Existing reviews and meta-analyses

To begin, it will be useful to consider existing reviews that have summarized literature relevant to quantifying the relative predictive validity of cognitive ability and parental SES on socioeconomic outcomes. One meta-analysis by Strenze (2007) [archived] shows that intelligence is a great predictor of future socioeconomic success, often better than many measures of parental SES (Table 1). Socioeconomic success was measured as educational attainment, occupational attainment, and income.

Even though cognitive ability has the highest correlation with each measure of socioeconomic status, cognitive ability does not have much greater correlations than parental SES. In fact, the confidence intervals for the correlation coefficients of cognitive ability often overlap with the confidence intervals for the coefficient for various measures of parental SES. Furthermore, because cognitive ability and parental SES are correlated, each of the correlations may partially be the result of confounding with other predictors. We should review individual studies that report the predictive validity of each predictor while controlling for the other predictors, i.e. studies that report the unique association of each of the predictors with the given outcome.

Kingston (2006) also conducted a review of studies comparing the predictive validity of parental SES, educational attainment, and cognitive ability on socioeconomic outcomes in the United States. He finds that whereas cognitive ability and educational attainment have “substantial” effects on economic outcomes after controlling for parental SES (page 119), parental SES have only “minor” effects on economic outcomes after controlling for cognitive ability and educational attainment (page 121). However, he considered many older studies and limited his review to studies in the United States. My post will consider some newer studies and many studies outside of the United States.

Roberts et al. (2007) [archived] also reviewed longitudinal studies to compare the predictive validity of personality traits, parental socioeconomic status, and cognitive ability on future important life outcomes. The authors focus on the predictive validity of these factors on mortality, divorce, and occupational attainment. For the purposes of this post, only occupational attainment is relevant. The authors considered studies that examined the independent associations of each of these factors on occupational attainment while holding fixed the other factors (Table 6). The following graph shows the standardized regression coefficients for each of the outcomes on occupational attainment (Figure 3):

Unfortunately, these findings are limited in a few ways (see the studies covered in Table 6):

• Most of the studies used were rather old, published in the 90s or earlier.
• Some of the studies controlled for either cognitive ability or parental SES, but not both.
• The measure of occupational attainment was highly heterogeneous, including factors such as erratic work life, age at entry in a stable career, occupational creativity, extrinsic career success, etc. (Table 6).
• It is not clear which studies relied on representative samples.
• It is not clear how old the participants were when their IQ and parental SES were measured.

Granted, I could have checked whether the last two points were genuine problems, but didn’t feel like doing so given the other issues. Regardless, I aim to avoid all of these concerns when reviewing the studies in this post.

Selection criteria

When searching for studies to review for this post, I tried to select studies that fulfilled the following conditions:

• Samples: The study should analyze large nationally representative longitudinal samples.
• Variables: The study should measure both cognitive ability and parental SES when participants were in their youth, ideally during childhood or adolescence. The study should measure socioeconomic outcomes (as educational attainment, occupational attainment, and/or income) for participants many years after measuring cognitive ability and parental SES. The study should ideally measure parental SES as a composite that includes parental educational attainment, occupational attainment/status, and income (although some studies will only have information on 2 of these measures). As I mentioned in my previous post, this measure of parental SES is in line with most existing research on parental SES. By occupational status/attainment, I have in mind something like occupational prestige. I’ll use the terms occupational “prestige”, “status”, and “attainment” interchangeably. Also, I will use the terms “cognitive ability”, “intelligence”, and “IQ” interchangeably.
• Quantify unique associations. The study should quantify the unique association of cognitive ability and parental SES with socioeconomic outcomes. That is, for each socioeconomic outcome, the study should quantify the association between each predictor and the outcome that is not associated with the other predictor. If we want to compare the predictive validity of parental SES and cognitive ability on a given outcome, then we must know the association between parental SES and outcomes that cannot be attributed to cognitive ability (and vice-versa). This avoids confounding bias as described in a previous post. This can be done using e.g. regression analysis.
• Avoid controls for mediators. The study should quantify the unique association of each predictor with socioeconomic outcomes without controlling for mechanisms, i.e. variables that mediate the association between the predictor(s) and the outcome. Controlling for intermediate variables leads to overadjustment bias as I have described in this post. For example, let’s say we want to compare the unique predictive validity of cognitive ability vs parental SES on educational attainment. Let’s also say that most of the effect of cognitive ability on educational attainment is mediated through high school grades, but very little of the effect of parental SES is so mediated. In that case, including high school grades in a regression model would reduce the coefficient for cognitive ability much more than it would reduce the coefficient for parental SES, resulting in a distorted picture of the relative predictive validities of parental SES and cognitive ability.

United States

I will analyze the relative predictive validity of cognitive ability and parental SES on socioeconomic outcomes in the United States using studies based on 3 large nationally representative longitudinal datasets: the National Longitudinal Survey of Youth 1979 (NLSY79), the National Longitudinal Survey of Youth 1997 (NLSY97), and the Project Talent dataset. Before covering the studies on the NLSY datasets, I will briefly describe both of the NLSY datasets here.

The National Longitudinal Surveys were 2 nationally representative longitudinal studies published by the U.S. Bureau of Labor Statistics for 2 cohorts of Americans: the 1979 National Longitudinal Survey of Youth [archived] (NLSY79) and the 1997 National Longitudinal Survey of Youth [archived] (NLSY97). Both surveys collected data on several thousand subjects from a representative sample of the noninstitutionalized civilian population in the United States. For both surveys, data was collected at the start of the survey when participants were in their youth (in 1979 or 1997). This data included information on the parental SES measures that are relevant for my purposes (e.g., parental income, parental education, and parental occupation). The participants also provided additional information in regularly scheduled follow-up interviews about important life outcomes, such as wages, educational attainment, marital status, etc. Some important differences between the two datasets are as follows:

• The NLSY79 includes about 10,000 participants born between 1957 to 1964. Participants were aged 14 to 22 years old when first interviewed in 1979. This cohort was interviewed in 28 separate rounds, with the most recent round being in 2018.
• The NLSY97 includes about 9,000 participants born between 1980 and 1984. Participants were aged 12 to 17 years old when first interviewed in 1997. This cohort was interviewed in 18 total rounds, with the most recent round being in 2021.

