PI Affiliation
Ludwig-Maximilians-Universität Munich
PI Affiliation
University of Wisconsin-Madison

Additional Trial Information

In development
Start date
End date
Secondary IDs
We investigate the connection between risk preferences and demand for insurance by asking three questions. First, are the identified behavioral constructs related to risk aversion—utility curvature, probability weighting, certainty effect and loss aversion—able to accurately predict an individual’s desired level of insurance? Second, given that there is some predictive power to these behavioral factors, which factor or combination of factors has the best predictive power? Third, which of the different decision models proposed in the literature can best predict the demand for insurance?. We will conduct an incentive compatible experiment in which we elicit these behavioral motives using lottery-based multiple price lists and ask subjects to make a number of insurance choices of varying prices and loss probabilities. We then conduct various analyses to answer the three questions stated above.
External Link(s)

Registration Citation

Jaspersen, Johannes, Marc Ragin and Justin Sydnor. 2018. "Predictors of insurance decisions." AEA RCT Registry. March 15.
Former Citation
Jaspersen, Johannes, Marc Ragin and Justin Sydnor. 2018. "Predictors of insurance decisions." AEA RCT Registry. March 15.
Sponsors & Partners

There is information in this trial unavailable to the public. Use the button below to request access.

Request Information
Experimental Details


There are no randomized conditions or interventions in this study. It is a measurement and correlation study and while the order of questions will be randomized, all subjects will answer the same sets of questions.
Intervention Start Date
Intervention End Date

Primary Outcomes

Primary Outcomes (end points)
Subjects will select the share of a $3 possible loss they wish to insure, from 0 to 100% in 12 scenarios that vary in the probability of the loss and the loading factor that determines the cost of the insurance.
Primary Outcomes (explanation)
Our dependent variable is the proportion of the $3 loss the subject has chosen to insure in each scenario.

Secondary Outcomes

Secondary Outcomes (end points)
Secondary Outcomes (explanation)

Experimental Design

Experimental Design
We plan to elicit risk preferences and simulate the insurance purchase decision in an experiment conducted online using the Qualtrics survey software. Our primary sample is subjects recruited via Amazon Mechanical Turk (“mTurk”), and we aim to recruit 1,350 mTurk subjects. As a robustness check, we will also have subjects complete the same Qualtrics study in-person at the University of Wisconsin-Madison BRITE lab. We are targeting 350 subject to complete the in-person version of the study. All subjects will see the same questions, though the order of questions and left/right display of choices will be randomized.

Subjects begin by earning $5.00 in virtual currency for completing a typing task. Each subject will then make 113 choices, one of which will be selected to play out at the end and determine their earnings. We determine risk preferences by presenting subjects with a series of 101 economic choices between uncertain outcomes. This includes 48 questions where subjects can only gain money (between $0.50 and $60.00), 32 questions where subjects can only lose money (between $0.10 and $4.00), and 21 questions where subjects may either gain money ($0.50 or $5.00) or lose money ($0.20 or $2.00). Subjects will use radio buttons to indicate their choice in these lotteries.

Subjects will also make 12 insurance decisions over a potential loss to their earnings of $3.00. In each of the scenarios, subjects may lose $3.00 of their $5.00 initial earnings and have the option to buy insurance to cover all or part of the loss. Subjects use a slider to choose an insurance amount ranging from 0% (no insurance) to 100% (full insurance) in 1% increments. With each click on a level of insurance, subjects can see the price of the insurance and both "loss" and "no loss" outcomes.

Prior to both the insurance and the lottery tasks, subjects have a set of instructions on how to make their choices and must answer several questions to show they understand the instructions. If they answer the questions incorrectly, they are asked to reread the instructions and try again (in-person subjects may ask clarification questions of the experimenter).

After completing both the lottery and insurance tasks, we draw a virtual “ball” from a simulated bucket filled with balls numbered 1-113. The number on the selected ball indicates the question number that will be played out. We then draw a second ball to determine the outcome of that question number (whether there is a gain, a loss, or nothing). The original $5.00 balance is adjusted up or down based on the outcome of the draw and the price of insurance selected (if applicable).
Experimental Design Details
Randomization Method
Randomization will be done by computer
Randomization Unit
Question order, left/right display of choices, question selected to play for real money using subject's choice, "ball" drawn to determine outcome
Was the treatment clustered?

