Spatial Variation in Hospitalizations for Cardiometabolic Ambulatory Care Sensitive Conditions Across Canada CRDCN Conference November 14, 2017 Martin Cooke Alana Maltby Sarah Singh Piotr Wilk
Today s Agenda Cardiometabolic disease Preventable hospitalizations Geographic variations Program of research Current study
Cardiometabolic disease Leading cause of death worldwide 5.8% of the adult population diagnosed with cardiovascular diseases (heart disease and/or stroke) 8.1% of the adult population diagnosed with diabetes Largely preventable
Economic costs of cardiometabolic disease Loss of economic productivity, time away from work, home care, physical care, pharmaceuticals, out-ofhospital therapies, etc. Second highest hospital expenditure nationwide; over 10% of all hospital expenditures The costs for hospitalization are expected to increase in the coming decades with the ageing demographic of the Canadian population
Preventable hospitalizations Not all hospitalizations can be prevented but the rates can be reduced Ambulatory Care Sensitive Conditions (ACSC): conditions for which hospitalization can be prevented in the presence of timely and effective ambulatory care A principal area of healthcare research in Canada and worldwide Close to 50% of all ACSC are attributed to cardiometabolic diseases
Spatial variation in ACSC hospitalizations Substantial geographic variation in hospitalization rates Intra-provincial variation: Lowest rates in British Columbia and Ontario Highest rates in Saskatchewan Inter-provincial variation: 5-fold regional differences across CMAs Similar differences have been reported internationally: 12-fold regional differences in Switzerland
Determinants of spatial variation Preventable hospitalizations are not resultant of a single cause, but rather are a culmination of contributors System-level factors: quality and accessibility of services across jurisdictions (primary care, disease management, preventative care) Individual-level factors: socioeconomic characteristics, health-related behaviours, severity and chronicity of illness Little is known about the causal factors for the existing geographic differences
Data limitations Lack of data linking individual characteristics with their hospitalization records Common usage of administrative data: Information on hospitalization events with limited data on patients Reliance on aggregate characteristics instead of individual level data Difficult to assess the magnitude of spatial variability while controlling for the effects of individual level determinants Difficult to assess interactions between system- and individual-level factors
Self-selection bias Without individual level data, it is impossible to conclude if the observed differences are due to: Quality and unequal access to healthcare services Compositional effects Assumption of exogeneity: the effects of the area-level factors on hospitalizations are independent of characteristics of individuals residing in a given geographic area This assumption may be unrealistic due to self-selection bias
Overall goals Generate new knowledge in the area of inter- and intraprovincial variation in cacsc hospitalizations Identify geographic regions and sub-populations with significantly high and low cacsc hospitalization rates Contribute to: Development of targeted prevention strategies and public policy Reduction of inequalities in Canadian healthcare system Development of inter-connected methodological approaches
Current study To assess the nature of spatial variations in rates of cacsc hospitalizations at policy relevant levels of geography: Province Health authority: policy relevant health regions defined by the provincial ministries of health Small geographic areas (forward sortation area) To assess the relative role various system- and individual-level factors
Linked data Statistics Canada has advanced Canada s health and social statistics systems by initiating a process of linking various sources of administrative data to census and health survey data 2006 Census of Population + 2006/07-2008/09 Discharge Abstracts Database (DAD) DAD: main diagnoses, length of stay and treatment information for inpatient and outpatient hospital visits
Additional data sources CCHS (e.g., area-based measures of quality of healthcare services) Census (e.g., area-based measures of socio-economic status and demographic composition) CIHI (e.g., physician rates) DMTI Spatial Inc. (e.g., density of and distance to health care services) GIS (e.g., shape files) Local sources
Sample Adults aged 18-75 years old who have participated in the 2006 Census Deaths that occur after the age of 75 are not considered premature or potentially preventable Only linked observations
Outcome variable Cardiometabolic ACSC Based on International Classification of Disease 10 ACSC indicators validated for use in Canada by CIHI: angina, congestive heart failure, hypertension, diabetes Exclusion criteria: certain hospital procedures
Analysis - Descriptive Crude and age-standardized hospitalization rates: National/provincial/Health authority/fsa Across time Spatial variation (autocorrelation): Global variation: Moran s I Local variation: LISA
Multilevel analysis Multilevel techniques can be used for spatially dependent data as they protect against: Ecological fallacy: making conclusions at the individual level from relationships studied at an aggregate level Atomistic (individualistic) fallacy: individual data used to model the effects of geographical context Inclusion of explanatory individual level variables that might account for the spatial differences Any significant unexplained variance at the higher level can be an evidence of a contextual effect
Problems with multilevel analysis Conceptualizes geography as vertical group dependence without accommodating for horizontal group dependence : interactions or spillovers among spatial entities due to their geographic proximity Potential concerns for modeling health data: Outcomes in one area could be affected by what happens in nearby areas Failing to account for spatial dependence may lead to unreliable estimates for variance parameters and regression coefficients
Spatial proximity analysis Has its own separate history and applications Considers if data are spatially autocorrelated and assesses spatial heterogeneity Can be performed within the frequentist approach (SAR models) and Bayesian approach (CAR models) Both approaches incorporate spatial autocorrelation coefficient as well as spatial weight matrix
Spatial weight matrix Weight matrix or spatial connectivity matrix : specifies the interaction structure among spatial units Geographic distance: based on centroid distances between each pair of spatial units Boundaries: based on shared boundaries between spatial units Combined distance and boundary weights
Combined models Multilevel techniques for spatially dependent data Combine features of traditional geographic techniques for spatially dependent data with statistical techniques for analysis of hierarchical data Multiple membership multiple classification Application of empirical Bayes procedure Application of spatial Gaussian processes
Modeling steps Assessment of contextual (spatial) pattern effects: Global clustering (Moran s I statistic) Local clustering (hot spot analysis; LISA) Fitting null model (no spatial or hierarchical structure) Fitting hierarchical (multilevel) model Fitting spatial model (CAR/SAR) Fitting combined hierarchical and spatial model