* Candidate for the Lee B. Lusted Student Prize Competition
Purpose: Cost-effectiveness analysis requires comparison of outcomes in treated and untreated populations. Data from randomized clinical trials (RCT) do not provide progression rates representative of the general population, while treatment effects in observational data may be biased due to non-randomization. We developed a novel approach for estimating untreated progression rates by using data from a population-based longitudinal survey, correcting for the effects of patients’ treatments as reported by pivotal trials.
Method: We used data from the 2000-2005 Sonya Slifka nationally representative MS cohort. Disease progression was characterized by disability-based disease states and relapses. We modeled probabilities of disease state transitions using a first-order annual Markov model that adjusted for demographics, disease duration, recent relapse rates, prior states, and the specific disease-modifying therapy (DMT). To estimate transitional probabilities, we developed an iterative multinomial logistic regression algorithm, constraining the effects of DMT to match those reported by RCTs as follows. We selected initial annual treatment factors and estimated first progression probabilities for controls. For those probabilities, using a numerical algorithm, we found new treatment factors that resulted in the same risk ratios of progression as reported by the trials. The new factors were used in the regression model to adjust for DMT effects and to re-estimate the probabilities for controls. We continued this process iteratively, until the identified factors for the final control probabilities matched published DMT effects from RCTs.
Result: After correcting for the DMT treatment effects and other observable risk factors, the probability of disability progression was greater for estimates based on all MS patients compared to the estimates based on untreated individuals only. The 95% confidence intervals using the entire cohort (including treated and untreated individuals) were narrower than the intervals based on the subsample of untreated patients.
Conclusion: Our results indicate that the untreated patients in our study had lower estimates of disease progression than the treated patients would have had if they remained untreated. This suggests that patients who forgo treatment are likely to have milder, slower progressing forms of MS. Correcting for treatment effects in a more inclusive group of patients likely provides a more realistic estimate of disease progression than simply characterizing progression in an untreated cohort. The use of a population-based cohort also improves the precision of disease progression estimates.
Purpose: Randomised controlled trials traditionally investigate treatment effects but can also be used to estimate selection effects (the self-selection of one treatment over another) and preference effects (the effect of receiving the preferred treatment). This study illustrates a method (Rucker 1989 Statist. Med.) to estimate treatment, preference and selection effects to investigate whether informed choice supported by a decision aid is beneficial compared to policy directed management (limited patient choice).
Method: The method is illustrated using data from the IMAP trial, which was designed to investigate the psychosocial outcomes over 1 year of an informed choice between HPV triage or usual care by repeat Pap smear compared to policy directed management of each option. We used a 3-arm trial design with patients randomised to either one of two treatments (limited choice) or to an informed choice arm. The method is unique in that it allows the effects of treatment, preference (i.e. choice) and selection (selection bias) to be estimated separately. Information from the choice arm is used to estimate effects within the randomised arms for those who did and did not receive their preferred treatment.
Results: With traditional analysis those in the HPV arm were more satisfied than those in the Pap arm, with little difference between informed choice and HPV. There was little difference in quality of life (SF36) scores between the three arms. The Rucker analysis showed weak evidence for an effect of preference on the SF36 scores: mental health score (6.0, 95% CI-0.6 to 12.9, P= 0.07) with choice associated with improved quality of life. There was evidence of a selection effect for the satisfaction of women with their health care in general and with the care of their abnormal Pap, with women who selected or would have selected HPV being less satisfied than those who selected or would have selected Pap triage (-2.1 95% CI-4.0 to -0.3, P=0.02 and -1.2, 95%CI-2.5 to -0.2, P=0.03).
Conclusions: The Rucker method should be used to estimate the effect of informed choice compared to policy or clinician directed management (ie. limited patient choice) as it brings important additional information to the interpretation of trial data.
Purpose: The optimal community-level approach to control pandemic influenza is unknown.
Method: We estimated the health outcomes and costs of combinations of 4 social distancing strategies (adult social distancing, child social distancing, school closure and household quarantine) and 2 antiviral medication strategies (treatment alone or treatment and prophylaxis) to mitigate an influenza pandemic for a demographically “typical” U.S. community. We used a social network, agent-based model to estimate strategy effectiveness. We used data from the literature to estimate clinical outcomes and health care utilization. Outcomes included cases averted, total quality-adjusted life years, total costs, cost per case averted and the incremental cost-effectiveness ratios (cost per quality-adjusted life-year saved) of alternative strategies. We tested sensitivity of the results to virus infectivity (Ro, the reproductive number), case fatality rate, population compliance, and costs.