Experiment Characteristics

Sample size: planned number of clusters
1,350 subjects on Amazon MTurk and 350 subjects at the University of Wisconsin BRITE Lab. For Amazon, we will post the task with the target recruitment as the limit and close the sample when either those targets are hit or 3 weeks have passed, whichever comes first. For the BRITE Lab, we will target 350 subjects for the study. To reach this target we will recruit subjects to the lab and close recruitment when 350 subjects have completed the survey (with complete defined as answering all questions in the incentivized survey elicitation).
Sample size: planned number of observations
1,350 subjects on Amazon MTurk and 350 subjects at the University of Wisconsin BRITE Lab.
Sample size (or number of clusters) by treatment arms
1,350 subjects on Amazon MTurk and 350 subjects at the University of Wisconsin BRITE Lab.
Minimum detectable effect size for main outcomes (accounting for sample design and clustering)
In the basic preference-scale correlations with insurance demand analysis number 1 using non-parametric preference scales, we should be able to detect a change of 0.05 standard deviations in the dependent variable due to a 1 standard deviation change in one of the 6 independent variables at a 1% significance level 95% of the time.

Institutional Review Boards (IRBs)

IRB Name
University of Georgia Institutional Review Board
IRB Approval Date
IRB Approval Number
Analysis Plan

There is information in this trial unavailable to the public. Use the button below to request access.

Request Information


Post Trial Information

Study Withdrawal

There is information in this trial unavailable to the public. Use the button below to request access.

Request Information


Is the intervention completed?
Intervention Completion Date
March 31, 2018, 12:00 +00:00
Data Collection Complete
Data Collection Completion Date
March 31, 2018, 12:00 +00:00
Final Sample Size: Number of Clusters (Unit of Randomization)
Was attrition correlated with treatment status?
Final Sample Size: Total Number of Observations
1,730 experiment participants
Final Sample Size (or Number of Clusters) by Treatment Arms
1,352 subjects completed the experiment online (Amazon MTurk), 378 subjects in university laboratory
Data Publication

Data Publication

Is public data available?

Program Files

Program Files
Reports, Papers & Other Materials

Relevant Paper(s)

We analyze an insurance demand experiment conducted in two different settings: in-person at a university laboratory and online using a crowdworking platform. Subject demographics differ across the samples, but average insurance demand is similar. However, choice patterns suggest online subjects are less cognitively engaged—they have more variation in their demand and react less to changes in exogenous factors of the insurance situation. Applying data quality filters does not lead to more comparable demand patterns between the samples. Additionally, while online subjects pass comprehension questions at the same rate as in-person subjects, they show more random behavior in other questions. We find that online subjects are more likely to engage in “coarse thinking,” choosing from a reduced set of options. Our results justify caution in using crowdsourced subjects for insurance demand experiments. We outline some best practices which may help improve data quality from experiments conducted via crowdworking platforms.
Jaspersen, J. G., Ragin, M. A., & Sydnor, J. R. (2022). Insurance demand experiments: Comparing crowdworking to the lab. Journal of Risk and Insurance, 89, 1077–1107.
Can measured risk attitudes and associated structural models predict insurance demand? In an experiment (n = 1730), we elicit measures of utility curvature, probability weighting, loss aversion, and preference for certainty and use them to parameterize seventeen common structural models (e.g., expected utility, cumulative prospect theory). Subjects also make 12 insurance choices over different loss probabilities and prices. The insurance choices show coherence and some correlation with various risk-attitude measures. Yet all the structural models predict insurance poorly, often less accurately than random predictions. This is because established structural models predict opposite reactions to probability changes and more sensitivity to prices than people display. Approaches that temper the price responsiveness of structural models show more promise for predicting insurance choices across different conditions.
Jaspersen, J. G., Ragin, M. A., and Sydnor, J. R. (2022). Predicting insurance demand from risk attitudes. Journal of Risk and Insurance, 89, 63–96.
The “general risk question” (GRQ) has been established as a quick way to meaningfully elicit subjective attitudes toward risk and correlates well with real-world behaviors involving risk. However, little is known about what aspects of attitudes toward financial risk are captured by the GRQ. We examine how answers to the GRQ correlate with different preference motives and biases toward financial risk using an incentivized choice task (n = 1,730). We find that the GRQ has meaningful correlation with loss aversion and attitudes toward variation in financial losses, but much weaker to non-existent correlations with attitudes toward variation in financial gains, likelihood insensitivity, and certainty preferences. These results suggest that practical applications using the GRQ as an index for financial risk preferences may be most appropriate in settings where decisions rest on attitudes toward financial losses.
Jaspersen, J.G., Ragin, M.A. & Sydnor, J.R. Linking subjective and incentivized risk attitudes: The importance of losses. J Risk Uncertain 60, 187–206 (2020).

Reports & Other Materials