Both cohorts completed the Armed Forces Qualification Test (AFQT) during the first interview. The AFQT comprises the following subsections of the Armed Services Vocational Aptitude Battery (ASVAB): mathematics knowledge, arithmetic reasoning, paragraph comprehension, and word knowledge. Now, the AFQT is technically not a cognitive ability test because it was not constructed with the goal of measuring cognitive ability or intelligence. However, AFQT test scores correlate very highly with other conventional IQ test scores, such as the Wechsler Adult Intelligence Scale (WAIS) (source [archived]) (r=0.8), even more than IQ tests often correlate with themselves. Thus, as Smith et al. (2000) note, “Although the AFQT is not an intelligence test, per se, it is highly correlated with IQ and is a widely used measure of adult aptitude” (page 812). Furthermore, the AFQT is “one of the most highly g-loaded tests in use (g refers to the general intelligence factor)” (Dickens and Flynn 2006 [archived], page 3). For these reasons, AFQT tests are often used as proxies for cognitive ability (e.g., Heckman et al. 2005, page 83; see Gjerde 2002, page 449; Shearer et al. 2002, page 238; Trevor 2001, page 628; Han et al. 2004, page 338). In line with existing literature, I also treat AFQT test scores as proxies for cognitive ability.

The National Longitudinal Survey of Youth 1979

The first analysis of the NLSY79 data that I consider comes from Herrnstein and Murray (1994) [archived]. These authors used this data to compare the predictive validity of youth cognitive ability and parental SES on important socioeconomic outcomes. Parental SES was measured as a composite of parental education, parental occupational status, and parental income. Youth cognitive ability was measured using AFQT scores, as described earlier.

The predictive validity of cognitive ability and parental SES was compared by analyzing the relationship between each predictor and an outcome while holding the other predictor fixed. First, the authors first plotted the relationship between IQ and the outcome in question while holding parental SES fixed at the mean. Then they performed the same analysis while swapping IQ and parental SES; that is, the authors then plotted the relationship between parental SES and the outcome in question while holding IQ fixed at the mean. These analyses were only conducted for white respondents to avoid possible conflations with race. The results in these analyses suggest that deviations in youth IQ matter much more than deviations in parental SES for predicting socioeconomic outcomes.

For example, consider the following findings on the relative associations between IQ vs parental SES on the probability of dropping out of high school (page 149).

As you can see, deviations in youth IQ have much greater associations with high school dropout rates than deviations in parental SES. At the higher levels of youth IQ and/or parental SES, there doesn’t seem to be much of a difference: having a high IQ or high parental SES seems to suggest similarly low chances of dropping out of high school. However, there are large differences at the lower levels. Individuals with low IQs (with mean parental SES) are much more likely to drop out of high school than individuals from low parental SES households (with mean IQ).

Similar findings emerge when analyzing the probability of attaining a bachelor’s degree (page 152).

The findings here are basically the same as the findings for high school dropout rates, except the behaviors at the tails are reversed: at low levels, individuals are similarly unlikely to attain a bachelor’s degree. However, individuals with high IQ (and mean parental SES) are much more likely to attain a bachelor’s degree than individuals with high parental SES (and mean IQ).

Finally, consider the results when analyzing the probability of ending up in poverty:

Again, deviations in IQ seem to be much more significant than deviations in parental SES, particularly at the lower ends of IQ and parental SES. Low IQ is a much bigger predictor of future poverty than low parental SES. If you want to know whether a young person will end up in poverty as an adult in the United States, it is better to know their current IQ than their parental SES.

The basic findings from Herrnstein and Murray were also analyzed in “A Reanalysis of The Bell Curve: Intelligence, Family Background, and Schooling” by Koreman and Winship (2000) [archived]. These authors compared the relative predictive validity of AFQT scores and family background after making a number of adjustments to the analysis by Herrnstein and Murray. For my purposes, I will focus on their analysis after introducing a broader set of variables into the measure of family background. Unlike Herrnstein and Murray’s SES index (parental education, occupation, and income), these authors created what they called a “detailed” family background (FB) index that included the following variables (see page 154):

• Family arrangement (e.g. two-parent, single-parent, etc.)
• Mother’s age at birth
• Number of siblings of the respondent
• Whether the respondent lived in an urban area
• Whether the respondent is the oldest child of the family
• Whether an adult female in the household worked outside the home
• Whether the family had a library card, received magazines regularly, and received newspapers regularly.

They also include controls for age, gender, and race/ethnicity (Koreman and Winship analyze all subjects in the sample rather than just non-Hispanic whites, unlike Herrnstein and Murray).

The results of the regression analysis after incorporating these detailed family background (FB) variables are reported in Table 7.4:

The first two columns simply repeat the analysis by Herrnstein and Murray. The 3rd and 4th columns perform the same analysis after incorporating controls for the detailed family background variables mentioned above. The 5th column displays the effects of a standardized measure of the family background effects and parental SES. The 6th column is the same as the 5th column except it includes race as well. All models include controls for age, gender, and race/ethnicity.

The z before the AFQT and SES variables indicate that the variables were standardized. Thus, each coefficient indicates the change in the outcome associated with a 1 standard deviation (SD) increase in the predictor (or a 1-unit increase in the z-score for the predictor). SES is measured using Herrnstein and Murray’s index (i.e. parental education, occupation, and income). Thus, for example, the first column shows that, without controlling for detailed family background, a 1 SD increase in AFQT scores is associated with a $6,975 increase in family income. The second column shows, again without controlling for detailed family background, a 1 SD increase in parental SES is associated with a $4,580 increase in family income.

