Monday, October 24, 2016: 2:00 PM - 3:30 PM
Bayshore Ballroom Salon D, Lobby Level (Westin Bayshore Vancouver)

Brian J. Zikmund-Fisher, PhD
University of Michigan
Associate Professor
Health Behavior and Health Education

2:00 PM

Jonathan R.G. Etnel1, Thomas H. Oostdijk2, Yvonne J.M. Licher2, Sanne Treep2, Peterke J. van der Zwaag2, Ad J.J.C. Bogers, MD, PhD1 and Johanna Takkenberg, MD, PhD1, (1)Department of Cardiothoracic Surgery, Erasmus University Medical Center, Rotterdam, Netherlands, (2)Faculties of Industrial Design, Physics and Mathematics, Delft University of Technology, Delft, Netherlands
Purpose: Information portals and decision aids often employ risk visualizations to help users better understand numerical risks. But are they indeed more effective than simple numerical risks alone? We assessed numeracy among the general Dutch population and investigated the effectiveness of various risk visualizations in improving the understanding of numerical risks.

Method: Randomly sampled from four public locations across the Netherlands and based on our power calculations, 187 members of the general Dutch population were included. Each respondent completed a questionnaire (4 questions) on basic risk conversions from percentages to natural frequencies(2 questions) and vice versa(2 questions), adapted from the Numeracy Scale. In each of these four questions, the numerical risk was portrayed by a different visualization: a pie chart, icon array or bar graph. Question–visualization combinations were varied across four different versions of the questionnaire. A fifth version of the questionnaire contained no risk visualizations. Allocation of the respondents to one of these five versions of the questionnaire was randomized and investigator-blinded. Respondent performance was scored ordinally from 0-4 corresponding with the number of questions answered correctly. Demographics and individual visualization preference were also recorded.

Result: Age(p=0.075), gender(p=0.849) and education level(p=0.069) were all comparable to the overall general Dutch population. Only 73 respondents(39.0%) answered all four questions correctly and 41(21.9%) answered three out of four questions correctly. Younger age(odds ratio[OR]/10 years of age=1.19;p=0.015) and higher education level(OR=1.27;p<0.001) were independently associated with a higher score, whereas male vs. female gender was not(OR 1.26;p=0.399). The use of risk visualizations was not associated with a higher score(OR=1.05;p=0.878). This finding held when considering each question individually(n=748 [4 per respondent];66.7% correctly answered): questions supported by a visualization were not more frequently answered correctly than questions that were not (OR=1.07;p=0.700). Each visualization considered separately, neither pie charts(OR=1.34;p=0.236), icon arrays(OR=1.12;p=0.590) nor bar graphs(OR=0.81;p=0.381) significantly improved the odds of answering a question correctly compared to no visualization. Pie charts were significantly more effective than bar graphs(p=0.031). Individual preference for a certain visualization was not associated with its effectiveness(OR=1.02;p=0.905).

Conclusion: The understanding of numerical risks among the general Dutch population is poor and is not improved by the use of risk visualizations. Pie charts are more effective than bar graphs in conveying risks. This provides helpful information for the development of patient information portals and decision aids.

2:15 PM

Raluca Cozmuta, MD, MHS, Emory University, Atlanta, GA, Evan A. Wilhelms, M.S., Vassar College, Poughkeepsie, NY, Valerie Reyna, PhD, Cornell University, Ithaca, NY, Julia Nolte, BSc, Heidelberg University, Heidelberg, Germany and Liana Fraenkel, MD, MPH, Yale University, West Haven, CT

Purpose: To examine whether an icon array (IA), an illustration of the gist of how medications regulate the immune system (a series of balance beams), or both influence willingness to start a medication.

Methods: Patients with a rheumatic disease were mailed a survey. They were asked to imagine that their symptoms had worsened and that their physician was recommending a new medication. We varied the probability of an adverse event (pneumonia requiring hospitalization): 2% or 0.2%, and the risk presentation format: numbers, numbers + IA, numbers + balance beams (BB), or numbers + both. Administration, benefit, and cost were held constant. Each subject responded to a single, randomly-assigned scenario. We controlled for socioeconomic status (SES) in a full-factorial model testing willingness to take the medication (5-point scale).

Results: Of 1453 surveys, 465 patients completed the survey. Overall, the mean (SD) age was 59.0 (14.8); 79.7% were female; 83.2% White and 39.1% were classified as having low SES. There were no statistical differences in patient characteristics across the risk presentation formats. Willingness to start the medication was predicted by the interaction between the risk presentation format and SES (F = 2.9, p = 0.03). Among low SES subjects, addition of an IA did not affect willingness compared to the numbers-only format. In contrast, addition of BB (mean difference = 0.47, p = 0.07), or both IA and BB increased willlingness (mean difference = 0.48, p = 0.04). Among high SES subjects, addition of an IA or BB or both did not influence willingness compared to the numbers only format. However, both formats including an IA increased willingness compared to the BB format among high SES subjects (mean difference IA vs BB= 0.53, p =0.01; mean difference IA vs IA + BB = 0.48, p = 0.02).

