New Zealand as a “Social Laboratory” [A James Cook Fellowship Proposal] COMPASS Seminar Series Monday, 3 August 2015, Fale Pacifika Professor Peter Davis Department of Sociology, COMPASS Research Centre
New Zealand as a “Social Laboratory”[A James Cook Fellowship Proposal]
COMPASS Seminar SeriesMonday, 3 August 2015, Fale Pacifika
Professor Peter Davis Department of Sociology, COMPASS Research Centre
New Zealand as a “Social Laboratory”[A James Cook Fellowship Proposal]
Preamble - Making Knowledge Claims• The Year of Evaluation – RCTs• Impact of societal inequality• Improving inference with better design• The simulation approach
The
Uni
vers
ity o
f Auc
klan
dN
ew Z
eala
nd
Pickett and Wilkinson,Soc Sci Med 2015
5
The
Uni
vers
ity o
f Auc
klan
dN
ew Z
eala
nd
Pickett and Wilkinson,Soc Sci Med 2015
6
The
Uni
vers
ity o
f Auc
klan
dN
ew Z
eala
nd
Avendano article,Soc Sci Med 2012
7
The
Uni
vers
ity o
f Auc
klan
dN
ew Z
eala
nd
Pickett and Wilkinson,Soc Sci Med 2015
8
Avendano 2012
9
New Zealand Correlation = -0.92
Source: Avendano data
France
Source: Avendano data
Correlation = +0.96
11
New Zealand as a “Social Laboratory”[A James Cook Fellowship Proposal]
Preamble - Making Knowledge Claims• The Year of Evaluation – RCTs• Impact of societal inequality• Improving inference with better design• The simulation approach
The
Uni
vers
ity o
f Auc
klan
dN
ew Z
eala
nd
Causal Inference in Observational Settings
7th Wellington ColloquiumStatistics NZ30 August 2013
Professor Peter DavisUniversity of Auckland, New Zealandand COMPASS Research Centrewww.compass.auckland.ac.nz
New Zealand as a “Social Laboratory”[A James Cook Fellowship Proposal]
The
Uni
vers
ity o
f Auc
klan
dN
ew Z
eala
nd
Assessing policy counterfactuals with a simulation-based inquiry system.
Peter Davis and ColleaguesCOMPASS Research CentreUniversity of Auckland New Zealandwww.compass.auckland.ac.nz
DISCLAIMER: Access to the data used in this study was provided by Statistics New Zealand under conditions designed to give effect to the security and confidentiality provisions of the Statistics Act 1975. The results presented in this study are the work of the author, not Statistics New Zealand.
17
The
Uni
vers
ity o
f Auc
klan
dN
ew Z
eala
nd
Assessing counterfactuals
Counterfactual paradigm of causal reasoningIf the putative causal factor had not been present, we would not have observed the recorded outcome.
• Randomised Controlled Trials (RCTs)
• Experimental and quasi-experimental methods
• Observational designs and statistical analysis
Simulation techniques
18
The
Uni
vers
ity o
f Auc
klan
dN
ew Z
eala
nd
Simulation techniques
Simulation – in our case, social simulationUse of computer models (computational techniques) to “mimic” social phenomena (e.g. social processes).
• Understand phenomena better in constructing the models
• Once understood and validated, one can alter features
• Particularly useful for sub-groups and future projections
• Ability to combine different data sources in a single model
• Overcome privacy and confidentiality issues
19
The
Uni
vers
ity o
f Auc
klan
dN
ew Z
eala
nd
Model Year Locality Type Life stage Domain Software Data Funder Collaborators End-usersMOSC 2005-8 NZ ABM/MSM Adults Marriage market,
residential segregation
NetLogoRepastJava
Census Marsden UOA
PCASO 2005-8 NZ Static discrete-time MSM
Older people Health care SAS NATMEDCANZHSANHS
HRC UOANatSem
BCASO 2009-12 NZ
Dynamic discrete-time MSM
Older people Health & social care
R NZHSNZDSCensus
HRC UOANatSem
MEL-C 2009-13 NZ Children Health, education, conduct
JavaR
CHDSDMDHSPIFSTHNRCensus2006
MBIE UOANatSemStatCan
MOEMOHMOJMSDTe Puni KokiriSUPERUChildren’s Commissioner
KNOW-LAB
2013-16 World Children & young people
Health, education, conduct, etc.