There are a number of different comparisons one might make to compare the relative predictive validity of cognitive ability and parental SES. The first two columns merely reproduce the findings from Herrnstein and Murray, i.e. AFQT scores are much more important than parental SES for main socioeconomic outcomes such as income and educational attainment. The 3rd through 5th columns show, respectively, the predictive validity of AFQT scores, parental SES, and the combination of parental SES and detailed family background on each outcome. We can summarize the main findings as follows:

• AFQT scores have greater predictive validity than even the broad measure of detailed family background for the main socioeconomic outcomes that I’m considering for this post (educational attainment, occupational prestige, and income). For example, a 1 SD deviation increase in AFQT scores is associated with about a 55% greater increase in annual earnings than a 1 SD increase in the FB + SES composite ($4,669 vs $3,007), and about a 3.6x greater increase than parental SES alone ($4,669 vs $1,285). Similarly, a 1 SD increase in AFQT is associated with about 2 times more years of schooling completed than a 1 SD increase in the FB + SES composite (0.58 vs 0.27), and about 3 times more years of schooling than a 1 SD increase in parental SES (0.58 vs 0.18).
• The table does show that the FB + SES composite predicts labor force participation and unemployment about as well as, or slightly better than, the AFQT scores. Interestingly, parental SES has a very weak correlation with these outcomes. This suggests that, whatever effect family background has on labor force participation and employment, the effect is not due to parental income, education, or occupational status.
• One measure of socioeconomic attainment where the FB + SES composite comes close to AFQT is family income ($6,516 vs $6,157). This is likely because family income (as opposed to individual earnings) depends heavily on whether one is in a dual- vs single-income earning household, which is likely influenced by marital norms that correlate heavily with one’s family background. For example, the table shows that marriage by age 30 has a strong association with the FB + SES composite (the association with AFQT scores is virtually nil).

Thus, for the main socioeconomic outcomes of concern in this post (education, occupation, income), cognitive ability is consistently a better predictor than both parental SES and the broad measure of family background. However, there are some other outcomes for which the detailed family background composite predicts as well or slightly better than cognitive ability (e.g., labor force participation, family income, etc.). The narrow measure of parental SES (parental education, occupation, and income) never predicts offspring socioeconomic outcomes better than does cognitive ability.

The National Longitudinal Survey of Youth 1997

The findings on the NLSY79 regarding educational attainment were reproduced for the NLSY97 by Belley and Lochner (2007) [archived]. These authors analyzed the importance of family income and cognitive ability in determining educational attainment using data from the NLSY79 and NLSY97. In addition to family income, the 1997 cohort also included data on net wealth. The following tables shows the relative effects of AFQT scores, family wealth, and family income in a regression model for high school completion and college attendance:

The figures above show the association of AFQT scores vs family income/wealth on the measures of educational attainment. The data shows that AFQT scores have a much larger effect on educational attainment than does family income/wealth. Focusing on college attendance, having an AFQT score in the top quartile is associated with a 50 percentage-point increase in the probability of attending college (relative to having an AFQT score in the bottom quartile). By contrast, coming from a family in the top quartile of family wealth or family income is associated with a 21 and 8 percentage point increase, respectively, in the probability of attending college (relative to coming from a family in the bottom quartile for those measures). In fact, having an AFQT score in the second-lowest quartile rather than the lowest quartile is associated with a 22 percentage-point increase in the chances of attending college, which is greater than the increase associated with coming from a family in the top quartile of wealth rather than the bottom quartile of wealth.

The study also included data on the predictors on completing 4+ years of college. The findings were similar. That is, AFQT scores were a much better predictor of educational attainment than were family income. Unfortunately, they did not provide data on family wealth for this analysis:

As you can see, in the NLSY97, having an AFQT score in the top quartile was associated with a 45 percentage-point increase in the chance of completing 4+ years of college by age 23 (compared to having an AFQT score in the bottom quartile). By contrast, coming from a family in the top quartile of family income was associated with only a 10 percentage-point increase in the chance of completing 4+ years of college (compared to coming from a family in the bottom quartile of family income).

Eid 2018 [archived] also replicated the findings from Herrnstein and Murray using data from the NLSY97. However, the study focused on adulthood poverty as the relevant outcome. Parental SES was measured by parental education, parental occupation, and parental income. In a regression model with poverty as the dependent variable and youth ability, parental SES, and age as independent variables, the effect of youth ability was 3 times greater than the effect of parental SES (Table 4):

In their controversial work The Bell Curve, Charles Murray and Richard Herrnstein hypothesize that intelligence plays a significant role in determining later life outcomes. Using data from the 1979 National Longitudinal Survey of Youth, Herrnstein and Murray specifically claim that Youth IQ is more important than Parental Socioeconomic Status in affecting future odds of living in poverty. We explore their hypothesis using the same statistical decisions but on a new data set: the 1997 National Longitudinal Survey of Youth. Our findings mirror those of Herrnstein and Murray; IQ has a larger effect than socioeconomic status on the likelihood of being in poverty, though the overall effect of both is diminished. We find that increasing Youth IQ by one standard deviation is expected to decrease the odds of ending up in poverty by 35 percentage points, compared to the 13 percentage point decrease from similarly increasing Parental Socioeconomic Status.

A more comprehensive analysis of the relative associations of parental SES and youth cognitive ability on socioeconomic outcomes in the NLSY79 and NLSY97 was performed by Marks (2022) [archived]. One difference between this study and the previous studies is that this study measures cognitive ability by extracting the general factor of intelligence – the g factor – from 10 ASVAB subtests instead of just using raw AFQT scores (which is just the summed score from 4 subtests of the ASVAB). First, let’s see the zero-order correlations involving offspring cognitive ability, various measures of parental SES, and offspring socioeconomic outcomes:

The chart is a bit difficult to see because it’s wide (open the image in a new tab for more space). But the data shows that cognitive ability has a much greater correlation with socioeconomic outcomes than do the parental SES measures. For example, in the NLSY79, the correlation between occupational prestige and cognitive ability (r = 0.54) is greater than the correlation between occupational prestige and father’s occupation (r = 0.36) or any of the other measures of parental SES.

To better compare the predictive validity of parental SES and cognitive ability, we should measure the unique associations of parental SES and cognitive ability with socioeconomic outcomes while holding the other predictor fixed. Marks performed this analysis by conducting various regression analyses for each of the socioeconomic outcomes shown in the table above (i.e. years of education, occupational prestige, income, and wealth). For each outcome, a series of regression models were conducted that involve a different combination of parental SES measures and cognitive ability. For the main analyses, he includes just parental income and education as the measure of parental SES. He later conducts supplementary analyses that incorporate parental occupation and parental wealth in the regression models as well. Introducing these additional variables do not substantially change the main results, as I will cover below.

