Inferring Delay Discounting Factors from Public Observables:

Applications in Risk Analysis and the Design of Adaptive Incentives

Adam Szekeres and Einar Snekkenes

Department of Information Security and Communication Technology,

Norwegian University of Science and Technology - NTNU, Gjøvik, Norway

Keywords:

Delay Discounting, Temporal Preferences, Psychological Proﬁle, Information Security Risk Analysis,

Stakeholder Behavior Prediction, Adaptive Incentives.

Abstract:

Decision-makers regularly need to make trade-offs between beneﬁts in the present and the future. Smaller

immediate rewards are often preferred over larger delayed rewards. The concept of delay discounting describes

how rewards further in the future lose their value in comparison to immediate or more proximal rewards.

Empirical evidence shows that people discount future rewards using a hyperbolic function, which gives rise to

preference reversals as the delay between a decision and receipt of the reward increases. People show great

differences in terms of their tendency to discount future beneﬁts. The extent of discounting is characterized

by each individuals’ discounting factor k. This study investigates the extent to which the discounting factor

k can be inferred from publicly observable pieces of information (i.e. ownership of items, habits) linked to

individuals. Data was collected from 331 respondents in an online questionnaire. The analyses show that

37% of the variance can be explained by public observables in the best case, and between 17-33%, when the

predictive model is tested on unseen data. The results contribute to the development of a risk analysis method

within the domain of information security, which currently lacks the temporal dimension when predicting

stakeholder behavior. Furthermore, the results have key implications for the emerging e-health sector, where

individuals’ immediate incentives need to be aligned with long-term societal beneﬁts.

1 INTRODUCTION

Information security-related decisions involve trade-

offs in the dimension of time. Resources need to be al-

located in the present, while their beneﬁts may mate-

rialize in the future. In order to enjoy greater beneﬁts

in the future, immediate, smaller rewards often must

be foregone. For example, a trade-off exists between

gaining immediate gratiﬁcation from various web ser-

vices (i.e. small immediate reward) and being pro-

tected from future privacy breaches (i.e. greater later

rewards) (Acquisti and Grossklags, 2003). E-health

initiatives aim at reaping the beneﬁts from digitiza-

tion within the health care sector (Eysenbach, 2001).

The health care eco-system is characterized by the

interaction of a large number of stakeholder groups

(e.g. citizens/patients, healthcare professionals, re-

searchers, data analytic and service providers, etc.),

where each group has speciﬁc incentives to interact

with the system. Future societal beneﬁts (e.g. en-

hanced drug and treatment research, predictive care,

etc.) are fundamentally dependent on the willingness

of primary data subjects (i.e. citizens or patients) in

the present to share their sensitive health data. The

situation requires that all stakeholder groups perceive

appropriate incentives to cooperate toward collective

goals instead of acting in their individual self-interest

which may result in tragedy of the commons (e.g.

degradation of common pool information resources

by overuse or distrust due to invasion of privacy) (Re-

gan, 2002).

Intertemporal choices are decisions involving

trade-offs among costs and beneﬁts at different times.

The concept of delay discounting refers to the phe-

nomenon where immediate rewards have a higher

value than delayed rewards, giving rise to preferences

which are biased toward the present (Acquisti and

Grossklags, 2003). Individuals can be characterized

by their unique discounting factor k, which governs

the rate at which future rewards lose value. Empir-

ical investigations revealed that the discounting fac-

tor k shows great inter-individual differences and that

such differences are associated with important and

varied health-related and economic outcomes (Kirby

and Marakovi

c, 1996; Frederick et al., 2002). There-

fore, the assessment of a decision-maker’s discount-

Szekeres, A. and Snekkenes, E.

Inferring Delay Discounting Factors from Public Observables: Applications in Risk Analysis and the Design of Adaptive Incentives.

DOI: 10.5220/0010663400003060

In Proceedings of the 5th International Conference on Computer-Human Interaction Research and Applications (CHIRA 2021), pages 70-80

ISBN: 978-989-758-538-8; ISSN: 2184-3244

ing factor k can have useful implications for several

purposes. On one hand it enables the prediction of

decisions across time for the purpose of risk analy-

sis; and on the other hand it can enable the design

of adaptive incentives which take into account inter-

individual differences regarding temporal preferences

in an emerging e-health ecosystem.

1.1 Problem Statement and Research

Question

This paper aims at contributing to the enhancement of

a risk analysis method which is to be applied within

the context of a democratic e-health eco-system in-

volving multiple stakeholders with conﬂicting incen-

tives. To date, the risk analysis method lacks the

temporal dimension regarding the stakeholder mod-

els, representing a limitation in its behavior predic-

tion capabilities. Thus, the primary objective is to

enhance the risk analysis method, while the second

objective is to propose a method for incentive de-

sign within a e-health context, which takes into con-

sideration stakeholders’ individual differences regard-

ing temporal preferences. Temporal preferences need

to be assessed unobtrusively based on publicly avail-

able pieces of information linked to decision-makers

to reach inaccessible or adversarial subjects. There-

fore, the main research question is as follows:

Research Question: To what extent is it possible to

infer individuals’ discounting factor k based on

publicly available and observable pieces of infor-

mation linked to decision-makers?

The paper is organized as follows: Section 2 pro-

vides an overview about the theoretical and empiri-

cal results related to delay discounting followed by

a presentation of the risk analysis method under de-

velopment. The section concludes with a description

of a democratic e-health ecosystem as an application

domain for the results. Section 3 describes the in-

struments used for data collection, procedures and the

composition of the sample. Section 4 presents the

ﬁndings and answers the main research question. Sec-

tion 5 discusses results and their relevance for risk

analysis and the democratic e-health ecosystem. Sec-

tion 6 concludes the paper.

