This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication on http://www.jmir.org/, as well as this copyright and license information must be included.
Risk communication involves conveying two inherently difficult concepts about the nature of risk: the underlying random distribution of outcomes and how a population-based proportion applies to an individual.
The objective of this study was to test whether 4 design factors in icon arrays—animated random dispersal of risk events, avatars to represent an individual, personalization (operationalized as choosing the avatar’s color), and a moving avatar—might help convey randomness and how a given risk applies to an individual, thereby better aligning risk perceptions with risk estimates.
A diverse sample of 3630 adults with no previous heart disease or stroke completed an online nested factorial experiment in which they entered personal health data into a risk calculator that estimated 10-year risk of cardiovascular disease based on a robust and validated model. We randomly assigned them to view their results in 1 of 10 risk graphics that used different combinations of the 4 design factors. We measured participants’ risk perceptions as our primary outcome, as well as behavioral intentions and recall of the risk estimate. We also assessed subjective numeracy, whether or not participants knew anyone who had died of cardiovascular causes, and whether or not they knew their blood pressure and cholesterol as potential moderators.
Animated randomness was associated with better alignment between risk estimates and risk perceptions (
Animated randomness may help people better understand the random nature of risk. However, in the context of cardiovascular risk, such understanding may result in lower healthy lifestyle intentions. Therefore, whether or not to display randomness may depend on whether one’s goal is to persuade or to inform. Avatars show promise for helping people grasp how population-based statistics map to an individual case.
Health risk communication is an inherently challenging proposition. People’s risk perceptions are shaped by powerful cognitive and affective biases [
Lack of alignment between actual and perceived risk may be partly due to barriers to comprehension, such as low health literacy or, in the case of communication about numbers, low numeracy [
Icon arrays (or pictographs) are graphical displays, often with 100 or 1000 icons arranged in rows and columns and in which each icon represents one unit in the population of interest. They have been shown to help people overcome natural tendencies toward misinterpretation [
One of the key challenges to such comprehension is adequately conveying the inherent uncertainty of risk statistics. In this study, we aimed to address the issue of aleatory or first-order uncertainty, which has been highlighted as an entrenched conceptual problem and a key challenge when communicating risk. First-order uncertainty arises from the “fundamental indeterminacy” of future events [
Randomness is conceptually challenging, especially for people with little training in statistics. For example, many people believe that their iPod’s shuffle feature does not actually choose songs randomly because the algorithm may play several songs from the same album in a row, or the user may not hear a new song within their expected time frame. These perceptions persist even though such behaviors on the part of the algorithm are perfectly reasonable within a random ordering [
In addition to the challenge of interpreting the meaning of background randomness, it is difficult for people to map population-based statistics, which are often proportions, onto individual circumstances, which are often whole numbers. No matter how average they might be, a family cannot, after all, actually have 2.3 children [
In this study, we considered another potential approach for helping people understand what population-based statistics mean for individual risk: the use of avatars. People understand avatars to represent individuals and react to them accordingly; for example, by putting more trust in more relatable avatars [
In this study, we evaluated 4 specific risk graphic design factors dealing with the display of randomness and the use of avatars that, in principle, might better convey these challenging concepts—the randomness of events and how population-based statistics apply to an individual—with the goal of helping people better understand the nature of a health risk.
In addition to these experimental factors, we also examined the potential moderating effects of 3 planned individual factors: (1) an individual’s actual level of risk, (2) his or her numeracy, and (3) whether or not she or he has known someone who has experienced the negative health event in question. Each of these may influence how people respond to different risk graphics by moderating the personal salience of the risk information or the way people understand risk numbers.
Regarding the first moderator, to our knowledge, no literature exists concerning the different responses to the same risk graphic formats at different levels of risk. However, it seems plausible that someone receiving a low risk estimate might respond to a particular risk presentation format differently than someone receiving a high risk estimate. Regarding the second moderator, different risk graphic formats have been shown to generate different responses in people at different levels of numeracy [
We further investigated the effects of 2 additional potential moderators whose importance emerged from our observed data: whether or not people know their (1) blood pressure and (2) cholesterol and are thus able to choose from a drop-down menu of potential ranges for such values. Although this information was not required in order for participants to receive a risk estimate, having entered more information relevant to one’s own individual health may well increase the salience of the risk information.
Ultimately, our aim is to improve understanding of risk estimates, which we operationalized in this study as alignment between subjective risk perception and an objective risk estimate along with accurate recall of the risk estimate. Therefore, we investigated the following specific research questions: (1) which design factors might help to increase alignment between perceptions of risk and actual risk, (2) which design factors might help to encourage intentions toward actions associated with healthy living, and (3) do any of the design factors affect recall of risk numbers? The first and third of these specifically addressed our primary goal of improving comprehension; the second addressed questions around the applicability of these design factors to different purposes.
We invited a random sample of US adults aged 35 to 74 years from a panel of Internet users administered by Survey Sampling International (SSI), stratified by gender, age, and race to ensure demographic diversity in similar proportions to the US population, to participate in an online survey. The number of email invitations sent to each stratum was dynamically adjusted to maintain demographic balance despite varying response rates. Participants who completed this experiment as well as another unrelated cross-randomized study contained within the same survey were entered into both an instant-win contest and a monthly drawing administered by SSI for modest prizes. The study was deemed exempt by the University of Michigan Health Sciences and Behavioral Sciences Institutional Review Board as anonymous survey research. All participants viewed a consent page before clicking to begin the study in which they were informed that the survey would involve learning about their personal risk of heart disease and stroke, and that if they did not want to learn about their risk, they should not participate in the study. At the conclusion of the survey, participants were provided with a list of resources for learning more about cardiovascular health and ways to prevent cardiovascular disease.
