Dr. David L. Pearl Dept. of Population Medicine, University of Guelph Guelph, Ontario, Canada

The role of multi-level modeling in The role of multi-level modeling in understanding risk factors for understanding risk factors for

reported rates of human cases of reported rates of human cases of Escherichia coliEscherichia coli O157:H7 O157:H7

Dr. David L. PearlDr. David L. Pearl

Dept. of Population Medicine, University of GuelphDept. of Population Medicine, University of Guelph

Guelph, Ontario, CanadaGuelph, Ontario, Canada

OutlineOutline

• Review epidemiology of E. coli O157:H7.

• Utility of multi-level modeling for zoonotic diseases in human populations.

• Multi-level study of E. coli O157:H7 in Alberta, Canada.

• Future of multi-level modeling for E. coli O157:H7 and other zoonoses.

Epidemiology of Epidemiology of E. coliE. coli O157:H7O157:H7

Nature of diseaseNature of disease

• Mild diarrhea to hemorrhagic diarrhea to HUS.• HUS – anemia, thrombocytopenia, renal failure (2-

10% of cases).• Highest incidence of illness in children less than

five and elderly.• Increased risk of exposure if rural vs. urban. • Greater incidence of disease in summer vs. other.• Asymptomatic and intermittent shedding in cattle

(major reservoir).

More than a “Hamburger More than a “Hamburger Disease”Disease”

• Meat• Waterborne• Cross-contamination in food

preparation• Contamination of produce

with ruminant feces • Person-to-person • Animal-to-person*• Other foods of animal origin

Canadian statisticsCanadian statistics

• 3-5 cases per 100 000 population.

• 365 hospitalizations per 1000 cases .

• 39 deaths per 1000 cases.

• 1 reported case for 4-8 symptomatic cases (Ontario).*

E. coliE. coli O157:H7 & Alberta O157:H7 & Alberta

• Crude rates higher than national average. • > 9 cases/100 000 from 2000-2002.

• Southern RHAs report highest rates.

• Province only reports community outbreaks to federal government (few and small).

Passive surveillance in AlbertaPassive surveillance in Alberta

• Alberta Health and Wellness collects Notifiable Disease Reports – (17 RHAs)

• Provincial Laboratory for Public Health (Microbiology) performs PFGE

Spatial clusteringSpatial clustering

• Spatial scan statistics show clustering in the south.

• Cluster location varies depending on use of sporadic vs. all cases.

• More complex than just issue of cattle density.

Pearl et al., 2006

Utility of multi-level modeling Utility of multi-level modeling for zoonotic diseasesfor zoonotic diseases

Multi-level modelingMulti-level modeling

• Hierarchical or random effects models.

• Statistical model including fixed and random effects.

• Fixed effects represent the mean effects typically seen in ordinary regression models.

• Random effects adjust for auto-correlated data while allowing the proper estimation of variables from different hierarchical levels.

Multi-level modeling – cont.Multi-level modeling – cont.

• Random effects correspond to effects (e.g., clusters) in the model being randomly selected from a population.

• With random effects, focus shifts from individual group to variability in the population of groups (σ2

group).• Can have random effects for multiple hierarchical

levels.• Variance components assist in identifying level where

intervention will have the most impact.

Multi-level modeling – cont.Multi-level modeling – cont.

• Diez-Roux (2000) emphasized multilevel modeling in public health research.

• Already common in veterinary epidemiology to account for farm-level clustering.

• Similar approaches used to account for spatial clustering among counties to investigate role of agriculture on rates of disease from E. coli O157 (Michel et al. 1999; Valcour et al. 2002).

Zoonoses and multi-level modelingZoonoses and multi-level modeling

• Individual-level factors – age, socio-economic factors, food-handling practices.

• Community-level factors – migration patterns, health policy and legislation, socio-economic factors as contextual variables.

• Regional factors – type of water-shed, density of animal reservoir.

Multi-level study of Multi-level study of E. coliE. coli O157:H7 in Alberta, CanadaO157:H7 in Alberta, Canada

(Pearl et al., 2009)(Pearl et al., 2009)

ObjectivesObjectives

1. Estimate the association between socio-economic and agricultural variables related to income, migration, degree of integration with urban core, aboriginal status, and cattle density on rates of E. coli O157:H7 in Albertan communities.