Let us start by examining the regressions for educational attainment. The following table shows the associations of parental education, parental income, and cognitive ability with educational attainment in different models:

I went over how to read these tables in my previous post that covered the same study when analyzing predictors of academic achievement. As I did in that post, there are a few ways to compare the independent associations between cognitive ability vs parental SES on educational attainment:

• Compare the R^2 (% of variance explained) of models 1 and 2. The model with just cognitive ability alone (model 2) explains about 60 to 70% more of the variance than does the model with both parental income and education (model 1) in both the NLSY79 (R^2 = 0.35 vs R^2 = 0.22) and the NLSY97 (R^2 = 0.32 vs R^2 = 0.19).
• Now, compare the R^2 of models 2 and 3. The R^2 for model 3 is only about 3 to 4 percentage points greater than the R^2 for model 2 in both of the datasets. This suggests that while parental education and family income (particularly parental education) provide some incremental validity over just cognitive ability in predicting educational attainment, the bulk of the variance explained can be encompassed by cognitive ability alone.
• Compare the standardized coefficients when they are entered into the same model (model 3). In model 3, the standardized coefficient of cognitive ability is about 2 times the coefficient for parental education in the NLSY79 (β = 0.48 vs β = 0.21) and 3 times the coefficient in the NLSY97 (β = 0.46 vs β = 0.18). The standardized coefficient for cognitive ability is about 16 times the coefficient for family income in the NLSY79 (β = 0.48 vs β = 0.03) and 8 times the coefficient for family income in the NLSY97 (β = 0.46 vs β = 0.06).

Similar analyses were performed for occupational attainment:

And income:

• Compare the R^2 of models 1 and 2. In both the NLSY79 and NLSY97, the model with just cognitive ability (model 2) explains about 2 to 3 times as much of the variance as the model with both parental education and family income (model 1). E.g. in the NLSY97, cognitive ability alone explains 17% of the variance in income whereas parental income and education together explain only 6% of the variance.
• Compare the R^2 of models 2 and 3. In both the NLSY79 and NLSY97, the model with cognitive ability, parental education, and family income (model 3) explains only about 1 to 2 percentage points more of the variance than does the model with cognitive ability alone (model 2). Thus, parental education and family income only provide marginal incremental validity over just cognitive ability in predicting occupational attainment and income.
• Compare the standardized coefficients when they are entered into the same model. The standardized coefficients are reported in model 4. For both occupational attainment and income, parental education has no statistically significant positive association with socioeconomic outcomes (in fact, parental education has a statistically significant negative association with income). So we only need to compare the effect sizes of parental income and cognitive ability. The standardized coefficients show that the effect of cognitive ability on later occupational attainment is about 6 to 7 times the effect of parental income, whereas the effect of cognitive ability on later income is about 2 to 3 times the effect of parental income.

There are two potential issues with the above analyses:

• One issue with using model 4 to compare the standardized coefficients of parental SES and cognitive ability is that years of education is also included as an independent variable. Years of education is likely to mediate the relationship between the other independent variables and the dependent variables. Thus, including years of education as an independent variable likely introduces overadjustment bias as mentioned in the beginning of the post.
• One might object that this analysis used an overly narrow measure of parental SES (just parental income and parental education).

Both of these issues are addressed with the supplementary analyses. The supplementary analyses reran the previous analyses using extended versions of model 3 for each outcome and dataset. The extended versions of model 3 contained the original independent variables (cognitive ability, parental education, family income) and introduced parental occupation (only for the NLSY79) and family wealth. Introducing these variables does not significantly change the results. The results are as follows:

Project Talent

Project Talent is a national longitudinal study that began following a large representative sample of high school students in 1960. The average age of the participants was 15.8 years at the first interview. After the initial wave, some of the participants were followed up 1, 5, 11, and 50 years after the original survey. Cognitive ability was measured by averaging the standardized scores of different scales measuring verbal, quantitative, and visualization and spatial abilities. Parental SES was measured as a composite based on home value, family income, father’s education, mother’s education, father’s job status, number of books, number of appliances, number of electronics, and whether the child had a private room.

Data on the socioeconomic outcomes of the participants was collected at the 11-year follow-up and 50-year follow-up. The original sample contained over 440,00 students. Damian et al. (2015) [archived] notes that it is “the only nationally representative longitudinal study in the U.S. of such large scale” (page 6). However, the sample size reduced to about 81,000 when considering only participants who provided data on socioeconomic outcomes in the 11-year follow-up. The sample size further reduced to about 1,952 when considering only participants who provided such data in the 50-year follow-up. Socioeconomic outcomes were assessed by measuring the participants’ highest level of education, occupational prestige, and annual income.

The first analysis of Project Talent data that I consider comes from Damian et al. (2015) [archived]. These authors analyzed the relationship between family background and individual differences (e.g. personality, cognitive ability) measured at the original survey and socioeconomic outcomes at the 11-year follow-up. Specifically, the researchers studied whether individual differences predict socioeconomic outcomes independently of family background. They were also interested in whether advantages in intelligence and/or personality could compensate for being born in a low SES household. There were correlations in the expected direction between cognitive ability, parental SES, and socioeconomic outcomes 11-years after the initial survey:

Consistent with previously presented data, both parental SES and intelligence are strong predictors of educational attainment and occupational prestige. The magnitudes of the correlation coefficients are similar to the coefficients presented in the meta-analysis by the aforementioned Strenze (2007).

Note that the correlation for intelligence is not substantially greater than the correlation for parental SES. But this may be due to the fact that intelligence and parental SES are themselves highly correlated (r = .44) in the table above. In order to tease apart the independent predictive validity of parental SES and cognitive ability, we must analyze the unique association between each predictor and the socioeconomic outcomes. These analyses were performed using regression analyses for each of the socioeconomic outcome variables. For each socioeconomic outcome, they conducted a separate regression for each individual variable (intelligence or personality trait) as an independent variable, along with parental SES, gender, race, and age. I will focus on the regressions involving intelligence.

