Medical Decision Making

Background. A common feature of most reviews or catalogues of health utilities has been their focus on adult health states or derivation of values from adult populations. More generally, utility measurement in or on behalf of children has been constrained by several methodological concerns. The objective of this study was to conduct the first comprehensive systematic review and meta-analysis of primary utility data for childhood conditions and descriptors, and to determine the effects of methodological factors on childhood utilities. Methods. The review followed PRISMA guidelines. PubMed, Embase, Web of Science, PsycINFO, EconLit, CINAHL and Cochrane Library were searched for primary studies reporting health utilities for childhood conditions or descriptors using direct or indirect valuation methods. The Paediatric Economic Database Evaluation (PEDE) Porject was also searched for cost-utility analyses with primary utility values. Mean or median utilities for each of the main samples were catalogued, and weighted averages of utilities for each health condition were estimated, by valuation method. Mixed-effects meta-regression using hierarchical linear modeling was conducted for the most common valuation methods to estimate the utility decrement for each health condition category relative to general childhood population health, as well as the independent effects of methodological factors. Results. The literature searches resulted in 272 eligible studies. These yielded 3,414 utilities when all sub-groups were considered, covering all ICD-10 chapters relevant to childhood health, 19 valuation methods, 12 respondent types, 8 modes of administration, and data from 36 countries. A total of 1,191 utility values were obtained when only main study samples were considered, and these were catalogued by health condition or descriptor, and methodological characteristics. 1,073 mean utilities for main samples were used for fixed-effects meta-analysis by health condition and valuation method. Mixed-effects meta-regressions estimated that 53 of 76 ICD-10 delineated health conditions, valued using the HUI3, were associated with statistically significant utility decrements relative to general population health, whereas 38 of 57 valued using a visual analog scale (VAS) were associated with statistically significant VAS decrements. For both methods, parental proxy assessment was associated with overestimation of values, whereas adolescents reported lower values than children under 12 y. VAS responses were more heavily influenced by mode of administration than the HUI3. Conclusion. Utilities and their associated distributions, as well as the independent contributions of methodological factors, revealed by this systematic review and meta-analysis can inform future economic evaluations within the childhood context.

Background. Discrete Choice Experiments including duration (DCE_TTO) can be used to generate utility values for health states from measures such as EQ-5D-5L. However, methodological issues concerning the optimum way to present choice sets remain. The aim of the present study was to test a range of task presentation approaches designed to support the DCE_TTO completion process. Methods. Four separate presentation approaches were developed to examine different task features including dimension level highlighting, and health state severity and duration level presentation. Choice sets included 2 EQ-5D-5L states paired with 1 of 4 duration levels, and a third “immediate death” option. The same design, including 120 choice sets (developed using optimal methods), was employed across all approaches. The online survey was administered to a sample of the Australian population who completed 20 choice sets across 2 approaches. Conditional logit regression was used to assess model consistency, and scale parameter testing investigated poolability. Results. Overall 1,565 respondents completed the survey. Three approaches, using different dimension level highlighting techniques, produced mainly monotonic coefficients that resulted in a larger disutility as the severity level increased (excepting usual activities levels 2/3). The fourth approach, using a level indicator to present the severity levels, has slightly more non-monotonicity and produced larger ordered differences for the more severe dimension levels. Scale parameter testing suggested that the data cannot be pooled. Conclusions. The results provide information regarding how to present DCE tasks for health state valuation. The findings improve our understanding of the impact of different presentation approaches on valuation, and how DCE questions could be presented to be amenable to completion. However, it is unclear if the task presentation impacts online respondent engagement.

Background. Existing methods to link preference-based and profile-based health-related quality of life (HRQoL) questionnaires have their limitations. Hence, we developed a new mapping method (the mean rank method, MRM) and applied it to map the World Health Organization Quality of Life Questionnaire (WHOQOL-BREF) to the EuroQoL 5 Dimensions 5 Levels (EQ-5D-5L). We then compared the new MRM with current methods; i.e., regression-mapped (OLS method) and equipercentile method (EPM). Methods. Singapore residents, aged ≥21 y, were recruited from the general population and 2 outpatient clinics in acute care hospitals. Performance of the MRM was evaluated using both simulation and split-sample validation (n = 658 in training and n = 657 in validation samples). Using the training sample, we derived 3 sets of mapped EQ-5D-5L utilities based on MRM, OLS method and EPM. Using simulation and the validation sample, we compared the performance of the mapping methods in terms of distribution parameters, mean utility by strata, association with health covariates, and prediction errors at the individual level, among others. Results. The WHOQOL-BREF Physical Health domain is the only domain significantly associated with EQ-5D-5L utilities. Simulation showed that MRM more accurately reproduced the variance and percentiles of the distribution of the observed utilities than did the OLS method or EPM. OLS method tended to underestimate the mean utility of good health states, overestimate the mean utility of poor health states, and underestimate the association with covariates. An analysis of validation sample gave similar results. Conclusion. In scenarios similar to the mapping of WHOQOL-BREF to the EQ-5D-5L, the MRM outperformed the OLS method and EPM in important—though not all—aspects. The simplicity and reproducibility of the MRM makes it an attractive alternative to current methods.