Conclusion: SES affects how subjects respond to risk presentation formats. IA marginally increases willingness in high SES subjects, while BB increases willingness in low SES subjects; when both IA and BB are present, SES differences disappear. BB, when not accompanied by an IA, may decrease willingness in high SES subjects. These results add to the literature demonstrating the differential effects of risk presentation formats, and highlight the need to identify mechanisms underlying their effects before implementing decision-support tools.


2:30 PM

Raluca Cozmuta, MD, MHS, Emory University, Atlanta, GA, Evan A. Wilhelms, M.S., Vassar College, Poughkeepsie, NY, Valerie Reyna, PhD, Cornell University, Ithaca, NY, Julia Nolte, BSc, Heidelberg University, Heidelberg, Germany and Liana Fraenkel, MD, MPH, Yale University, West Haven, CT

Purpose: Effectively communicating risk information to patients is difficult. We examined whether addition of an icon array (IA), a series of 3 consecutive balance-beam (BB) illustrations depicting how medications regulate the immune system, or both resulted in patients being better able to differentiate between an uncommon (2%) and rare (0.2%) adverse event (AE).

Methods: Patients currently being treated for a chronic inflammatory rheumatic disease were mailed a survey in which they were asked to imagine that their symptoms had recently worsened and that their physician was recommending a new medication. The medication was described using 8 scenarios (manipulated using a 2x4 design). We varied the probability of a serious AE (pneumonia requiring hospitalization): 2% vs 0.2% and the risk presentation format: numbers only, numbers + IA, numbers + BB, and numbers + IA + BB. Route of administration, benefit, and cost were held constant. Each subject responded to a single, randomly-assigned scenario. Dependent variables included perceived riskiness, worry, global gist related to the acceptability of the side effect, and willingness to take the medication (all measured on 5-point ordinal scales).

Results: We mailed 1453 surveys. 465 patients completed and mailed the survey back (32% response rate). Overall, the mean age of responders was 58.99 (SD=14.85); 79.7% of were female; 83.2% White and 39.1% had a low socioeconomic status. There were no statistical differences in demographic or clinical characteristics across the four risk presentation formats. Mean (SD) perceived riskiness, worry, global gist, and willingness to take the medication for 2% versus 0.2% chance of the AE, by socioeconomic status (SES) level, are presented in the Table.

Perceived riskiness was lower for a 0.2% versus 2% risk of the AE in the numbers + IA condition in higher SES subjects. Lower SES subjects who viewed both IA and BB were more worried about the AE and found the AE to be less acceptable in the .2% versus 2% condition.

Conclusion: With the exception of the IA's impact on perceived riskiness, the risk formats used did not enable subjects to correctly differentiate between a 0.2% and a 2% risk of a serious AE.

2:45 PM

Angie Fagerlin, PhD1, Megan Knaus, MPH2, Brian J. Zikmund-Fisher, PhD2, Enny Das, PhD3 and Aaron M. Scherer, PhD2, (1)University of Utah / VA Salt Lake City, Salt Lake City, UT, (2)University of Michigan, Ann Arbor, MI, (3)Radboud University Nijmegen, Nijmgen, Netherlands
Purpose: With infectious diseases increasingly in the headlines (e.g., Ebola, Zika), it is important to identify methods to communicate disease risks that influence perceived risk and vaccination intentions. Currently, emotional reactions are often not concordant with the actual risk. For example, worry is often either greater than the actual risk (e.g., Ebola) or less than the actual risk (e.g., influenza). As part of a larger, multi-country study, we evaluated the effects of different risk graphics and disease terms on people’s reactions to an infectious disease.

Method: Using Internet panels in 11 countries (United States, United Kingdom, Germany, Netherlands, Hungary, Poland, Finland, Sweden, Norway, Italy, Spain; N=16,037) we tested two factors on participants’ knowledge, risk perceptions, and intended behavior in an influenza scenario: 1) graphical communication of prevalence and death via either a heat map, dot map, or trendline and 2) different terms used to describe a hypothetical type of influenza (H11N3 Influenza, Horse Flu, Yarraman Flu). Materials were translated to each country’s language.

Result: The heat map display produced the highest vaccination intentions (F=4.67, p=0.01), highest perceived likelihood of contracting influenza (F=7.47, p<0.001), and the greatest interest in seeking more information about the outbreak (F=4.04, p=0.02). It was also most preferred (F=248.81, p<0.001). Using Yarraman Flu or H11N3 Influenza produced greater perception of likelihood of contracting the disease (F=5.95, p=0.003) and greater interest in vaccination (F=5.23, p=0.005). Yarraman Flu and Horse Flu resulted in a greater interest in seeking more information (F=16.13, p<0.001).