2 RELATED WORK

2.1 Delay Discounting

Patience, self-control, willpower are similar concepts

describing one’s ability to postpone immediate grati-

ﬁcation for later, better outcomes. Psychological ex-

periments were conducted about delayed gratiﬁcation

using marshmallows as rewards for preschool chil-

dren (Mischel et al., 1972). Signiﬁcant individual dif-

ferences were found among children in their ability to

delay gratiﬁcation. Follow-up studies with the same

subjects revealed that self-control in preschool chil-

dren was a useful predictor of later outcomes such as

scholastic performance, skills to cope with stress, so-

cial competences, etc. Willpower has been conceptu-

alized as a cognitive skill which can be enhanced and

trained with simple strategies to regulate emotions,

overcome temptations and to become more future-

oriented (Mischel et al., 1989). The concept has been

also incorporated into behavioral economic theories

to improve decision-maker models by including the

temporal dimension. The concept is known as de-

lay discounting characterizing a decision-maker’s im-

pulsivity or present-orientedness. “Delay discount-

ing is a behavioral phenomenon wherein reinforcers

become devalued as a function of their delay to re-

ceipt” (Kaplan et al., 2016). Two models have been

proposed to capture decision-makers’ temporal pref-

erences: exponential discounting and hyperbolic

discounting. Exponential discounting refers to a

constant-rate discounting (constant across delays and

reward amounts), described by the following equa-

tion:

V = Ae

−kD

where V is the present value of the delayed reward,

A is the amount of the delayed reward, k is the dis-

counting rate parameter, and D is the delay (Kirby

and Marakovi

c, 1996). In contrast, hyperbolic dis-

counting assumes that discounting rates are not con-

stant across delays (higher for small delays and lower

for long delays). Empirical investigations showed that

real-world decision-makers’ behavior is best approx-

imated by a hyperbolic function of the form (Kirby

and Marakovi

c, 1996):

V =

1 + kD

The key implication of hyperbolic discounting is that

it gives rise to temporal preferences for smaller im-

mediate rewards over larger later rewards, but these

preferences change as the delay between the choice

and receipt of rewards increases. Thus, a preference

reversal occurs, such that individuals make choices

in the present that their future-self would prefer not

to have made (Kirby and Marakovi

c, 1996). Fig-

ure 1 demonstrates how two rewards (i.e. a Smaller

Earlier Reward (SER) and a Larger Delayed Reward

(LDR)) are discounted across time according to expo-

nential and hyperbolic functions. Preferences remain

stable over time (i.e. SER > LDR for both proximal

Inferring Delay Discounting Factors from Public Observables: Applications in Risk Analysis and the Design of Adaptive Incentives

Exponential discounting

smaller earlier reward (SER) larger delayed reward (LDR)

LDR

SER

LDR

V >

SER

LDR

V >

SER

LDR

PROXIMAL REWARDS

DISTANT REWARDS

Value

Time

(a) Exponential discounting.

Hyperbolic discounting

smaller earlier reward (SER) larger delayed reward (LDR)

LDR

SER

LDR

V >

SER

LDR

V <

SER

LDR

PROXIMAL REWARDS

DISTANT REWARDS

Value

Time

(b) Hyperbolic discounting.

Figure 1: Exponential and hyperbolic discounting functions adapted from (Kalenscher and Pennartz, 2008). The delay (d)

between options (i.e. SER or LDR), is identical for exponential and hyperbolic functions as well as for proximal and distant

rewards. The amounts of reward (A) and discounting factors (k) are identical for both functions.

and distant rewards) with the exponential function as

shown in Figure 1a. With the hyperbolic function

(Figure 1b), immediate or proximal rewards get dis-

counted more steeply, compared to distant rewards,

resulting in preference for the SER, whereas the LDR

is preferred when a signiﬁcant delay is introduced be-

tween the choice and receipt of the reward (i.e. pref-

erence reversal occurs: SER > LDR for proximal re-

wards, but SER < LDR for distant rewards).

Delay discounting shows signiﬁcant inter-

individual differences, and the concept has been

used to explain procrastination (Steel and König,

2006), various addictions (e.g. heroin, alcohol,

tobacco, gambling) where immediate short-term

CHIRA 2021 - 5th International Conference on Computer-Human Interaction Research and Applications

rewards are chosen at the expense of larger delayed

rewards (i.e. better health, longevity) (Kirby et al.,

1999). Empirical evidence also shows that people

use different discounting rates in different contexts.

For example, health-related future beneﬁts are dis-

counted at a higher rate than rewards in the monetary

domain (Chapman and Elstein, 1995). Results about

the intra-individual stability of the delay discounting

construct are mixed. While some data suggests that

delay discounting can be assumed as a relatively

stable, enduring trait (Odum, 2011), a more recent

systematic review (Scholten et al., 2019) identiﬁed

several studies reporting interventions which success-

fully decreased discounting rates on the short-term.

These results suggest that delay discounting may

better be conceptualized as a state variable.

2.2 Conﬂicting Incentives Risk Analysis

The Conﬂicting Incentives Risk Analysis (CIRA) was

developed within the domain of information security

and privacy to simplify the risk analysis procedure by

focusing on human stakeholders and their perceived

incentives (Rajbhandari and Snekkenes, 2013). Risks

within CIRA result from the interdependent relation-

ship between stakeholders, where one person is ex-

posed to the actions or inactions of another person.

