Medical Decision Making

Background. The SF-36 and the Quality of Well-being index (QWB) both quantify health status, yet have very different methodologic etiologies. The authors sought to develop an empirical equation allowing prediction of the QWB from the SF-36. Data. They used empirical observations of SF-36 profiles and QWB scores collected in in terviews of 1,430 persons during the Beaver Dam Health Outcomes Study, a com munity-based population study of health status, and 57 persons from a renal dialysis clinic. Method. The eight scales of the SF-36, their squares, and all pairwise cross- products, were used as candidate variables in stepwise and best-subsets regressions to predict QWB scores using 1,356 interviews reported in a previous paper. The re sulting equation was cross-validated on the remaining 74 cases and using the renal dialysis patients. Results. A six-variable regression equation drawing on five of the SF- 36 components predicted 56.9% of the observed QWB variance. The equation achieved an R² of 49.5% on cross-validation using Beaver Dam participants and an R² of 58.7% with the renal dialysis patients. An approximation for computing confidence intervals for predicted QWB mean scores is given. Conclusion. SF-36 data may be used to predict mean QWB scores for groups of patients, and thus may be useful to modelers who are secondary users of health status profile data. The equation may also be used to provide an overall health utility summary score to represent SF-36 profile data so long as the profiles are not severely limited by floor or ceiling effects of the SF-36 scales. The results of this study provide a quantitative link between two important measures of health status. Key words: health status; SF-36; Quality of Well- being index; quality of life; health-state utility; population study. (Med Decis Making 1997;17:1-9

This paper examines how values should be assigned to health states when policy decisions must be made about who should receive treatment. The paper demonstrates that, if priority were to be assigned to those people who would benefit most from treatment, standard health-state utilities might fail to identify resource allocations that would maximize total health-related well-being in society. A new measurement instru ment is proposed that is based on the direct comparison of the well-being achieved by different people in various health states and thus captures such community priori ties. A sample of 72 health administration students used the instrument to evaluate speech and mobility dysfunctions as they afflicted hypothetical people who differed by gender, family status, and occupational type. This preliminary analysis indicates that the instrument is feasible to use, and that the valuations of respondents did, for some health conditions, significantly depend on the type of person afflicted. Key words: util ities ; interpersonal comparisons; ethics; health policy; extended-sympathy instrument. (Med Decis Making 1997;17:10-20)

Pliskin, Shepard, and Weinstein identified three preference conditions that ensure that quality-adjusted life years (QALYs) represent preferences over gambles over chronic health profiles. This paper presents an experimental test of the descriptive validity of two of these preference assumptions: utility independence and constant proportional tradeoff. Eighty students at the Stockholm School of Economics and 92 students at Erasmus University Rotterdam participated in the experiment. The results of the ex periment support the descriptive validity of constant proportional tradeoff: both within groups and between groups constant proportional tradeoff could not be rejected. The results are less supportive of the descriptive validity of utility independence. Within- groups utility independence was rejected. Between-groups utility independence could not be rejected, but this may have been due to a lack of statistical power. Analysis of the individual responses revealed that without adjustment for imprecision of preference, 39 respondents (22.8%) satisfied constant proportional tradeoff. Twenty-three respon dents (13.4%) satisfied utility independence without adjustment for imprecision of pref erence. However, because of the relative unfamiliarity of the respondents with both the health states to be evaluated and the methods of health-state-utility measurement, it is likely that the respondents' preferences were imprecise. Adjusted for imprecision of preference, the upper estimates of the proportions of respondents who satisfied constant proportional tradeoff and utility independence, respectively, were 90.1% (155 respondents) and 75.6% (130 respondents). Pliskin et al. further derived that if an individual's preferences satisfy both constant proportional tradeoff and utility indepen dence, then these preferences can be represented by a more general, risk-adjusted QALY model. Without adjustment for imprecision of preference, ten respondents (5.8%) satisfied both constant proportional tradeoff and utility independence. Adjusted for imprecision of preference, the upper estimate of the proportion of respondents who satisfied both constant proportional tradeoff and utility independence was 68.6% (118 respondents). The results of this study indicate that constant proportional tradeoff holds approximately. The evidence is much weaker for utility independence, however. This has important implications for the use of QALY-type measures in medical decision making. Key words: QALYs; health utility measurements; medical decision making; individual preferences. (Med Decis Making 1996;17:21-32)