To explore our research questions, we chose the clinical context of general cardiovascular disease because of the availability of a robust simple model [
Estimates returned by the model range from a risk of less than 1% to a risk greater than 30%. Risk estimates between the lower and upper limits are returned as integer percent values. In other words, the vast majority of risk estimates were numbers such as 4% or 21%, but results at the upper and lower ends of the range were not simple integers. For the risk graphics, we described these less specific values in the legend and introductory text as being “less than 1%” and “more than 30%,” and used 1 and 31 event rectangles, respectively, in the array of 100. Similarly, in our analyses, we used the values .999 and 31 as conservative estimates of these cases, respectively. Throughout the experiment, rather than the term, “general cardiovascular disease,” we used the more familiar terminology “heart disease and stroke.”
The model applies only to people who have not already experienced general cardiovascular disease, so we screened out those who had a history of cardiovascular disease (448/4124, 11%). Remaining participants were asked to enter their information to calculate their personal risk estimate, which they subsequently viewed in an animated risk graphic randomly assigned from 10 possible versions created from the 4 risk graphic design factors (see Design of Risk Graphics section). We only presented absolute numbers and did not provide any context of expected levels of risk. In other words, we did not give participants any indication of whether their risk was higher or lower than might be expected for their age and gender before we assessed their risk perceptions.
For all designs, we used a 10×10 matrix of rectangle icons and animated the construction of the icon array (as previously tested [
In keeping with our goal of conveying 2 key concepts (ie, the underlying randomness of events and the mapping of population-based statistics onto individuals), we used 2 main experimental design factors in our graphics. The first of these was animated randomness, a factor that showed promise when previously explored in combination with a different set of design factors [
The second main design factor was the use of an avatar. We designed this factor to give more explicit signals about how such risks apply to a single individual. Graphics that included a standard avatar had a generic avatar shape animated to drop into the icon array and disappear, then emerge with a question mark at the conclusion of the animation. The disappearance and re-emergence with a question mark were intended to convey that we do not know which event out of the 100 will apply to a single individual. Within the avatar design factor, we also had 2 other nested factors. The first of these (avatar moves) specified that after the avatar was dropped into the array, it would move within the array, randomly landing on either event rectangles or nonevent rectangles to further emphasize how the randomness inherent in the risk statistic applies to a single individual. The second nested factor (color choice), designed to help participants identify with the avatar, offered participants the chance to choose a different color from a palette of Web colors, instead of the default color, which was standard Web black (#000000).
These factors created a 2×2 factorial design nested within another 2×2 factorial design. See
Randomization and graphics factors.
Sample risk graphic (random, avatar moves, color choice).
We programmed the risk graphics and an avatar color-chooser in ActionScript 3.0 and integrated them into a custom survey system programmed in Ruby on Rails. The number of event rectangles in the icon array were set dynamically via JavaScript, using participants’ risk numbers as calculated by the model algorithm implemented in the survey system.
There were 4 independent dichotomous variables. The random variable describes whether or not the event rectangles were dispersed randomly in the pictograph (random condition) or whether they are grouped together at the bottom of the page (standard condition). The avatar variable describes whether or not an avatar was used in the risk graphic. The avatar moves variable indicates whether or not the avatar moved around randomly within the pictograph, randomly landing on event rectangles or not, as the animation proceeded. The color choice variable refers to whether or not participants were asked to select a color for the avatar that they felt best represented themselves.
The primary outcome variable for this study, risk perception, was created from 3 questions, all asked together on the same page immediately after viewing the risk graphic. These questions were intended to capture people’s immediate reactions to the risk number and graphic presentation. We first asked participants to answer the question, “How big or small does this risk feel to you?” on a 10-point Likert scale with anchors “extremely small” on the left and “extremely big” on the right. We then asked people to indicate, “How worried do you feel about your chance of getting heart disease or stroke in the next 10 years?” on a 10-point Likert scale anchored by “not at all worried” on the left and “extremely worried” on the right. Values for 10-point Likert scales were assigned as 0-9 but survey responses were not labeled numerically, meaning that participants did not see any numbers, only a horizontal visual array of equally-spaced radio buttons. Finally, we asked them, “How likely does it feel to you that you will actually get heart disease or stroke in the next 10 years?” which we assessed on a horizontal slider. The slider recorded integer values between 0 (label “extremely unlikely”) and 100 (label “extremely likely”). Participants saw only the visual position of the slider, not the numeric values representing their response. Because we wanted to capture participants’ subjective risk sense, we used this measure rather than asking for a numeric estimate of their risk. We surmised that if we asked for a numeric estimate, many participants would simply return the risk estimate they had been given. We further suspected that this would be most likely to occur among participants with higher numeracy; thus, this measure could bias the potential effects of numeracy on the subjective feeling of being at risk. To combine these 3 measures with equal weight accorded to each, we rescaled the likelihood question by multiplying values by 9/100. We then averaged responses to the 3 questions (Cronbach alpha=.88) to calculate risk perception.
We considered 3 secondary outcome variables in this study: 2 behavioral intention measures and a recall task.