Two different stakeholder categories are distinguished

in the game-theoretic framework: risk owner and

strategy owner. Each stakeholder is modelled by

their overall utility using multi-attribute utility the-

ory. Incentives refer to the beneﬁts or losses expected

by a stakeholder when interacting with a system and

other stakeholders. Incentives may be aligned or mis-

aligned. When incentives are misaligned, there is

a risk and every risk is represented by another per-

son’s incentive. Risks are subjective to the person (i.e.

risk owner) being exposed to the conscious choices of

other stakeholders (i.e. strategy owners). Two types

of incentive misalignment are possible. Threat risk

refers to undesirable outcomes for the risk owner and

a potential gain for the strategy owner which resem-

bles the traditional notion of risk referring to unde-

sirable consequences. A ransomware attack on pa-

tient health records can be considered a threat risk

where the patient, or hospital personnel are risk own-

ers, and the hackers motivated by monetary gains are

strategy owners. Incentives can also be misaligned in

a way that results in opportunity risk, where strat-

egy owners lack incentives to act in a desirable way

for the risk owner. For example citizens of e-health

system may lack incentives to share their data (Spil

and Klein, 2014; Sunyaev, 2013), which may result in

suboptimal societal outcomes on the long-term (i.e.

missed opportunities for better treatments, decreased

overall efﬁciency, etc.). The CIRA method assumes

adversarial and inaccessible stakeholders during the

risk analysis procedure, therefore the method relies

on unobtrusive (i.e. indirect) assessment of personal

attributes of stakeholders to decrease the possibility of

motivated misrepresentation or cheating by the stake-

holders under investigation. Previous work has inves-

tigated the extent to which publicly observable fea-

tures are useful for inferring stakeholder motivational

proﬁles for the purpose of risk analysis (Szekeres and

Snekkenes, 2020). Another study demonstrated how

unobtrusive psychological proﬁling can be conducted

using publicly available interviews for the improve-

ment of the CIRA method (Szekeres and Snekkenes,

2019).

2.3 Health Democratization

Healthcare is undergoing radical changes due to dig-

itization. The domain is characterized by a large

number of stakeholders including patients, healthcare

professionals, researchers, industrial players (e.g.

pharmaceutical companies, equipment manufactur-

ers), the authorities, national health insurance, etc.,

each having distinct goals and incentives for interact-

ing with the system (Direktoratet for e-helse, 2018).

Incentive conﬂicts are inherent is such complex sys-

tems, therefore it is important to identify and miti-

gate risks, so that patients get a favourable deal. The

Norwegian Health Democratization project aims at

reinforcing the health data infrastructure in mobility

and assurance through data democratization (N.A.,

2019). While democracy is a broad concept and sev-

eral ideas can be included, a key democratic aspect

in the project is that all stakeholders will be rep-

resented as equal entities in the protocol, such that

their unique distinguishing features (e.g. market in-

ﬂuence, administrative power, proﬁtability) are dis-

regarded when parties prove, negotiate and conﬁgure

their rights w.r.t. health data (N.A., 2019). As the

primary data subjects will have increased possibili-

ties to inﬂuence outcomes related to how their data

is used, their willingness to contribute with sensitive

health data is crucial for the expected societal bene-

ﬁts (e.g. cost reduction, improved drug and treatment

discoveries, predictive healthcare, etc.). Another im-

portant democratic aspect is related to the possibility

of choices. The system needs to implement various

opportunities to incentivize data trading for beneﬁt or

proﬁt depending on several factors (e.g. risks, bene-

ﬁts, temporal preferences, etc.). Democratic e-health

initiative’s build on citizen’s active participation in

the decision-making, where patients are treated as

Inferring Delay Discounting Factors from Public Observables: Applications in Risk Analysis and the Design of Adaptive Incentives

partners in health-related decision-making (Aaviksoo,

2015). This approach represents a move away from

the traditional paternalistic model of medicine, where

specialized service providers assist both doctors and

patients in a cooperative decision-making. The en-

visaged system will utilize autonomous agent-based

solutions and smart contracts for data sharing, where

agents can represent people, software, or other ap-

plications. Various agents (e.g. patient agent, GP

agent, ambulance agent) will interact with each other

on behalf of their principals. The agents need

to be equipped with negotiation mechanisms, rules

and protocols, strategies and decision-making mod-

els (Boudko and Leister, 2019).

2.4 Summary of Related Work

In a democratic e-health ecosystem patients or citi-

zens considering sharing their data with other entities

can be conceptualized both as risk owners and strat-

egy owners in terms of CIRA using a two-step pro-

cess model. In the ﬁrst step, a citizen takes the role

of the risk owner and conducts an implicit risk as-

sessment considering risks (e.g. data breaches, ran-

somware attacks, data misuse, etc.) and beneﬁts (e.g.

improved treatment, health monitoring, monetary in-

centives, etc.) associated with sharing sensitive health

data. In the second step, citizens take the role of the

strategy owner and set sharing options, terms and con-

ditions, rules and access rights, etc. depending on the

results of the risk assessment. Such decisions have

a high level of complexity and relative rarity; there-

fore, people are not expected to be skilled in making

these decisions. In such situations people can beneﬁt

from Nudges or other choice architecture approaches

implemented in the system to make socially optimal

decisions (Thaler and Sunstein, 2009). On the larger

scale the main objective is to mitigate the opportunity

risk (i.e. beneﬁts foregone) at the societal level re-

sulting from reluctance of citizens to share their data

which may be due to lack of trust, too high risk or a

lack of incentives. Future societal beneﬁts can only

get realized if decisions in the present are made ac-

cording to long-term preferences, thus adaptive incen-

tive designs need to be developed to match individ-

ual’s discounting proﬁles with a variety of incentives

offered.

3 METHODS

This section describes the data collection procedures,

the sample and the instruments used for collect-

ing data from participants about personal attributes

(i.e. discounting proﬁles) and public observables (i.e.

habits, items owned by respondents).