Published literature
MBIE UOAStatCan
SOCLAB 2015-17 NZ Whole society
Social Modgen/OpenM++
NZLC Royal Society
UOA Open-source modelling community
TPM 2015-20 NZ Whole society
Social Census TEC UOAMOTU
LEGENDModel Data Funder
MOSC: Modelling social changePCASO: Primary care in an ageing societyBCASO: Balance of care in an ageing societyMEL-C: Modelling the early life courseKNOW-LAB: Knowledge laboratorySOC-LAB: NZ social laboratoryTPM CORE: Te Punaha Matatini
NATMEDCA: National Primary Medical CareNZHS: NZ Health SurveyNZDS: NZ Disability SurveyANHS: Australian National Health SurveyCHDS: Christchurch Health & Development StudyDMHDS: Dunedin Multidisciplinary Health & Development StudyPIFS: Pacific Island Families StudyTHNR: Te Hoe Nuku RoaNZLC: NZ Longitudinal Census
HRC: Health Research CouncilMBIE: Ministry of Business, Innovation & EmploymentRoyal Society: James Cook FellowshipTEC: Tertiary Education Commission
CollaboratorsUOA: University of AucklandNatSem: National Centre for Social & Economic Modelling, University of CanberraStatCan: Statistics CanadaMOTU: Motu Economic and Public Policy Research
Type End-usersABM: Agent based modelMSM: Micro-simulation model
MOE: Ministry of Education, MOH: Ministry of HealthMOJ: Ministry of Justice,MSD: Ministry of Social Development
Simulation at COMPASS
20
The
Uni
vers
ity o
f Auc
klan
dN
ew Z
eala
nd
Micro-simulation approach.
We start with a sample of individualsReal (studies) / synthetic (derived from Census)
We derive statistical rules to create a “virtual cohort” that mimics the “real” one
Derive rules best able to reproduce real dataApply these rules to the base file to create a synthetic sample of typical biographies through life course
We then simulate what might happen if policywere to change, by altering parameters
Using software application to test counterfactuals21
Virtual versus real cohort: family doctor visits, reading ability, and conduct problems, by year of age
22
Year Real cohort (CHDS)n=1017
Virtual cohort (simulated) n=1017
Absolute error Absolute error / CHDS mean
Family doctor visits (mean (95% CI))1 5.82 5.82 - -2 5.34 5.28 0.06 -3 3.31 3.18 0.13 -4 3.13 3.15 0.02 -5 3.22 3.12 0.10 -6 3.35 3.32 0.03 -7 2.43 2.41 0.02 -8 2.14 2.15 0.01 -9 1.96 1.90 0.06 -
10 1.65 1.68 0.03 -All years 3.24 3.20 (3.15-3.25) 0.04 1.2%
Reading ability: BURT score (mean (95% CI))8 45.3 45.3 - -9 54.4 54.7 0.3 -
10 64.1 63.7 0.4 -11 72.8 71.9 0.9 -12 79.5 78.9 0.6 -13 85.2 84.6 0.6 -
All years 66.9 66.5 (65.7-67.4) 0.4 0.6%Conduct problems (mean (95% CI))
6 10.6 10.6 - -7 24.6 24.8 0.2 -8 24.4 25.0 0.6 -9 24.7 25.3 0.6 -
10 24.9 25.6 0.7 -All years 21.8 22.3 (22.1-22.4) 0.5 2.3%
The
Uni
vers
ity o
f Auc
klan
dN
ew Z
eala
nd
Inquiry Tool
23
New Zealand as a “Social Laboratory”[A James Cook Fellowship Proposal]
ANY QUESTIONS AT THIS POINT!?
