Abstract
Intergenerational educational mobility reflects a welfare state's ability to provide citizens with opportunities to climb the social ladder. Representing two distinct welfare state types, studies have contrasted mobility patterns in Scandinavia and the United States but have provided no consistent answer as to who achieves the highest level of intergenerational educational mobility. We conduct a meticulous examination of intergenerational educational mobility in Denmark, Norway, Sweden, and the United States for cohorts born between 1958 and 1987 using comparable operationalizations and methods and the best available data (administrative data in Scandinavia and eight surveys in the United States). We focus on methods that capture relative mobility. Across models, we find that inequality in Scandinavia is 20 percent to 30 percent lower than in the United States. A multiverse analysis, which can run a large number of models within a single framework, shows that our results are robust to alternative variable specifications.
Keywords
Intergenerational educational mobility captures the extent to which children's education is associated with that of their parents. As an indicator of the capacity of societies to provide equal opportunities for their citizens, it is a crucial component in research on inequality. Studies have long contrasted mobility patterns in Scandinavia (Denmark, Norway, and Sweden) with those in the United States because these cases exemplify different welfare state types. Such comparisons may seem straightforward, yet findings are mixed: Some studies report higher mobility rates in Scandinavia (Chevalier, Denny, and McMahon 2009; Esping-Andersen 2015; Hertz et al. 2008; OECD 2018), but others find mobility levels to be similar (Karlson and Landersø 2025; Landersø and Heckman 2017; Pfeffer 2008).
These mixed findings pose at least two challenges. First, some level of consensus on descriptive accounts of intergenerational mobility is necessary to increase our understanding of patterns of intergenerational inequality and inform hypotheses about the mechanisms driving these patterns. As such, descriptive accounts of an extensive array of topics (e.g., educational, occupational, income, wealth inequality) are not only informative in themselves but also lay the groundwork for causal analysis (Gerring 2012:733). Second, contrasting findings lead researchers to reach different conclusions about the role of the welfare state in providing opportunities for citizens. For example, Esping-Andersen (2006) finds a high level of intergenerational mobility in education in Scandinavia, attributing this to a “welfare state effect.” Low economic inequality, high-quality daycare, universal social security, and tuition-free education systems reduce the risk of children from all backgrounds not pursuing higher education (Esping-Andersen 2006, 2015). In contrast, Landersø and Heckman (2017:220) argue that the Scandinavian welfare state (using Denmark as a proxy), with its relatively high wage compression (low returns to education) and high levels of welfare benefits, provides insufficient incentives for citizens to invest in education, leading to a mobility pattern for both income and education that resembles that of the United States. These mixed findings surface in media coverage as either a showcase of the successful Scandinavian welfare state or signs of its failure, adding confusion to public debates on societal reforms and policy strategies.
To address these mixed findings, we perform a meticulous examination of intergenerational educational mobility—the statistical association between children's and parents’ education—in Denmark, Norway, Sweden, and the United States. We include all three Scandinavian countries to strengthen our analysis and highlight any potential differences within Scandinavia. Drawing on major approaches from sociology and economics, comparable operationalizations, and administrative data from Scandinavia alongside the best available U.S. survey data, we highlight the implications of making different methodological choices. Our overall finding is that intergenerational educational inequality is 20 percent to 30 percent higher in the United States than in Scandinavia. We perform a multiverse analysis to gauge how sensitive our results are to the choice of data sources, sample restrictions, and alternative variable specifications (Engzell and Mood 2023; Steegen et al. 2016). Our multiverse analyses show that our results are robust to alternative variable specifications.
Measuring Educational Mobility
Intergenerational mobility research often distinguishes between absolute and relative mobility (Breen and Müller 2020; Erikson and Goldthorpe 1992). Absolute educational mobility captures whether children achieve more or less education than their parents. It is an intuitive and important measure of, for example, whether a general rise in living standards has also led to a rise in educational attainment. However, absolute mobility reflects both structural changes, such as extended compulsory schooling or educational expansion, and relative changes in the likelihood of achieving, for instance, higher education for different social groups. In contrast, relative mobility, the primary focus in this article, is margin-insensitive; it accounts for structural changes in the educational distribution (Erikson and Goldthorpe 1992; Xie 1992).
The methodological approach applied in educational mobility research will determine the ability to account for marginal changes. Figure 1 illustrates this in a schematic outline of our set of choices. The first choice concerns the conceptualization of education as either a continuous (e.g., years of education) or categorical (e.g., degrees) variable. The second choice pertains to the units of the education variable, which can be measured on an interval scale (e.g., years of education), transformed from an interval scale (e.g., standardized years of education), or using a nominal scale (e.g., degrees). These choices result in models with different levels of sensitivity to marginal changes: Ordinary least squares (OLS) models are the most margin-sensitive, Pearson correlations less so, and rank correlations and odds ratio–based measures are margin-insensitive. 1
Measuring intergenerational educational mobility.
The OLS linear regression model is the first way of analyzing the association between children's and parents’ education. OLS models assume that qualitatively different types of education can be placed on a monotonic scale, implying that a one-unit increase carries the same meaning across the entire distribution (see Blanden 2013). In addition, model estimates are directly scaled on years of education and hence sensitive to distributional differences between generations (Jäntti and Jenkins 2015:838), such as changes driven by the timing of educational expansion or educational reforms (Rauscher 2016). A straightforward example illustrates the mechanism. In Denmark, the 1972 educational reform increased compulsory schooling from seven to nine years. As a result, right after the reform, all children mechanically obtained more years of education than their parents. Because the parents of later cohorts obtained at least nine years of education themselves, fewer children achieved more education than their parents. All else being equal, in an OLS model, this will result in higher and higher beta coefficients because truncating the variable range affects the variance (for an example, see S5 in the online supplement).