Behavioral intentions were all collected together on 1 page. Participants were given the text, “There are ways to improve your heart health and reduce your risk of heart disease and stroke. How likely are you to do the following things in the next 30 days?” This was followed by a list of 4 or 5 potential actions: quit smoking (presented only to participants who indicated in the risk calculator they had smoked in the last month); exercise 30 to 60 minutes a day, at least 5 days a week; eat a diet that is low in salt, low in fat, and has at least 5 to 10 servings of fruits and vegetables each day; start a weight loss program; and make an appointment to see a doctor about your heart health. Responses were collected for each action on a 10-point Likert scale with anchors “not at all likely” on the left and “extremely likely” on the right. Again, responses were not labeled with their numeric value, meaning that participants did not see any numbers, only a horizontal visual array of equally-spaced radio buttons. Participants were not provided with details about these behaviors beyond their verbal label, nor were they given any information about the extent to which engaging in such behaviors might lower their risk. At the conclusion of the survey, after the study was complete, participants were provided with links to webpages by reputable sources about healthy lifestyles for reducing cardiovascular risk. (See also
The first 4 behavioral statements (3 for nonsmokers) are typical behavioral outcomes in interventions addressing cardiovascular health. We averaged them to form the lifestyle intentions scale (nonsmokers: Cronbach alpha=.68; smokers: Cronbach alpha=.70). We added the final variable, see a doctor, because we postulated that intentions to see a doctor for personalized counsel would be a more appropriate measure of the effects of a brief online risk calculator. In other words, an increased understanding of one’s risk may not be sufficient to provoke behavior change, but it may prompt people to seek more information via a medical consultation.
Recall was collected on the last page of the survey (participants had been presented with 14 to 22 pages since receiving their risk estimate) by asking participants, “Please answer the following question based on your memory of the numbers you were given by the risk calculator earlier in this study. If you are unsure, please take your best guess. Please do not go back to check your answers. If there were 100 people exactly like you, how many of them would have heart disease or stroke in the next 10 years?” Participants entered their recalled value in a text box. To analyze the effects of experimental and moderating factors on recall, we defined the dependent variable recall as the absolute difference between the recalled estimate and the correct value. However, to maximize clarity for readers when tabulating descriptive results about recall in this paper, we define correct recall as a recalled risk within 5 percentage points in either direction of the risk estimate.
We planned for the inclusion of 3 attributes in our model that might moderate participants’ responses. First, we considered the impact of the participant’s actual estimated 10-year risk of cardiovascular disease as presented to them (actual risk). We used the original quasi-continuous variable in our analyses, but to facilitate readers’ interpretation of descriptive statistics in this paper, we present data according to whether a participant’s risk was below the median risk (8%) or not. We further distinguish participants at either end of the spectrum of risk estimates for whom the model provided a less precise numeric risk estimate. Thus, the levels for reporting are very low risk (<1%), lower risk (1%-7%), higher risk (8-30%), and very high risk (>30%). We emphasize that these labels were not shown to study participants nor were they used for analysis; they are simply for readers’ comprehension. We further note that because so few participants were in the very low risk group (n=7; see
We collected 2 self-report individual difference measures, selected because of their potential moderating effect on individuals’ responses to different ways of presenting risk numbers and graphics about cardiovascular disease. Participants completed a validated measure of numeracy, the Subjective Numeracy Scale, which asks people how confident they feel with numbers and how much they prefer information be presented numerically [
In addition to these planned moderators, we noted in our data that a sizeable proportion of participants indicated that they did not know their blood pressure and/or cholesterol. Given that the input of such personal information might affect the salience of the risk, we also included 2 additional variables, blood pressure known and cholesterol known, in our analyses. For the latter, we classified participants who knew either 1 or both of their total or high-density lipoprotein (HDL) cholesterol as knowing their cholesterol.
The effects of risk graphic factors and individual difference measures on outcomes were examined via nested factorial ANOVA. All main effects were analyzed, as were all possible interactions. For the primary outcome, we used an alpha level of .05. For the 3 secondary outcomes, to control for Type I error, we applied a Bonferroni correction, yielding an alpha level of .017. All tests were 2-tailed. To present results, we give
Of the 4859 people who received an invitation email and clicked the link to the survey, 4124 (85%) completed the survey. Of these, 3676 (89%) were eligible for this study, meaning that they were between ages 35 and 74 years and had neither been diagnosed with heart disease nor had a stroke. For analysis, we included participants who completed the full survey, which included this study, a second unrelated study, demographic questions, and other measures of individual differences. The median time to complete the full survey was 16 minutes and the interquartile range (IQR) was 11 minutes. We excluded participants who completed the full survey in less than 6 minutes from analysis because this speed suggested that they may not have been paying attention to the content. Thus, the final sample for analysis comprised responses from 3630 people (99% of eligible respondents).
Participants were diverse in terms of gender, age, ethnicity, race, and level of education. The median 10-year risk of general cardiovascular disease was 8% (IQR 11%). See
Study participant characteristics (N=3630).