Table 6 shows the regression coefficients of each of the independent variables in the regressions including intelligence as a predictor:

Note that the parental SES and intelligence have been standardized, so it is sensible to compare the coefficients for parental SES and intelligence. For example, the above figure shows that a 1 standard deviation increase in intelligence is associated with a 0.49, 0.05, and 8.49 unit increase in educational attainment, annual income, and occupational prestige. On the other hand, a 1 standard deviation increase in parental SES is associated with a 0.30, 0.06, and 4.21 standard deviation increase in the same outcomes. Thus, compared to parental SES, youth intelligence appears to be a much better predictor of educational attainment (63% greater regression coefficient) and occupational prestige (102% greater regression coefficient) roughly 11 years after high school.

The authors conclude by noting that high intelligence (unlike personality traits) stands out as the key trait that can compensate for low parental SES.

In the case of personality traits, neither the main effects nor the interactive effects were large enough to compensate for low parental SES and this can be seen on the response surface figures…Therefore, personality traits, while important in the prediction of attainment outcomes, did not suffice to make up for low parental socioeconomic status. However, the story was different for intelligence…In other words the smartest but least wealthy people were close to getting a college degree, whereas their least smart but wealthiest counterparts were close to getting an associate’s degree. In sum, even though we have evidence that certain personality traits may compensate for background disadvantage (in the absence of intelligence controls), the effects were not large enough to overcome the main effect of SES. The only individual difference that seemed to be able to do that was intelligence. Thus, we would conclude that the American Dream, as manifest through personality, is more myth than fact. On the other hand, the American Dream manifest through intelligence is still alive and well.

However, intelligence does not appear to be a greater predictor of annual income. This may be due to the fact that annual income was measured fairly early in life (just 11 years after the initial high school survey, meaning the participants would have been about 26.8 years old on average). Previous data suggests that intelligence becomes a better predictor of annual income measured later in life as individuals have greater opportunity to use their intelligence to increase their earnings. For example the meta-analysis by Strenze 2007 notes the following regarding the predictive validity of intelligence on income by age (page 413):

The correlation between intelligence and income undergoes the most dramatic changes: the correlation is barely above zero in the 20–24 group but jumps to the value of .20 in the 25–29 group, and then takes another jump to the value of .27 in the 30–34 group; after the age of 40, the correlation appears to decline again but not as low as the values it had before the age of 30.

Therefore, it would be useful to compare the predictive validity of parental SES and cognitive ability in the 50-year follow-up. Spengler et al. (2018) [archived] did just this. These authors investigated the same dataset as the previous study, but included data on both the 11-year and 50-year follow-ups. Before providing the results of the regression analyses. Let us analyze the correlation coefficients involving parental SES, youth IQ, and socioeconomic outcomes.

The correlations here are similar in magnitude to the correlations reported in the previous study. One interesting finding is that the correlation coefficients for income increased substantially between the 11-year follow-up and the 50-year follow-up. The correlation between income and youth IQ increased by about 84% (from r = 0.19 to r = 0.35), and the correlation between income and parental SES increased by about 115% (from r = 0.13 to r = 0.28). However, the correlations with educational attainment and occupational prestige decreased slightly over this period.

As with the previous study, IQ does not have a substantially greater correlation with socioeconomic outcomes than does parental SES. But, again, parental SES and IQ are highly correlated (r = 0.44). So we should analyze the unique associations that parental SES and IQ have with socioeconomic outcomes. The authors performed such analyses by conducting various regression analyses on each of the main socioeconomic outcomes. I’ll focus on the regressions performed on the main socioeconomic outcomes during the 50-year follow-up.

Here are the results of the regressions for educational attainment in the 50-year follow-up:

In order to compare the predictive validity of parental SES and youth cognitive ability (IQ), it is best to focus on Model Set A.2 in order to avoid overadjustment bias as mentioned at the beginning of the post. This is because models A.3 through A.5 include traits that plausibly mediate the relationship between parental SES or cognitive ability and socioeconomic outcomes.

That being said, Model Set A.2 shows that youth IQ is a much better predictor of educational attainment than is parental SES. The standardized coefficient for IQ is about 70% greater than the coefficient for parental SES (β = 0.42 vs β = 0.25).

Similar findings emerge when analyzing the regressions for occupational prestige:

Again, in Model Set B.2, the standardized coefficient for youth IQ is about twice the coefficient for parental SES (β = 0.29 vs β = 0.16).

Finally, consider the regressions for annual income:

Again, youth IQ is a substantially better predictor of annual income than parental SES. The standardized coefficient for youth IQ is over twice the coefficient for parental SES (β = 0.30 vs β = 0.14, although the confidence intervals do overlap).

Note that these findings for income are much different than the findings for income reported in the previous study by Damian et al. (2015), which found that youth IQ and parental SES were roughly equivalent predictors of annual income. This is likely due to the fact that the previous study analyzed outcomes only 11 years after the initial study, where high-IQ individuals have less time and opportunity to take advantage of their potential. In fact, the authors of this study also find that parental SES and intelligence have similar unique associations with annual income in the 11-year follow-up (Table 5). It is only in the 50-year follow-up that IQ outdoes parental SES. The authors speculate that one reason that the predictors had larger correlations with the 50-year follow-up than the 11-year follow-up is that “it took time for people to achieve their maximum amount of education and thus reveal the effect of student behaviors on achievement” (page 631).

Summary

Before moving on from the data on the United States, it will be useful to summarize the main findings for the 3 main U.S. datasets (NLSY79, NLSY97, Project Talent) in one place. I will focus on the studies that reported regression coefficients for parental SES and cognitive ability on each of the socioeconomic outcomes (Koreman and Winship 2000, Marks 2022, Spengler et al. 2018). From those studies, here are the ranges of the standardized coefficients of cognitive ability and the parental SES when they were simultaneously entered in regressions for each outcome:

Educational attainment

Occupational attainment

Income

Benner et al. (2016) also examined the predictors of educational attainment in a large, nationally representative sample of high schoolers in the United States. The study analyzed roughly 15,000 10th grade students and measured their educational attainment 8 years after their expected high school graduation. Parental SES was measured as a composite of parental education, parental occupational prestige, and household income. The study did not include an explicit measure of cognitive ability, but it did include performance on a standardized achievement test of mathematics and reading from the 10th grade. I’ve presented studies in previous posts showing that academic achievement tests such as the SAT and ACT are acceptable measures of general cognitive ability. Thus, performance on the standardized achievement test in this study is likely a good proxy of general cognitive ability. Even if not, the relative predictive validity of standardized test scores vs parental SES on educational attainment is interesting in its own right.