The second method in Figure 1 is the Pearson correlation coefficient (equaling a linear regression with standardized variables; transformed to have a mean of 0 and a variance of 1). Correlations are less sensitive to marginal changes because they equalize the variance of parents’ and children's educational distribution, providing a measure that is not purely mechanically affected by distributional differences between generations or countries (Björklund and Jäntti 2011; Blanden, Doepke, and Stuhler 2023). The third method is also a correlational measure, but here, education is transformed into ranks, as in Spearman's rank correlation or the quantile rank correlation. The Spearman's rank correlation coefficient is a correlation of ranked variables (it equals the regression coefficient of ranked y on ranked x). Quantile rank correlations are known from intergenerational relative income inequality studies, where income is ranked into, for example, deciles (Chetty et al. 2014). For a granular variable like income, Spearman and quantile correlations will yield the same results, abstracting fully from marginal changes (Jäntti and Jenkins 2015). However, unlike income, education comprises discrete categories with a lumpy distribution (see S4 in the online supplement), posing challenges to the application of quantile rank correlations. Researchers have used various techniques to address the “lumpiness” of education (for an overview, see Shavit and Park 2016). Following Narayan et al. (2018), we approximate sufficient granularity for quantile rank correlations by randomly allocating individuals who cannot be placed in a single decile to the nearest deciles.
The fourth method encompasses odds ratio–based models, where education is measured using nominal categories that carry distinct values (e.g., diplomas, certificates, degrees). The odds ratio expresses, for example, the odds of obtaining higher education relative to no higher education among children of highly educated parents relative to children whose parents have no higher education. This measure captures relative mobility net of structural mobility because it is insensitive to changes in the marginal distribution (Hout 1983). Odds ratio–based models have been widely used in intergenerational mobility research since Goodman’s (1969) influential work. 2 We focus on the uniform difference (UniDiff) parameter, providing a single summary of the level of educational mobility.
Previous Research
We review comparative research on intergenerational educational mobility that includes the United States and at least one Scandinavian country. Table 1 provides a systematic overview of the research. These studies vary in at least three respects: (1) They use different data sources, (2) their operationalization of education varies, and (3) they apply different methods, with some studies favoring correlations and OLS regression models, and others preferring odds ratio–based methods. We begin by reviewing studies that use correlations and OLS regression models, followed by those that apply odds ratio–based models.
Studies on Intergenerational Educational Mobility That Include the United States and at Least One Scandinavian Country.
Note: NELS_88 = National Education Longitudinal Study of 1988; PIIAC = Programme for the International Assessment of Adult Competencies; ESS = European Social Survey; PSID = Panel Study of Income Dynamics; ISSP = International Social Survey Programme; NLSY = National Longitudinal Survey of Youth; UniDiff = uniform difference; OLS = ordinary least squares; IALS = International Adult Literacy Survey; GSS = General Social Survey.
Replication package available.
One frequently cited study is Hertz et al. (2008), which measures intergenerational educational mobility across 42 countries. Their preferred metric, the Pearson correlation, indicates that Denmark has the highest mobility, followed by Norway, Sweden, and the United States. However, the results differ somewhat when using OLS regressions, with Norway showing the highest mobility, followed by the United States, Denmark, and Sweden. Liu and Ding (2020) revisit Hertz and colleagues’ analysis for 20 OECD countries, finding that mobility is higher in Scandinavia than in the United States based on both correlations and regressions. Chevalier et al. (2009) investigate mobility among 26-, 44-, and 65-year-olds using Spearman's rank correlation, revealing similar persistence in Denmark and the United States for the oldest group but greater mobility in Scandinavia than in the United States for the youngest group. An OECD (2018:257) report, “A Broken Social Elevator?,” applies correlations and regressions and finds slightly more mobility in Scandinavia than in the United States for individuals ages 30 to 55. Using OLS regression, Karlson and Landersø (2025:202) observe that educational mobility in Denmark has sharply declined since the 1950s, reaching levels comparable to the United States for the 1983 to 1988 cohorts. In summary, correlations produce a fairly consistent picture, and OLS regression coefficients show mixed results regarding mobility across these four countries.
Turning to studies that use odds ratio–based models, Hout and Dohan (1996) analyze mobility across four cohorts (1905–1919, 1920–1934, 1935–1949, and 1950–1964), reporting less inequality in Sweden than in the United States for the most recent cohorts. Using a UniDiff model, Pfeffer (2008) examines adults born between 1929 and 1972 (ages 26–65) from 20 industrialized countries. Pfeffer finds the highest relative mobility in Denmark, followed by the United States and Sweden, with Norway showing the lowest mobility estimates. Esping-Andersen (2015) finds that for individuals born in the 1970s, the odds of attaining upper-secondary education for children of low-educated fathers are 3 to 5 times higher in Scandinavia than in the United States. In contrast, Reizel (2011), who examines college attainment and parental background in Norway and the United States, finds only moderate differences between the two countries. Another OECD (2014:93) report, Education at a Glance, reports ratios of the likelihood of participating in tertiary education for individuals ages 20 to 34 whose parents have upper-secondary or tertiary education relative to those whose parents lack upper-secondary education. They find more inequality in the United States than in Scandinavian countries. Landersø and Heckman (2017) compare transition matrices for Denmark and the United States, finding similar educational transition levels by parents’ education in both countries. Using the same data sources, Andrade and Thomsen (2018, 2021) turn to logistic regression models and find more inequality in the United States than in Denmark. Finally, the World Bank (Narayan et al. 2018) uses a positional method, comparing the share of individuals from the 1980s cohort who were born into the bottom half of the national distribution of educational attainment and reached the top quartile. Denmark and Sweden are among the countries with the highest share of upwardly mobile individuals, and the United States is among the bottom 50 economies (Narayan et al. 2018:107).