3.1 Sample and Procedure

As the primary purpose of the study was to assess the

usefulness of a large set of publicly observable pieces

of information for the construction of stakeholder dis-

counting proﬁles (i.e. discounting factor k), it was

necessary to reach a high number of respondents from

the working age population (above 18 years). There-

fore, an online survey was selected as the most ap-

propriate data collection method and invitations were

distributed on several online channels: ﬁrst, a pilot

study was conducted on Amazon Mechanical Turk

(AMT) to test the feasibility of data collection. Based

on the results of the pilot study, some modiﬁcations

were implemented and links to the online survey were

distributed on university mailing lists, and key so-

cial media platforms (Norwegian Facebook groups,

the biggest Norwegian reddit group). The survey was

available in English and Norwegian, and the Norwe-

gian version was proof-read by a professional edito-

rial service. The survey was implemented in the open-

source Limesurvey tool and was hosted on internal

servers provided by the Norwegian University of Sci-

ence and Technology (NTNU). The questionnaire was

completely anonymous, and participants had to read

and accept a consent form before the questionnaire

started upon visiting the link. The number of fully

completed questionnaires is shown in Table 1, orga-

nized according to distribution channels.

Table 1: Number of completed surveys by distribution chan-

nels. AMT: Amazon Mechanical Turk.

Distribution

channel

Number of

completed surveys

AMT 9

Social media 25

University e-mail lists 332

Total 366

Respondents who completed the survey under 10

minutes (average completion time: 23 minutes) were

removed to increase the validity of the dataset. Thus,

the ﬁnal convenience sample consisted of (n = 331)

respondents (173 males, 153 females, and 5 with un-

known sex). The mean age was 40.28 years (SD =

13.27). Most respondents were from Norway (75%),

while other countries represented 25% of the sam-

ple. Most subjects had a completed Master’s degree

(53%), followed by a PhD (24%), Bachelor’s (16%),

and completed secondary education (7%). Most re-

spondents were married or in a long-term relation-

CHIRA 2021 - 5th International Conference on Computer-Human Interaction Research and Applications

ship (71%), followed by singles (24%) and divorced

or separated individuals (5%).

3.2 Measures

The online survey consisted of three main parts fol-

lowing the introduction explaining the purpose of the

data collection:

1. Basic demographic information (age, sex, level of

education, nationality).

2. Behavioral responses for deriving individuals’ de-

lay discounting rates.

3. Publicly observable features linked to the individ-

uals.

3.2.1 Delay Discounting - MCQ-21

The validated 21-item Monetary Choice Question-

naire (MCQ-21) instrument was used for collecting

responses from participants to compute each individ-

ual’s overall discounting factor k. The MCQ-21 is a

self-reporting questionnaire comprising of a set of 21

questions requiring participants to make a choice be-

tween a smaller, immediate reward (SIR) or a larger,

delayed reward (LDR) with monetary values (Kaplan,

2016). The original instructions for the question-

naire: “For each of the next 21 choices, please in-

dicate which reward you would prefer: the smaller

reward tonight, or the larger reward in the speciﬁed

number of days. Although you will not actually re-

ceive any of the money, pretend that you will actually

be receiving the amount that you indicate. Therefore,

please answer each question honestly and as if you

will actually receive the amount chosen either tonight

or after a speciﬁed number of days. To indicate your

choice, please clearly circle the amount and time as

shown in following example: 0. Would you prefer

$100 tonight, or $100 in 45 days?” (Kaplan, 2016)

were modiﬁed so they suit better for the online survey

format. For each question two radio buttons were pro-

vided to make the choice task clear: e.g. $30 tonight

or $85 in 14 days.

Discounting metrics were computed for each re-

spondent using the Excel-based automated scoring

tool, which facilitates the complex computations to

derive the discounting factor k from MCQ-21 (Ka-

plan et al., 2016). The tool reports summary statistics

for the whole sample, checks consistency and outputs

several discounting metrics on the individual level:

overall k, small k, medium k, large k, geomean k (tak-

ing the geometric mean of the small, medium, and

large k values), as well as the log and ln for each of

the k scores. The following analyses use the "overall k

factor" measuring the daily rate at which rewards lose

their value. Rearranging the equation of the hyper-

bolic function gives the formula for the discounting

factor k:

k =

− 1

where V is the smaller, immediate reward; A is the

larger, delayed reward; and D is the delay associated

with A. For a more detailed explanation on deriving

the overall k factor see: (Kaplan et al., 2016).

3.2.2 Publicly Observable Attributes

This section of the questionnaire aimed at collecting

information linked to respondents, which can be eas-

ily observed in most public settings (e.g. work) with-

out direct interaction with the stakeholder. Two cat-

egories of data can be distinguished: ownership of

items and habits. Ownership questions focused on

the presence of attributes, while questions related to

habits were concerned with the frequency of various

actions.

A single choice response format was used to as-

sess the presence of the attributes, and for certain

attributes, additional questions were included to ob-

tain a more detailed description. Question cate-

gories were as follows: real estate (number, loca-

tion, size), car (number, brand, model, type, color, en-

ergy source, unique license plate), motorcycle (num-

ber, brand, type), bicycle (brand, type), boat (brand,

type), phone (brand, model, color, cover, cover color),

laptop (brand, OS, size, camera cover, decoration),

tablet (brand, size), watch (type, brand), headphones

(brand), sunglasses (brand), backpack (brand), brief-

case (brand), jewellery (type, material), wallet (ma-

terial), sport equipment (17 items), pets (7 species

+ other), style description (15 categories), cosmetic

surgery, hair dye, hair length, facial hair, dietary

lifestyle (7 categories), tattoo (general categories,

place of tattoo), social media (existing accounts), pre-

ferred browser, preferred search engine.