Characteristic | Statistic | ||
Age (years), mean (SD) | 53 (10) | ||
|
|
||
|
Female | 2000 (55) | |
|
Male | 1630 (45) | |
|
|
||
|
Hispanic | 404 (11) | |
|
Middle Eastern | 44 (1) | |
|
|
||
|
White or Caucasian | 2827 (78) | |
|
Black or African American | 514 (14) | |
|
American Indian or Alaska Native | 48 (1) | |
|
Asian or Asian-American | 145 (4) | |
|
Pacific Islander or Native Hawaiian | 10 (<1) | |
|
Other | 124 (3) | |
|
|
||
|
None | 1 (<1) | |
|
Elementary school | 3 (<1) | |
|
Some high school, but no diploma | 73 (2) | |
|
High school (diploma or GED) | 681 (19) | |
|
Trade school | 216 (6) | |
|
Some college, but no degree | 975 (27) | |
|
Associate’s degree (eg, AA, AS) | 384 (11) | |
|
Bachelor’s degree (eg, BS, BA) | 871 (24) | |
|
Master’s degree (eg, MA, MPH) | 335 (9) | |
|
Doctoral/professional degree (eg, PhD, MD) | 88 (2) | |
General cardiovascular disease 10-year risk, median (IQR) | 8 (11) | ||
|
|
||
|
Very low risk (<1%) | 7 (0.2) | |
|
Lower risk (<median risk or 1-7%) | 1714 (47) | |
|
Higher risk (≥median risk or 8-30%) | 1630 (45) | |
|
Very high risk (>30%) | 279 (8) | |
|
|
||
|
|
|
|
|
|
<35 | 160 (4) |
|
|
35-44 | 304 (8) |
|
|
45-49 | 218 (6) |
|
|
50-59 | 231 (6) |
|
|
≥60 | 321 (9) |
|
|
I don’t know | 2396 (66) |
|
|
|
|
|
|
<160 | 566 (16) |
|
|
160-199 | 622 (17) |
|
|
200-239 | 368 (10) |
|
|
240-279 | 70 (2) |
|
|
≥280 | 24 (1) |
|
|
I don’t know | 1980 (55) |
|
|
|
|
|
|
<120 | 989 (27) |
|
|
120-129 | 1095 (30) |
|
|
130-139 | 478 (13) |
|
|
140-149 | 186 (5) |
|
|
150-159 | 59 (2) |
|
|
≥160 | 36 (1) |
|
|
I don’t know | 789 (22) |
|
Currently taking medication to treat high blood pressure, n (%) | 1182 (33) | |
|
Has diabetes, n (%) | 469 (13) | |
|
Has smoked in the past month, n (%) | 974 (27) | |
|
|
|
|
|
|
<18 (underweight) | 30 (1) |
|
|
18-24.9 (normal weight) | 690 (19) |
|
|
25-29.9 (overweight) | 794 (22) |
|
|
≥30 (obese) | 932 (26) |
|
|
Height and/or weight not givenb | 1184 (33) |
|
|
||
|
Subjective numeracy (out of possible 6-48), median (IQR) | 35 (10) | |
|
Knows someone who died because of heart problems, n (%) | 2702 (75) |
aHDL: high-density lipoprotein.
bHeight and weight were only asked of participants who did not know their cholesterol counts.
We first explored relationships between actual risk and risk perception via Pearson correlations, stratifying by all possible combinations of design factors as shown in
Testing the risk graphic factors and moderators for their effects on risk perception, we observed an interaction between the actual risk and the random variables in their association with risk perception. Adding the element of randomness resulted in lower risk feeling smaller, higher risk feeling slightly larger, and very high risk feeling larger (see details in
Correlations between actual risk and risk perception by study arm.
Type of Avatar | Standard | Random | ||
|
|
|
|
|
No avatar | .13 | .02 | .25 | <.001 |
Avatar moves: no; color choice: no | .25 | <.001 | .30 | <.001 |
Avatar moves: no; color choice: yes | .23 | <.001 | .18 | <.001 |
Avatar moves: yes; color choice: no | .13 | .01 | .28 | <.001 |
Avatar moves: yes; color choice: yes | .11 | .03 | .21 | <.001 |
Summary of findings for primary outcome risk perception.
Effects | Mean valuesa (SD) |
|
|
|||
|
|
|
|
|||
|
|
|
6.12 | .01b | ||
|
|
|
|
|
||
|
|
|
Standard | 3.2 (2.1) |
|
|
|
|
|
Random | 3.0 (2.2) |
|
|
|
|
|
|
|
||
|
|
|
Standard | 3.7 (2.0) |
|
|
|
|
|
Random | 3.8 (2.1) |
|
|
|
|
|
|
|
||
|
|
|
Standard | 4.1 (2.1) |
|
|
|
|
|
Random | 4.6 (1.9) |
|
|
|
|
|
4.61 | 0.03b | ||
|
|
No avatar | 3.3 (2.1) |
|
|
|
|
|
Avatar | 3.5 (2.2) |
|
|
|
|
|
|
5.88 | .02 | ||
|
|
|
|
|
|
|
|
|
|
No avatar | 2.7 (2.0) |
|
|
|
|
|
Avatar | 3.2 (2.1) |
|
|
|
|
|
|
|
|
|
|
|
|
No avatar | 3.5 (2.1) |
|
|
|
|
|
Avatar | 3.6 (2.1) |
|
|
|
|
4.57 | .03 | |||
|
|
|
|
|
|
|
|
|
|
No avatar | 3.5 (2.2) |
|
|
|
|
|
Generic avatar | 3.3 (2.2) |
|
|
|
|
|
Avatar with color choice | 3.6 (2.2) |
|
|
|
|
|
|
|
|
|
|
|
|
No avatar | 3.3 (2.1) |
|
|
|
|
|
Generic avatar | 3.6 (2.1) |
|
|
|
|
|
Avatar with color choice | 3.5 (2.1) |
|
|
|
|
|
|
|||
|
|
|
166 | <.001 | ||
|
|
Lower risk | 3.1 (2.2) |
|
|
|
|
|
Higher risk | 3.7 (2.1) |
|
|
|
|
|
Very high risk | 4.4 (2.0) |
|
|
|
|
|
|
86.2 | <.001 | ||
|
|
Low numeracy | 3.8 (2.1) |
|
|
|
|
|
High numeracy | 3.2 (2.1) |
|
|
|
|
|
|
28.3 | <.001 | ||
|
|
No familiarity | 3.1 (2.1) |
|
|
|
|
|
Familiarity | 3.6 (2.1) |
|
|
|
|
|
7.56 | .006 | |||
|
|
|
|
|
|
|
|
|
|
Blood pressure unknown | 3.3 (2.2) |
|
|
|
|
|
Blood pressure known | 3.0 (2.1) |
|
|
|
|
|
|
|
|
|
|
|
|
Blood pressure unknown | 3.5 (2.1) |
|
|
|
|
|
Blood pressure known | 3.8 (2.0) |
|
|
|
|
|
|
|
|
|
|
|
|
Blood pressure unknown | 4.3 (2.0) |
|
|
|
|
|
Blood pressure known | 4.4 (2.0) |
|