Scaling instruments for the measurement of quality-adjusted life years (QALYs) incor porate either a personal or an impersonal perspective on the benefits of a health intervention and either do or do not incorporate considerations of equity. This paper sets out three hypotheses concerning perspective and equity: 1) that more equally distributed benefits will be preferred to less equally distributed benefits; 2) that the preference value elicited for a health benefit will be greater when the respondent to a QALY questionnaire is a potential beneficiary; and 3) that, by comparison with personal preferences, individuals will be more concerned with the quantity than the quality of life in other people. These hypotheses were tested using two existing instruments and two other instruments that were created for this study. Results gave no support to the third hypothesis; some support to the first hypothesis, and strong support for the sec ond hypothesis. It is concluded that perspective can significantly alter the values in corporated in a QALY instrument. The policy implications of the results are twofold. First, they give some support to the view that distributional consequences of health programs are of importance to the population and that they should be included in the evaluation of any health program. Second, they indicate that an evaluation should consider whether the health-state values to be obtained should incorporate an imper sonal perspective reflecting the purely "social" judgment of a health planner or a per spective reflecting self-interest. Key words: quality-adjusted life years; scaling instru ments ; equity; preferences; policy. (Med Decis Making 1997;17:33-41)

The authors conducted a study exploring whether preferences for sequences of events can be approximated by preferences for component discrete states. Visual-analog- scale (VAS) and standard-gamble (SG) scores for a subset of the possible sequences of events (path states) and component temporary and chronic outcomes (discrete states) that can follow prenatal diagnostic decisions were elicited from 121 pregnant women facing a choice between chorionic villus sampling and amniocentesis. For in dividuals, preference scores for path states could not be predicted easily from discrete- state scores. Mean path-state VAS scores, however, were predicted reasonably ac curately by multiple regression models (R² = 0.85 and 0.82 for two different anchoring schemes), with most measured scores lying within the 95% confidence intervals of the derived scores. It is concluded that, for individual patient decision making, preferences for path states should be elicited. When mean preference values for a population are sought, however, it may be reasonable to derive regression weights from a subset of respondents and then to apply those weights to preferences for discrete states elicited from a larger group. Key words: utility measurement; patient preferences; multiple re gression ; standard gamble; visual analog scaling; prenatal diagnosis. (Med Decis Mak ing 1997;17:42-55))

Objective. To determine the influence of alternative explanations by physicians of the purpose of a medical intervention (intubation and ventilatory support, IVS) on three types of patient preferences: desire for IVS, the length of time patients would find IVS acceptable, and the minimum probability of a good medical outcome patients would require before assenting to continued support with IVS. Methods. Structured interviews were conducted with patients followed in a continuity care general medicine clinic at the Department of Veterans Affairs Medical Center, Portland, Oregon. Patients were asked to consider whether they would accept IVS at a future time. Patients were randomly assigned to one of two explanation conditions that differed in terms of their future medical contexts, that is, the degrees of specification of the nature of the medical condition patients were asked to consider. The general-explanation group was asked to consider the future medical context of "an unspecified medical condition"; the spe cific-explanation group was asked to consider the future medical context of "a severe pneumonia." Patients were asked three questions: 1) Would you accept IVS (yes or no)? 2) How long would you allow your physician to continue IVS?; and 3) After being on IVS for 2-3 days, what would be the minimum chance of recovery from the con dition you would require before agreeing to continued IVS? "Chance of recovery" was defined for both groups as the probability that the patient would be able to leave the hospital and be able to take care of activities of daily living unassisted with minimal change in his or her mental state from the pre-hospitalization status. Results. Of 186 patients (mean age = 66.5 years; mean education = 12.7 years), 97 received the general explanation and 89 received the specific explanation about a severe pneu monia. Significantly fewer (p = 0.03) patients receiving the general explanation wanted physicians to intervene with IVS (general 94% vs specific 100%). Patients receiving the general explanation were willing to accept significantly fewer (p = 0.009) days of intubation (general 65 days vs specific 96 days). Significantly fewer (p < 0.0001) pa tients receiving the general explanation wanted physicians to continue IVS when the probability of a successful outcome was less than 50% (general 30% vs specific 64%). Conclusion. Physician explanations—based on general (unspecified medical condi tion) vs specific (severe pneumonia) explanations—have a marked influence on the duration of IVS patients would permit and the probability of a good outcome required to continue IVS. Key words: advance directives; life support; patient preferences; patient-physician decision making. (Med Decis Making 1997;17:56-60)