We find no clear consensus regarding mobility patterns in the surveyed literature. Of the five studies using OLS regressions or correlations, correlational estimates tend to show higher mobility levels in Scandinavia than in the United States, and findings from regressions are mixed. Of the eight studies using categorical/positional measures (two) or odds ratio–based approaches (six), three report similar levels of inequality, and five studies find lower levels in Scandinavia compared to the United States.
Education Systems in The United States and Scandinavia
The educational systems in the four countries vary in several ways. Most notably, the United States has a less standardized system compared to Scandinavia. Another difference is the near-universal high school attendance in the United States, where schooling is mandatory until ages 16 to 18, whereas in Scandinavia, compulsory education usually ends after Grade 9 or 10. The systems also differ financially: Education is tuition-free throughout Scandinavia, and students 18 or older receive a monthly government grant and options for a low-interest loan. Figure 2 provides a stylized overview of each country's educational system.
Stylized picture of education systems in Denmark, Norway, Sweden, and the United States (as of 2020).
Denmark
Danish children start with 10 years of compulsory education (Grades 0–9, ages 6–16). Afterward, they can choose between two upper-secondary paths: the higher education preparatory academic track, “gymnasium” (3 years), or the vocational education and training track (3–4 years), which qualifies students for a wide range of skilled trades, such as construction, electrical work, and automotive repair. Students with a gymnasium diploma can continue into higher education, pursuing one of three options: (1) an academy professional degree (2- to 3-year programs for private-sector jobs), (2) a university college degree (3- to 4-year applied bachelor's programs, mainly training teachers, nurses, and social workers), or (3) a university degree in various fields (a 3-year bachelor's, a 2-year master's, or a 3-year doctoral degree [i.e., PhD]). The majority of students earning a bachelor's degree go on to complete a master's degree. Access to many higher education programs, especially at the university level, is competitive and primarily based on the student's high school grade point average (GPA).
Norway
In Norway, children spend 10 years in compulsory schooling (Grades 1–10, ages 6–16). After completing basic education, all students are entitled to 3 to 4 years of upper-secondary education, where they can choose between vocational or academic tracks. Students with a vocational diploma must complete an additional year of academic coursework to qualify for higher education. Admission to the academic track often depends on students’ GPAs from lower-secondary school. Completing this track allows students to apply to universities and university colleges, where admission is based on their grades. As in Denmark, students with the highest GPAs are admitted when there are more applicants than available spots. Typically, students finish a 3-year bachelor's program at universities and colleges, with many continuing to the master's level. PhD students are usually offered 4-year contracts, which include 1 year of teaching duties.
Sweden
Most Swedish children start in introductory school at age 6 (Grade 0). Compulsory education lasts until Grade 9. Upper-secondary education offers two tracks: academic and vocational. In the early 1990s, educational reforms aimed to reduce differences between these tracks. However, gaps still exist, and requirements for tertiary education often include courses not available in vocational programs. Similar to Norway and Denmark, completing the academic track allows students to apply for higher education, with GPA being the main factor for admission. The most common degree is a bachelor's (3 years). After earning a bachelor's degree, students can pursue either a 1-year or 2-year master's degree. Ultimately, PhD students can choose between a licentiate degree (2 years) or a doctorate (4 years).
The United States
Compared to Scandinavia, education in the United States is more diverse. Most U.S. children begin elementary education with kindergarten at age 5 or 6, continuing through Grade 5 in primary school, and then moving to middle or junior high school, typically Grades 6 to 8. Because U.S. children must be enrolled until age 16 to 18, depending on the state, nearly all students continue through high school (i.e., Grades 9–12). The U.S. system does not separate high school into academic and vocational tracks. Instead, students typically apply to higher education institutions before choosing a specific degree program. U.S. colleges and universities assess applicants based on multiple criteria, such as standardized test scores (SAT/ACT), GPA, extracurricular activities, motivational letters, family legacy, athletic achievements, and diversity factors. Tuition costs vary widely between schools and depending on residency status. After acceptance at a junior college, community college, or university, high school graduates can pursue an undergraduate degree, usually a 2-year associate's or a 4-year bachelor's. Postgraduate studies generally require a 4-year bachelor's degree and include master's degrees (typically 2 years), doctoral degrees (usually 4 years), or professional postgraduate programs, such as medical or law school.
Scandinavian and U.S. Educational Systems Compared
Due to substantial differences between the four educational systems, comparisons of educational levels across the four countries should be made with caution. For example, qualifications for working in skilled manual or service occupations are typically obtained at the high school level (through a vocational track) in Scandinavia. In contrast, in the United States, vocational qualifications may be acquired in high school or community colleges. For example, an auto mechanic's skills may be acquired at the (community) college level in the United States but at the high school level in Scandinavia. Another significant difference is at the postgraduate level. Among Danes graduating with a university bachelor's degree (approximately half of all bachelor's degree holders), most continue to a master's degree program because they are legally entitled to do so without applying. Denmark has never had a de facto labor market for holders of university bachelor's degrees. 3 In contrast, in the United States (and in Sweden), most college students graduate with a bachelor's degree. These differences render it challenging to translate educational categories into years of education and vice versa.
Data and Coding
For Scandinavia, we use administrative data from Statistics Denmark, Statistics Norway, and Statistics Sweden. Administrative data cover the entire population in each country and have historically been collected to aid governments in tasks such as tax collection and social benefit distribution. Data typically date back to the early 1980s and cover a wide range of yearly information on demographics, education, employment, and income. Each citizen has a unique identification number linked to their parents. For the United States, we include surveys containing intergenerational education information on one or more of our birth cohorts of interest. Table 2 provides an overview of the data sources.
Data Sets.
