Linking Genes and Political Orientations: Testing the Cognitive Ability As Mediator Hypothesis

Linking Genes and Political Orientations: Testing the Cognitive Ability as Mediator Hypothesis

Online Appendix

Sven Oskarsson

Uppsala University

David Cesarini

New York University

ChristopherDawes

New York University

James H. Fowler

University of California, San Diego

Magnus Johannesson

Stockholm School of Economics

Patrik KE Magnusson

Karolinska Institutet

Jan Teorell

Lund University

Items Included in the Indices on Political Orientations

The five indices measuring political orientations are based on a principal component analysis of 34 survey items on attitudes toward different policy issues. The principal component analysis with varimax rotation yielded 8 factors with eigenvalues greater than unity.However, upon closer inspection of the scree plot and the pattern of variables loading on each factor we decided to retain only the first five of these. Based on these results we created five additive indices including variables with loadings greater than 0.5 on the first five factors. The same items have been used in several other large scale Swedish surveys such as the Swedish Election Studies, and earlier studies based on the same or related item batteries have found similar factor structures (SCB 2008).The following items are included in the five indices (the response alternative range from (1) “very good proposal” to (5) “very bad proposal”):

Economic policy opinions

Decrease the public sector

Decrease taxes

Sell public companies to private buyers

Have more private companies in health care

Have more private schools

Give private companies more freedom.

Redistribution policy opinions[1]

Increase the economic support to rural areas

Introduce 6-hour working day for all employees

Immigration policy opinions

Introduce much harder punishments for criminals

Increase labor immigration to Sweden

Introduce a language test to become a Swedish citizen

Decrease foreign aid

Accept fewer refugees in Sweden

Increase the economic support to immigrants so that they can preserve their own culture

Environmental policy opinions

Invest more to prevent environmental damages

Decrease carbon dioxide emissions

Foreign policy opinions

Sweden should leave the EU

Sweden should introduce the EURO

Sweden should become members in NATO

[Table A1 about here]

The Univariate ACE Model

In the classical twin design the variance in a trait is decomposed into environmental and genetic influences (Neale et al. 2004; MedlandHatemi 2009). The departure point is a partitioning of the variance into that which is shared by siblings in a twin pair and that which is unique to each twin. Nonshared or unique environmental influences (E) represent the unshared experiences each twin has been exposed to. In the classical twin design E also includes measurement errors. Shared variance can be further partitioned into two sources: common environmental influences shared by the siblings in a twin pair (C) and additive genetic effects (A).[2]

Figure A1 displays the path diagram for a univariateACE model. The squares denote the measured trait (y) for each sibling (i=1, 2) in a twin pair (j). y is expressed as deviations from zero with unit variance. The circles denote the three latent variables A, C, and E. Since common environment by definition is shared by and nonshared environment unique to the twins in a twin pair,and . Furthermore, since monozygotic (MZ) twins share 100% of their genes while dizygotic (DZ) twins share on average half of their segregating genes, for MZ twins and for DZ twins.

[Figure A1 about here]

The model leaves us with three parameters to be estimated (a, c, and e) using three moments from the data (the population variance and the two covariances between MZ and DZ twins, respectively). Using the tracing rules for path analysis (Wright 1934) it can be shown that:

(A1)

(A2)

(A3)

wherea2 is the heritability, defined as the proportion of the total variance accounted for by additive genetic factors, c2 is the share accounted for by common environmental factors, and e2 the proportion accounted for by unique environmental factors.[3]

Some scholars have objected to the so called equal environments assumption (EEA) underlying the classical twin model (Beckwith and Morris 2008; Charney 2008). Above all, if MZ twins in fact are more strongly affiliated compared to DZ twins, then greater concordance among identical twins might as well reflect the fact that MZ twins are treated more similarly compared to DZ twins (an environmental effect). That is, violations of the equal environments assumption might lead to upward bias in the estimates of the heritability of a trait.