Questions related to habits asked the frequency of

various activities on a 9-point response format where

each point had a textual label ranging from 0 - never in

the last 12 months to 8 - every day or nearly every day.

Questions assessed the frequency of: wearing certain

clothes (23 items), doing various sports (17 sports),

listening to music (14 genres), consuming drinks (11

drink types), consuming other products (6 items), en-

gaging in various other activities (26 activities).

4 RESULTS

The ﬁnal dataset contained valid responses from a

total 331 subjects. The key dependent variable for

Inferring Delay Discounting Factors from Public Observables: Applications in Risk Analysis and the Design of Adaptive Incentives

the analysis was individuals’ overall k score. Based

on the automated scoring tool for MCQ-21, descrip-

tive statistics were as follows: Mean overall k scores

= 0.0115, (SD = 0.0235). Overall consistency of

choices was high: 95.6% (SD = 6.21%) showing the

dataset had a high validity, while the overall pro-

portion of LDR (larger delayed reward) chosen was

67.33% (SD = 26.34%), indicating a low general ten-

dency to discount future beneﬁts. Overall k scores

in the present sample were smaller (i.e. evidence of

greater self-control) than the same overall k scores

(Mean = 0.0727, SD = 0.0886) found in a similar-

sized sample (n = 328) with a gambling disorder di-

agnosis (Steward et al., 2017). The computed dis-

counting scores were fed back to the master database,

and independent categorical (i.e. nominal) variables

were dummy coded into indicator variables (where

0 = no/attribute is not present; 1 = yes/attribute is

present). This procedure allows categorical variables

to be included in regression models. The analyses

were performed in SPSS 25 and scikit-learn library

for Python.

The forward selection algorithm was used for con-

structing multiple linear regression models with over-

all k as the single dependent variable and the set of

publicly observable features as predictors in SPSS.

The algorithm is a stepwise feature selection proce-

dure which enters variables into the equation based

on their strength of correlation with the dependent

variable. Criterion for probability of entry was set

to: p ≤ 0.05 and p ≥ 0.1 for exclusion. The pro-

cedure terminated when no more variables met the

criterion of entry (IBM, 2016). Model performance

was evaluated by two metrics provided by SPSS: R

- coefﬁcient of determination or the proportion of the

variance in the dependent variable explained by the

set of independent variables in the model; and the ad-

justed R

score which penalizes each additional pre-

dictor, providing a more conservative estimate about

the model’s goodness-of-ﬁt. Following the feature se-

lection and model construction procedures the best re-

gression model (F(21, 309) = 6.125, p < 0.00) with an

= 0.371, adjusted R

= 0.311 was found, as shown

in Figure 2. The complete model with the best ﬁt for

predicting the overall k score is provided in Table 2.

Based on the formula for multiple linear regression,

an individual’s overall discounting factor k can be pre-

dicted by:

= β

+ β

+ ... + β

+ ε

summing the unstandardized β

− β

coefﬁcients of

the predictors multiplied by the unobtrusively as-

sessed raw scores X

− X

(using 0-8 for frequency

of habits, and 0-1 for indicator variables) with the rel-

evant ε error terms. All predictors were signiﬁcant at

p ≤ 0.05. The model can be considered a best-case

scenario, since the metrics only provide information

about the model’s ﬁt, but the error of prediction for

unobserved data is not assessed in this step.

0.371

0.311

0 0.2 0.4 0.6 0.8 1

R-squared

Adj. R-squared

Prediction accuracy of Overall k

Best-case scenario

Upper bound

(95 % CI),

0.33

Mean R-squared,

0.25

Lower bound

(95 % CI),

0.17

0.0 0.2 0.4 0.6 0.8 1.0

Prediction accuracy of Overall k

5-fold cross validation

Figure 2: Prediction accuracy for the overall k discount-

ing factor. Goodness-of-ﬁt metrics (R

and Adjusted R

)

provide best case scenarios, since error of predicting un-

observed data is not assessed. An R

= 1 would indicate

perfect ﬁt of the model.

In order to assess the model’s expected perfor-

mance on unseen data, a 5-fold cross-validation pro-

cedure was conducted. Cross-validation makes it pos-

sible to quantify how well the model performs on un-

seen data (i.e. how well the model generalizes be-

yond the sample used for training the model) (Yarkoni

and Westfall, 2017). Due to the relatively small num-

ber of subjects, a train-test split was performed where

each model was trained on 80% of the original dataset

and performance was tested on the remaining 20% of

data. The dataset was randomized for each run. The

results of the 5-fold cross-validation are presented in

Figure 3. Compared to the best-case scenario the ex-

pected performance of the model on unseen data re-

duces to R

: 0.253 +/- 0.079 (with CI 95%), using

the mean of the R

scores derived from 5 independent

runs.

4.1 Illustrative Scenarios

In order to illustrate the utility of inferring individu-

als’ discounting factor (k) two simple cases relevant

to the paper’s topic (i.e. prediction of stakeholder be-

havior for risk analysis and adaptive incentives) are

presented. The following examples only focus on dif-

ferences in sensitivity to delayed rewards, while the

security of the e-health system and the risks of data

sharing, etc. are not considered. These critical factors

need to be addressed carefully during the design and

implementation of the system.

In a CIRA-type scenario a CEO of a small or

medium-sized enterprise (strategy owner) needs to

make a choice between taking out an immediate div-

idend or investing in security controls with delayed

beneﬁts. The discounting factor k of the stakeholder

CHIRA 2021 - 5th International Conference on Computer-Human Interaction Research and Applications

Table 2: Regression model for predicting the overall k discounting factor. Predictors are sorted in order of importance from

most important to least important based on the Standardized β coefﬁcients. Variables assessing frequency of activity are

marked with (freq), dummy variables are marked with (y/n).