|
|
|
|
4.36 | .04 | ||
|
|
|
|
|
|
|
|
|
|
Blood pressure unknown | 3.6 (2.1) |
|
|
|
|
|
Blood pressure known | 3.8 (2.1) |
|
|
|
|
|
|
|
|
|
|
|
|
Blood pressure unknown | 3.2 (2.3) |
|
|
|
|
|
Blood pressure known | 3.2 (2.1) |
|
|
aAssessed on scale of 0 (lowest risk perception) to 9 (highest risk perception).
bNo longer significant when participants at very low or very high risk were removed from the sample.
We observed a main effect for the design factor avatar, with slightly higher overall risk perceptions when an avatar was used, as well as another interaction between avatar and familiarity on this outcome. This interaction suggested that for people who knew someone who had died of cardiovascular problems, the use of an avatar was associated with a minimal increase in risk perceptions, but for people who lacked such familiarity, an avatar significantly increased their risk perceptions.
As expected, all 3 planned moderators had significant main effects, with risk perception increasing with actual risk and decreasing with increasing numeracy. People who knew someone who had died of cardiovascular problems (familiarity) perceived their risk as larger. Neither additional moderator (blood pressure known and cholesterol known) had a significant main effect on risk perception. There was, however, a significant interaction between blood pressure known and actual risk on this outcome. Among participants at lower risk, knowing one’s blood pressure was associated with lower risk perception whereas the reverse was true for participants at higher risk and, to a certain extent, those at very high risk. We also observed an interaction between blood pressure known and numeracy. For participants with higher numeracy, knowing one’s blood pressure did not appear to affect risk perception whereas for those with lower numeracy, knowing one’s blood pressure was associated with somewhat higher risk perception.
Finally, we observed an interaction between blood pressure known, avatar, and color choice in their association with this outcome. Among people who knew their blood pressure, the presence of a generic avatar was associated with somewhat higher risk perception but no additional increase was observed for a personalized avatar. However, among people who did not know their blood pressure, a generic avatar was associated with a small decrease in risk perception whereas a personalized avatar was associated with a small increase.
When we explored these analyses on the middle 2 subsets of participants, removing all participants with risk estimates less than 1% or greater than 30%. Findings remained similar overall; however, the observed interaction between actual risk and random was no longer significant (
Examining the effects of different variables on lifestyle intentions, we observed that the factor random had a main effect: participants who received randomly dispersed events were less likely to indicate intentions toward healthy behaviors in the next 30 days (see details in
In addition, nearly all moderating variables had significant main effects. Greater intentions toward healthy lifestyles were observed among participants with lower actual risk, those with higher numeracy, and those who knew their blood pressure and cholesterol.
No significant interactions were observed on this outcome. When we explored these analyses within the subgroup of participants that remained after removing all participants with risk estimates less than 1% or greater than 30%, all findings remained similar.
Summary of findings for secondary outcome lifestyle intentions.
Effects | Mean valuesa (SD) |
|
|
||
|
|
|
|
||
|
|
|
11.1 | <.001 | |
|
|
Standard | 5.2 (2.2) |
|
|
|
|
Random | 4.9 (2.2) |
|
|
|
|
|
|
||
|
|
|
17.4 | <.001 | |
|
|
Lower risk | 5.2 (2.2) |
|
|
|
|
Higher risk | 5.0 (2.3) |
|
|
|
|
Very high risk | 4.7 (2.1) |
|
|
|
|
|
25.4 | <.001 | |
|
|
Lower numeracy | 4.9 (2.2) |
|
|
|
|
Higher numeracy | 5.2 (2.2) |
|
|
|
|
30.8 | <.001 | ||
|
|
Blood pressure unknown | 4.7 (2.4) |
|
|
|
|
Blood pressure known | 5.2 (2.2) |
|
|
|
|
|
34.9 | <.001 | |
|
|
Cholesterol unknown | 4.8 (2.3) |
|
|
|
|
Cholesterol known | 5.4 (2.1) |
|
|
aAssessed on scale of 0 (lowest intentions) to 9 (highest intentions).
We observed a significant interaction between actual risk and avatar in which the use of an avatar appeared to increase the spread, making those at lower risk less likely to plan to see a doctor, and those at higher risk more likely (see details in
The 3 moderating variables having to do with medical data all had significant main effects in expected directions on participants’ intentions to see a doctor in the next 30 days. Participants with higher actual risk indicated stronger intentions, as did those who knew their blood pressure and cholesterol.