However, a handful studies have concluded that EEA violations have not seriously biased heritability estimates reported for political attitudes.Smith et al. (2012) attempted to test whether mutual influence within twin pairs inflated heritability estimates of political attitudes by analyzing measures of frequency of contact. However, the authors found no evidence that this was the case. Using longitudinal data Hatemi et al. (2009) showed that genetic influences on political attitudes are absent during childhood, but differences emerged in adulthood. The authors concluded that such a pattern is incompatible with a violation of the EEA.Hatemi et al. (2010) tested the validity of the EEA assumption for politically relevant traits. Using an extended family design (including twins and their parents, spouses, and non-twin siblings), the authors found no twin-specific environmental effects across a large set of political and social attitudes, thus disconfirming the charge raised by critics of the EEA assumption.[4]

A recently developed method of estimating heritability that avoids the EEA utilizes genotyped single nucleotide polymorphisms (SNPs) to estimate a lower-bound estimate of heritability (Yang et al. 2010, Yang, Lee, Goddard & Visscher 2011, Visscher, Yang & Goddard 2010). Unlike twin studies, this method relies on individuals who are not in the same extended families and therefore environmental similarity is uncorrelated with genetic relatedness.

Using a subsample of the individuals analyzed in this study, Benjamin et al. (2012) apply this technique to estimate the heritability of political and economic attitudes. The estimated share of the variation in these traits explained by SNP arrays is about half of the heritability estimated using twin and family samples. These results, therefore, partially corroborateevidence of significant heritability in political attitudes from earlier behavior genetic studies. The differences may be due to the fact that the genotyped SNPs do not represent all of the heritable variation in these traits or that estimates from twin studies are inflated upwards due to a violation of the EEA.

The Bivariate Cholesky Model

The logic underlying univariate variance decomposition can be extended to the bivariate (or multivariate) case using so called Cholesky decomposition(Loehlin 1996; Evans, Gillespie & Martin 2002; Neale et al. 2004; MedlandHatemi 2009). A schematic example of a bivariate Cholesky decomposition is provided in Figure A2.Corresponding to the univariateACE model and is equal to 1 for MZ twins and 0.5 for DZ twins andand for both MZ and DZ twins.The model leaves us with nine parameters (a11, a21, a22, c11, c21, c22, e11, e21, e22) that can be solved for using nine moments from the data[5]:

(A4)

(A5)

(A6)

(A7)

(A8)

(A9)

(A10)

(A11)

(A12)

is the sole genetic cause of y via path a11 and a partial cause of z via path a21. , then, represents the genetic variance in z via path a22 independent of .[6]

[Figure A2 about here]

The parameter estimates generated by this initial Cholesky decomposition can be used to construct other quantities of interest. Above all, estimates of the correlations across the latent components (ra, rc, and re) are provided. The genetic correlation (ra) is a measure of the degree to which the genetic endowments of trait y and zcovary and is obtained by dividing the genetic covariance between the two traits by the square root of the product of their genetic variances (rc and re are defined analogously):

(A13)

A correlation of 0 means that the two traits are influenced by completely different genes and a correlation of 1 (or –1) means that the same genes influence both traits.

A second quantity of interest is the share of the total phenotypic correlation accounted for by genetic factors. Similar to the univariate case where we decompose the phenotypic variance, we can also decompose the covariance between two traits into additive genetic, shared environmental, and unique environmental components. Above all we are here interested in the share or proportion of this covariation that can be explained by additive genetic factors (share genetic). The share of the total phenotypic correlation accounted for by genetic factors is defined as:

(A14)

where the phenotypic correlation (r) can be written as:

(A15)

Large estimates of share genetic tell us that the covariation between cognitive ability and political orientations mainly is driven by genetic factors.

The Direction of Causation Model

Evidence of a common genetic source for two traits (y and z) does not prove causal mediation. Instead, significant estimates of the genetic correlation may be consistent with any of four different scenarios. First, it may be the case that genes influence y which in turn influences z. Second, the reverse may be true: genes influence z which in turn affects y. The third possibility is reciprocal causation, where y influences z but z also influences y. Finally, it is also possible that the same set of genes is influencing y and z independently – a relationship known as pleiotropy (Posthuma et al. 2003).