In a regression model for educational attainment 8 years after high school graduation, the standardized coefficient of the 10th grade achievement test was about 60% greater than the coefficient of parental SES (β = 0.31 vs β = 0.19). The following figure shows the relation between parental SES, parental involvement, and performance on the standardized achievement test.

The findings show that, for a given level of parental involvement, high-achieving, low-SES 10th graders go on to earn higher levels of education attainment than low-achieving, high-SES 10th graders. Even high-achieving students with low-SES parents with low-involvement acquired similar or slightly higher levels of education than low-achieving students with high-SES parents with high-involvement.

For other studies that allow comparing the predictive validity of parental SES and cognitive ability on socioeconomic outcomes, see Nagoshi et al. (2008). Nagoshi et al. (2008) found that “family background had a trivial influence and own cognitive ability had a substantial influence on educational attainment” in a sample of subjects in the Hawaii Family Study of Cognition (HFSC).

United Kingdom

Now, let us consider studies on the relative predictive validities of cognitive ability and parental SES in the United Kingdom. I’ll focus on two large, nationally representative longitudinal cohort studies. Both of these studies were conducted by collecting information on virtually all newborns who were born in England, Scotland, or Wales within a given week. The participants were regularly followed up in later sweeps to collect information on their childhood environment and life trajectory. These datasets allow us to analyze the relationship between youth cognitive ability, parental SES, and offspring socioeconomic outcomes in the United Kingdom. The two datasets in more detail are as follows:

• The 1958 National Child Development Study (NCDS): this study follows the lives of an initial 17,415 people born in England, Scotland, and Wales in a particular week of 1958. When the participants were 11 years of age, cognitive ability was measured with a test consisting of verbal and non-verbal items.
• The 1970 British Cohort Study (BCS70): this study follows the lives of an initial 16,568 people born in England, Scotland, and Wales born in a particular week of 1970. When participants were 11 years of age, cognitive ability was measured using a modified version of the British Ability Scales (BAS). The test involved four subscales: word definitions and word similarities as measures of verbal ability, and digit recall and matrices as measures of non-verbal ability.

The U.K. datasets have some advantages and disadvantages over the U.S. datasets. One advantage is that the participants have their cognitive ability tested in the middle of childhood (age 11), which is considerably younger than the participants in the U.S. datasets. One disadvantage is that parental social status does not include a measure of family income (though I address this concern below). Another disadvantage is that the studies below do not report the total unique associations of parental SES and offspring cognitive ability on occupational attainment; instead, the studies report the associations of those predictors on occupational attainment after controlling for offspring educational attainment.

U.K. psychologists Helen Cheng and Adrian Furnham have conducted numerous studies on these datasets to analyze the associations between parental SES and childhood cognitive ability on various life outcomes (e.g., Cheng and Furnham 2012, Furnham and Cheng 2013, Cheng and Furnham 2013, Furnham and Cheng 2017, Furnham and Cheng 2017). All studies report similar findings regarding the relative influence of parental SES vs childhood cognitive ability on socioeconomic outcomes. I will cover two of these studies below, one that focuses on the NCDS and the BCS70.

The 1958 National Child Development Study

Cheng and Furnham (2013) analyzed the NCDS to study the relationship between childhood intelligence, parental social status at birth, personality traits, educational attainment, occupational prestige, and mental well-being. There were 5,090 participants who contained all of the necessary data. Parental social status was measured as a composite of occupational social status and parental education. Educational attainment and occupational prestige were measured when participants were 50 years of age. Educational attainment was measured based on highest academic or vocational qualifications on a six-point scale ranging from 0 = “No qualifications” to 5 = “University Degree”.

I chose this study to represent the NCDS data because the authors analyzed the association between parental SES, childhood intelligence, and socioeconomic outcomes in a model with no intermediate variables. Most other studies I found included a plethora of intermediate variables that likely mediate the effect of parental SES and childhood intelligence on socioeconomic outcomes (which leads to overadjustment bias, as described above). That said, here are the zero-order correlations among the relevant variables:

As you can see, the measures of childhood cognitive ability had much greater associations with adulthood socioeconomic outcomes than the measures of parental SES. For educational attainment, the correlation with childhood cognitive ability was about 60 to 80% greater than the correlation with either parental social class or parental education. For occupational prestige, the correlation with childhood cognitive ability was about 50% greater than the correlation with parental social class and about 2 times the correlation with maternal or paternal education.

The authors constructed a structural equation model to represent the relationship between childhood intelligence, parental social status, gender, educational attainment, occupational prestige, and mental well-being (Figure 1). This allows us to quantify the unique associations of childhood intelligence vs parental social status with the socioeconomic outcomes:

The main figures to note are the path coefficients connecting childhood intelligence and parental social status to education and occupational prestige. The path coefficients are all standardized so they can be meaningfully compared. As you can see, the path coefficients from childhood intelligence to both education and occupational prestige are over twice that of the corresponding path coefficients from parental social status (0.40 vs 0.19 for educational attainment, and 0.19 vs 0.08 for occupational prestige). In other words, if you want to predict a child’s education or occupational prestige, it is more important to know their intelligence than it is to know their parents’ education or occupations.

Another study was conducted by Furnham and Cheng (2013) to analyze how well these same predictors predict earnings at adulthood. I didn’t post the model from this study because it includes a large number of intermediate variables (personality traits) which can distort the relative influence of childhood intelligence and parental social status. However, the basic findings are similar to those reported above even when personality traits are included as controls: childhood intelligence has about twice the association standardized path coefficient with educational attainment and occupational prestige as does parental social status. However, when controlling for personality, educational attainment, and occupational prestige, childhood intelligence and parental social status have roughly equal associations with adult earnings (see Figures 1 and 2). This implies that any advantage that childhood intelligence has over parental social status in predicting adult earnings is completely mediated by educational attainment and occupational prestige.