We observed another interaction between numeracy, avatar, and color choice. Among those with higher numeracy, personalization via color choice was associated with somewhat increased intentions to see a doctor, whereas this difference was not observed for those with lower numeracy.
When we explored these analyses within the subgroup of participants that remained after removing all participants with risk estimates <1% or >30%, all findings described previously remained similar; however, an interaction that did not reach significance in the analysis of the full dataset (
Summary of findings for secondary outcome see a doctor.
Effects | Mean valuesa (SD) |
|
|
|||
|
|
|
|
|||
|
|
|
6.38 | .01 | ||
|
|
|
|
|
|
|
|
|
|
No avatar | 4.7 (3.0) |
|
|
|
|
|
Avatar | 4.4 (3.0) |
|
|
|
|
|
|
|
|
|
|
|
|
No avatar | 4.8 (2.9) |
|
|
|
|
|
Avatar | 5.2 (3.0) |
|
|
|
|
|
|
|
|
|
|
|
|
No avatar | 4.9 (3.4) |
|
|
|
|
|
Avatar | 5.8 (2.6) |
|
|
|
|
|
10.1 | .001 | ||
|
|
|
|
|
|
|
|
|
|
No avatar | 4.8 (2.9) |
|
|
|
|
|
Generic avatar | 4.9 (3.0) |
|
|
|
|
|
Avatar with color choice | 4.8 (3.0) |
|
|
|
|
|
|
|
|
|
|
|
|
No avatar | 4.7 (3.0) |
|
|
|
|
|
Generic avatar | 4.8 (3.0) |
|
|
|
|
|
Avatar with color choice | 5.1 (3.1) |
|
|
|
|
|
|
|||
|
|
|
81.6 | <.001 | ||
|
|
Lower risk | 4.5 (3.0) |
|
|
|
|
|
Higher risk | 5.1 (3.0) |
|
|
|
|
|
Very high risk | 5.7 (2.8) |
|
|
|
|
|
|
44.5 | <.001 | ||
|
|
Blood pressure unknown | 4.2 (3.1) |
|
|
|
|
|
Blood pressure known | 5.0 (3.0) |
|
|
|
|
|
|
63.0 | <.001 | ||
|
|
Cholesterol unknown | 4.4 (3.1) |
|
|
|
|
|
Cholesterol known | 5.4 (2.9) |
|
|
aAssessed on scale of 0 (lowest intentions) to 9 (highest intentions).
We observed an interaction between random and actual risk in their association with recall. Participants at lower risk demonstrated a slight increase in correct recall in the random condition, those at higher risk showed a slight decrease, and those at very high risk had a larger decrease (see details in
Differences in numeracy were also associated with differences in recall. Participants with lower numeracy had more trouble accurately recalling their risk estimate. We also observed a similar main effect for blood pressure known. Participants who knew their blood pressure were also more able to recall their risk estimate.
Rerunning these analyses after removing the participants who had received a less precise estimate in the text (ie, those who received an estimate of “less than 1%” or “more than 30%” but who nonetheless received a risk graphic with a discrete number of event rectangles), we found that the main effects of moderating variables remained similar, but the interaction between random and actual risk was no longer significant (
Summary of findings for secondary outcome recall.
Effects | Participants with correct recalla |
|
|
|||
|
|
|
|
|||
|
|
7.06 | .008b | |||
|
|
|
|
|
|
|
|
|
|
Standard | 83% |
|
|
|
|
|
Random | 85% |
|
|
|
|
|
|
|
|
|
|
|
|
Standard | 79% |
|
|
|
|
|
Random | 76% |
|
|
|
|
|
|
|
|
|
|
|
|
Standard | 76% |
|
|
|
|
|
Random | 64% |
|
|
|
|
|
|
|||
|
|
|
75.7 | <.001 | ||
|
|
Lower numeracy | 74% |
|
|
|
|
|
Higher numeracy | 86% |
|
|
|
|
|
10.6 | .001 | |||
|
|
Blood pressure unknown | 74% |
|
|
|
|
|
Blood pressure known | 82% |
|
|
aAnalysis used quasi-continuous difference between recalled value and actual risk. Correct recall for reporting purposes defined as within 5 percentage points.
bNo longer significant when participants at very low or very high risk were removed from the sample.
Percent correct recalla by study arm.
Type of Avatar | All data included (n=3597) | <1 and >30 removed (n=3312) | ||
|
Standard | Random | Standard | Random |
No avatar | 85% | 80% | 85% | 82% |
Avatar moves: no, color choice: no | 86% | 82% | 86% | 82% |
Avatar moves: no, color choice: yes | 78% | 77% | 78% | 78% |
Avatar moves: yes, color choice: no | 78% | 83% | 79% | 84% |
Avatar moves: yes, color choice: yes | 77% | 75% | 78% | 78% |
aCorrect recall for reporting purposes defined as recall within 5 percentage points of given estimate.
Our results demonstrate several key findings. First, consistent with our earlier work [
The observed effect of animated randomness was driven by the strength of the effect for the 8% of participants who were at very high risk and whose risk estimate was presented in text as “more than 30%.” These participants had to wait as 31 event rectangles appeared randomly in the graphic one by one, never quite knowing where the next one would appear or when the process would stop. The combined uncertainty of the text statement, random positioning of the event rectangles, and uncertainty around how many event rectangles would appear may well compound each other and lead to a heightened sense of being at risk. Because it is common for models of health risk to be mathematically convergent only within certain boundary conditions and/or to generate ranges of risk estimates rather than point estimates, this design technique of animated randomness combined with the temporal signaling of one event appearing at a time may be broadly applicable. Nonetheless, further research will be needed to determine its effectiveness—or lack thereof—across a range of situations.