Background. Although there has been increasing interest in patient engagement, few measures are publicly available and suitable for patients with limited health literacy. Objective. We sought to develop a Consumer Health Activation Index (CHAI) for use among diverse patients. Methods. Expert opinion, a systematic literature review, focus groups, and cognitive interviews with patients were used to create and revise a potential set of items. Psychometric testing guided by item response theory was then conducted among 301 English-speaking, community-dwelling adults. This included differential item functioning analyses to evaluate item performance across participant health literacy levels. To determine construct validity, CHAI scores were compared to scales measuring similar personality constructs. Associations between the CHAI and physical and mental health established predictive validity. A second study among 9,478 adults was used to confirm CHAI associations with health outcomes. Results. Exploratory factor analyses revealed a single-factor solution with a 10-item scale. The CHAI showed good internal consistency (alpha = 0.81) and moderate test–retest reliability (ICC = 0.53). Reading grade level was found to be at the 6^th grade. Moderate to strong correlations were found with similar constructs (Multidimensional Health Locus of Control, r = 0.38, P < 0.001; Conscientiousness, r = 0.41, P < 0.001). Predictive validity was demonstrated through associations with functional health status measures (depression, r = −0.28, P < 0.001; anxiety, r = −0.22, P < 0.001; and physical functioning, r = 0.22, P < 0.001). In the validation sample, the CHAI was significantly associated with self-reported physical and mental health (r = 0.31 and 0.32 respectively; both P < 0.001). Conclusions. The CHAI appears to be a valid, reliable, and easily administered tool that can be used to assess health activation among adults, including those with limited health literacy. Future studies should test the tool in actual use and explore further applications.

Background. Variation in the intensity of acute care treatment at the end of life is influenced more strongly by hospital and provider characteristics than patient preferences. Objective. We sought to describe physicians’ mental models (i.e., thought processes) when encountering a simulated critically and terminally ill older patient, and to compare those models based on whether their treatment plan was patient preference-concordant or preference-discordant. Methods. Seventy-three hospital-based physicians from 3 academic medical centers engaged in a simulated patient encounter and completed a mental model interview while watching the video recording of their encounter. We used an “expert” model to code the interviews. We then used Kruskal–Wallis tests to compare the weighted mental model themes of physicians who provided preference-concordant treatment with those who provided preference-discordant treatment. Results. Sixty-six (90%) physicians provided preference-concordant treatment and 7 (10%) provided preference-discordant treatment (i.e., they intubated the patient). Physicians who intubated the patient were more likely to emphasize the reversible and emergent nature of the patient situation (z = −2.111, P = 0.035), their own comfort (z = −2.764, P = 0.006), and rarely focused on explicit patient preferences (z = 2.380, P = 0.017). Limitations. Post-decisional interviewing with audio/video prompting may induce hindsight bias. The expert model has not yet been validated and may not be exhaustive. The small sample size limits generalizability and power. Conclusions. Hospital-based physicians providing preference-discordant used a different mental model for decision making for a critically and terminally ill simulated case. These differences may offer targets for future interventions to promote preference-concordant care for seriously ill patients.

Objective. Many patients have low numeracy, which impedes their understanding of important information about health (e.g., benefits and harms of screening). We investigated whether physicians adapt their risk communication to accommodate the needs of patients with low numeracy, and how physicians’ own numeracy influences their understanding and communication of screening statistics. Methods. UK family physicians (N = 151) read a description of a patient seeking advice on cancer screening. We manipulated the level of numeracy of the patient (low v. high v. unspecified) and measured physicians’ risk communication, recommendation to the patient, understanding of screening statistics, and numeracy. Results. Consistent with best practices, family physicians generally preferred to use visual aids rather than numbers when communicating information to a patient with low (v. high) numeracy. A substantial proportion of physicians (44%) offered high quality (i.e., complete and meaningful) risk communication to the patient. This was more often the case for physicians with higher (v. lower) numeracy who were more likely to mention mortality rates, OR=1.43 [1.10, 1.86], and harms from overdiagnosis, OR=1.44 [1.05, 1.98]. Physicians with higher numeracy were also more likely to understand that increased detection or survival rates do not demonstrate screening effectiveness, OR=1.61 [1.26, 2.06]. Conclusions. Most physicians know how to appropriately tailor risk communication for patients with low numeracy (i.e., with visual aids). However, physicians who themselves have low numeracy are likely to misunderstand the risks and unintentionally mislead patients by communicating incomplete information. High-quality risk communication and shared decision making can depend critically on factors that improve the risk literacy of physicians.