Behavior genetics scholars have developed the Direction of Causation (DoC) model which attempts to empirically distinguish between these potential causal pathways (Heath et al 1993: Duffy & Martin 1994; Gillespie et al. 2003).[7] The DoC model predicts that y and z are each determined by independent genetic and environmental sources and that the covariation between them is accounted for by (i) a unidirectional causal influence of y on z, (ii) a unidirectional influence of z ony, or (iii) reciprocal causation. An illustration of the three directional hypotheses is provided in Figure A3.

[Figure A3 about here]

Under certain conditions, cross-sectional analysis of genetically informative data can offer important information on the direction of causation between two traits. The leverage of the DoC model is provided by the different expected cross-twin cross-trait covariances under the three directional hypotheses (y causes z (eq. A16), z causes y (eq. A17), and dual causation (eq. A18):

(A16)

(A17)

(A18)

It is apparent that the two unidirectional models are nested in the dual causation model. Moreover, it can be shown that all three directional models are submodels of the fully saturated Cholesky model (Heath et al 1993). Thus, we can compare the goodness of fit of the two unidirectional DoC models against the dual causation model and all three DoCmodels against the Cholesky model using a likelihood-ratio chi-square test.[8]

The three DoCmodels and the fully saturated Cholesky model correspond to the four scenarios described above. For example, a best fitting Cholesky model would suggest that the relationship between y and z is most likely due to common underlying genetic and/or environmental influences and thus correlational rather than causal.

It is well known that DoC models are only interpretable under very restrictive conditions (Heath et al. 1993; Duffy & Martin 1994; Gillespie et al. 2003). First, the parameters of a model which simultaneously allows for directional causality and pleiotropy cannot be identified with only twin data (Heath et al. 1993). Thus, the pattern of cross-twin cross-trait correlations can only falsify strong hypotheses about the direction of causation. Second, testing causal hypothesis using cross-sectional data requires significantly different heritabilities in the variables (Gillespie et al. 2003). For example, if y and z have identical modes of inheritance then the cross-twin cross-trait covariances will be equivalent under the two unidirectional hypotheses and distinguishing them will not be possible. Third, the directional models are sensitive to measurement errors (Heath et al. 1993; Duffy & Martin 1994). Ignoring measurement error will lead to biased estimates of the parameters of the DoC model and therefore increased risk of incorrect conclusions about the direction of causation.[9]

DoC results for the relationship between cognitive ability and political orientations are presented in Table A2. For each ideological dimension with a phenotypic correlation greater than 0.15 we estimate the three DoC models (two unidirectional and one bidirectional). Columns 1 to 4 provide the relevant fit information for comparisons between the Cholesky model and the three DoC models. Columns 5 and 6 report the estimated causal effects under the three DoC scenarios.[10] The results for the redistribution and foreign policy dimensions are consistent with the mediation hypothesis. In both cases the best fitting model suggests that cognitive ability is causing political attitudes. A reciprocal causation model fits best for the relationship between cognitive ability and immigration policy opinions. According to the coefficient estimates the moderate phenotypic correlation reflects a combination of a strong and expected positive effect of cognitive ability on immigration policy attitudes combined with a much unexpected negative but substantially weaker effect in the opposite direction.

[Table A2 about here]

Figures

Figure A1: The Univariate ACE Model

Note: In squares are the measured traits. In circles are the three latent variance components additive genetic (A), common environment (C) and unique environment (E). Single-headed paths denote hypothesized causal relationships. Double-headed arrows denote covariances between variables.

Figure A2: Bivariate Cholesky Decomposition

Note: In squares are the two observed variables y and z. In circles are the three latent variance components additive genetic (A), common environment (C) and unique environment (E). The latent factors underlying trait y (Ay, Cy, and Ey) are assumed to also influence trait z, however the latent factors underlying trait z (Az, Cz, and Ez) do not affect trait y. To simplify the path diagrams, the model is presented for one twin only.

