Cognitive skills, non-cognitive skills, and family background: evidence from sibling correlations

This paper estimates sibling correlations in cognitive and non-cognitive skills to evaluate the importance of family background for skill formation. Based on a large representative German dataset including IQ test scores and measures of non-cognitive skills, a restricted maximum likelihood model indicates a strong relationship between family background and skill formation. Sibling correlations in non-cognitive skills range from 0.22 to 0.46; therefore, at least one-fifth of the variance in these skills results from shared sibling-related factors. Sibling correlations in cognitive skills are higher than 0.50; therefore, more than half of the inequality in cognition can be explained by shared family background. Comparing these findings with those in the intergenerational skill transmission literature suggests that intergenerational correlations capture only part of the influence of family on children’s cognitive and non-cognitive skills, as confirmed by decomposition analyses and in line with previous findings on educational and income mobility.

1. See, for example, Heckman et al. (2006) and Heineck and Anger (2010). An extensive overview can be found in Almlund et al. (2011).

2. These circumstances comprise both genetic endowment and environmental factors, such as parental income, social networks, or parenting style, and hence differ in their degree to which they can be targeted by policy makers to increase equality of opportunity.

3. It is hard to judge which specific value of family influence should be considered as “fair.” Which unfavorable environmental factors should be offset by social policies “is a value judgment that different societies may well make differently.” (Corak, 2013, p. 9).

4. Although economic research on non-cognitive skill formation is rather scarce, intergenerational correlations have been analyzed by psychologists for decades (e.g., Loehlin 2005). However, the data used in most psychological studies are based on a small number of observations or lack representativeness.

5. Björklund and Jäntti (2012) call this partial effect the ”tip of the iceberg.”

6. This includes shared family background and community factors. Among others, Solon et al. (2000), Page and Solon (2003), Leckie et al. (2010), and Nicoletti and Rabe (2013) and (Lindahl 2011) show that shared family factors are more important than shared neighborhood factors for education and earnings. Bügelmayer and Schnitzlein (2014) present results on German adolescents suggesting that although the influence of shared neighborhood factors are not negligible in Germany, shared family background is the predominant factor for education, cognitive ability, and physical and mental health outcomes. Thus, in the following sections, when we speak of shared family background, this discussion includes shared community factors.

7. For example, using brother correlations, Schnitzlein (2014) reports that approximately 45 % of the variance in permanent earnings can be attributed to shared family or neighborhood factors in the USA and Germany, whereas the corresponding estimate for Denmark is only 20 %.

8. Nicoletti and Rabe (2013) report sibling correlations on exam scores, which are similar in size to sibling correlations in cognitive skills but refer to educational achievement.

9. In our study, we cannot actually identify causal effects of the family on skill formation with the data at hand. Hence, any family influences discussed in this study relate to statistical correlations and not to causation.

10. Whereas around 50 % of non-cognitive skills are shaped by genetic factors (e.g., Krueger et al. 2008), it has been shown that genes are the predominant determinant of cognitive skills (e.g., Plomin et al. 1994; Toga and Thompson 2005). Nevertheless, there is also evidence from the economic literature that cognitive skills are shaped by environmental factors, such as educational activities in the family or parenting style (e.g., Sacerdote 2002; Plug and Vijverberg 2003; Ermisch 2008; Fiorini and Keane 2014). For a recent discussion on the role of genetic versus environment for non-cognitive skills, see Fletcher and Schurer (2015).

11. We use SOEPv29 (DOI: 10.5684/soep.v29). For more information, see http://www.diw.de/soep.

12. Although CAPIs are standard for newer SOEP subsamples, the initial subsamples are still interviewed using PAPI (paper and pencil interviewing).

13. The SOEP data provide different parental identifiers. In this study, we use the identifiers provided in the SOEP file BIOPAREN. These parental identifiers are mainly based on cohabitation at age 17 (or older if the respondent is older in the first interview). In the few cases, in which either the mother or the father are absent from the household, BIOPAREN provides a parental identifier from earlier waves, in which the missing parent was still present in the household. Although the SOEP also provides information on biological children for all women in the survey, information on the biological children of men has been recorded only since 2000. As our sample includes children primarily from the initial SOEP households, which were sampled before 2000, using the biological identifier for men would significantly reduce our sample size. However, we know that for approximately 95 % of the mother-child pairs in our sample, the social mother is also the biological mother. Thus, nearly all of the siblings studied share at least a biological mother. If genetics are an important factor, then considering social instead of biological parents would result in underestimating the estimated sibling correlations. In this case, our estimates could be considered to be a lower bound.

14. More specifically, singletons contribute to the identification of the family component (see Section 4 for details). In our sample, about two-thirds of the children are singletons.

15. We do not impose restrictions on the age difference of siblings within families. On average the age difference between siblings in our sample is 4.5 years. When restricting our analysis to families with age differences of 5 years or less (71 % of our sample), the estimated sibling correlations are very similar to those reported in Section 5 (available upon request).

16. The share of women in our sample is 48 %. Because there is no theoretical reason to expect differences between sons and daughters with respect to family background effects, we do not separate the analysis by gender.

17. Since performance in the word fluency test depends on the skill level in the language in which the test is administered in, we exclude all non-native Germans in the analysis of cognitive skills.

18. Lang et al. (2007) conduct reliability analyses and find test–retest coefficients of 0.7 for both the word fluency and symbol correspondence tests.

19. Using average test scores is expected to reduce the error-in-variable bias by diminishing the random component of measured test scores. Furthermore, average test scores could be interpreted as an extract of a general ability type, which captures both, coding speed and verbal fluency.