Objectives. To assess the diagnostic values of single and combined data from the history, physical examination, and medical record with regard to peripheral arterial occlusive disease (PAOD) in patients with leg complaints; to construct a multivariable model for the clinical diagnosis of PAOD by primary care physicians. Setting. 18 gen eral practice centers in The Netherlands. Design. Cross-sectional comparison of signs, symptoms, and data from the medical record with the independently assessed ankle- brachial systolic pressure index (ABPI; cutoff point < 0.90); analysis: bivariate, multiple logistic regression (MLR). Population. 2,455 individuals with leg complaints, aged 40.7-78.4 years; ABPI < 0.90 present in 9.2% of legs (11.7% of individuals). Outcome measures. Clinical variables: sensitivity, specificity, positive and negative predictive values (PV+, PV-), diagnostic odds ratio (OR); models: likelihood ratio test, area under the receiver operating characteristic curve (AUC). Results. Bivariate analysis: highest sensitivity: age more than 60 years (77.3%); highest specificity: wounds or sores on toes and foot (99.7%); highest PV+: typical intermittent claudication (IC) (45.0%) (abnormal foot pulses 41.3%); highest PV-: strong pulses of both foot arteries (97.7%). MLR: the best-performing model (AUC 0.89) consisted of ten clinical varia bles : gender (OR 1.5), age more than 60 (OR 2.2); IC (OR 3.5); palpation of the skin temperature of the feet (OR 2.5), palpation of both foot pulses [OR 16.4 (abnormal) and 7.0 (doubtful)], auscultation of the femoral artery (OR 3.5); previous diagnosis of IHD (OR 1.7) or diabetes (OR 1.6), history of smoking (OR 2.1), and elevated blood pressure (OR 1.5). The range of predicted probabilities was 0.4-98%. The Hosmer- Lemeshow goodness-of-fit test indicated good overall fit (p = 0.52). Conclusions. Pal pation of both foot pulses is the key procedure for the clinical diagnosis of PAOD. Traditional clinical evaluation enables the general practitioner to exclude the diagnosis of PAOD in many individuals with a high degree of certainty, to establish the diagnosis in a small group of patients, and to define a limited group of patients where supple mentary noninvasive testing is appropriate. The MLR model can be used as a diag nostic checklist and as a reference for the physician's clinical hypothesis. Key words: peripheral arterial disease, intermittent claudication, diagnosis, general practice, mul tiple logistic regression. (Med Decis Making 1997;17:61-70)

This study illustrates the use of consensus theory to assess the diagnostic perform ances of raters and to estimate case diagnoses in the absence of a criterion or "gold" standard. A description is provided of how consensus theory "pools" information pro vided by raters, estimating rater competencies and differentially weighting their re sponses. Although the model assumes that raters respond without bias (i.e., sensitivity = specificity), a Monte Carlo simulation with 1,200 data sets shows that model esti mates appear to be robust even with bias. The model is illustrated on a set of elbow radiographs, and consensus-model estimates are compared with those obtained from follow-up data. Results indicate that with high rater competencies, the model retrieves accurate estimates of competency and case diagnoses even when raters' responses are biased. Key words: clinical competence; interobserver variation; diagnostic evalu ation ; models—mathematical; consensus theory. (Med Decis Making 1997;17:71- 79)

This research investigated the effect of computer-assisted test interpretation (CATI) on physicians' readings of electrocardiograms (ECGs). The authors used an experimental method based on direct observations of 22 cardiologists, each reading 80 ECGs, for a total of 1,760 (of which 1,745 were used in the study). There were 40 sets of clini cally-matched pairs of ECGs, one with CATI and one without. Reading time was ob served and interpretation accuracy was measured by criterion-referenced aggregate scoring. To control for potential biases, the findings were subjected to multivariate analyses using ordinary least-squares regressions. The impact of CATI on cardiolo gists' readings of ECGs is demonstrably beneficial: the main empirical conclusion of this study is that, compared with conventional interpretation, the use of computer- assisted interpretation of ECGs cuts physician time by an average of 28% and signif icantly improves the concordance of the physician's interpretation with the expert benchmark, without increasing the false-positive rate. Moreover, CATI is the most ac curate and saves the most time when the ECGs have many unambiguous diagnoses. Given that computers alone cannot perform the task of cardiovascular diagnosis, and that cardiologists' ECG interpretations are greatly enhanced by ubiquitous CATI tech nology, it appears that the best approach is one that combines person and machine. Key words: electrocardiogram interpretation; computer-assisted test interpretation; physician decision making; diagnostic concordance; technology utilization. (Med Decis Making 1997;17:80-86)

