Extreme Makeover -- Data Edition: Extreme Makeover -- Data Edition: Outside the Box Outside the Box Presentation at the CityMatch Conference, August 2007 Michael Kogan, Ph.D. U.S. Department of Health and Human Services (DHHS) Health Resources and Services Administration (HRSA) Maternal and Child Health Bureau (MCHB) Director, Office of Data and Program Development
Extreme Makeover -- Data Edition: Outside the Box. Presentation at the CityMatch Conference, August 2007 Michael Kogan, Ph.D. U.S. Department of Health and Human Services (DHHS) Health Resources and Services Administration (HRSA) Maternal and Child Health Bureau (MCHB) - PowerPoint PPT Presentation
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Extreme Makeover -- Data Edition: Extreme Makeover -- Data Edition: Outside the BoxOutside the Box
Presentation at the CityMatch Conference, August 2007
Michael Kogan, Ph.D.U.S. Department of Health and Human Services
(DHHS)Health Resources and Services Administration
(HRSA)Maternal and Child Health Bureau (MCHB)
Director, Office of Data and Program Development
Multilevel Modeling: Multilevel Modeling: A Real Life ExampleA Real Life Example
True or False statement?True or False statement?
““Before I had children, I didn’t have any Before I had children, I didn’t have any gray hairs. NONE. Now I have a lot.”gray hairs. NONE. Now I have a lot.”– MK (speaking to his children) MK (speaking to his children)
Multilevel Modeling – Real Life Multilevel Modeling – Real Life ExampleExample
TRUE!TRUE!
But what’s wrong with the statement?But what’s wrong with the statement?
Linear and Logistic Regression – Linear and Logistic Regression – ReviewReview
linear model review:linear model review:
logistic model review:logistic model review:
Y = β0 + β1X1 + β2X2…+ εβ0 = intercept β1X1 = beta associated with exposure
β2X2 = beta associated with first covariate + …
ε = error term
ln [P(X) / (1-P((X))] = α + β1X1 + β2X2…α = constant β1X1 = beta associated with exposureβ2X2 = beta associated with first covariate
First ModelFirst Model
Y = β0 + β1X1 + Y = β0 + β1X1 + εε
Where Y = # of gray hairs on MK’s headWhere Y = # of gray hairs on MK’s head
β1X1 = The presence of children (the β1X1 = The presence of children (the exposure of interest)exposure of interest)
Where Y = # of gray hairs on MK’s headWhere Y = # of gray hairs on MK’s head
β1X1 = The presence of children (the β1X1 = The presence of children (the exposure of interest)exposure of interest)
β2X2 = The number of children (the β2X2 = The number of children (the covariate or a variable that needs to be covariate or a variable that needs to be controlled for)controlled for)
β3X3 = His age (a covariate)β3X3 = His age (a covariate)
β4X4 = Marital status (a covariate)β4X4 = Marital status (a covariate)
BUT…BUT…
… … notice that all those factors are notice that all those factors are individual-level factors (age, number of individual-level factors (age, number of children, marital status).children, marital status).
What if there are broader factors What if there are broader factors influencing the number of gray hairs influencing the number of gray hairs between 1989 and 2007?between 1989 and 2007?
Multilevel Modeling:Multilevel Modeling:Definition and SynonymsDefinition and Synonyms
Multilevel modeling: a method that allows Multilevel modeling: a method that allows researchers to investigate the effect of group or researchers to investigate the effect of group or place characteristics on individual outcomes place characteristics on individual outcomes while accounting for non-independence of while accounting for non-independence of observationsobservations
synonyms:synonyms: different models:different models:– multilevel modelsmultilevel models - fixed effects- fixed effects– contextual modelscontextual models - random effects- random effects– hierarchical analysishierarchical analysis - generalized estimating equations- generalized estimating equations
Why Use Multilevel Models?Why Use Multilevel Models?