Note: ADDH = National Longitudinal Study of Adolescent to Adult Health; GSS = General Social Survey; IALS = International Adult Literacy Survey; NLSY79 = the National Longitudinal Survey of Youth 1979; NLSY97 = National Longitudinal Survey of Youth 1997; NSFH = National Survey of Families and Households; PSID = Panel Study of Income Dynamics; SIPP = Survey of Income and Program Participation.
N captures persons with information on education for both them and their parents.
At the time of writing, administrative data for Norway and Sweden were only available up to 1984 and 1985, respectively.
The U.S. surveys include the National Longitudinal Study of Adolescent to Adult Health (ADDH), the General Social Survey (GSS), the International Adult Literacy Survey (IALS), the National Longitudinal Survey of Youth 1979 (NLSY79), the National Longitudinal Survey of Youth 1997 (NLSY97), the National Survey of Families and Households (NFHS), the Panel Study of Income Dynamics (PSID), and the Survey of Income and Program Participation (SIPP). We constructed cross-sectional samples for the U.S. longitudinal surveys following the individual survey guidelines where available (see replication package). The longitudinal surveys are either panel data from a representative sample of the entire population (PSID, SIPP, NFHS, ADDH) or panel studies that focus on specific cohorts of young adults (NLSY79, NLSY97). The GSS is the only repeated cross-sectional survey that spans the entire period we examine; we thus pay special attention to the GSS in our analysis. We used official probability and frequency weights. When probability weights were unavailable, we calculated them from frequency weights. For the pooled data, we constructed normalized probability weights (for details on surveys and weights, see S1 in the online supplement). We do not report confidence intervals for the Scandinavian countries because we are using the full population.
Comparing administrative data and survey data is not straightforward. Survey data may be subject to measurement errors, including recall bias, attrition, nonresponse, and sample selection bias (Engzell and Jonsson 2015). Weights may address some of these challenges, but weights may disproportionately represent self-selected individuals and hence not accurately account for misrepresented people. To the extent that underprivileged groups have lower response rates and more educational immobility, survey-based data likely overestimate intergenerational mobility (for the case of PSID, see Schoeni and Wiemers 2015). This alone can make mobility seem higher in the United States than in Scandinavia. S19 in the online supplement indicates that survey data underestimate inequality compared to administrative data in Scandinavia. Administrative data tackle most of these problems. However, the representation of immigrants may be incomplete because background information on immigrants’ parents is often missing, and if available, it may be imputed based on surveys. 4 To assess how excluding immigrants and descendants affects mobility estimates, S10, S17, and S18 in the online supplement show mobility estimates for Denmark with and without immigrants and descendants. We find only minimal estimate differences.
Education as a Continuous Variable
We use years of education as our continuous measure. For parents, we use information on the parent with the longest education (models using the mean of parents’ education are in S11 in the online supplement). 5 For Scandinavia, years of education are calculated as the cumulative stipulated time to complete a degree (the minimum time required). For example, in Denmark, a PhD equals 21 years of education (10 years of compulsory school + 3 years of high school [academic track] + 3 years for a bachelor's degree + 2 years for a master's degree + 3 years for a PhD). 6 For the United States, we use the highest grade attended (years in the education system) as reported in the individual surveys (GSS, PSID, NSFH, NLSY79/97), or if not provided (as in SIPP and ADDH), we infer grades from diplomas and degrees (as stipulated time to degree). We use the stipulated time to degree to measure parental grade in SIPP, PSID, IALS, and ADDH, which only ask for parental educational degree (for details, see S1 in the online supplement). Grade attended differs qualitatively from the Scandinavian way of measuring years of education. In contrast to Scandinavia, U.S. years of education are (mostly) self-reported and intended to capture actual years spent in the education system regardless of whether they lead to a degree. The different ways of obtaining information on years of education are significant noncomparability sources and, in themselves, a strong argument for preferring categorical (where degrees are compared to degrees) over continuous measures.
Education as a Categorical Variable
We use the highest level of education obtained as our categorical measure. For the United States, we use information on high school graduation and degrees reported in the individual surveys. For children, degrees in the NSFH, NLSY79, and NLSY97 are derived from grades attended. Parental degrees are provided in all surveys, except NSFH, where they are derived from the information on parental grades. We apply a five-level categorical variable, representing the highest level of disaggregation permitted by the combined U.S. surveys. These five levels correspond to International Standard Classification of Education (ISCED) 2011 levels 1+2, 3, 4+5, 6, and 7+8 (for our ISCED classification, see S2 in the online supplement) and cover (1) no high school diploma/no GED, (2) high school diploma/GED, (3) associate degree/junior college degree, (4) bachelor's degree, and (5) postgraduate degrees (master’s degree, professional degrees, or PhD). For parents, we use the highest degree obtained by either parent in our categorical models.
Cohorts
We use the widest birth cohort range possible, split into five groups: (1) 1958–1963, (2) 1964–1969, (3) 1970–1975, (4) 1976–1981, and (5) 1982–1987. For Scandinavia, we use the information on the highest level of education obtained at age 33. For the United States, we increase the age range to 28 to 38 due to the small sample size of U.S. surveys relative to Scandinavian administrative data. 7 To reduce bias introduced by different age points, we exploit the panel structure of several of our data sets and identify for each individual with multiple interviews the interview closest to age 33. Analyses with the Scandinavian cohorts coded with an age range of 28 to 38, similar to the United States, yield results similar to those presented in this article (see S7 in the online supplement). For the Scandinavian countries, we measure parents’ education when their children are 16 years old in Norway, 23 years old in Denmark, and 25 years old in Sweden. 8 S8 in the online supplement shows that mobility estimates are nearly identical whether we measure parents’ education when their child is 5, 16, 23, 25, or 30.