Cross country comparisons revealed few differences in the risk graphic used. However, cross-country differences were observed in response to the disease term manipulation: Knowledge was highest when the flu was called H11N3 influenza among subjects in Germany, Hungary, and Italy, but lowest among subjects in Finland, Spain, and Sweden. The flu label seemed to have little impact on knowledge among subjects in the Netherlands, Poland, the UK, and the US.

Conclusion: When communicating to the public about infectious diseases, both heat map displays and use of either a scientific term (H11N3 Influenza) or an exotic sounding term (Yarraman flu) to label the disease increased risk perceptions and interest in vaccinations. Using these methods in future pandemics could improve the public’s interest in reading information about novel diseases and getting necessary vaccinations.

3:00 PM

Pete Wegier, PhD and Victoria A. Shaffer, PhD, University of Missouri, Columbia, MO

Diagnostic screening tests require an understanding of conditional probabilities, specifically the positive predictive value—PPV—of a test (i.e., "given a positive test result, what is the probability that you have the disease?"). Evidence exists for a "collective statistical illiteracy", even amongst highly educated individuals (Gigerenzer et al., 2008). We investigated whether presenting statistical properties of screening tests by simulating test outcomes—simulated experience—improves understanding of conditional probabilities, compared to a descriptive presentation of statistical information.


Across two studies (Experiment 1, undergraduate students, N=64; Experiment 2, Internet sample recruited via Amazon Mechanical Turk, N=176), participants were presented with a vignette describing a person receiving a positive result from a full integrated screening test for Down syndrome. Participants were asked to estimate the PPV of the test. In the description condition, participants were provided the prevalence of Down syndrome and the sensitivity and specificity of the test—the PPV could be derived using Bayes' theorem. In the simulated experience condition, participants were presented with grids of 100 coloured squares denoting a representative sample of disease statuses and screening test results (see figure), blue representing false positives and orange representing true positives. Participants could sample up to 5000 fictitious test results and PPV could be estimated by tracking experienced outcomes. Experiment 2 also included three items measuring attitudes toward the test before and after providing PPV estimates.


Across the two experiments, participants' PPV estimates in the simulated experience condition were significantly more accurate than the description condition, (79% of the experience condition provided an estimate within 5 percentage points of the underlying PPV, compared with 14% of the description condition), p<.001. Experiment 2 revealed participants' attitudes towards screening significantly decreased after providing PPV estimates in the experience condition, but not the description condition, p<.05. Participants in the experience condition rated their likelihood of undergoing screening as 4.66 (6-point scale) 95%CI [4.37, 4.95] before and only 3.81 [3.47, 4.16] after estimating PPV. No differences were observed in the description condition.


Simulated experiences can significantly improve PPV estimates while decreasing interest in screening. This intervention has the potential to dramatically improve patient understanding of diagnostic screening tests and may also be used to reduce reliance on questionable screening tests, and this reducing overtreatment.

3:15 PM

Alexander Davis, PhD1, Baruch Fischhoff, PhD1, Tamar Krishnamurti, PhD1 and Janel Hanmer, MD, PhD2, (1)Carnegie Mellon University, Pittsburgh, PA, (2)The University of Pittsburgh, Pittsburgh, PA
Purpose: We conducted a longitudinal cohort study examining physician use of probabilities to assess patient risk of deterioration.

Method: Over a six month period, residents, interns, overnight interns, and attending physicians from the general internal medicine ward of a single academic tertiary care center were asked to assess: 1) the probability that each of their patients would experience a clinical deterioration in the next 24 hours (defined as a cardiac arrest, call to a rapid response team, or transfer to the ICU), and 2) the probability of just a rapid response team call for their last patient for the day. This latter probability was compared to the physician's assessed probability of the extensional (inclusive) event (cardiac arrest, rapid response team call, or transfer to the ICU).

Result: 135 unique physicians made 6,310 judgments across the entire study. Physicians remained in the study for either two or four weeks with four to eight data collection points. Judges were initially insensitive to the base rate (~2%) when predicting 24 hour decompensation risk for their patients (median = 10%). Over the duration of their rotation, their probability assessments significantly decreased (median = 5%) but remained consistently higher than the base rate. This pattern held for all cohorts of physicians (interns, residents, and attendings). Attendings violated the laws of probability by giving a lower probability to the extensional (inclusive) event 8% of the time, compared to 10% for residents, 12% for interns, and 19% for overnight interns. For the day teams, these violations tended to strongly cluster within physicians. For the overnight team, violations were more evenly distributed across physicians.

Conclusion: Physicians initially overestimated the probability of their patients deteriorating by a factor of 5, but over time gave probabilities about 2 times the base rate. The pattern held for cohorts of very different experience levels (interns, residents, attendings), suggesting that this learning was due to being explicitly asked to assess the probabilities. A minority of physicians consistently violated the extensional laws of probability. However, rates were higher (and more uniform) for overnight physicians, who may have been more cognitively taxed (due to sleep deprivation).