outcomes may be clustered by some unit outcomes may be clustered by some unit of aggregation (contextual unit)of aggregation (contextual unit)
individuals within contexts may be similar individuals within contexts may be similar in ways that are unmeasuredin ways that are unmeasured
to take into account clustering / non-to take into account clustering / non-independence of observationsindependence of observations
to partition the observed variability into to partition the observed variability into within-context and between- context within-context and between- context variablesvariables
Why Context MattersWhy Context Matters
empirically, individual outcomes can’t be empirically, individual outcomes can’t be explained exclusively by individual-level explained exclusively by individual-level exposuresexposures
persistent contextual effects are persistent contextual effects are observed in all (?) outcomes across observed in all (?) outcomes across populationspopulations
exposures are structured; distributions exposures are structured; distributions are differentialare differential
Beyond Individual Determinants Beyond Individual Determinants of Healthof Health
Health Status
Demographics, health behaviors, socioeconomic position, support, etc
Ind
ivid
ual-
level
Health Status
Demographics, health behaviors, socioeconomic position, support, etc
Different from Ecological Different from Ecological AnalysesAnalyses
Area and HealthArea and Health
Area—states, counties, cities, neighborhoods--Area—states, counties, cities, neighborhoods--acts as a source of adverse or protective acts as a source of adverse or protective exposures and factors impacting health, such as:exposures and factors impacting health, such as:– PoliciesPolicies– Economic well-being—jobs, unemployment, economic Economic well-being—jobs, unemployment, economic
Other ConsiderationsOther Considerations Unit of analysis: states, counties, zip codes, Unit of analysis: states, counties, zip codes,
census tracts, census block groups, etc. census tracts, census block groups, etc. – Should “neighborhood” be defined using to Should “neighborhood” be defined using to
geographic boundaries? If so, which one?geographic boundaries? If so, which one?
Data sources—going beyond censusData sources—going beyond census
Characteristics to examine, need rationale for Characteristics to examine, need rationale for each indicatoreach indicator
Modeling approachesModeling approaches
Sources of DataSources of Data
Direct MechanismsDirect Mechanisms Community Social EnvironmentCommunity Social Environment
Social relationshipsSocial relationships transmit information transmit information Neighborhood cohesionNeighborhood cohesionsocial controlsocial control Shared cultural norms and valuesShared cultural norms and values Civic participationCivic participation demand services demand services Access to education and employmentAccess to education and employment
Community ServiceCommunity Service Grocery storesGrocery stores Recreational opportunitiesRecreational opportunities Health care facilitiesHealth care facilities Retail storesRetail stores
Alarm, resistance, exhaustion.Alarm, resistance, exhaustion. Repeated cycles lead to cumulative damage to Repeated cycles lead to cumulative damage to
organism.organism.
McEwen & Stellar, 1993—Allostatic LoadMcEwen & Stellar, 1993—Allostatic Load The cost of maintaining stability through changeThe cost of maintaining stability through change
Mental HealthMental Health Negative EmotionsNegative Emotions DepressionDepression Anger/hostilityAnger/hostility
Community context consistently has aCommunity context consistently has a““modest association” with numerous healthmodest association” with numerous healthoutcomes.outcomes.
25 studies reviewed25 studies reviewedDeveloped countriesDeveloped countriesIndividual-level attributes controlled forIndividual-level attributes controlled for23/25 had significant neighborhood 23/25 had significant neighborhood
effectseffects
Reviewed in Pickett & Pearl,Reviewed in Pickett & Pearl, J. Epidemiol Community Health J. Epidemiol Community Health, 2001; 55, 2001; 55
What We Know About What We Know About Neighborhoods and Child Neighborhoods and Child
Well-BeingWell-Being ProtectiveProtective
– AffluenceAffluence– ResourcesResources– Social capitalSocial capital– CohesionCohesion– Collective efficacyCollective efficacy– Urban renewal?Urban renewal?– Political activity?Political activity?
Are We Studying the Right Are We Studying the Right Factors for Policy & Program Factors for Policy & Program
Purposes?Purposes?