Predictors

Standardized β

Coefﬁcients

Unstandardized β

Coefﬁcients

Std.

Error

Sig.

Constant 5.516 0.032 0.006 0.00

style: wearing tattoo (y/n)

0.264 4.566 0.025 0.005 0.00

gambling (freq)

0.221 4.453 0.004 0.001 0.00

wearing shorts (freq)

-0.220 -4.478 -0.003 0.001 0.00

going to party (freq)

-0.206 -4.052 -0.003 0.001 0.00

style: facial hair (y/n)

0.173 3.57 0.019 0.005 0.00

listening to blues music (freq)

-0.171 -2.894 -0.002 0.001 0.00

listening to jazz music (freq)

0.163 2.992 0.001 0.000 0.00

ownership of SUV (y/n)

0.156 3.176 0.013 0.004 0.00

going ﬁshing (freq)

0.147 3.046 0.002 0.001 0.00

listening to electronic music (freq)

-0.143 -2.777 -0.001 0.000 0.01

drinking coffee (freq)

-0.139 -2.877 -0.001 0.000 0.00

home location: countryside (y/n)

-0.133 -2.781 -0.012 0.004 0.01

wearing baseball cap (freq)

-0.129 -2.654 -0.001 0.001 0.01

ownership of boat (y/n)

0.127 2.668 0.015 0.005 0.01

watch type: digital (y/n)

0.126 2.636 0.010 0.004 0.01

listening to country music (freq)

0.114 2.168 0.001 0.000 0.03

brand of sunglasses: Ray-Ban (y/n)

0.112 2.334 0.006 0.003 0.02

search engine: other than Google (y/n)

-0.112 -2.358 -0.009 0.004 0.02

no account on Instagram (y/n)

0.104 2.088 0.005 0.002 0.04

phone color: white (y/n)

0.103 2.144 0.010 0.004 0.03

playing football (freq)

0.101 2.047 0.002 0.001 0.04

0.371

0.311

0 0.2 0.4 0.6 0.8 1

R-squared

Adj. R-squared

Prediction accuracy of Overall k

Best-case scenario

Upper bound

(95 % CI),

0.33

Mean R-squared,

0.25

Lower bound

(95 % CI),

0.17

0.0 0.2 0.4 0.6 0.8 1.0

Prediction accuracy of Overall k

5-fold cross validation

Figure 3: Prediction accuracy of overall k based on 5-fold cross-validation in terms of the mean R

metric.

is assessed as 0.01 % /day according to MCQ-21. The

bonus (smaller earlier reward - SER) is worth 300k,

while the losses avoided by the investment in controls

(larger delayed reward - LDR) is worth 500k. Us-

ing knowledge about the individual’s discounting rate

k and the amounts of rewards involved, it is possi-

ble to predict the stakeholder’s choices in time. In

this speciﬁc case, when a delay of more than 66 days

is present between the SER and the LDR, the person

would chose the smaller earlier reward (i.e. bonus)

and discount the later beneﬁts of protection, meaning

that U

SER

> U

LDR

. If the same choice was to be made

1 day later with the same amount of delay, the prefer-

ences would be reversed so that U

SER

< U

LDR

, where

U refers to the utility of a choice.

Within a democratic e-health system, citi-

zens/patients will exhibit great heterogeneity regard-

ing their delay discounting factors. Based on the as-

sumption, that distant societal beneﬁts fundamentally

depend on the population’s willingness to share their

data in the present, it is reasonable to design adap-

tive incentives, which can be customized based on

Inferring Delay Discounting Factors from Public Observables: Applications in Risk Analysis and the Design of Adaptive Incentives

the needs of different individuals. For example, the

willingness to share data with commercial data ana-

lytic providers may be increased by matching incen-

tives based on each individual’s discounting factor. A

citizen with a high overall discounting factor k of 0.2

is more likely to share data in exchange for smaller

immediate rewards (e.g. immediate social approval),

while a patient with a low overall discounting factor

k of 0.01, is more likely to wait for a greater later re-

ward (e.g. a valuable travel voucher). Thus, a demo-

cratic health ecosystem system can be designed which

enables people to exercise freedom of choice with re-

spect to various incentives / business models match-

ing their preferences.

5 DISCUSSION AND FURTHER

WORK

The purpose of this study was to investigate the ex-

tent to which individual’s discounting factor k can

be inferred from publicly observable pieces of infor-

mation. The discounting concept has been shown to

play a signiﬁcant role in various socially important

behaviors (e.g. substance use disorders, obesity, envi-

ronmental concerns, sexual risks, technology depen-

dence, etc.) (Kaplan et al., 2016). To the authors’

knowledge this is the ﬁrst study to investigate the ex-

tent of predictability of the discounting factor k from

publicly observable pieces of information.

This online survey-based study collected data

about a broad range of items possessed by subjects

and about their habits. The survey utilized the vali-

dated instrument MCQ-21 and the accompanying au-

tomated scoring tool which facilitates the computa-

tion of several discounting metrics (Kaplan et al.,

2016). The results indicate that the best model is ca-

pable of explaining around 31-37% of the variance in

the overall k scores, and between 17-33% when a 5-

fold cross-validation technique is used to assess how

well the model generalizes beyond the training set.

The results can be applicable in two domains for

different purposes. The primary use of the results

is within the CIRA method, which can be enhanced

with the temporal dimension to predict stakeholder

behavior. Inferring the discounting factor of inac-

cessible and/or adversarial stakeholders from publicly

observable pieces of information can increase the an-

alyst’s capability to assess the action-desirability as

perceived by the stakeholder allowing the prediction

of behavior in the temporal dimension. However,

there is a need to decrease uncertainties in the predic-

tion accuracy which may be achieved by increasing

the number of respondents in subsequent studies and

by including other sets of publicly available pieces of

information.