20. Our earnings measure is based on mothers’ and fathers’ average observed earnings (in 2007 euros) between 25 and 60 years of age in order to reduce measurement error resulting from transitory fluctuations. We include years with zero earnings and use (earnings + 1) in our calculations. On average, the earnings measure includes approximately 16 years of parental earnings information.

21. Because of the low number of parental observations with IQ test scores, we cannot include parental cognitive skill measures in the analysis. The effect of cognitive skills will to some extent be captured by parental education.

22. The displayed means of the skills (particularly those for crystallized intelligence) deviate slightly from zero, as our sample consists of (adult) children who rated some of their personality traits differently and performed better in the cognitive tests than the relatively older generations in the SOEP. This result can be partially explained by age-related cognitive decline and by the so-called Flynn effect, which indicates a rise in average cognitive ability test scores for the last three generations (Flynn 1994).

23. This is in line with findings of Cobb-Clark and Schurer (2012, 2013) who showed that personality traits and locus of control are relatively stable within 4-year windows for all adult age groups.

24. The correlation between mothers’ and fathers’ Locus of Control is 0.49. Parental correlations in reciprocity are 0.37 (positive reciprocity) and 0.39 (negative reciprocity), and for the Big Five the correlations are 0.31 (Openness), 0.29 (Conscientiousness), 0.12 (Extraversion), 0.29 (Agreeableness), 0.20 (Neuroticism) in the respective subsamples.

25. Moreover, because family background is identified based on siblings in our analysis, the question arises as to whether children with siblings and singletons have different cognitive and non-cognitive skills. However, apart from emotional stability and fluid intelligence, which seem slightly lower for children without (identified) siblings in our dataset, both personality traits and cognitive abilities appear to be fairly equal for all family types.

26. For the analysis of non-cognitive skills, we estimate the linear multilevel model as presented in Eq. 6 using all available observations from the survey years 2005 and 2009/2010.

27. Note that Anger (2012) does not report results for reciprocity.

28. We use the sum of the mother’s and father’s average individual labor earnings as defined above. Families above the median are labeled as high-income families.

29. Mothers with at least 12 years of education are defined as highly educated, including all mothers who have at least an intermediate secondary degree plus a vocational school degree.

30. See Table 8 for the decomposition results, when only the father’s or the mother’s characteristics are included. The separate decompositions yield similar results for the inclusion of the father’s and mother’s characteristics. However, including both parents’ characteristics clearly best explains the influence of shared family background on skill formation.

31. For example, we have information on whether an individual’s parents divorced during childhood or whether childhood was spent in a rural or an urban area. However, these factors may differ between—and hence would not be shared by—siblings of different ages.

32. However, due to differences in data availability and methods, we have to interpret any cross-country differences with caution.

33. In the USA and Germany, approximately 45 % of the variance in earnings can be attributed to family factors, whereas this share is only 20 % in Denmark based on brother correlations (Schnitzlein 2014). Cross-country differences in the importance of family background are also found for educational attainment. In Nordic countries, approximately 45 % of the variance in education can be attributed to shared family and neighborhood (Raaum et al. 2006; Lindahl 2011), whereas this share is more than 60 % in Germany (Schnitzlein 2014) and up to 70 % in the USA (Mazumder 2011).

34. As Mazumder (2008) uses a different measure of cognitive skills (AFQT test scores surveyed in the NLSY between 1978 and 1998), the results may not be directly comparable.

35. As shown in the last row of Table 6, the variance of the transitory component is of substantial size in all estimations.

36. In these cases, the model in Eq. 6 is estimated without the transitory component.

We would like to thank Anders Björklund, Markus Jäntti, Matthew Lindquist, Shelly Lundberg, Bhashkar Mazumder, and Catherine Weinberger; seminar participants of SOFI in Stockholm, UC Santa Barbara, ISER at the University of Essex, RWI Essen, the University of Hamburg, the University of Bath, the University of Bristol and The Danish National Centre for Social Research; and conference participants at the Annual Conference of the Scottish Economic Society 2013, SOLE 2013, ESPE 2013, IWAEE 2013, SMYE 2013, the 2013 Annual conference of the German Economic Association, and EALE 2013 for their useful comments and discussions. Moreover, we are grateful to the editor of this journal and to three anonymous referees for their valuable comments and helpful suggestions.

Authors and Affiliations

1. Research Department Education and Employment over the Lifecourse, IAB Nuremberg, Regensburger Str. 100, 90478, Nuremberg, Germany

   Silke Anger

2. Faculty of Social Sciences, Economics, and Business Administration, University of Bamberg, Feldkirchenstr. 21, 96052, Bamberg, Germany

   Silke Anger

3. Research Fellow, IZA Bonn, Schaumburg-Lippe-Str. 5-9, 53113, Bonn, Germany

   Silke Anger

4. Institute of Labour Economics, Leibniz University Hannover, Königsworther Platz 1, 30167, Hannover, Germany

   Daniel D. Schnitzlein

5. SOEP Department, DIW Berlin, Mohrenstrasse 58, 10117, Berlin, Germany

   Daniel D. Schnitzlein

Corresponding author

Correspondence to Daniel D. Schnitzlein.

Conflict of interests

The authors declare that they have no conflict of interest.

Responsible editor: Junsen Zhang

Appendix

Distribution of cognitive skills; full SOEP sample vs. sample with parent identifiers. Source: SOEPv29

Full size image

Distribution of non-cognitive skills; full SOEP sample vs. sample with parent identifiers. Source: SOEPv29

Full size image

Reprints and permissions