New Zealand as a “Social Laboratory”[A James Cook Fellowship Proposal]
Assessing Counterfactuals about Society• Background and concept• Central ingredients of project• COMPASS team contribution• Building blocks• Book proposal
Outline
• Background• Central ingredients of the James Cook• COMPASS team contribution• Building blocks
– New Zealand Longitudinal Census (1981-2013)– Synthetic base file– Estimating equations– Open-source micro-simulation platform
• Book proposal
Background –assessing counterfactuals
• New Zealand, an early policy pioneer– 1890-1920 seen by observers as “social laboratory”– Social policies tried out by reforming governments– A “natural experiment” in a new, fluid society
• How to draw credible inferences about policy?– RCTs, experimental and quasi-experimental designs– Non-experimental work (e.g. case studies)– Virtual “experiments”, using simulation techniques– Any precedents? Think of climate change scenarios
Background –assessing counterfactuals
• New Zealand, an early policy pioneer– 1890-1920 seen by observers as “social laboratory”– Social policies tried out by reforming governments– A “natural experiment” in a new, fluid society
• How to draw credible inferences about policy?– RCTs, experimental and quasi-experimental designs– Non-experimental work (e.g. case studies)– Virtual “experiments”, using simulation techniques– Any precedents? Think of climate change scenarios
Central Ingredients of James Cook• Three aims
– Create model of NZ pop via synthetic cohort– Statistical model from NZLC to generate cohorts– Conduct experiments, “virtual counterfactuals”
1. Constructing smaller, synthetic cohortsNeed synthetic starting file for each cohort, 1981
2. Estimating statistical model driving cohortsMethod for reproducing biographical trajectories
3. Testing “virtual counterfactuals”Particular interest in impact of social assets
Three “Virtual Counterfactuals”• Health impact of 1980s/1990s restructuring
– Blakely et al. use a cross-comparative counterfactual (Norway)– We can try returning key exogenous parameters to the long-
term pre-disruption trend line– Assess impact on health inequalities
• Long-term impact of “Working For Families”– What would have happened had “normal” settings applied?– Did the in-work tax credit work against the poor?
• Impact of social assets on valued goals and outcomes – Determine relationship of assets (e.g. social and cultural capital)
to achievement of valued goals and outcomes– Alter distribution of these non-monetary assets to assess impact
The
Uni
vers
ity o
f Auc
klan
dN
ew Z
eala
nd
Senior RF -Barry Milne
Research-Policy Collaboration – Published 2014
34
The
Uni
vers
ity o
f Auc
klan
dN
ew Z
eala
nd
Senior RF –Roy Lay-Yee
Determinants and Disparities – Published 2015
35
The
Uni
vers
ity o
f Auc
klan
dN
ew Z
eala
nd
Statistician –Jessica McLay
Regression Estimation for Dynamic Microsimulation (McLay et al.)
• ACCEPTED WITH REVISIONS (International Journal of Microsimulation)
Abstract: Microsimulation models seek to represent real-world processes and can generate extensive amounts of synthetic data. Most often, the parameters that drive the data generation process are estimated by statistical modelling techniques, such as regression. But which techniques are best suited to this purpose? We assess the performance of five regression-style estimation techniques: ordinary least squares regression with a lagged dependent variable, random effects with and without an autoregressive order 1within-unit error structure, a hybrid model combining features from both econometric fixed effects and random effects models, and a dynamic panel model estimated with system generalised method of moments. The criterion for good performance was the proximity of fit of simulated data to empirical data on various characteristics. It was found that ordinary least squares regression with a lagged dependent variable out-performed the other techniques. Random effects with autoregressive errors of the first order was the next best, followed by standard random effects. The dynamic panel model came fourth followed by the hybrid model. This empirical assessment provides practical guidance to those contemplating dynamic microsimulation and other applications using regression-style techniques of synthetic data generation.
36
Outline
• Background• Central ingredients of the James Cook• COMPASS team contribution• Building blocks
– New Zealand Longitudinal Census (1981-2013)– Synthetic base file– Estimating equations– Open-source micro-simulation platform
• Book proposal
Cohort Estimating Equations
Open-Source Simulation Software
Book Proposal
Central Ingredients of James Cook• Three aims
– Create model of NZ pop via synthetic cohorts– Statistical model from NZLC to generate cohorts– Conduct experiments, “virtual counterfactuals”
1. Constructing smaller, synthetic cohortsNeed synthetic starting file for each cohort, 1981
2. Estimating statistical model driving cohortsMethod for reproducing biographical trajectories
3. Testing “virtual counterfactuals”Particular interest in impact of social assets
New Zealand as a “Social Laboratory”[A James Cook Fellowship Proposal]
QUESTIONS, COMMENTS!