Health Econometrics in International Contexts

Einleitung / Introduction

Chronic non-communicable diseases (CNCD) pose a major challenge in the 21st century. Worldwide, over 40 million people pass away due to a CNCD each year, whereas more than 75% of the deaths occur in low- and middle income countries (LMICs). National health systems in LMICs are particularly struggling to provide full coverage of health services. Therefore, some burden of healthcare costs related to CNCDs are shifted to the individuals through out-of-pocket expenditures (OOPE). In LMICs, OOPE accounts for 93% of private spending and over 60% of total health spending. In this paper, we identify individuals’ determinants of incurring OOPE as well as specify those individuals at risk of a relatively higher expenditure given a reported CNCD in Burkina Faso.

Methode / Method

The data originates from a cross-sectional endline household survey, pursued in the framework of a performance-based impact evaluation. It was conducted in person between March and June 2017 in Burkina Faso and entails a wide range of individual, household and healthcare system characteristics. In the survey, individuals were able to report up to three CNCDs, whether they sought treatment of any kind and their level of OOPE. Overall, the data consists of 7,947 households with 52,562 members. We use a two-part model to estimate what explains OOPE for individuals with CNCDs: The model accounts for the zero healthcare expenditures first and then, conditional on a non-zero value, the absolute level of OOPE of an individual is estimated. The model is estimated both with a logit model as well as a generalised linear model with an inverse Gaussian family and a log link specification.

Ergebnisse / Results

We observe 1107 individuals who report a CNCD, 363 thereof seek medical treatment of some kind, and 214 of those incur positive OOPE. We identify low-risk, high-frequency cost drivers such as medication as expenditure drivers. The two-part model finds that being young, living in a relatively smaller household and reporting more than one chronic condition decreases the odds of incurring OOPE. Considering the expenditure magnitude, being female, living close to a hospital and not being Christian or Muslim is negatively associated with expenditure. Being the household head and perceiving a chronic condition as highly severe is positively associated.

Zusammenfassung / Conclusion

OOPE remain a significant issue in Burkina Faso, hence calling for adequate social protection systems given the increasing prevalence of chronic non-communicable diseases. By identifying the current composition of OOPEs and their determinants for individuals, we point out subgroups which may particularly benefit from an effective social health protection.

AutorInnen

Jan Köhler, ZEW – Leibniz-Zentrum für Europäische Wirtschaftsforschung

Julia Lohmann, London School of Hygiene and Tropical Medicine & Heidelberg Institute of Global Health

Paul Jacob Robyn, World Bank

Saidou Hamadou, World Bank

Serge M. A. Somda, Centre MURAZ; Université Nazi Boni; West African Health Organization

Jean-Louis Koulidiati, Université Nazi Boni

Stephan Brenner, Heidelberg Institute of Global Health

Volker Winkler, Heidelberg Institute of Global Health

Manuela De Allegri, Heidelberg Institute of Global Health

Einleitung / Introduction

A hospital’s capacity, i.e. infrastructure and personnel, determines how many patients can be treated at any given time. Evidence from the literature suggests that hip and knee replacement patients benefit from a rapid recovery path (mobilization within 6h post-surgery) compared with a conventional recovery path (mobilization after 6h post-surgery). However, the rapid recovery path is more personnel intensive; thus patients qualifying for rapid recovery might not be put on this path due to capacity constraints. Therefore, assuming that capacity is fixed, a hospital aiming to maximize health outcomes needs to identify the patients that benefit the most from the rapid recovery path. We aim at determining the optimal assignment rule for the rapid recovery path under capacity constraints, and at identifying the welfare effects of its implementation. To achieve this, we developed a model to identify heterogeneous treatment effects of the two post-treatment care paths. Building on these results, we developed assignment rules with and without capacity constraints to estimate the potential welfare gains on the patient-reported outcome measures (PROMs).

Methode / Method

Patient-level observational data from nine German hospitals from the years 2020 and 2021 from the Innovation Fund study “PROMoting Quality” were used. Data contains personal information, previous treatments, comorbidities, the post-treatment path, and PROMs, including the EQ-5D-5L. 3’697 total hip replacement and 3’110 total knee replacement patients are used for the analysis. We used a causal forest to estimate the double-robust propensity scores of the individual treatment effects, controlling for patient characteristics, and, subsequently, a policy tree was built to develop the optimal treatment assignment rules.

Ergebnisse / Results

47% (44%) of hip (knee) replacement patients follow the rapid recovery path. The expected results are that rapid recovery increases the EQ-5D-5L improvement on average for all patients. We expect to see heterogeneous effects that will allow the model to optimally assign different patient profiles to the two alternative recovery paths, to help physicians in the patient pathway assignment choice. We expect to improve welfare by deciding which patients should follow the rapid recovery path based on the optimal treatment assignment.