Analysis
We apply a straightforward setup to analyze the association between parents’ and children's education. We do this to provide a robust, descriptive account of patterns of educational mobility in Scandinavia and the United States using a range of alternative specifications. It is important to note that U.S. estimates can vary substantially depending on the survey chosen. To illustrate how much effect sizes vary across U.S. surveys, we present estimates for the individual surveys in addition to the pooled U.S. data in select figures. Our setup enables us to construct large three-way contingency tables for each Scandinavian country, sufficiently anonymized and aggregated to allow for public use. A link at the end of the article provides access to a complete replication package.
We divide our analysis into five parts. First, we outline educational expansion as a background for interpreting our findings. Second, we examine educational mobility using OLS models. Third, we apply correlational measures, and fourth, we use odds ratio–based models. Finally, we perform a multiverse analysis, running a large number of models within a single framework to assess the sensitivity of our results to alternative variable specifications.
The Educational Expansion
As a backdrop for our analysis, Figure 3 offers a brief overview of educational expansion in the four countries over the 30-year period studied. For the United States, we include both our combined U.S. survey data and U.S. census data for comparison. All cohorts benefited from increasing educational levels across these countries, with an ever-growing proportion of each cohort attaining postsecondary education. The share of higher education graduates in the latest cohort is fairly similar across nations: Sweden saw an increase from 22 percent to 47 percent, Denmark from 25 percent to 48 percent, Norway from 25 percent to 51 percent, and the United States from 34 percent to 52 percent (census data).
Educational expansion in Denmark, Norway, Sweden, and the United States; completed education for individuals born 1958 to 1987.
In the latest cohorts, the percentage of people without an upper-secondary education is about twice as high in Scandinavia as in the United States. As mentioned earlier, education is compulsory in the United States up to age 16 to 18, which may partly explain the lower percentage of individuals without upper-secondary education in the United States.
Continuous Interval Measures (OLS Linear Regression Models)
Figure 4 displays beta coefficients from cohort- and country-specific OLS models that regress children's years of schooling on parents’ years of schooling. The coefficient denotes the increase in children's schooling years when parents gain one additional year of education, with a higher coefficient reflecting more persistence (less mobility). For transparency, we show weighted estimates for the individual U.S. surveys and the combined U.S. data in the right graph.
The association between educational origin and destination for Denmark, Norway, Sweden, and the United States; beta coefficients from ordinary least squares regression models.
Norway has seen a slight decrease in immobility (0.40 to 0.37), whereas Denmark and Sweden have seen increases, moving from 0.21 to 0.33 and 0.20 to 0.37, respectively (the Danish increase is similar to the one reported by Karlson and Landersø 2025:185). The U.S. trend is relatively constant. We note that U.S. estimates based on degrees obtained (similar to the Scandinavian way of measuring years of education) yield estimates at a higher level than when using grades, moving from 0.46 to 0.47. Estimates differ substantially by individual U.S. surveys, ranging from 0.30 to 0.52 for the 1958–1963 cohort and 0.29 to 0.48 for the 1982–1987 cohorts. In the later cohorts, the PSID estimates are much lower than the GSS, NLSY97, and ADDH estimates. 9 The GSS, the only repeated cross-sectional survey that covers all five cohorts in our data, exhibits a pattern similar to that of the pooled U.S. data (grades). We now turn to correlations that transform (standardize or rank) years of education, thereby adjusting for differences in variance in education between parents and children (Black and Devereux 2011:1490; Jäntti and Jenkins 2015:908).
Continuous Transformed Measures (Correlations)
The Spearman correlation differs from the Pearson correlation in measuring the association between ranked years of education. Because of this, Jäntti and Jenkins (2015:908) consider it “arguably the preferred scalar index of persistence (as it most clearly abstracts from differences in marginal distributions).” They further note that the wide use of beta coefficients and Pearson correlations (r) is “surprising because it is the positional mobility concept that has been of the greatest interest in this context, and yet Beta and r reflect structural as well as exchange mobility.” For this reason and because Pearson and Spearman correlations yield very similar results, we relegate the Pearson estimates to S11 in the online supplement and present the Spearman rank correlation in Figure 5.
The association between educational origin and destination for Denmark, Norway, Sweden, and the United States; spearman rank correlations.
As the right graph in Figure 5 shows, we observe much less survey variation than found with the unstandardized linear regression coefficients in the previous section (Figure 4). For the United States, estimates decrease slightly across cohorts, indicating somewhat more mobility (0.49 to 0.46). Estimates for the United States are nearly identical whether based on “theoretical” years of education (degrees obtained) or years spent in the education system (highest grade attended). The GSS estimates move from 0.47 to 0.43. In the latest period, the least mobile country, the United States, is at 0.46, and the most mobile countries, Denmark and Sweden, are at 0.35. Norway is the least mobile of the Scandinavian countries. Unlike with the OLS regression coefficients, Sweden, Denmark, and to a lesser extent, Norway have substantially lower correlation estimates than the United States. 10
For a highly granular variable such as income, Spearman and quantile rank correlations yield similar results, both focusing on positional change. Researchers may, however, prefer the quantile measure if they wish to examine, for example, the likelihood of individuals in the bottom quantiles reaching the top quantiles of a distribution. For education, these methods, in principle, circumvent the problem of comparing educational categories that are nominally similar but qualitatively different across countries and cohorts. Unfortunately, the educational distribution is not granular. To deal with the lack of granularity, we apply Narayan et al.’s (2018) approach, which allows us to calculate quantile correlations and odds ratios using a positional metric. Calculating odds ratios from quantiles may be preferable to using nominally similar groups (e.g., parents with no upper-secondary education) because such groups can differ greatly in composition across cohorts and countries. For example, the group of unskilled parents may be more negatively selected in later cohorts.
