Heritability of trust and distrust remains unknown

In “Trust is heritable, whereas distrust is not,” Reimann et al. (1) analyze data from monozygotic (MZ) and dizygotic (DZ) twins and conclude that “the disposition to trust is explained to some extent by heritability but not by shared socialization, [whereas] the disposition to distrust is explained by shared socialization but not by heritability.” We believe that the evidence does not justify either inference.

As illustrated by the title of the article (1), the study claims to demonstrate that heredity has no effect on distrust. This claim confuses a lack of evidence for an effect with confirmation that no effect exists. The study’s data show only that one should not reject the null hypothesis that distrust is not inherited. In fact, estimates from the study’s own data with respect to the heritability of distrust also reveal that any null hypothesis between 0% and 31% should not be rejected.

The Reimann et al. (1) article also fails to provide evidence that the effect of heredity and shared socialization can be separated. To test the effect of heredity on trust, correlations in trust scores between MZ and DZ twins are compared, but the difference fails to meet customary significance thresholds (∆R = 0.16; Z = 1.34; P = 0.18). ACE structural modeling is then used to separate the effect of heredity (a²) and shared socialization (c²) on both trust and distrust. The ACE models fit the data poorly: seven of eight models fail χ² fit tests, and Akaike Information Criterion (AIC) and Bayesian Information Criterion scores suggest that no single best model can be selected with confidence. As a result, Reimann et al. are not justified in picking one reduced model (AE or CE) as the only basis for inference. Models with all ACE factors reveal that confidence intervals (CI) for the effect of heredity and shared socialization are large and overlapping (for trust, a²: 0.17–0.42, c²: 0.00–0.31; for distrust, a²: 0.00–0.31, c²: 0.07–0.30). The only justified inference is that more research is needed.

We propose that future research should supplement ACE modeling with other statistical approaches. ACE modeling involves strong assumptions about the data-generating process. As shown by Bencheck and Morris (2), ACE estimates and intervals can be biased if these assumptions are violated even slightly. Indeed, our own replications of the study’s estimates using different software resulted in large and varying confidence intervals (Table 1). Given the potential for varying estimates, we believe that future research on twins should triangulate to reliable inferences by supplementing ACE models with other statistical approaches. In our case, we chose to use Tobit regression to test the prediction that trust scores would be more similar between MZ than DZ twin pairs (Tables 2 and 3). The Tobit results reinforced our inference that the Reimann et al. (1) study’s data are very noisy and thus do not provide reliable evidence that trust is inherited.

Table 1.

ACE reanalysis using published study data and two software programs

Model	Mean estimated proportions of total variance (95% CI)
	a2	c2	e2	AIC
Trust
Reported estimates	0.30 (0.17, 0.42)	0.00 (0.00,0.31)	0.70 (0.58,0.83)	1809.92
Replication with OpenMX	0.30 (0.00, 0.42)	0.00 (0.00,0.31)	0.70 (0.58,0.84)	1809.92
Replication with STATA	0.30 (0.21,0.42)	0.00 (0.00,1.00)	0.70 (0.66,0.74)	1809.92
Distrust
Reported estimates	0.00 (0.00,0.31)	0.19 (0.07,0.30)	0.81 (0.70,0.93)	1928.97
Replication with OpenMX	0.00 (0.00,0.31)	0.19 (0.00,0.30)	0.81 (0.70,0.93)	1928.97
Replication with STATA	0.00 (0.00,1.00)	0.19 (0.10,0.30)	0.81 (0.79,0.84)	1928.97

OpenMX warns of uncertainty in estimates; STATA warns of estimate uncertainty.

Table 2.

Tobit estimates of absolute differences in trust scores for twin pairs

Variable	Coefficient	SE	t	P > |t|	Low CI	High CI
MZ twin	−0.46	0.57	−0.80	0.42	−1.58	0.67
Race	0.08	1.19	0.06	0.95	−2.26	2.41
Age	−0.01	0.02	−0.70	0.48	−0.05	0.02
College difference	2.53	0.74	3.42	0.00	1.07	3.99
Marriage difference	−0.08	0.60	−0.13	0.90	−1.25	1.10
Constant	3.05	1.49	2.05	0.04	0.12	5.98

Model significant at P < 0.05 level. About 1% of variance explained.

Table 3.

Tobit estimates of absolute differences in distrust scores for twin pairs

Variable	Coefficient	SE	t	P > |t|	Low CI	High CI
MZ twin	0.92	0.84	1.09	0.28	−0.73	2.57
Race	2.90	1.81	1.61	0.11	−0.66	6.46
Age	0.01	0.03	0.50	0.62	−0.04	0.07
College difference	2.66	1.07	2.48	0.01	0.55	4.77
Marriage difference	−0.54	0.87	−0.62	0.54	−2.25	1.17
Constant	−1.51	2.28	−0.66	0.51	−6.00	2.98

Model not significant at P < 0.05 level. Less than 1% of variance explained.

We applaud Reimann et al. (1) for gathering and disclosing valuable twin data. We conclude, however, that not much can be inferred from their data about the origins of trust or distrust.