Another study by Cheng and Furnham (2012) on the same data showed that childhood intelligence was the second best predictor of advancement of occupational prestige between age 33 and 50. The best predictor was educational qualifications. Thus, educational qualifications and childhood intelligence were both better predictors of occupational advancement than parental social status or personality:

The strongest correlates of current occupational levels were educational qualifications, followed by childhood cognitive ability, parental social class, and personality traits. Structural equation modelling showed that for the change of occupation over 17 years, the strongest predictor was education, followed by childhood intelligence. Personality traits (extraversion, conscientiousness, and openness) had modest but significant influence in the upgrading of occupational attainment over the period of time, and parental social status predicted occupational change mediated through education and initial occupational levels. Education and childhood intelligence are more powerful predictors of current occupational prestige than personality factors or family social background.

The 1970 British Cohort Study

Furnham and Cheng (2017) [archived] also analyzed the relationship between parental social status, childhood intelligence, and financial well-being at age 38 using data from the BCS70. They also examined how various non-cognitive skills (locus of control and self-esteem), educational attainment, and occupational attainment mediated this relationship. Unfortunately, including self-esteem and locus of control in the models likely contributes to overadjustment bias, but I was unable to find studies for the BCS70 that did not also include many intermediate variables. Financial well-being was measured based on the weekly net income, number of rooms in one’s home, and home ownership status of the participant.

The zero-order correlations do not show the measures of childhood cognitive ability to have much greater associations with socioeconomic outcomes than do the measures of parental social status (Table 1). However, it is important to analyze the associations of childhood intelligence vs parental social status with socioeconomic outcomes while holding the other variables fixed. Similar to the previous studies by the same authors, they performed this analysis using structural equation modeling. Their first model includes childhood intelligence, parental social status, self-esteem, locus of control, and financial well-being. They conducted separate models for the male and female sample, but I’ll just present the male sample since the results are similar for both:

As you can see, the standardized path coefficient from childhood intelligence to financial well-being is about 45% greater than the corresponding association from parental social status (0.29 vs 0.20).

However, due to overadjustment bias, this likely underestimates the advantage of childhood intelligence over parental social status in predicting financial well-being. This model includes locus of control and self-esteem as intermediate variables that predict financial well-being. In particular, locus of control is a modest predictor of financial well-being, and is strongly associated with childhood intelligence but is not associated with parental social status (0.41 vs 0.01). Thus, it is highly likely that some of the association between cognitive ability and financial well-being is mediated through locus of control, but this association is not represented by the coefficient for the direct path from childhood intelligence to financial well-being. Thus, merely comparing the direct paths for childhood intelligence and parental social status misses some of the advantage of childhood intelligence.

Nevertheless, the authors amended the model by including educational attainment by age 34, occupational attainment at age 38, and malaise (a strong correlate of mental instability) at age 30. Here are the results of this updated path model for the male sample:

Here we see findings in line with the data from the NCDS dataset. That is, the path from childhood intelligence to occupational attainment is 3 times the same path from parental social status (0.20 vs 0.06). The path from childhood intelligence to educational attainment is also about 3 times the same path from parental social status (0.46 vs 0.16).

I want to again emphasize that overadjustment bias likely underestimates the advantage of childhood intelligence over parental social status. Insofar as the intermediate variables (in particular, locus of control) mediates the relationship between childhood intelligence and the socioeconomic outcomes, the direct paths from childhood intelligence to the socioeconomic outcomes will miss some of the effect of childhood intelligence. In fact, the authors themselves explicitly mention how childhood intelligence, but not parental social status, influence locus of control which in turn predicts educational success:

Figure 1 and Figure 2 show the results of exploring the possibility of moderator variables. They also show that the direct effect of intelligence is stronger than that of parental social class in predicting adult financial well-being. There are four important points resulting from that analysis. First, that almost a third of the variance could be accounted for by these four factors alone, all of which were measured at least 20 years before mid-life adult well-being. Second, that the pattern for men and women was almost identical. Third, that intelligence and social class both influence self-esteem, but that it is only intelligence that influences locus of control which is directly related to the outcome variable. Fourth, that the strongest relation was between intelligence and locus of control, suggesting that more intelligent children (aged 10) have a more instrumentalist, internal locus of control at age 16 which is a significant predictor of educational success.

Nevertheless, the paper concludes by noting that childhood intelligence is the best predictor of educational attainment:

The present study confirms the strong and direct effects of parental social class and intelligence on education, occupational and financial well-being in middle age. Current occupation is the strongest predictor of financial well-being, followed by educational achievement, and childhood intelligence is the best predictor of educational attainment. There is a continuous and persistent effect of parental social class on adult financial well-being yet intelligence and locus of control shared a substantial amount of the variance.

Summary

Standardized path coefficients from the NCDS and BCS70 datasets

Other studies

Before moving on from the U.K. data, I should note some other U.K. studies and explain why I did not include them.

von Stumm et al. (2009) also analyzed the predictors of socioeconomic outcomes in the BCS70. Their findings were similar to those reported above. However, I excluded this study because it used a more narrow measure of parental social status (only parental occupational class was used).

Schoon (2008) [archived] analyzed data in both the NCDS and BCS70 that might seem to contradict the findings from the studies above in some respects. This study reports that parental social status is a better predictor of education than is childhood intelligence. See the following structural equation models:

This is consistent with the above studies in that childhood cognitive ability is a better direct predictor of adult social status than is parental social status. But parental social status seems to be a better direct predictor of education in both datasets. Similar findings were reported in another study by Schoon and Polek (2011) [archived] which included net yearly income in the measure of adult social status.

Why the inconsistency? Ignoring the fact that motivation is included as an intermediate variable which may lead to overadjustment bias, one key difference is the measure of education used. The Schoon studies measure education as age finishing full-time education, whereas the previous studies by Furnham and Cheng measured education as highest academic or vocational qualifications. In fact, the Schoon studies did consider highest qualifications, but as a measure of adult social status rather than education. These studies find that parental social status is a better predictor of age finishing full-time education than of highest qualification, whereas the reverse is true for childhood cognitive ability.

For example, the above table shows that whereas father’s age leaving school is a better predictor of “education” (i.e. age finishing full-time education) than of highest qualification for male participants (r = 0.33 vs r = 0.28), cognitive ability is a better predictor of highest qualification than of “education” (r = 0.55 vs r = 0.43).