We further note that although there was no main effect of animated randomness on recall, those at higher risk demonstrated a small decrease in their ability to precisely recall the risk estimate they were given whereas those at lower risk demonstrated a small increase. As with the interaction discussed previously, this relationship was driven by the strength of effect in those at very high risk, who were not given a precise numerical risk estimate in text and thus had the more challenging task of recalling the number of event rectangles in their risk graphic. This additional difficulty may have contributed to the lower recall in this group. We further speculate that people who are reassured by a risk estimate may find it somewhat easier to remember the number whereas those who are alarmed may be less likely to remember an exact number because of distracting emotions, such as fear.
Despite the overall welcome finding that animated randomness may help better align risk perceptions with actual risk, we note that in the study context of lifestyle-preventable disease, emphasizing the haphazard and random distribution of negative outcomes led to lower intentions of behavior that might prevent such events. Visually depicting randomness may cause people to focus on the role of chance in health outcomes, drawing their attention to the fact that one’s behavior does not completely determine one’s health outcomes. We note that we did not present visual depictions of the potential for health behaviors to change the risk estimate. Doing so might possibly have reduced the negative effects of animated randomness on healthy behavior intentions. Further research will be necessary to fully understand this aspect of our findings. Such research should explore the role of potential moderators of health behavior, such as beliefs about the efficacy of lifestyle changes as well as self-efficacy and fatalism in health.
Ultimately, these findings suggest that the value of explicitly showing randomness may depend on whether one’s goal is to persuade or inform. Helping people understand randomness may be less useful in persuasive contexts such as promoting lifestyle change. However, it may be more useful in cases in which the primary objective is to inform; for example, in preference-sensitive decisions or when informing people about the risk of a side effect. Importantly, aiming to fully inform people is arguably more ethical than aiming to persuade them. This design factor will need to be tested in other contexts and may also require some unpacking to determine the effects of design choices that did not vary across experimental conditions, such as the fact that higher risks took a longer time to appear.
Second, using an avatar increased overall risk perceptions, and showed promise particularly for people who did not personally know anyone who had experienced grave outcomes in this context. Thus, avatars may be especially useful for drawing attention to risks related to rare or hidden conditions. This interaction, combined with the fact that the main effects of familiarity were similar to the effects of an avatar in the absence of familiarity, suggests that the avatar achieved our design goal of helping people better grasp how population-based statistics can apply to an individual. This conclusion is bolstered by the fact that use of an avatar significantly shifted intentions to see a physician in sensible directions, with people at lower and higher risk indicating, respectively, lower and higher intentions to see a doctor in the next 30 days.
However, other design factors related to the presentation of the avatar showed mixed results. Having the avatar move around randomly in the risk graphic appeared to simply confuse most participants. Allowing people to choose the color of their avatar may have increased identification with the avatar among those with higher numeracy, as they were more likely to indicate intentions to see a doctor when they were encouraged to choose the color. We speculate that those with higher numeracy may be better equipped to understand the risk graphic and thus, adding an extra element that draws their attention can be helpful. By contrast, the extra factor may have added a level of confusion or overwhelm for those with lower numeracy. Allowing people to choose the color of their avatar also appeared to make up for the loss of personal salience among participants who did not know their blood pressure and thus, slightly encouraged intentions to engage in health lifestyle behaviors and see a doctor. However, for those who knew their blood pressure, it tended to have the opposite effect. Taken together, these results suggest that when risk salience may be low, using a personalized avatar may help people feel like the risk applies to them, individually. However, these effects were small; moreover, if risk salience is higher, such basic attempts at personalization may backfire. Therefore, using color choice as a method of personalization, although efficient and quick, appears to be insufficient to allow all people to identify with their avatar and may even lead to undesirable results. Further research will be required to investigate the effects of different forms of personalization.
Third, we note that all 3 planned moderating factors had significant effects. For example, people at higher risk perceived their risk as higher and indicated stronger intentions to see a doctor in the next 30 days; people scoring lower on numeracy indicated higher overall risk perceptions, lower intentions toward healthy behaviors, and lower recall, and participants who knew someone who had died of cardiovascular causes had higher risk perceptions. We also observed that people at higher risk tended to indicate lower intentions toward healthy behavior. This latter finding—that people at higher risk have lower intentions to do such things as quit smoking, exercise, and eat well in the next 30 days—may reflect that those at lower risk may already be engaging in those behaviors and thus can easily indicate higher intentions toward such behaviors in the next 30 days. It also aligns with findings about how negative feedback can discourage people, whereas positive feedback can motivate people in a success breeds success cycle [
Analyses of participants who did or did not know their blood pressure and/or cholesterol suggested that these were important moderators, particularly the former. People who knew their blood pressure had greater alignment between their actual and perceived risk, overall higher intentions toward healthy lifestyle actions and seeing a doctor, and more accurately recalled their risk estimate. It may be that this latter finding is reflective of an underlying ability to recall numbers; however, in such a case, we would expect to see an interaction with numeracy, which was not present. Numeracy did interact with knowledge of one’s blood pressure when it came to risk perceptions and behavioral intentions. For people with higher numeracy, risk perception was consistent whether they knew their blood pressure or not, whereas for those with lower numeracy, their overall higher risk perceptions were further increased with knowledge of their blood pressure. In addition, for those with lower numeracy, knowing one’s blood pressure was more influential in increasing behavioral intentions than it was for those with higher numeracy. We speculate that people with lower numeracy may accord more importance to their blood pressure number. Similar to the results for blood pressure, people who knew their cholesterol were more likely to indicate intentions to engage in healthy behaviors and see a doctor in the next 30 days. These findings support the idea that risk estimates are likely to be more impactful when they are more individually tailored.