ProtectiveProtective– AffluenceAffluence– Social capitalSocial capital– CohesionCohesion– Collective efficacyCollective efficacy– Safe public spacesSafe public spaces– Services & Services &
resourcesresources– Political activity & Political activity &
supportsupport
RisksRisks– Concentrated deprivationConcentrated deprivation– UnemploymentUnemployment– Residential mobilityResidential mobility– Incivilities Incivilities – Poor availability of services Poor availability of services
(health & social)(health & social)– Norms (health and other)Norms (health and other)– Areas for play/social interactionAreas for play/social interaction– SegregationSegregation– Poor housing qualityPoor housing quality
problemproblem: making cross-level inferences : making cross-level inferences [drawing inferences regarding factors [drawing inferences regarding factors associated with variability in outcome at one associated with variability in outcome at one level based on data collected at another level]level based on data collected at another level]
e.g., making individual inferences based on e.g., making individual inferences based on group-level associationsgroup-level associations
Yij = β0 + β1ijX1 + β2ijX2…+ jGj + εij
Yij = outcome for individual i in context j β1ijX1 = beta associated with exposure for individual i in context j
βj Gj = observed community variable ε ij = error term
When are Observations Not When are Observations Not IndependentIndependent??
when data are collected by cluster / when data are collected by cluster / aggregating unitaggregating unit– children within schoolschildren within schools– patients within hospitalspatients within hospitals– drug users within neighborhoodsdrug users within neighborhoods– cholesterol levels within a patientcholesterol levels within a patient
why care about clustered data?why care about clustered data?– two children / observations within one school are two children / observations within one school are
probably more alike than two children / observations probably more alike than two children / observations drawn from different schoolsdrawn from different schools
– does knowing one outcome inform your does knowing one outcome inform your understanding about another outcome?understanding about another outcome?
Back to the Example…Back to the Example…
Potential aggregate factors to consider:Potential aggregate factors to consider:
In 2000, MK moved to a neighborhood with In 2000, MK moved to a neighborhood with a very high percentage of lawyers;a very high percentage of lawyers;
In 2003, the Red Sox lost in the 7In 2003, the Red Sox lost in the 7thth game of game of a playoff series to the Yankees, blowing a a playoff series to the Yankees, blowing a three run lead in the 8three run lead in the 8thth inning. (MK inning. (MK threw pillow at the tv, and was properly threw pillow at the tv, and was properly chastised for that).chastised for that).
Final Outcome of Real-Life Example Final Outcome of Real-Life Example (Assume Two-Level Outcome)(Assume Two-Level Outcome)
Adjusted Odds RatioAdjusted Odds Ratio
Presence of ChildrenPresence of Children 1.15 (1.0, 1.3)1.15 (1.0, 1.3)
Number of ChildrenNumber of Children 1.06 (.80, 1.4)1.06 (.80, 1.4)
AgeAge 1.79 (1.5, 2.1)1.79 (1.5, 2.1)
Marital StatusMarital Status 1.22 (1.1, 1.4)1.22 (1.1, 1.4)
% of Lawyers in % of Lawyers in Neighborhood Neighborhood
1.47 (1.3, 1.6)1.47 (1.3, 1.6)
Red Sox Loss to Yankees Red Sox Loss to Yankees in 2003 Playoffsin 2003 Playoffs
1.75 (1.4, 2.3)1.75 (1.4, 2.3)
AcknowledgmentsAcknowledgmentsLynne Messer, PhD, University of North Lynne Messer, PhD, University of North
CarolinaCarolinaPat O’Campo, PhD, University of TorontoPat O’Campo, PhD, University of TorontoJennifer Culhane, PhD, Drexel UniversityJennifer Culhane, PhD, Drexel University
Contact InformationContact Information
Michael Kogan, Ph.D.Michael Kogan, Ph.D.HRSA/MCHBHRSA/MCHBDirector, Office of Data and Program Director, Office of Data and Program