Several Biomedical Data Science faculty presented research at the 2016 Joint Statistical Meetings (JSM) in Chicago, IL, July 30th-August 4th. JSM is the largest gathering of statisticians held in North America.
Title: Hierarchical Regression and Variance-Function Modeling to Estimate the Inter-Rater Intraclass Correlation Coefficient in Assessments of Shared Decision Making
Author(s): James O'Malley and Paul J. Barr and Glyn Elwyn
Abstract: We describe a novel methodological approach to assessing the inter-rater reliability of assessments of the amount of shared decision-making in a patient-physician clinical encounter. Two-raters each assess recordings of patient-physician encounters across three clinical sites using the Option5 shared-decision-making tool. The desired output is the interrater intra-class correlation coefficient (ICC) of the OPTION5 scores, accounting for heterogeneity between studies and heteroscedasticity of ratings with respect to the true amount of shared decision-making. We describe a three-level hierarchical model with random effects for patient-physician encounter and clinical site, covariates for rater and other predictors, and a variance-mean function and a Bayesian estimation method. We show that the ICC varies widely depending on whether the encounters being distinguished are restricted to the same site and to the amount of shared decision making in the encounter. Because current applied practice in shared decision often ignores these subtleties, we argue for the introduction of standards regarding the definition of ICC and other measures of inter-rater reliability in this field.
Title: Bayesian Covariance Analysis of Geographical Variation in Medicare Service Use
Author(s): Alan M. Zaslavsky and James O'Malley and Bruce E. Landon
Abstract: Geographical variation in patterns of service use has been well studied in Traditional Medicare (TM), using data from claims. However, less is known about the extent of such variation in Medicare Advantage (MA) contracts, which do not submit claims to CMS. In particular, it is not known to what extent the MA contracts follow local patterns of utilization in TM, except at a very aggregated "per member per month" level. We investigated this question using a relatively new data source for MA, the HEDIS utilization measures, and similarly defined measures constructed from TM claims. We used a Bayesian approach to covariance matrix estimation in multilevel data (O'Malley and Zaslavsky, JASA 2008) to yield posterior probabilities for descriptive statements about correlations of measures, summarized through factor analysis and regressions. In general MA utilization patterns are well aligned with those in TM and show similar levels of variation across Healthcare Referral Regions (HRRs).
Title: A Semiparametric Joint Model for Terminal Trend of Quality of Life and Survival in Palliative Care Research
Author(s): Zhigang Li and H. Rob Frost and Lihui Zhao and Lei Liu and Kathleen Lyons and Huaihou Chen and Bernard Cole and David Currow and Marie Bakitas and Tor Tosteson
Abstract: Palliative medicine is an interdisciplinary specialty focusing on improving quality of life (QOL) for patients with serious illness and their families. Palliative care programs are available or under development at over 80% of large US hospitals (300+ beds). Palliative care clinical trials present unique analytic challenges relative to evaluating the palliative care treatment efficacy which is to improve patients' diminishing QOL as disease progresses towards end of life (EOL). A unique feature of palliative care clinical trials is that patients will experience decreasing QOL during the trial despite potentially beneficial treatment. Often longitudinal QOL and survival data are highly correlated which, in the face of censoring, makes it challenging to properly analyze and interpret longitudinal QOL trajectory. To address these issues, we propose a novel semiparametric statistical approach to jointly model longitudinal QOL and survival data. There are two sub-models in our approach: a semiparametric mixed effects model for longitudinal QOL and a Cox model for survival. We assess the model through simulation and application on a recently completed palliative care clinical trial.
Title: P, Sorry, D-Value
Author(s): Eugene Demidenko
Abstract: Statistical hypothesis testing and the associated P-value are at the heart of empirical evidence sciences and yet several prominent statisticians recently warn the widespread abuse of the P-value and its contribution to false positive findings. The current discussion follows up a paper of the same author "The P-value you can't buy" and suggests an alternative, called the D-value, which has a more clear interpretation. Unlike the P-value, the new measure does not decrease to zero when the sample size, n, goes to infinity. The D-value is computed by the same formula as the P-value but uses n=1, and therefore may be viewed as the n-of-1 P-value. The D-value is at the crossroads of major statistical concepts such as the area under ROC curve, Mann-Whitney U test, and effect size. It has a clear interpretation as the probability that a randomly chosen patient from the treatment group gets worse than a randomly chosen patient from the placebo group. Thus, unlike P-value with the emphasis on the average, D-value reflects the individual comparison. The D-value is in unison with the voices from medical doctors: "We treat not a group of patients but an individual."
Title: Constrained Community Detection and the External Field
Author(s): Weston Viles and James O'Malley
Abstract: The literature on community detection is rife with diverse methods. Included among them, modularity optimization has found widespread application and success and is a special case of the present article. From a statistical physics perspective, modularity optimization and its variants are equivalent to ground state determination of the Potts model. We extend this notion to include the external field in the Potts energy function to impose constraints on the derived communities. Furthermore, we retain the probabilistic aspect of the Potts model with external field to define a MCMC sampling regime for the community labels. We apply this novel method to a network of hospitals, some of which are specialized. Our constraint is that each community must contain at least one specialized hospital. For one to make meaningful comparisons, a level of homogeneity is necessary. The specialized hospitals are equipped to implant intra-cardiac devices (ICDs). The hospitals are partitioned into heath referral regions (HRRs) and we require that each HRR contain at least one ICD capable hospital.