Result: The most effective strategies are multilayered interventions. At 1% influenza mortality, moderate infectivity (Ro 2.1 or greater), and 90% population compliance, the preferred strategy is adult and child social distancing, school closure, and antiviral treatment and prophylaxis. This strategy reduces cases in the population from 35% to 4%, averts 3,100 cases per 10,000 population, costs $2,700 per case averted and $22,000 per quality-adjusted life-year gained compared to the same strategy without school closure. If antivirals are either ineffective or unavailable, then a strategy of adult and child social distancing and school closure reduces cases to 8% of the population, averts 2,700 cases per 10,000 population, costs $4,200 per case averted and $17,300 per quality-adjusted life-year gained compared to the same strategy without school closure. The preferred strategies are robust to varied assumptions. The addition of school closure to adult and child social distancing and antiviral treatment and prophylaxis, if available, is not cost-effective for viral strains with low infectivity (Ro 1.6 and below) and low case fatality rates (1% and below). High population compliance lowers costs to society substantially when the pandemic strain is severe (Ro 2.1 or greater).
Conclusion: Multilayered mitigation strategies that include adult and child social distancing, use of antivirals and school closure are effective and, for a severe pandemic, cost effective. Choice of mitigation strategy should be driven by the severity of the pandemic, as defined by the case fatality rate and infectivity.
Purpose: Properly designed experiments have five advantages over observational studies: (1) Estimates are un-confounded (e.g. between race/ethnicity and socioeconomic status); (2) Cause and effect can be assessed directly; (3) They are cost efficient due to smaller sample size requirements; (4) They are free of selection biases (known and unknown); and (5) Observed covariates can be added to the analysis. Variations in medical decision making, such as discrimination by patient gender or race/ethnicity, are often examined using observational studies of massive data sets (e.g. claims data). However, results are confounded and compromised by selection biases and can only estimate possible associations. An experiment in which patient characteristics (but not medical details) are systematically varied allows unbiased cause and effect conclusions to be drawn with smaller samples.
Method: We used data from five factorial experiments concerning the diagnosis of coronary heart disease (CHD), depression, and diabetes. Between 256-384 randomly selected primary care physicians viewed video vignettes of patients. The patients varied by gender, age, race/ethnicity and socioeconomic status (depicted by dress and current or former occupation) with 16-24 different vignettes per experiment. Physicians were stratified by country, gender, and level of experience (measured by year of graduation from medical school). The patient characteristics were experimentally varied.
Result: We found that physicians were inconsistent in their attention to base rates of disease for various patients. Consistent with base rates, their certainty for a CHD diagnosis was lower for younger women (p <.01), and the probability of a diabetes diagnosis was lower in Whites compared to Blacks (p<.05). However, they were inconsistent with base rates in that their certainty of a depression diagnosis did not vary by gender and their probability of a diabetes diagnosis did not vary by age or socioeconomic status. When examining physician characteristics, we found (consistently across studies) that women would ask more questions. We also found that women and less experienced physicians would offer more lifestyle advice.
Conclusion: By experimentally manipulating patient characteristics we found significant non-medical (patient characteristics) influences on clinical decision making, with a relatively small number of physicians. We also found consistent physician effects. Our experimental design ensures these effects are un-confounded, free of selection biases and directly attributable to the manipulated causes.
Purpose: Previous research (e.g. Klein, Windschitl, Lipkus, Fagerlin) has shown that providing people with comparative risk information (i.e., whether their risk is above or below average) can change risk perceptions and subsequent health behaviors. As part of a larger study on contextual risk information, we explored whether comparative risk status would also affect people’s memory of the statistics themselves.
Method: 884 adults ages 30-60 completed an Internet-administered survey vignette about a hypothetical medication to prevent colon cancer. All participants received information about their own 20-year risk (4%) and the average person’s risk, but were randomized to have the average risk be higher (8%) or lower (2%) than their own. In addition, half of participants were explicitly told that they were at “high risk” or “low risk” of developing colon cancer while the remainder just received the numerical risk information. The primary outcome measure for this analysis was recall of (a) one’s own risk and (b) the risk reduction of the hypothetical medication, each coded as correct or incorrect.
Results: Participants’ whose personal risk was said to be higher than the provided average risk statistic were significantly more likely to accurately recall both their own risk (67% vs. 49%, p<0.001) and the absolute risk reduction provided by the medication (65% vs. 45%, p<0.001) than participants whose own risk was below average. Among participants who were at higher than average risk, being explicitly told that they were at “high risk” decreased recall of the personal risk statistic (62% vs. 71%, p=0.03). Such “high/low risk” statements were not significantly associated with recall of the risk reduction.
Conclusions: Recall of a personal risk statistic was significantly better if it represented a comparatively high risk versus a comparatively low risk, suggesting that learning that one is above average may lead to more detailed cognitive processing. This effect is weaker when people were explicitly told that their risk was “high.” Our findings suggest that providing comparative risk information does not simply change people’s emotional and cognitive evaluations of personal risk information, but also influences their ability to remember such statistics in the first place.