The other domain where the results are to be uti-

lized is within the Health Democratization project,

where the existence of the right incentives for data

subjects (i.e. citizens, patients) in the present, is

crucial for generating desirable societal outcomes in

the future. People face different risks based on their

needs, expectations and health conditions. In a demo-

cratic healthcare ecosystem where patients are both

risk owners (i.e. facing the risk of data breaches) and

strategy owners (i.e. able to inﬂuence outcomes and

make informed decisions), it is important to match in-

centives for data sharing with various personal pref-

erences (e.g. temporal, risk tolerance, etc.). Health

data is being generated by more and more devices and

organizations (e.g. ﬁtness devices and apps, medical

devices in outpatient care settings). Currently the key

incentives for citizens is to access health-related in-

formation generated by wearable devices; receive ser-

vices (e.g. data analytics, health-related recommen-

dations) or get improved care in case of medical mon-

itoring sensors (Baig et al., 2017). The willingness

of data subjects to share or trade health data for other

purposes (e.g. research, proﬁt) needs to be explored

and various incentive mechanisms may be necessary

to facilitate data portability beyond current possibil-

ities. Trust in organizations which generate and col-

lect personal health data is necessary which can be

established by making organizational practices trans-

parent and easy to understand. Allowing people to

make informed choices and opting out of undesirable

data processing activities (e.g. local storage of data)

without negative consequences is important to give

people control over their data and facilitate portabil-

ity (Garmin, 2021). In order to avoid a tragedy of the

commons scenario regarding privacy (Regan, 2002)

in the e-health sector, incentives for all stakeholder

groups (e.g. citizens, commercial actors, regulators)

should be designed so that self-interests are aligned

with societal interests. Future work also needs to in-

vestigate how negotiation agents (Boudko and Leis-

ter, 2019) can be enhanced with information about the

principal’s temporal preferences and other character-

istics to mitigate risks. Future work will also inves-

tigate how various health conditions can be used to

augment risk owner proﬁles using international clas-

siﬁcations such as the International Statistical Clas-

siﬁcation of Diseases and Related Health Problems

(ICD) (WHO, 2021).

It is important to mention that the study is not

without limitations. The sample was a convenience

sample, therefore probabilistic sampling methods

with more subjects would be desirable for future stud-

CHIRA 2021 - 5th International Conference on Computer-Human Interaction Research and Applications

ies to increase generalizability. It should be noted

that the instrument (MCQ-21) used for collecting dis-

counting scores is restricted to monetary rewards. As

evidence shows, people tend to use different discount-

ing rates for different contexts (Chapman and El-

stein, 1995), thus the utility of the derived discount-

ing rates outside of the monetary domain is question-

able. Additionally, the MCQ-21 uses relatively small

hypothetical rewards (highest reward is $85), there-

fore the validity of the discounting scores for much

larger rewards needs careful considerations. Other in-

struments focusing on rewards in different domains

and with higher reward values can be used in future

studies as well. Considering that emotions and moral

concerns play a key role in human decision-making,

it is also important to investigate how non-monetary

incentives (e.g. emotional, moral, social) are dis-

counted, and which non-monetary incentives are ap-

plicable in the e-health domain. These questions are

especially relevant since the commercial exploitation

of human biological materials and personal health

data is often seen as morally problematic. However,

the promotion of communal beneﬁt sharing and the

establishment of appropriate regulations (which hin-

der the commodiﬁcation of the human body) could

render the commercial use of biobanks and health

data more acceptable (Steinsbekk et al., 2013).

Finally, certain principles and ethical standards

need to be considered for incentive-designers. The

potential to abuse information gained from unobtru-

sive measures needs careful attention (e.g. unfair in-

centive schemes). Vulnerable groups (e.g. high dis-

counters) may get offered lower pay-outs and false

incentives if unethical actors infer discounting scores.

Regulations need to be designed with the expectation

that certain actors will try to maximize their beneﬁts

at the expense of other stakeholders in an unfair man-

ner.

6 CONCLUSIONS

Despite increasing levels of digitization in several

domains of life, people are fundamentally respon-

sible and accountable for the decisions which af-

fect themselves and others. E-health represents an

emerging domain where incentive conﬂicts will be

highly prevalent due to the large number of stake-

holder groups in an interdependent relationship. Un-

derstanding how people overvalue the present at the

expense of the future has key implications for be-

havior prediction within risk analysis using the CIRA

method, and for designing incentives based on indi-

vidual differences in a democratic e-health system,

where individual’s will be active participants with a

desire to make informed decisions related to how their

data is used. Aligning incentives in the present so that

people make decisions which will generate beneﬁcial

outcomes in the future is key to a successful demo-

cratic health eco-system.

ACKNOWLEDGEMENTS

This work was supported by the Health Democratiza-

tion project, funded by the Research Council of Nor-

way, IKTPLUSS program, grant number 288856.

REFERENCES

Aaviksoo, A. (2015). E-Health Brings Democracy to

Healthcare. [Online; accessed 26. May 2021].

Acquisti, A. and Grossklags, J. (2003). Losses, gains, and

hyperbolic discounting: An experimental approach

to information security attitudes and behavior. In

2nd Annual Workshop on Economics and Information

Security-WEIS, volume 3, pages 1–27. Citeseer.

Baig, M. M., GholamHosseini, H., Moqeem, A. A., Mirza,

F., and Lindén, M. (2017). A systematic review

of wearable patient monitoring systems–current chal-

lenges and opportunities for clinical adoption. Journal

of medical systems, 41(7):1–9.

