Session 1 Slides Genetic Variation and Economic Behavior David - PDF document

Session 1 Slides

Genetic Variation and Economic Behavior David Cesarini CESS, Economics Department New York University Workshop to Explore SSGAC • 12 February 2011

Heritability of Social Science Outcomes • Socioeconomic Outcomes – Educational attainment (Behrman et al., 1975; Miller et al., 2001; Scarr and Weinberg, 1994; Lichtenstein et al., 1992) – Income (Björklund, Jäntti and Solon, 2005; Sacerdote, 2007; Taubman, 1976) • Economic Preferences – Risk preferences (Cesarini et al., 2009; Zhong et al. 2009; Zyphur et al. 2009) – Bargaining behavior, altruism and trust (Wallace et al., 2007; Cesarini et al., 2008) • Economic Behaviors – Financial decision-making (Barnea et al., 2010; Cesarini et al, 2010) – Susceptibility to decision-making anomalies (Cesarini et al., 2011) David Cesarini, NYU 2

An Example: Educational Attainment NICE GRAPHICS HERE. David Cesarini, NYU 3

Evidence from the SALTY survey David Cesarini, NYU 4

Concluding Thoughts • Variance decomposition subject to a number of important issues of interpretation – Environmental mediation of genetic effects (Dickens and Flynn, 2005; Jencks, 1979; Ridley, 2003) • Suggests a need to understand a need to understand why genotype correlates with economic outcomes and behaviors • Heritable variation in these complex traits likely explained by a heterogeneous collection of mechanisms – But many of the precursors of socioeconomic outcomes, for example risk preference, are measured with noise. David Cesarini, NYU

The Case for a Social Science Genetic Association Consortium Daniel J. Benjamin Economics Department Cornell University Workshop to Explore SSGAC • 12 February 2011

Collaborators for Results in the Talk Craig Atwood (University of Wisconsin-Madison) Jonathan Beauchamp (Harvard University) Christopher F. Chabris (Union College) Jeremy Freese (Northwestern University) Edward L. Glaeser (Harvard University) Vilmundur Guðnason (Icelandic Heart Association) Tamara B. Harris (National Institute on Aging) Robert M. Hauser (University of Wisconsin-Madison) Taissa S. Hauser (University of Wisconsin-Madison) Benjamin M. Hebert (Harvard University) David I. Laibson (Harvard University) Lenore J. Launer (National Institute on Aging) Shaun Purcell (Massachusetts General Hospital, Broad Institute) Albert Vernon Smith (Icelandic Heart Association) We gratefully acknowledge NIA for financial support Daniel Benjamin - Cornell University 2

Some Payoffs from “Genoeconomics” 1. Genes as instrumental variables 2. Understanding market and behavioral mediation of genetic effects – Genes are measures of (until-now latent) parameters of economic models: abilities and preferences. 3. Biological mechanisms for social behavior – Could decompose crude concepts like “risk aversion” and “patience.” 4. Policy implications of genetic information – Effects of public release on, e.g., market prices and allocations of health insurance. – Do the benefits of private release (anticipatory behaviors, reduced uncertainty) outweigh the costs? – Targeting social-science interventions – E.g., children with dyslexia-susceptibility genotypes could be taught to read differently from an early age. Daniel Benjamin - Cornell University 3

Challenge #1: Phenotype selection • Want high-reliability phenotypes, consistently measured across many datasets. – E.g., height, g , years of education. • Want proximate biological pathway for effect. – If pathway too distal, effect will likely be small, so low power. – If different pathways in different local environments, few datasets available to replicate. – Proximate pathway more likely for phenotypes shared with animal models. – E.g., aggression? Risk aversion? Impulsivity? Daniel Benjamin - Cornell University 4

Challenge #2: Causal inference • Confounds, e.g.: – Ethnicity – Gene-environment correlation – Gene-gene correlation • Need convergent evidence from: – Large family samples – Modeling and estimation of environmental effects – Knock-out experiments with animal models – Biological evidence on protein products of genes • Will take a long time to accumulate evidence. Daniel Benjamin - Cornell University 5