Figure A3: Direction of Causation Models

Note: Three directional causation hypotheses between y and z, measured on a pair of twins: (i) y causes z, (ii) z causes y, and (iii) reciprocal causation. In the boxes are given the expected cross-twin cross-trait covariances for MZ and DZ twins under each directional hypothesis.

1

Tables

Table A1: Principal Component Analysis Results
Item
Decrease the public sector / 0.67 / -0.18 / 0.20 / -0.06 / 0.10
Decrease taxes / 0.62 / 0.07 / 0.29 / -0.08 / 0.10
Sell public companies to private buyers / 0.73 / -0.12 / -0.03 / -0.02 / 0.23
Have more private companies in health care / 0.76 / -0.06 / -0.02 / -0.09 / 0.14
Have more private schools / 0.73 / 0.07 / -0.08 / -0.07 / 0.02
Give private companies more freedom / 0.57 / 0.10 / 0.10 / 0.08 / 0.12
Increase the economic support to rural areas / -0.06 / 0.53 / 0.03 / 0.12 / -0.21
Introduce 6-hour working day for all employees / -0.25 / 0.61 / 0.07 / 0.03 / -0.21
Introduce much harder punishmentsfor criminals / 0.13 / 0.32 / 0.65 / 0.00 / -0.04
Increase labor immigraion to Sweden / 0.20 / -0.05 / -0.66 / 0.01 / 0.19
Introduce a language test to becomea Swedish citizen / 0.15 / 0.04 / 0.67 / 0.00 / 0.05
Decrease foreign aid / 0.14 / 0.03 / 0.69 / -0.19 / -0.05
Accept fewer refugees in Sweden / 0.04 / 0.04 / 0.82 / -0.05 / -0.05
Increase the economic support to immigrants so / 0.01 / 0.17 / -0.67 / 0.05 / 0.03
that they can preserve their own culture
Invest more to prevent environmental damages / -0.07 / 0.03 / -0.06 / 0.90 / -0.00
Decrease carbon dioxide emissions / -0.04 / 0.05 / -0.07 / 0.89 / 0.02
Sweden should leave the EU / -0.15 / 0.23 / 0.28 / -0.05 / -0.69
Sweden should introduce the EURO / 0.16 / -0.07 / -0.10 / -0.00 / 0.79
Sweden should become members in NATO / 0.27 / 0.06 / 0.23 / -0.08 / 0.58
Decrease defense expenses / 0.01 / 0.08 / -0.08 / 0.03 / -0.03
Decrease welfare benefits / 0.47 / -0.22 / 0.43 / -0.03 / 0.10
Keep property taxes / -0.32 / -0.07 / -0.12 / 0.03 / 0.02
Decrease income inequality in society / -0.35 / 0.43 / -0.04 / 0.13 / -0.10
Decrease the influence offinancial markets on politics / -0.17 / -0.00 / -0.00 / 0.09 / -0.04
Keep the maximum daycare fee / 0.02 / 0.13 / -0.11 / 0.05 / -0.03
Introduce grades earlier in school / 0.46 / -0.15 / 0.30 / -0.30 / 0.15
Forbid all kinds of pornography / -0.01 / 0.49 / 0.02 / 0.00 / -0.00
Limit the right to free abortion / 0.21 / 0.19 / 0.15 / -0.06 / -0.10
Strengthen animal rights / -0.03 / 0.55 / 0.12 / 0.31 / 0.05
Sweden should in the long run carry througha nuclear phase-out / -0.10 / 0.22 / -0.30 / 0.35 / -0.31
Stop motoring in the inner city / -0.02 / 0.14 / -0.10 / 0.44 / -0.09
Remit debt to developing countries / -0.10 / 0.09 / -0.47 / 0.25 / 0.01
Sweden should work for increased free tradeall over the world / 0.18 / 0.04 / -0.09 / 0.20 / 0.33
Sweden should actively supportthe US war on terrorism / 0.20 / 0.32 / 0.34 / -0.03 / 0.37
Note: Rotated factors from principal component analysis of 34 issue items. Items with factor loadings greater than 0.5 are shaded. Only the first five factors are displayed.