Boudko, S. and Leister, W. (2019). Building blocks of ne-

gotiating agents for healthcare data. In Proceedings

of the 21st International Conference on Information

Integration and Web-based Applications & Services,

pages 635–639.

Chapman, G. B. and Elstein, A. S. (1995). Valuing the

future: Temporal discounting of health and money.

Medical decision making, 15(4):373–386.

Direktoratet for e-helse (2018). Konseptvalgutredning for

helseanalyseplattformen. [Online; accessed 2. Jun.

2021].

Eysenbach, G. (2001). What is e-health? Journal of medical

Internet research, 3(2):e20.

Frederick, S., Loewenstein, G., and O’donoghue, T. (2002).

Time discounting and time preference: A critical re-

view. Journal of economic literature, 40(2):351–401.

Garmin (2021). Privacy policy for Garmin Connect, Garmin

Sports and compatible Garmin devices. [Online; ac-

cessed 16. Jun. 2021].

IBM (2016). Linear Regression Variable Selection Meth-

ods. [Online; accessed 4. Jun. 2021].

Kalenscher, T. and Pennartz, C. M. (2008). Is a bird in the

hand worth two in the future? the neuroeconomics of

intertemporal decision-making. Progress in neurobi-

ology, 84(3):284–315.

Kaplan, B. (2016). 21-item-mcq.pdf. [Online; accessed 2.

Jun. 2021].

Inferring Delay Discounting Factors from Public Observables: Applications in Risk Analysis and the Design of Adaptive Incentives

Kaplan, B. A., Amlung, M., Reed, D. D., Jarmolowicz,

D. P., McKerchar, T. L., and Lemley, S. M. (2016).

Automating scoring of delay discounting for the 21-

and 27-item monetary choice questionnaires. The Be-

havior Analyst, 39(2):293–304.

Kirby, K. N. and Marakovi

c, N. N. (1996). Delay-

discounting probabilistic rewards: Rates decrease as

amounts increase. Psychonomic bulletin & review,

3(1):100–104.

Kirby, K. N., Petry, N. M., and Bickel, W. K. (1999). Heroin

addicts have higher discount rates for delayed rewards

than non-drug-using controls. Journal of Experimen-

tal psychology: general, 128(1):78.

Mischel, W., Ebbesen, E. B., and Raskoff Zeiss, A. (1972).

Cognitive and attentional mechanisms in delay of

gratiﬁcation. Journal of personality and social psy-

chology, 21(2):204.

Mischel, W., Shoda, Y., and Rodriguez, M. I. (1989). Delay

of gratiﬁcation in children. Science, 244(4907):933–

938.

N.A. (2019). Health Democratization – IKTPLUSS NFR

project - NTNU. [Online; accessed 2. Jun. 2021].

Odum, A. L. (2011). Delay discounting: trait variable? Be-

havioural processes, 87(1):1–9.

Rajbhandari, L. and Snekkenes, E. (2013). Using the con-

ﬂicting incentives risk analysis method. In IFIP Inter-

national Information Security Conference, pages 315–

329. Springer.

Regan, P. M. (2002). Privacy as a common good in the

digital world. Information, Communication & Society,

5(3):382–405.

Scholten, H., Scheres, A., De Water, E., Graf, U., Granic, I.,

and Luijten, M. (2019). Behavioral trainings and ma-

nipulations to reduce delay discounting: A systematic

review. Psychonomic bulletin & review, 26(6):1803–

1849.

Spil, T. and Klein, R. (2014). Personal health records suc-

cess: why google health failed and what does that

mean for microsoft healthvault? In 2014 47th Hawaii

International Conference on System Sciences, pages

2818–2827. IEEE.

Steel, P. and König, C. J. (2006). Integrating theories of mo-

tivation. Academy of management review, 31(4):889–

913.

Steinsbekk, K. S., Ursin, L. Ø., Skolbekken, J.-A.,

and Solberg, B. (2013). We’re not in it for the

money—lay people’s moral intuitions on commercial

use of ‘their’biobank. Medicine, Health Care and Phi-

losophy, 16(2):151–162.

Steward, T., Mestre-Bach, G., Fernández-Aranda, F.,

Granero, R., Perales, J. C., Navas, J. F., Soriano-Mas,

C., Baño, M., Fernández-Formoso, J. A., Martín-

Romera, V., et al. (2017). Delay discounting and im-

pulsivity traits in young and older gambling disorder

patients. Addictive Behaviors, 71:96–103.

Sunyaev, A. (2013). Evaluation of microsoft healthvault

and google health personal health records. Health and

Technology, 3(1):3–10.

Szekeres, A. and Snekkenes, E. A. (2019). Predicting ceo

misbehavior from observables: Comparative evalua-

tion of two major personality models. In E-Business

and Telecommunications. ICETE 2018. Communica-

tions in Computer and Information Science, volume

1118, pages 135–158. Springer, Cham.

Szekeres, A. and Snekkenes, E. A. (2020). Construction

of human motivational proﬁles by observation for risk

analysis. IEEE Access, 8:45096–45107.

Thaler, R. H. and Sunstein, C. R. (2009). Nudge: Improving

decisions about health, wealth, and happiness. Pen-

guin.

WHO (2021). International Statistical Classiﬁcation of Dis-

eases and Related Health Problems (ICD). [Online;

accessed 4. Jun. 2021].

Yarkoni, T. and Westfall, J. (2017). Choosing prediction

over explanation in psychology: Lessons from ma-

chine learning. Perspectives on Psychological Sci-

ence, 12(6):1100–1122.

CHIRA 2021 - 5th International Conference on Computer-Human Interaction Research and Applications