Challenge #3: Statistical power • Low power is due to small effect sizes. – COMT has R 2 = .1% for cognitive ability. – Largest height association is R 2 = .3%. • Low power exacerbated by: – Multiple hypothesis testing + publication bias. – Inconsistent or low-reliability phenotypes. – Search for G x E or G x G interaction. • Evidence for low power: – Many published associations not reproducible. Daniel Benjamin - Cornell University 6

Calibration: Power Analysis • Two alleles: High and Low. • Equal frequency of High and Low. • Phenotype distributed normally. • Either there is a true association or not. • If associated, R 2 = .1% (large for behavior). • Sample size for 80% power: 7,845. • Now suppose significant association at α = .05. Daniel Benjamin - Cornell University 7

Posterior probability of a true association Sample size N = 100 N = 5,000 N = 30,000 (power = .06) (power = .61) (power = .99) Prior .01% .01% .12% .20% prob- 1% 1% 11% 17% ability 10% 12% 58% 69% Calculated by Bayes’ Rule: Daniel Benjamin - Cornell University 8

Case Study: My Experience • We developed SNP panel and applied to large, ethnically homogeneous, well-characterized longitudinal dataset: AGES-Reykjavik Study. • We conducted association analysis with 415 SNPs and 8 “economic” phenotypes. ( N ≈ 2300) Daniel Benjamin - Cornell University 9

Case Study: My Experience • We developed SNP panel and applied to large, ethnically homogeneous, well-characterized longitudinal dataset: AGES-Reykjavik Study. • We conducted association analysis with 415 SNPs and 8 “economic” phenotypes. ( N ≈ 2300) • We found 3 associations with .001 significance threshold. • One replicated in a non-overlapping sample from the same dataset: SSADH rs2267539 associated with “human capital” (composed of years of schooling and number of languages learned). ( N ≈ 1750) Daniel Benjamin - Cornell University 10

Hum an Capital by Genotype 0.6 0.5 Mean Hum an Capital I ndex 0.4 0.3 0.2 0.1 0 -0.1 G/G A/G A/A n = 2824 n = 1081 n = 108 (8.3 years) (8.8 years) (8.9 years) -0.2 Genotype Daniel Benjamin - Cornell University 11

Case Study: My Experience • We developed SNP panel and applied to large, ethnically homogeneous, well-characterized longitudinal dataset: AGES- Reykjavik Study. • We conducted association analysis with 415 SNPs and 8 “economic” phenotypes. ( N ≈ 2300) • We found 3 associations with .001 significance threshold. • One replicated in a non-overlapping sample from the same dataset: SSADH rs2267539 associated with “human capital” (composed of years of schooling and number of languages learned). ( N ≈ 1750) • We found the association was mediated by cognitive function. Daniel Benjamin - Cornell University 12

Cognitive Function by Genotype 0.2 0.15 Mean Cognitive Function I ndex 0.1 0.05 0 -0.05 G/G A/G A/A n = 2282 n = 893 n = 90 -0.1 Genotype Daniel Benjamin - Cornell University 13

Case Study: My Experience • We developed SNP panel and applied to large, ethnically homogeneous, well-characterized longitudinal dataset: AGES- Reykjavik Study. • We conducted association analysis with 415 SNPs and 8 “economic” phenotypes. ( N ≈ 2300) • We found 3 associations with .001 significance threshold. • One replicated in a non-overlapping sample from the same dataset: SSADH rs2267539 associated with “human capital” (composed of years of schooling and number of languages learned). ( N ≈ 1750) • We found the association was mediated by cognitive function. • The association failed to replicate in 3 other samples. Daniel Benjamin - Cornell University 14

Are we alone? • We could not replicate a promising candidate gene result. – Even though the result survived initial replication attempts. – Even though there seemed to be a reasonable physiological story connecting the gene to the variable. • Does the social science genetics literature contain many false positives? – Beauchamp et al (forthcoming) find 20 promising, biologically plausible SNPs in an education GWAS in Framingham ( N = 7,574). – In replication attempt with Rotterdam Study ( N = 9,535), none significant at .05 level, and only 9 of 20 had same sign. • Candidate gene associations with social science variables seem to be especially vulnerable to being false positives. – Using WLS data, we could not replicate any of 13 SNPs with published g associations. – We had good power, positive controls (APOE4–parental AD). Daniel Benjamin - Cornell University 15