1

Table A2: Model Fitting Results and Parameter Estimates for DoC models
Coefficients
Model / diff χ2 / Df / p / AIC / Cog Ab→Att / Att→Cog Ab
Redistribution policy opinions / Full Cholesky / 4290.32
Reciprocal Causation / 3.62 / 1 / 0.06 / 4287.94 / 0.40 / 0.01
[0.21, 0.77] / [-0.44, 0.20]
Cognitive Ability / 3.64 / 2 / 0.16 / 4285.96 / 0.42
Causes Attitudes / [0.36, 0.47]
Attitudes Causes / 12.36 / 2 / 0.00 / 4294.68 / 0.41
Cognitive Ability / [0.34, 0.47]
Immigration policy opinions / Full Cholesky / 7844.00
Reciprocal Causation / 2.34 / 1 / 0.13 / 7840.34 / 0.44 / -0.15
[0.31, 0.57] / [-0.30, -0.03]
Cognitive Ability / 7.93 / 2 / 0.02 / 7843.93 / 0.29
Causes Attitudes / [0.24, 0.35]
Attitudes Causes / 34.02 / 2 / 0.00 / 7870.01 / 0.24
Cognitive Ability / [0.18, 0.29]
Foreign policy opinions / Full Cholesky / 6284.56
Reciprocal Causation / 2.44 / 1 / 0.12 / 6281.00 / 0.17 / 0.04
[0.01, 0.33] / [-0.11, 0.19]
Cognitive Ability / 2.73 / 2 / 0.26 / 6279.29 / 0.21
Causes Attitudes / [0.15, 0.27]
Attitudes Causes / 6.60 / 2 / 0.04 / 6283.16 / 0.18
Cognitive Ability / [0.13, 0.24]
Note: Column 1 presents the chi square distributed differences in -2LL between the Cholesky model and each directional model. Columns 2 and 3 present the corresponding degrees of freedom and p-values. Column 4 reports the Akaike Information Criterion. Columns 5 and 6 report the estimated causal effect of cognitive ability on political attitudes (5) and of political attitudes on cognitive ability (6) under each DoC scenario. The best fitting models are bolded. All estimates are corrected for measurement attenuation. The results are based on models using male twins.

1

References

Beckwith, J. and C. Morris. 2008. “Twin Studies of Political Behavior: Untenable Assumptions?” PS: Political Science & Politics 41:785–791

Benjamin, D., D.Cesarini, M. van der Loos, C. Dawes, P.Koellinger, P.Magnusson et al. 2012. “The Genetic Architecture of Economic and Political Preferences.” Proceedings of the National Academy of Sciences of the United States of America 109:8026–8031

Charney, E. 2008.“Genes and Ideologies.”PS: Political Science & Politics 41:299–319

Duffy, D. and N. Martin. 1994. “Inferring the Direction of Causation in Cross-Sectional Twin Data: Theoretical and empirical Considerations.” Genetic Epidemiology 11:483–502

Evans, D., N. Gillespie, and N. Martin. 2002. ‘Biometric Genetics.’ Biological Psychology61: 33–51

Gillespie, N., G. Zhu, M. Neale, A. Heath, and N. Martin. 2003. “Direction of Causation Modeling Between Cross-Sectional Measures of Parenting and Psychological Distress in Female Twins.” Behavior Genetics 33:383–396

Hatemi, P., C. Funk, S.Medland, H.Maes, J.Silberg, N. Martin, and L. Eaves. 2009. “Genetic and Environmental Transmission of Political Attitudes Over a Life Time.” Journal of Politics 71:1141–1156

Hatemi, P., J. Hibbing, S.Medland, M. Keller, J. Alford, K. Smith et al. 2010. “Not by Twins Alone.” American Journal of Political Science 54:798–814