Our quantile correlations provide estimates similar to the Spearman correlations (see S12 in the online supplement). For odds ratios calculated from quantiles, S13 in the online supplement presents the odds ratios of being in the top versus bottom of the child educational distribution for children of parents in the top versus bottom of the parent educational distribution. Whether using tertiles, quartiles, or quintiles, the odds ratios are 2 to 3 times higher in the United States than in Scandinavia.
Categorical Measures (Odds Ratio–Based Models)
We present the results of our odds ratio–based models using a categorical variable for completed education with five levels: (1) no high school/GED, (2) high school/GED, (3) associate's degree/junior college, (4) bachelor's degree, and (5) postgraduate degrees. We focus on comprehensive measures that capture relative mobility across all destinations in a single estimate, relegating odds ratios for obtaining specific educational destinations to S14 in the online supplement. We use the UniDiff parameter from a UniDiff (or log-multiplicative layer effects) model, which summarizes the multiplicative difference between a given pattern of odds ratios and a reference set, estimated in a model that accounts for changes in the marginal distribution (Erikson and Goldthorpe 1992; Xie 1992).
Figure 6 shows results for the four countries in the first and last cohorts, with the first cohort in the United States as a reference. Because inequality may differ depending on the level of aggregation, we include nine different levels of educational aggregation (for our ISCED classifications, see S2 in the online supplement). The estimates represent the difference in educational inequality for each country cohort relative to the reference category of Americans born between 1958 and 1963. Values larger than 1 indicate more inequality (less relative mobility), and estimates below 1 indicate less inequality (more relative mobility).
UniDiff parameters for different categorizations of education, Denmark, Sweden, Norway, and the United States; first and latest cohorts shown.
We observe four important patterns in Figure 6. First, inequality in the United States has declined, but the picture for the Scandinavian countries is more mixed (an increase for Denmark, a decrease for Norway, and stability for Sweden). Second, the more we disaggregate education, the larger the differences in inequality between the United States and Scandinavia become (i.e., point estimate distances between the United States and Scandinavia are larger in the bottom-right five-level graph than in the top-left two-level graph). Third, and in line with our earlier results, relative to the United States, Scandinavian countries display fairly similar levels of inequality, with Sweden being the most mobile, followed by Denmark and Norway. Fourth, in most levels of educational disaggregation, the most recent cohort in Scandinavia is more educationally mobile than the U.S. cohort. The only exception is the top-left panel, which compares high school versus higher education, where we find similar levels of inequality between Scandinavia and the United States (noting that high school attendance is more universal in the United States than in Scandinavia).
Looking at the most detailed classification of education into five categories (Figure 6, bottom-right panel), parameter differences between the United States and Scandinavia in the last cohort are 0.29 for Sweden, 0.23 for Denmark, and 0.19 for Norway. The difference between the least and most mobile Scandinavian countries, Norway and Sweden, is 0.10, whereas the difference between the United States and Norway is 0.19. In other words, the difference between the United States and the least mobile Scandinavian country is about 2 times larger than the largest inter-Scandinavian difference. These parameter differences translate into inequality being 26 percent lower in Denmark, 21 percent lower in Norway, and 33 percent lower in Sweden than in the United States. In summary, although inequality has decreased in the United States over the period, the most detailed level of educational classification reveals that inequality remains significantly higher in the United States than in Denmark, Sweden, and Norway.
The UniDiff parameter cannot show the absolute levels of inequality because models are calculated relative to a reference set that is set to 1. Additionally, the UniDiff model assigns equal weight to each origin-destination pair in contingency tables, making it sensitive to small cell sizes (Hout and Guest 2013; Xie and Killewald 2013; Zhou 2015). Building on Altham (1970) and Goodman (1996), Bouchet-Valat (2022) proposed a normalized version of the Altham statistic, called the intrinsic association coefficient (IAC), which provides a more straightforward interpretation than Altham and UniDiff parameters (for recent applications, see Andrade and Thomsen 2021; Berger et al. 2023; Breen and In 2023).
The IAC offers several advantages. First, in its normalized form, the IAC can be interpreted similarly to a correlation, ranging from 0 (no association) to 1 (complete association). Second, whereas the Unidiff parameter is relative to a reference, the IAC also provides estimates for the reference itself, enabling assessment of both the trend (as in the UniDiff case) and the overall level of inequality. Third, each pairwise comparison in a contingency table can be weighted by the marginal proportions of the corresponding rows and columns, which may be preferable to the UniDiff models’ method of assigning equal weight to each comparison. The IAC can be computed either nonparametrically (equivalent to the Altham index) or from models fitted to the data (e.g., the UniDiff model).
Figure 7 shows the nonparametric and normalized IAC estimates for the most disaggregated five-level educational classification. The IAC estimates, based on a categorical measure of education, reflect the overall patterns observed in the rank correlations in Figure 5, with the important exception that we do not observe the same rise in immobility for Denmark and Norway. The IAC estimates are closer to the UniDiff trends in Figure 6. The United States (0.49 to 0.47), Norway (0.41 to 0.37), and Sweden (0.35 to 0.32) have seen some decline in intergenerational inequality, whereas Denmark has experienced a very slight increase in inequality (0.33 to 0.34). In the latest period, the United States is the most unequal country, with an IAC at 0.47, with Denmark, Norway, and Sweden at 0.34, 0.37, and 0.32, respectively. S16, S17, and S18 in the online supplement contain alternative specifications for IAC and Unidiff, all of which show the same pattern as described here.
Intrinsic association coefficient estimates for Denmark, Sweden, Norway, and the United States; five-level educational classification.