2. Determine the impact of using all available data or only sporadic data on the associations observed.

Study Data - Alberta 2000-2002Study Data - Alberta 2000-2002

• 875 cases recorded by passive surveillance• 826 human cases examined using PFGE• NDR provided spatial, temporal, and epidemiological data• 2001 Canada Census location of CSD centroids and socio-

agricultural data • Based on review of NDR 78.9 % of cases were sporadic

(14 community outbreaks & 55 household outbreaks)

Hierarchical structure of dataHierarchical structure of data

C e n su s S u b d iv is io n (C S D )

C o n so lid a te d C e n sus S u b d iv isio n (C C S )

C e n su s D iv is io n (C D )

Variables available through the Variables available through the 2001 Canada Census2001 Canada Census

1. Cattle density (CCS) – forced into all models

2. Metropolitan influenced zones (CSD)

3. Percent movers (CSD)

4. High aboriginal population (CSD)

5. Percent low income households (CSD)*

* Non-significant

Statistical model buildingStatistical model building1. Models: Negative binomial models & Poisson models

with random effects.2. Outcome and exposure - case counts and expected

counts, respectively.3. Methods: Adaptive quadrature for multi-level models

(GLLAMM procedure).4. Levels: 2 & 3.

ln E(Y) = ln η + β0 + β1 X1i j+….+ βk Xkij + u(ccs(j))

u(ccs(j)) ~ N(0, σ2 u)

Statistical model building – cont.Statistical model building – cont.5. Data used: All or sporadic.6. Assessed linearity assumption for continuous

variables.7. Tested interactions among cattle density and socio-

economic factors.8. Likelihood-ratio tests to assess significance of fixed

effects.9. Additional random effect at CSD-level assessed for

residual overdispersion.10. Used AIC and BIC scores to compare different models

for quality of fit.

Best fitting modelsBest fitting models

1. Poisson models built with sporadic cases alone: 2-level model with a random intercept for CCS.

2. Poisson models built with outbreak and sporadic cases: 2-level model with random intercepts for CSD and CCS.

Fixed effects – Poisson 2-level (AFixed effects – Poisson 2-level (A)) VariablesVariables IRR (95% CI)IRR (95% CI)

Cattle/km2 (Q4 v. Q1)* 2.40 (1.16 – 4.98)

MIZ – Strong_Mod 1.45 (0.82 – 2.55)

MIZ – Weak 1.38 (0.77 – 2.50)

MIZ – Non-influenced 4.59 (2.10 – 10.03)

Percent movers

Percent movers squared

1.26 (1.12 – 1.43)

0.996 (0.993-0.999)

High aboriginal pop 0.34 (0.18 – 0.66)

* Statistically non-significant in model using sporadic data.

Major findingsMajor findings1. Economic link to urban centers associated with

community risk.2. Population stability associated with community

risk.3. High aboriginal population appears to have a

sparing effect (reporting issues).4. Statistical significance of cattle density influenced

by models (power issues or type of industry).5. Quality of surveillance system may impact our

understanding.

Study limitationsStudy limitations

• Specificity of agricultural and socio-economic variables limited due to confidentiality issues with Census.

• Limited to ecological studies without data concerning individual-level exposures.

• Inferences should be limited to contextual effects.

Future of multi-level modeling Future of multi-level modeling for for E. coliE. coli O157:H7 and other O157:H7 and other

zoonoses in human populationszoonoses in human populations

Requirements for future studiesRequirements for future studies

• Access to high resolution census data (Research Data Centres (RDC) Program).

• Incorporating contextual variables with case-control data at the individual level.

• Further development of statistical methodologies to integrate hierarchical approaches with spatial and social-network effects.

AcknowledgementsAcknowledgements

Scott McEwen and Wayne Martin (University of

Guelph)

Kathryn Dore, Karen Grimsrud, and Pascal Michel

(Public Health Agency of Canada)

Marie Louie and Linda Chui (Alberta Provincial

Laboratory for Public Health)

Larry Svenson (Alberta Health and Wellness)