Health Services and Outcomes Research Methodology

Asthma and obesity are both prevalent conditions that appear related, but the etiology for this association remains unclear. This study examines whether asthma is associated with obesity and physical activity limits 10 years later among a subsample from the National Longitudinal Survey of Youth 1979 who were age 40 at baseline. We addressed selection bias using inverse-propensity score weighting (N = 5077), and confirmed the results with full matching (N = 5041), and with both methods we estimated new sampling weights so that the sample would remain nationally representative. Both matched sampling methods balanced adults with asthma versus those without asthma on all 7 covariates: baseline obesity, sex, race/ethnicity, family income, poverty status, general health status and physical activity limits. Before matching, baseline asthma was significantly associated with developing obesity 10 years later in an unadjusted model [OR = 1.44 (1.10–1.90)], but not in the multivariable model [OR = 1.15 (0.80–1.67)]. Baseline asthma was not associated with obesity 10 years later after inverse propensity weighting [OR (95% CI = 1.03 (0.69–1.53)] and full matching [1.16 (0.75–1.80)]. Results remained similar after excluding subjects with baseline obesity. In a cumulative logistic model using complex survey and full matching weights, those with baseline asthma had 83% greater odds of reporting physical activity limits compared to those without asthma, OR = 1.83 (1.21–2.76). Baseline asthma was not associated with obesity among either a nationally representative sample of middle-aged adults or a non-obese subset. However, asthma was associated with physical activity limits in the full matched sample. Asthma disease management programs should communicate that asthma does not imply obesity and also encourage exercise within the physical limitations of their populations. Selection bias on factors such as low socioeconomic status may explain previous asthma-obesity associations.

Objective: Illustrate Ward's clustering algorithm, an approach that creates small geographic areas based on similarity in the pattern of hospital use. Compare areas resulting from Ward's algorithm using all hospital discharges to 1) areas resulting from the plurality rule, the more widely used approach, and 2) areas resulting from Ward's algorithm using subsets of discharges. Findings: Compared to areas from Ward's algorithm, the plurality rule resulted in many more single zip code areas and somewhat more large zip code areas. R2, a measure of within-area variation in the pattern of hospital use, did not differ much by clustering method. When areas from Ward's algorithm were paired with areas from the plurality rule, about 81% of the discharges from Ward's algorithm areas were in the paired plurality rule areas. Though in many cases there were large errors when zip code discharge rates were estimated by area rates, the distribution of errors was similar whether Ward's algorithm or the plurality rule was used to create the areas. When areas from different clustering methods were used in a study to identify those areas with large differences in utilization patterns by payor type, between 70% and 83% of zip codes and discharges were flagged in common. Conclusions: Though for several summary measures there was little difference by clustering method, the actual areas created were different. There is no gold standard to determine which set of areas is best. However, Ward's clustering algorithm has conceptual appeal compared to the more widely used plurality rule.

Survival extrapolation plays a key role within cost effectiveness analysis and is often subject to substantial uncertainty. Use of external data to improve extrapolations has been identified as a key research priority. We present findings from a pilot study using data from the COU-AA-301 trial of abiraterone acetate for metastatic castration-resistant prostate cancer, to explore how external trial data may be incorporated into survival extrapolations. External trial data were identified via a targeted search of technology assessment reports. Four methods using external data were compared to simple parametric models (SPMs): informal reference to external data to select appropriate SPMs, piecewise models with, and without, hazard ratio adjustment, and Bayesian models fitted with a prior on the shape parameter(s). Survival and hazard plots were compared, and summary metrics (point estimate accuracy and restricted mean survival time) were calculated. Without consideration of external data, several SPMs may have been selected as the ‘best-fitting’ model. The range of survival probability estimates was generally reduced when external data were included in model estimation, and external hazard plots aided model selection. Different methods yielded varied results, even with the same data source, highlighting potential issues when integrating external trial data within model estimation. By using external trial data, the most (in)appropriate models may be more easily identified. However, benefits of using external data are contingent upon their applicability to the research question, and the choice of method can have a large impact on extrapolations.

In health intervention research, epidemiologists and economists are more and more interested in estimating causal effects based on observational data. However, collaboration and interaction between both disciplines are regularly clouded by differences in the terminology used. Amongst others, this is reflected in differences in labeling, handling, and interpreting the sources of bias in parameter estimates. For example, both epidemiologists and economists use the term selection bias. However, what economists label as selection bias is labeled as confounding by epidemiologists. This paper aims to shed light on this and other subtle differences between both fields and illustrate them with hypothetical examples. We expect that clarification of these differences will improve the multidisciplinary collaboration between epidemiologists and economists. Furthermore, we hope to empower researchers to select the most suitable analytical technique from either field for the research problem at hand.