This study was limited by the fact that we used an Internet survey panel to recruit participants. Although this recruitment choice allows us to ensure a more diverse sample, it necessarily introduces selection bias in that participants are those who registered on a panel to take surveys; thus, they may not be representative of the broader population.
In addition, we observed small effects. This is expected in a study with outcomes such as risk perceptions and behavioral intentions because such outcomes have significant variation in individual responses. Even the actual risk estimate was associated with only a 1.4-point difference on a 10-point scale of risk perception between those at high and very low risk, suggesting that there is little room on the scale within which to work. This limits the overall utility of these design factors, as it may not be worth the additional design complexity and development time to make small gains. Such small gains, however, may be worth pursuing when one considers cumulative effects within a population.
The underlying risk model also limited this study in 2 ways. First, because the risk estimate for cardiovascular disease uses age as an important predictor, we were not able to isolate the effects of age on participants’ reactions to the different risk communication designs. It is possible, for example, that older adults might have different reactions to randomization or avatars. Further research will be needed to explore the effects of these kinds of design factors in older versus younger adults. Second, because the model does not provide numerical risk estimates below 1% or above 30%, approximately 8% of the sample population in this study did not receive a precise numerical text estimate. Instead, these participants received a risk estimate of “less than 1%” (7 participants) or “more than 30%” (279 participants) along with 1 or 31 event rectangles in their risk graphic, respectively. This additional ambiguity in the textual risk estimate appeared to amplify findings regarding randomization in the risk graphic. As discussed previously, this is a realistic portrait of many risk calculators, as many risk analysis models are mathematically convergent only within certain boundary conditions. However, the observed interaction between risk level and animated randomness may not translate to calculators that yield precise estimates across a full range of potential risk.
Finally, it is important to note that because this study was conducted in the United States before the introduction of the Affordable Care Act (“ObamaCare”), findings about participants who did not know their blood pressure or cholesterol and findings concerning intentions to see a doctor may reflect an underlying issue of lack of access to medical care rather than effects that would translate to other settings.
Previous work has suggested that randomly displaying events in an icon array can increase understanding of the random nature of such events, but at the expense of comprehension of the numerical risk estimate [
A previous study by Ancker and colleagues [
Another study conducted by Han and colleagues [
In comparing our study to previous work, we note that the studies by both Ancker et al [
To the best of our knowledge, ours is the first study to test the effects of including an avatar as a design factor in risk communication graphics. Previous work in other contexts has suggested that people identify strongly with their avatar [
More recent research has also suggested that graphical displays may not always outperform simple percentages or absolute frequencies in risk communication [
The present study continues a program of research by our group in which we previously noted that 2 animated displays side by side were problematic [
An animated display of risk that adds events one at a time in a randomly dispersed icon array and where they settle at the bottom of the display at the conclusion of the animation may help align risk perceptions with actual risk estimates without sacrificing number sense. This method shows promise for helping people better understand the random nature of risk. Such understanding may come at a cost of discouraging behavioral intentions, suggesting that the use of this method may depend on whether the goal of the risk communication is to persuade or to inform.
The use of an avatar in a risk graphic also shows promise for helping people to grasp how population-based statistics can apply to an individual, particularly in cases when the person does not know anyone who has experienced the outcome under consideration. An avatar that is animated to move randomly within the graphic does not appear to be helpful. Personalization via color choice shows mixed effects, suggesting that personalization of an avatar may be an interesting avenue for further study, but that this particular method of personalization does not appear to be optimal.
Video of risk graphic (.mp4 version).
Video of risk graphic (.avi version).
Secondary study: effects of heart age message.
Associate of Arts
Associate of Science
Bachelor of Arts
Bachelor of Science
General Educational Development
high-density lipoprotein
Master of Arts
Doctor of Medicine
Master of Public Health
Doctor of Philosophy
sample standard deviation
Survey Sampling International
This work was funded by a grant from The Foundation for Informed Medical Decision Making (FIMDM), now the Informed Medical Decisions Foundation (IIG 0126-1, PI: Zikmund-Fisher). Dr Zikmund-Fisher was supported by a career development award from the American Cancer Society (MRSG-06-130-01-CPPB). The funding agreements ensured the authors’ independence in designing the study, in collecting, analyzing, and interpreting the data, in writing the manuscript, and in the decision to submit the manuscript for publication. The authors thank Mark Swanson for programming the Flash-based animations for this study and the manuscript reviewers for their constructive comments.
All authors contributed to the conception and design of the study. HOW, AFF, MD, NE, VK, and BJZ-F contributed to technical design and HCW contributed clinical insight about use of the model and wording. MD programmed the application and survey. HOW, MD, LH, NE, and BJZ-F were involved in data acquisition. HOW analyzed and interpreted data with insight from HCW and BJZ-F. HOW wrote the first draft of the article; all authors revised it critically for important intellectual content, saw, and approved the final version submitted.
None declared.