Summary
We conclude our analyses by briefly recapping our results, summarizing the different model estimates for the latest cohorts (1982–1987) in Table 3. The first four rows present results from models based on education as a continuous measure, and the last two rows present results from models based on categorical measures. Table 3 paints a uniform picture of intergenerational educational mobility in Scandinavia and the United States. The only anomaly is the OLS model (which factors in structural changes to a higher extent), where estimates for the four countries are relatively close. However, showcasing how sensitive OLS estimates are to alternative specifications, if U.S. years of education are coded from degrees (as in Scandinavia), the United States is substantially less mobile than the Scandinavian countries.
Summary of Estimates: Latest Cohorts (1982–1987).
Note: OLS = ordinary least squares; UniDiff = uniform difference; IAC = intrinsic association coefficient.
All other models show higher intergenerational educational mobility in Scandinavia than in the United States, with inter-Scandinavian mobility being somewhat higher for Denmark and Sweden than for Norway. The IAC estimates, which can be interpreted as a correlation, show that inequality is 20 percent to 30 percent lower in Scandinavia than in the United States. For the Unidiff and IAC estimates, the gap between the United States and the least mobile Scandinavian country (Norway) is 2 to 3 times larger than the biggest inter-Scandinavian difference.
How Robust Are Our Estimates to Alternative Variable Specifications?
To further evaluate the robustness of our results to alternative specifications, we turn to the multiverse approach (Engzell and Mood 2023; Steegen et al. 2016). Researchers select samples and code variables in different ways when analyzing mobility, for example, by choosing the mean or maximum of parents’ years of education (see Thaning and Hällsten 2020), selecting different cohorts and age groups, or including or excluding immigrants. Such decisions reflect considerable researcher discretion and involve multiple “forking-path” choices. The multiverse approach exposes the implications of these choices within a single framework, providing us with the full range of coefficient estimate sizes. The multiverse approach is therefore a valuable tool for testing the robustness of our findings. Our multiverse framework includes alternative specifications that we consider theoretically and conceptually justified. For example, it is hard to find any substantive reason for analyzing intergenerational educational mobility for extreme age ranges, such as 25- to 75-year-olds, because this would conflate cohort and period effects. Scholars do, however, debate the inclusion of immigrants in mobility analyses because immigrants may have been educated outside the focal country and thus not represent educational mobility within the national education system. S3 in the online supplement lists all the different specifications included in our multiverse framework.
An intuitive way to present the results of a multiverse analysis is through density plots for each country and method (Engzell and Mood 2023; Strömberg and Engzell 2025). These plots offer an overview of the range of coefficient estimates and the frequency (density) of these estimates. To provide such an overview, Figure 8 displays eight panels: four illustrating the estimate ranges for OLS models and Pearson, Spearman, and quantile rank correlations; two showing the estimate ranges for our UniDiff model; and two presenting the estimate ranges for our IAC calculations.
Multiverse analysis.
The x-axis in the graphs in Figure 8 represents coefficient estimates, and the y-axis indicates the frequency with which these estimates are returned in models included in the multiverse analysis. As expected, the OLS model is most sensitive to alternative variable specifications. The extreme case here is the United States, where estimates, depending on specification choices, range from 0.3 to 0.62. The Pearson, Spearman, and quantile rank measures are much less sensitive to alternative variable specifications, and the median estimates are fairly similar. For Spearman, the median is 0.44 for the United States, 0.37 for Denmark, 0.35 for Sweden, and 0.40 for Norway, medians close to the estimate summary reported in Table 3. For the Spearman and quantile rank measures, estimate sizes for Denmark and Sweden are lower than for the United States across almost the entire range of alternative specifications. Next, the UniDiff graphs show the first (Figure 8, left) and last (Figure 8, right) period. Because the oldest cohort in the United States is the reference (1 in the left graph of Figure 8), the U.S. estimate ranges are less dispersed than those in Scandinavia in the right graph. Looking at the medians in the right graph, we find median estimates around 0.63 to 0.70 for Scandinavian countries and 0.83 for the United States. Finally, IAC estimates for Scandinavia and the United States are far apart, with a median of 0.47 for the United States, 0.34 for Denmark, 0.36 for Norway, and 0.32 for Sweden, similar to the IAC estimates reported in Table 3. In summary, the multiverse analysis supports our main findings. Regardless of variable specification, we observe a general pattern of higher educational mobility in Scandinavia compared to the United States.
Conclusions
Intergenerational educational mobility gauges the extent to which children's educational attainment is associated with that of their parents. Scandinavia and the United States, representing two distinct welfare systems, have often been central in country comparisons. Yet research has not provided a consistent answer as to whether intergenerational educational mobility in Scandinavia differs substantially from that in the United States. We have provided robust evidence for higher mobility in Scandinavia than in the United States. The exception is estimates from the OLS model—which also reflect structural changes—showing smaller differences between the countries. However, as Hout (2015: 28) notes, the fact that children receive more education than their parents may be due to growth, a marginal change, whereas genuine improvements in opportunity or fairness are captured by accounting for marginal change. For measures that more effectively capture the level of educational mobility net of marginal changes, the difference between Scandinavia and the United States is substantial. For example, Spearman correlations and IAC estimates, which can be interpreted similarly to correlations (Bouchet-Valat 2022), show that inequality in educational mobility is 20 percent to 30 percent lower in Scandinavia than in the United States. Because correlations capture how closely parents’ and children's educational attainment are linked, we can say that intergenerational transmission of education is 20 percent to 30 percent stickier in the United States than in Scandinavia. The IAC estimates further show that inequality decreased by 4 percent in the United States (from 0.49 to 0.47), 10 percent in Norway (from 0.41 to 0.37), and 9 percent in Sweden (from 0.35 to 0.32), whereas it increased by 3 percent in Denmark (from 0.33 to 0.34). Our multiverse analyses demonstrate that these results are robust to alternative variable specifications.
