Monday, Aug 5: 10:30 AM - 12:20 PM
5050
Contributed Papers
Oregon Convention Center
Room: CC-B111
Main Sponsor
Health Policy Statistics Section
Co Sponsors
Caucus for Women in Statistics
Presentations
The literature for Cluster-Randomized Trials (CRTs) has emphasized that Restricted CRTs are superior to CRTs which utilize simple randomization for a single allocation (AKA randomization scheme). The four types of Restricted CRTs most-often utilized are Matched-Pair Cluster-Randomization (MPCR), Stratified Cluster-Randomization (SCR), Minimization Cluster-Randomization (MCR), and Covariate-Constrained Cluster-Randomization (CCCR). While recent literature seems to land on either M-CR or CC-CR as being superior to both MPCR and SCR, there is a lack of either empirical or theoretical evidence to support this claim. In fact, some of the putative disadvantages of MPCR compared to CCCR have been shown to be theoretically doubtful. It is also a fact that MPCR has both a conceptual and intuitive appeal to our non-statistician colleagues, as well as close analogues to matched cohort and case-control designs with which they are often intimately familiar. As such, we present empirical evidence to better understand the conditions under which MPCR is not inferior to CCCR, both via a simulation study and through several real-world examples.
Keywords
Cluster-Randomized Trials
Population-level interventions often face practical constraints that require a non-randomized and staggered implementation. In this work, we provide a motivating example of a population-based payment intervention program that was implemented in a staggered fashion by a health provider. Leveraging patient demographic information, clinical registry data, and medical claims data, we illustrate how one could evaluate the program's impact on quality-of-care indicators such as cancer screening, diabetes control, and hypertension control. To our advantage, significant progress has been made recently in the development of statistical methods to estimate the effects of non-randomized staggered interventions. We showcase how these state-of-the-art methods can be used to estimate the effects of the program and under which settings they provide unbiased effect estimates. Moreover, we discuss how these methods approach additional challenges encountered in our real-world example, including heterogeneity across clinics, time-varying confounding, treatment switching, and spatial correlation. Finally, this work includes a comparative assessment of the considered methods based on simulation studies.
Keywords
staggered adoption
medical claims
causal inference
Medicare is a federally funded insurance program that enables essential health services for 60 million older and chronically disabled US adults. The fastest-growing care program in Medicare is Medicare Advantage (MA), whose enrollment surpassed 30 million Americans in 2023. Enrollees in MA buy insurance from contracted private insurers who are reimbursed by the federal government. Reimbursement amounts are determined by a regression that predicts per-patient spending as a weighted risk score of a patient's diagnoses and demographic information. This approach aims to disincentivize insurers from avoiding high-cost enrollees. In practice, this prediction function incentivizes insurers to make their enrollees appear sicker, commonly termed "upcoding." Upcoding has been estimated to cost the government $12-25 billion annually with no clinical benefit to patients. However, a challenge in addressing upcoding is that no formal definition of such behavior exists to evaluate current prediction functions. We address this by developing a formal and operational definition of MA upcoding that can serve as an evaluation metric, so such behavior can be more reliably monitored and corrected.
Keywords
metrics
evaluation metrics
healthcare
policy
risk adjustment
Medicare
Many public health interventions are conducted in settings where individuals are connected to one another and the intervention assigned to randomly selected individuals may spill over to other individuals within their network.
Evaluating such interventions in spillover settings involves assessing both the average individual effect and spillover effect.
To estimate these effects, we propose an Egocentric Network-based Randomized Trial (ENRT) design, wherein a set of index participants is recruited from the population and randomly assigned to the treatment group.
Additionally, recognizing that certain individuals are more likely to influence their peers due to their social connectedness and their individual characteristics, intervening on these individuals can lead to more effective treatment strategies.
Multiple Comparison with the Best (MCB) is modified to identify key influencers by examining heterogeneity of the spillover effect.
The proposed methods are applied in a study of network-based peer HIV prevention education study, providing insights into strategies for selecting peer educators in peer education interventions.
Keywords
Casual Inference
Interference
Social Networks
Key Influencers
Multiple Comparisons
Introduction: The 2014 rescheduling of hydrocodone aimed to reduce opioid misuse, but its impact on cancer patient prescriptions is under-researched. This study examines the effects of this policy change on opioid prescriptions in older breast cancer patients.
Methods: We analyzed SEER-Medicare data for female patients aged 65+ diagnosed with invasive breast cancer (2011-2018), focusing on hydrocodone and other opioid prescriptions within one year post-diagnosis. Logistic regression models were used, adjusting for demographics, clinical characteristics, and time trends.
Results: Of 79,899 patients, 45.3% received hydrocodone and 47.7% other opioids. The multivariable logistic regression showed that post-rescheduling, hydrocodone prescriptions decreased significantly (OR=0.78, p<0.001), while non-hydrocodone opioid prescriptions increased (OR=1.23, p<0.001).
Conclusion: The rescheduling led to reduced hydrocodone prescriptions among older breast cancer patients but increased non-hydrocodone opioid prescriptions, indicating a shift in prescribing patterns. This highlights the effects of policy changes on prescription behavior in oncology.
Keywords
policy change
breast cancer
SEER-Medicare
opioid use
oncology
large observational data
Pediatric cancers represent a group of heterogeneous and serious diseases with apparent
variations in disease etiology, treatment response, and mortality risk. Histology-based cancer types are the
primary drivers of survival disparities, with patient characteristics at diagnosis such as race-ethnicity, sex,
age, and geographic location playing secondary roles. Variations in cancer-specific survival based on age
at diagnosis could enhance overall survival by facilitating precise age-specific treatments. We utilized the
frailty model on a SEER dataset of 101,328 pediatric cancer patients diagnosed and monitored between
1975 and 2016 to ascertain age-specific survival rates for all cancer types, while adjusting for
heterogeneity in mortality risk across patient characteristics and diagnosis years. Cancer types including
ALL, AML, Brain, CNS, Liver, Endocrine, Genital system, Lymphoma nodal, Oral cavity, Respiratory
system, Skin, and Soft tissue showed substantial variation in mortality risk. The use of a large dataset and
application of appropriate methods provided estimations with enhanced precision.
Keywords
Pediatric Cancers
Age-Specific
Precision
Heterogeneous
Frailty Model
SEER
It is challenging to perform analysis of social network data including community detection when there are missing values including node covariates, entire nodes, and edges. Node covariates provide an additional resource for network community detection in addition to the structure of the network. We propose an iterative method to simultaneously update missing covariates using imputation and perform covariate assisted community detection for networks modelled using Exponential Random Graph Models (ERGMs).
The proposed model is assessed using simulated network data with known communities and covariate values. In addition to simulated networks, time series of networks are generated based on human movement between Oregon cities that participated in a wastewater surveillance program run by a team at Oregon State University since mid-2020. The COVID wastewater data along with demographic and other COVID metrics are considered node covariates. Some of these covariates are assumed missing at random(MAR).Wastewater-based epidemiology is an effective approach to monitor the presence, prevalence, and trend of diseases, and understanding their spread through human movement networks.
Keywords
Network community detection
Wastewater based epidemiology
Time series networks
Network Missing data imputation
Human movement network
Disease spread in human movement network