Machine learning (ML) has the potential to revolutionize healthcare, allowing healthcare providers to improve patient-care planning, resource planning and utilization. Furthermore, identifying key-risk-factors and interaction-effects can help service-providers and decision-makers to institute better policies and procedures. This study used COVID-19 electronic health record (EHR) data to predict five crucial outcomes: positive-test, ventilation, death, hospitalization days, and ICU days. Our models achieved high accuracy and precision, with AUC values of 91.6%, 99.1%, and 97.5% for the first three outcomes, and MAE of 0.752 and 0.257 days for the last two outcomes. We also identified interaction effects, such as high bicarbonate in arterial blood being associated with longer hospitalization in middle-aged patients. Our models are embedded in a prototype of an online decision support tool that can be used by healthcare providers to make more informed decisions.

Studies show increasing use of urgent care centers (UCCs) and there is interest in evaluating their potential for cost savings. Previous research provides limited information on generalizable methods of identifying urgent care centers and does not validate these methods. The objective of this study is to describe and validate two claims-based UCC definitions. We used FAIR Health insurance claims from 444,263 organization National Provider Identifiers (NPIs) with at least 10 claims, January 2016–March 2019 and merged this data with National Plan and Provider Enumeration System data. The first definition required (1) a UCC place of service code (POS), (2) ≥ 10% Current Procedure Terminology (CPT) codes specific to UCCs, or (3) a UCC taxonomy code in the primary field. The second definition relaxed these criteria. A random sample of 5% of NPIs identified as UCCs were validated through internet searches. Prevalence and positive predictive value (PPV) were calculated for both definitions. The first definition identified 6669 (1.5%) of NPIs as UCCs resulting in a PPV of 92%. The second definition identified 8261 (1.9%) of NPIs as UCCs and had a PPV of 87%. Out of NPIs identified under the first definition, 96% were identified using POS codes, 50% were identified using taxonomy codes, and 46% using CPT codes, with 62% of NPIs meeting multiple criteria. Findings suggest that these methods may be used by researchers to identify UCCs in studies of cost or utilization in different healthcare settings.

This article compared standard regression (logistic), propensity score weighting, propensity score matching, and difference-in-difference (DID) methods in determining the impact of second-generation antidepressant (AD) use on mania-related visits among adult patients with bipolar disorder. Using a large managed care claims database, a logistic regression was developed as a standard approach to predict the likelihood of having mania-related visits after receiving various types of treatments (AD monotherapy, mood stabilizer (MS) monotherapy, and AD-MS combination therapy) controlling for individual baseline characteristics. The propensity score method predicted the propensity to be with one treatment type versus another in the first-stage. Both weighting and greedy matching approaches were applied in the second-stage outcome model. For the DID method, a logistic regression was applied to predict the differential likelihood of having mania-related visits in post-baseline versus baseline periods on different treatments. Both full sample and propensity score-matched sample were applied for the DID method. Except DID with full sample, the results from all other methods suggested no higher likelihood of mania-related visits for second-generation AD-related therapies compared to MS monotherapy. We concluded that standard regression, propensity scoring, and DID methods may produce inconsistent outcomes in a logistic regression framework, when patient baseline characteristics are different between comparison groups and/or not all potential confounders can be correctly measured and fully controlled. Researchers need to be cautious of the basic assumptions and sensitivities of various methods before making a final conclusion. The DID method may be considered in outcome studies when pre-and-post data are available.

At the 2018 International Conference on Health Policy Statistics (ICHPS) held in Charleston, South Carolina, Anirban Basu was awarded the Mid-Career Excellence Award from the American Statistical Association Section on Health Policy Statistics (HPSS). Anirban was exceptionally and uniquely qualified for this award. Highlights include his providing outstanding service to the HPSS, advancing statistical methodology, advancing methodology in other domains of health policy, and performing extensive and highly impactful applied work in medicine and health care. In this interview, we trace Anirban’s upbringing, schooling, early career, and mid-career phases to gain insights into his success. We also sought his opinions on salient topics or issues.

Sally C. Morton is internationally recognized in the use of statistics in health policy, and has had a career with incredible impact as evidenced by her many leadership roles. She was awarded the ASA Health Policy Statistics Section’s (HPSS) Long-Term Excellence Award in January 2018 at the 12th International Conference on Health Policy Statistics. Morton is currently the Dean of the College of Science at Virginia Tech. This article is conversation with Morton about her career.