In this article, we placed weight on margin-insensitive estimates because they capture the openness of a society net of any distributional changes. However, we should not forget that educational expansion can have real effects on the lives of disadvantaged families. In fact, these families are likely more occupied with seeing their children better off than themselves rather than comparing their children's outcomes to those of more advantaged families. For this reason, it may be relevant to zoom in on the opportunities for disadvantaged groups, not least because one of the foremost responsibilities of a welfare state is arguably to support those at the lower end of the educational distribution. Here, we find that compared to children in Scandinavia, children of less educated parents in the United States face greater barriers to achieving higher levels of education than their parents do. As reported in S14 in the online supplement, U.S. children whose parents did not complete high school have a 10 percent chance of obtaining a bachelor's degree or higher compared to 16 percent to 19 percent in Scandinavia. For children whose parents have high school or less, the chance increases to 19 percent in the United States, but it remains higher in the Scandinavian countries: 29 percent in Denmark, 28 percent in Norway, and 24 percent in Sweden. Presented as relative risk measures, in the United States, children of parents with at least a bachelor's degree are a little over 3 times as likely to obtain at least a bachelor's degree themselves compared to children of parents with no more than a high school diploma. In the Scandinavian countries, inequality is lower; children of parents with at least a bachelor's degree are a little over 2 times as likely to obtain at least a bachelor's degree compared to children of parents with no more than a high school diploma (see S14 in the online supplement).
Why do our estimates consistently show that Scandinavia is more equal than the United States compared to previous research? We suggest three main reasons. First, we observed substantial disparities across individual U.S. surveys. This is noteworthy given that most existing studies rely on only one data source for the United States (see Table 1). Survey differences have been noted in other studies reporting survey-specific estimates for the United States (Hertel and Pfeffer 2020; Jackson and Holzman 2020). Because most of the reviewed studies rely on survey data, we suspect that Scandinavian estimates from surveys may differ from those based on administrative data. In S19 in the online supplement, we compare results based on administrative data with results based on the European Social Survey. These results indicate that survey data tend to underestimate inequality compared to administrative data because odds ratios are lower when using survey data for the Scandinavian countries than when using administrative data. The differences between survey and administrative data would be an important area for future research. Second, differences in variable choice, sample restrictions, and cohort or age specifications may contribute to divergent results. In particular, studies should be cautious when comparing estimates derived from samples with heterogeneous cohort compositions. Multiverse analysis may provide a sound backdrop for assessing how sensitive estimates are to alternative variable specifications. Third, coding differences and coding errors may also lead to differing results. However, because only 1 of the 13 studies reviewed provides replication packages (Andrade and Thomsen 2021), it is difficult to assess the extent to which diverging coding practices account for the differences in results.
Our findings carry important implications for both researchers and society. First, there is a pressing need for robust and transparent accounts of intergenerational mobility to capture overall trends accurately. These accounts must clearly state how they conceptualize mobility. For researchers looking for a margin-insensitive, universal measure of relative mobility based on categorical measures, we have made a case for the IAC proposed by Bouchet-Valat (2022). The IAC offers a more intuitive interpretation than the UniDiff parameter, it can be interpreted as a correlation, and it enables assessment of both the trend and the overall level of inequality. Together with a multiverse analysis that can provide a powerful robustness check, the IAC may arguably be the go-to metric for universal, margin-insensitive mobility estimates based on categorical measures.
Second, the media and policy implications of drawing conclusions from single estimates are far-reaching, especially when such estimates are prematurely interpreted as evidence of the Scandinavian welfare state's failure to provide equal opportunities (for such an example, see Landersø and Heckman 2017:220). For these reasons, we urge researchers to explicitly state the assumptions underlying their chosen measures and openly share coding files to enable replication. In this article, we presented compelling evidence of higher relative mobility in Scandinavia compared to the United States. Educational inequality is a problem in all four countries, but much more so in the United States than in Scandinavia.
Supplemental Material
sj-docx-1-soe-10.1177_00380407261424502 – Supplemental material for Intergenerational Educational Mobility in Scandinavia and the United States
Supplemental material, sj-docx-1-soe-10.1177_00380407261424502 for Intergenerational Educational Mobility in Scandinavia and the United States by Jens-Peter Thomsen, Stefan Bastholm Andrade, Florian R. Hertel, Max Thaning and Øyvind Nicolay Wiborg in Sociology of Education
Footnotes
Acknowledgements
This article was presented at the Spring 2023 meeting of the ISA Research Committee 28 on Social Stratification and Mobility in Paris and at seminars at the Danish Center for Social Science Research (VIVE). We thank the participants of these meetings for their valuable comments. We also appreciate Asta Breinholt, Per Engzell, Kristian B. Karlson, and Hans Henrik Sievertsen for extensively commenting on earlier drafts of our article and providing excellent feedback. Finally, we are grateful for the valuable comments from the three anonymous reviewers.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: Jens-Peter Thomsen and Stefan B. Andrade acknowledge funding from NORFACE - New Opportunities for Research Funding Agency Cooperation in Europe (Grant No. 462-16-073). Florian Hertel acknowledges funding from the German Federal Ministry of Research, Technology and Space (Grant No. NWGWIHO01). Max Thaning acknowledges funding from the Swedish Research Council for Health, Working Life and Welfare – Forte (Grant No. 2024-01411). Øyvind Wiborg acknowledges access to Norwegian data from HISTCLASS - Paradoxes of wealth and class: historical conditions and contemporary configurations”, Funded by The Research Council of Norway, Project number: 275249.
Research Ethics
This study uses public survey data and de-identified data from Statistics Denmark, Statistics Norway, and Statistics Sweden, which have been sufficiently aggregated to ensure anonymity, in compliance with all relevant regulations concerning publication.
ORCID iDs
Supplemental Material
Supplemental material for this article is available online.
Notes
Author Biographies
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