The Markov process is a useful tool for modeling the natural history of disease, which is becoming increasingly important as new diagnostic tests increase the detectability of early-stage disease. The accuracy of a Markov model, however, depends on the accuracy of the estimates for the transition probabilities between different stages of disease. Because these estimates are usually based on "expert opinion" or small cohort studies, they are subject to imprecision and bias. The authors describe an alternative method of estimating transition probabilities from the stage distribution of disease observed at the time of death and age-specific mortality rates from other causes. In addition, they prove that the transition probabilities are unique given certain assumptions about how they change with age. Finally, they illustrate the method using population-based data for prostate cancer. Key words: Markov process; transition prob abilities ; cohort studies; selection bias; precision; population-based mortality rates; stage distribution at autopsy. (Med Decis Making 1996;16:87-93)

Receiver operating characteristic (ROC) analysis, which yields indices of accuracy such as the area under the curve (AUC), is increasingly being used to evaluate the performances of diagnostic tests that produce results on continuous scales. Both par ametric and nonparametric ROC approaches are available to assess the discriminant capacity of such tests, but there are no clear guidelines as to the merits of each, particularly with non-binormal data. Investigators may worry that when data are non- Gaussian, estimates of diagnostic accuracy based on a binormal model may be dis torted. The authors conducted a Monte Carlo simulation study to compare the bias and sampling variability in the estimates of the AUCs derived from parametric and nonparametric procedures. Each approach was assessed in data sets generated from various configurations of pairs of overlapping distributions; these included the binormal model and non-binormal pairs of distributions where one or both pair members were mixtures of Gaussian (MG) distributions with different degrees of departures from bi- normality. The biases in the estimates of the AUCs were found to be very small for both parametric and nonparametric procedures. The two approaches yielded very close estimates of the AUCs and of the corresponding sampling variability even when data were generated from non-binormal models. Thus, for a wide range of distributions, concern about bias or imprecision of the estimates of the AUC should not be a major factor in choosing between the nonparametric and parametric approaches. Key words: ROC analysis; quantitative diagnostic test; comparison, parametric; binormal model; LABROC; nonparametric procedure; area under the curve (AUC). (Med Decis Making 1997;17:94-102)

Test-indication curves (TiCs) are tools for determining whether a test is indicated for a given patient. They apply the threshold approach of Pauker and Kassirer in graphic form. These curves are composed of two parts: the raw curve, which plots posttest probability versus pretest probability (given values for specificity and sensitivity); and the final curve, in which three straight lines are added to the raw curve by the clinician to generate a TIC for a given treatment threshold. In the final curve, the complete range of pretest probability is segregated into three zones, corresponding to the three groups described by Pauker and Kassirer: those patients in whom disease is assumed to be present and who are thus best treated empirically; at the other extreme, those who require neither testing nor treatment; and, finally, those in the middle, for whom the test is indicated, since the decision to treat would be based on the test result. Thus the clinician could consult the TIC and determine with certainty whether the test should be employed for a given patient. It also could be modified with ease for a different patient, with a different set of threshold values. TICs provide a complete, visual inter pretation of a test's diagnostic power, in the context of a given treatment threshold. They foster an intuitive comprehension of Pauker and Kassirer's method, and offer the clinician a facile means to prove that a test is indicated in a given setting. By promoting the use of exactly those tests that are indicated, TICs can help spare the patient the cost, burden, and risk of unnecessary testing, and help spare the physician the cost, burden, and risk of interpreting inconclusive test results. Key words: test-indication curves; treatment thresholds; laboratory tests. (Med Decis Making 1997;17:103-106)

Cost-effectiveness (CE) ratios vary considerably, not only across interventions, but within single interventions. Using a simple decision-tree model of the treat-vs no-treat decision to organize the analysis, four potential errors leading to these within-treatment differences in CE ratios are identified. These errors arise from estimates relating to 1) prior probabilities of disease; 2) treatment efficacies; 3) costs of treatment; and 4) patient preferences. Systematic biases, where present, suggest overuse of medical interventions. For diagnostic tests, two additional potential sources of error are con sidered (using a simple decision tree incorporating both test and treat decisions). These involve 5) sensitivity and specificity of the diagnostic test and 6) inappropriate choice of "cutoff" to determine abnormal patients, in part arising from errors in esti mating prior probability of disease. Key words: cost-effectiveness; biases; errors; re source use; utilities. (Med Decis Making 1997;17:107-117)