Background. Previous research has shown that format effects, such as the “1-in-X” effect—whereby “1-in-X” ratios lead to a higher perceived probability than “N-in-N*X” ratios—alter perceptions of medical probabilities. We do not know, however, how prevalent this effect is in practice; i.e., how often health professionals use the “1-in-X” ratio. Methods. We assembled 4 different sources of evidence, involving experimental work and corpus studies, to examine the use of “1-in-X” and other numerical formats quantifying probability. Results. Our results revealed that the use of the “1-in-X” ratio is prevalent and that health professionals prefer this format compared with other numerical formats (i.e., the “N-in-N*X”, %, and decimal formats). In Study 1, UK family physicians preferred to communicate prenatal risk using a “1-in-X” ratio (80.4%, n = 131) across different risk levels and regardless of patients’ numeracy levels. In Study 2, a sample from the UK adult population (n = 203) reported that most GPs (60.6%) preferred to use “1-in-X” ratios compared with other formats. In Study 3, “1-in-X” ratios were the most commonly used format in a set of randomly sampled drug leaflets describing the risk of side effects (100%, n = 94). In Study 4, the “1-in-X” format was the most commonly used numerical expression of medical probabilities or frequencies on the UK’s NHS website (45.7%, n = 2,469 sentences). Conclusions. The prevalent use of “1-in-X” ratios magnifies the chances of increased subjective probability. Further research should establish clinical significance of the “1-in-X” effect.

Background. Although people are likely to underestimate the frequencies of risks to health from common diseases and overestimate those from rare diseases, we still do not know much about reasons for this systematic bias, which is also referred to as “primary bias” in the literature. In this study, we take advantage of a series of large epidemics of mosquito-borne diseases to examine the accuracy of judgments of risk frequencies. In this aim, we assessed the perceived v. observed prevalence of infection by Zika, chikungunya or dengue fever during these outbreaks, as well as their variations among different subpopulations and epidemiological settings. Methods. We used data drawn from 4 telephone surveys, conducted between 2006 and 2016, among representative samples of the adult population in tropical regions (Reunion, Martinique, and French Guiana). The participants were asked to estimate the prevalence of these infections by using a natural frequency scale. Results. The surveys showed that 1) most people greatly overestimated the prevalence of infection by arbovirus, 2) these risk overestimations fell considerably as the actual prevalence of these diseases increased, 3) the better-educated and male participants consistently yielded less inaccurate risk estimates across epidemics, and 4) these biases in the perception of prevalence of these infectious diseases are relatively well predicted by the probability weighting function developed in the field of behavioral decision making. Conclusions. These findings suggest that the primary bias, which has been found in laboratory experiments to characterize a variety of probabilistic judgments, equally affects perception of prevalence of acute infectious diseases in epidemic settings. They also indicate that numeracy may play a considerable role in people’s ability to transform epidemiological observations from their social environment to more accurate risk estimates.

Patients and doctors often need to make decisions based on the results of medical tests. When these results are presented in the form of conditional probabilities, even doctors find it difficult to interpret them correctly. There is over 20 y of research supporting the finding that people are better able to calculate the correct positive predictive value of a test when given information in natural frequencies, as opposed to conditional probabilities. Natural frequencies are one of a few psychological tools that have made it into evidence-based medicine. Recently, Pighin and others (Med Decis Making 2016;36:686–91) argued that natural frequencies could hinder informed decision making, a critique based on a single task and a crude scoring criterion we refer to as the 50%-Split. Our commentary addresses these criticisms based on three analyses. First, we show how the 50%-Split scoring used by Pighin and others misclassifies known errors, such as solely attending to the hit rate (true-positive rate) of the test, as strategies that support understanding. Second, we reanalyze data from 21 additional problems completed by various participant groups to show that their scoring criterion does not support their results in 19 out of 21 cases. Third, we apply the mean deviation scoring method and show that, when given information in natural frequency formats, participants provide estimates that are closer to the correct Bayesian solution than for conditional probability formats. In each analysis, natural frequencies lead to more correct judgements and therefore promote informed decision making relative to conditional probabilities. We welcome further discussions of performance metrics that can provide insight into how the public and therefore patients understand the implications of medical test results.

Microsimulation models are becoming increasingly common in the field of decision modeling for health. Because microsimulation models are computationally more demanding than traditional Markov cohort models, the use of computer programming languages in their development has become more common. R is a programming language that has gained recognition within the field of decision modeling. It has the capacity to perform microsimulation models more efficiently than software commonly used for decision modeling, incorporate statistical analyses within decision models, and produce more transparent models and reproducible results. However, no clear guidance for the implementation of microsimulation models in R exists. In this tutorial, we provide a step-by-step guide to build microsimulation models in R and illustrate the use of this guide on a simple, but transferable, hypothetical decision problem. We guide the reader through the necessary steps and provide generic R code that is flexible and can be adapted for other models. We also show how this code can be extended to address more complex model structures and provide an efficient microsimulation approach that relies on vectorization solutions.