Zusammenfassung / Conclusion

We present an approach to determine the optimal assignment of the rapid recovery path to hip and knee replacement patients under capacity constraints, and we show the welfare effects, in terms of the EQ-5D-5L, of different assignment rules. If our expected results are correct, it would show that welfare can be increased without the need to change the hospital capacity, but simply through a better assignment of which patients get to follow the rapid recovery path.

AutorInnen

Johannes Cordier, University of St. Gallen

Irene Salvi, University of St. Gallen

Viktoria Steinbeck, TU Berlin

Justus Vogel, University of St. Gallen

Alexander Geissler, University of St. Gallen

Einleitung / Introduction

A substantial share of Indian patients at risk of developing cardiovascular disease fail to achieve control of blood pressure, blood sugar, and blood cholesterol. As health systems in India are already under pressure from providing care to the growing number of aging patients, clinicians and health administrators need approaches to target additional attention and efficiently allocate limited resources to those patients that would be benefit the most.

Methode / Method

We use machine learning models built on the unique clinical and socioeconomic data available from two studies with diabetic patients from urban areas to predict which patients are least likely to achieve clinically meaningful improvements in cardiovascular health (measured through HbA1c, systolic blood pressure and LDL). We compare boosted linear models, boosted trees and support vector machines and sequentially increase the number of predictors from basic clinical indicators (baseline CVD biomarkers) to socioeconomic factors (e.g., education and household income). For each outcome, we assess the model with the largest median area under the curve (AUC) across ten cross-validation folds (“best” model) on a held-out validation set of 20% of the sample.

Ergebnisse / Results

All models predict equally well whether patients fail to achieve blood glucose control (HbA1c>=8) after one year, with an AUC between 0.69 and 0.71. The models vary slightly more for failing to achieve blood pressure control (systolic blood pressure>=140, AUC between 0.67 and 0.72), and much more for lipids (LDL>=130, AUC between 0.53 and 0.76), due to the poor performance of the support vector machines in both cases. For all outcomes and algorithms, the small models using only baseline clinical data perform similar to the large models including socioeconomic information. The best model for blood glucose correctly detects 62% of all patients in the validation set who fail to achieve control (sensitivity), and correctly classifies 73% of the patients achieving control (specificity). The best model for blood pressure [lipids] reaches a sensitivity of 62% [50%] and a specificity of 66% [59%] for the validation set.

Zusammenfassung / Conclusion

We show that models based on few clinical parameters can predict well the failure of achieving CVD control for diabetic patients in urban India. Such models could be easily integrated into routine care and would enable efficient delivery of health resources for improving cardiovascular health in urban India.

AutorInnen

Anna Reuter, Heidelberg Institute of Global Health

Einleitung / Introduction

Estimating causal effects based on administrative health care data comes with challenges due to the lack of randomization, unmeasured covariates and a wide range of variables with complex and unknown dependencies. To target these problems, many methodological approaches have been developed, e.g., propensity scores, difference-in-differences models, and doubly-robust semi-parametric approaches. These methodological approaches address some of the challenges mentioned above and have different strengths and weaknesses. Researchers working with administrative health data lack guidelines for choosing the best approach for their analysis. By comparing and evaluating different approaches, simulation studies can provide important guidance on this question. However, many of the simulation studies published to date do not take sufficient account of the complexity and character of real-world administrative health data. Therefore, it remains unclear which approach to estimating causal effects is best suited for this particular type of data. The aim of the study is to conduct a simulation study based on real-world administrative health data to assess and compare approaches estimating causal effects.

Methode / Method

We compare methods frequently used by researchers as well as methods recommended by previous simulation studies: AIPW (augmented inverse probability weighting), TMLE (targeted maximum likelihood estimation), Propensity Score Matching with differences-in-differences analysis (DiD), IPTW (inverse probability of treatment weighting) with DiD and entropy balancing with DiD. We further estimate the nuisance parameters like the propensity score with regression models and an ensemble learner approach. We are interested in treatment effects on health care costs and utilization. To generate data allowing for the complex data structures given in administrative health care data, we resample the covariates from real-world administrative health care data of three German sickness funds without modification and simulate a true treatment effect of choice based on the covariates (plasmode data set). To replicate the problem of unobserved heterogeneity, a subset of covariates is omitted from analysis after estimation of the outcome variables.

Ergebnisse / Results

Preliminary results of 72 simulation scenarios drawing data from 49.348 subjects suggest differences in performance between the methodological approaches used to estimate causal effects.

Zusammenfassung / Conclusion

We run a simulation based on real-world administrative health data to compare approaches estimating causal treatment effects. Results will inform researchers and policy makers on which approach to take when estimating causal effects (e.g., when evaluating health care interventions) using administrative data.

AutorInnen

Vanessa Ress, Hamburg Center for Health Economics (HCHE)/Universtität Hamburg

Christina C. Bartenschlager, Lehrstuhl für Health Care Operations / Health Information Management, Universität Augsburg

Eva Wild, Hamburg Center for Health Economics (HCHE)/Universtität Hamburg