Similar findings hold for the BCS70 dataset (see Table 2):

Thus, it seems that, in the United Kingdom, parental SES is a better predictor than childhood cognitive ability of time spent in full-time education. However, childhood intelligence appears to be a better predictor of one’s highest academic qualifications in the long-term. The likely explanation is that, among 10- to 11-year-olds in the U.K., having high-SES parents is more important than having high intelligence in determining whether a child will be able to attend university full-time after secondary school. However, having high intelligence is more important than having high-SES parents in determining the highest educational qualifications that the child ultimately attains.

Similar studies were conducted on different datasets in the United Kingdom. Staff et al. (2017) and von Stumm et al. (2010) studied the relationship between childhood intelligence, parental social class, and adult social class in birth cohorts of children born in Aberdeen, Scotland in 1936 and the 1950s. In both studies, childhood intelligence was measured when children were around 11 years old. Socioeconomic outcomes were measured when participants were well into their late 40s. The measures of parental social class were different in the two studies. Both studies included parental occupational class and various other indicators, such as number of rooms in the home, whether the family had a car, etc. Unfortunately, there were no measures of parental education or parental income. Nevertheless, both studies found childhood intelligence to be a better predictor of adult social class than was parental social class. For example, von Stumm et al. (2010) concluded by noting that childhood intelligence had twice the effect on educational attainment as did parental social class:

The current results confirmed that social class of origin, childhood intelligence, behavior disturbance, and educational qualifications predict social status at midlife. Overall, the predictor variables accounted for almost 50% of the variance in social class of destination. Intelligence, behavior disturbance and social class of origin were inter-correlated predictors of educational and status attainment. After controlling for these associations, the effects of intelligence on education were twice as strong as those of social class of origin (standardized path weights of .46 and .23, respectively). Also, intelligence had a greater impact on status attainment than social class of origin, albeit the difference was not quite as large compared to their respective effects on education. Education had the greatest effect on social class at midlife and partially mediated the effects of intelligence and social class of origin on status attainment. The effects of behavior disturbance on status attainment at midlife were fully mediated by education.

For other papers concerning roles of parental social status and childhood cognitive ability on adult social status in the United Kingdom, see debates by Saunders (1997), Breen and Goldthorpe (1999), Saunders (2002), Breen and Goldthorpe (2002), Bukodi et al. (2013), Bourne et al. (2018), Betthäuser et al. (2020), Marks (2020), and Betthäuser et al. (2020). However, these papers don’t all directly address the relative influence of parental social status and childhood cognitive ability. Many of them concern other questions, such as the degree to which the influence of parental social status is mediated through childhood cognitive ability, whether the United Kingdom is a meritocracy, etc.

Other Countries

Sweden

Sorjonen et al. (2012) analyzed data from a population-representative of nearly 50,000 Swedish males born between 1949 and 1951 who were conscripted for compulsory military service in 1969 or 1970 (~age 20). This includes almost all Swedish men of this cohort, as only 2-3% of Swedish men were exempted from conscription. The variables are described as follows:

• Socioeconomic background (SEB) was measured based on father’s income from tax records, father’s occupational status (e.g. unskilled worker, farmers, etc.), and subject-reported rating of family economy during upbringing (5-point scale from “very poor” to “very good”).
• Cognitive ability was measured at conscription via various IQ tests measuring logical intelligence, verbal intelligence, spatial intelligence, and mechanical ability.
• Emotional capacity was also measured at conscription via interactions with a psychologist and medical records.
• Socioeconomic outcomes were measured in terms of educational attainment in 1990 and occupational status and income in 1985 and 1990 when participants were in their late 30s to early 40s.

The authors constructed a model where socioeconomic outcomes are predicted from intelligence, SEB, and emotional capacity, with direct paths from the predictors and indirect paths mediated through educational attainment. The model ended up explaining 42% of variation in education, 51% of variation in occupational position, and 13% of variation in income.

The path coefficients here show the estimated direct associations between the variables. For example, the path from intelligence to occupation estimates the direct effect of intelligence on occupation, i.e. the effect that is not mediated through education. As you can see from the graph, intelligence has a much stronger direct effect on both educational attainment and occupational status than socioeconomic background did. As the authors note, “Intelligence has a stronger effect on level of education and occupational position compared with socioeconomic background and emotional capacity” (page 273).

When incorporating the indirect effects of intelligence and socioeconomic background on occupational position and income (i.e., incorporating the effects mediated through education), intelligence again has a much stronger effect. The magnitudes of the direct and indirect effects of intelligence and socioeconomic background on socioeconomic outcomes are presented in Table 2.

As you can see, the total effects of intelligence and socioeconomic background on education, occupational position, and income are far greater than the total effects of socioeconomic background. The total effect of intelligence is 2 to 3 times greater than the total effect of socioeconomic background for education and occupational position. For income, socioeconomic background actually has no total effect on income. Thus, the patterns of success for Swedish males born in the middle of the 20th century seem similar to that in the United States and the United Kingdom: in order to predict the future socioeconomic success of Swedish male youth, it is more important to their intelligence than the socioeconomic success of their father.

The authors note the following when discussing the main results:

The model in Fig. 2 explains 42.0% of the variation in education, 50.7% of the variation in attained occupational position, and 12.8% of the variation in income. The direct, indirect and total effects of the four predictors – intelligence, socioeconomic background, emotional capacity, and level of education – on level of education, attained occupational position, and income are presented in Table 2. Intelligence has a stronger effect on level of education and occupational position compared with socioeconomic background and emotional capacity. Because intelligence has a substantial indirect effect on attained occupational position via level of education, the total effect of intelligence is on par with the total effect of education, although education has a stronger direct effect.

The effects on income are substantially weaker than those on attained occupational position, except for emotional capacity (Table 2). Emotional capacity is actually a more important direct predictor of income than intelligence. However, as intelligence has a stronger indirect effect on income, via level of education, this difference ceases to be significant when looking at the total effects. Socioeconomic background has a weakly negative direct and a weakly positive indirect effect on income, which results in a non-significant total effect. Level of education has the strongest total effect on income, although the effect is not significantly different from the total effect of emotional capacity.

Norway

Kristensen et al. (2009) analyzed the relative influences of parental income, parental education, and youth cognitive ability on educational attainment in young adulthood in a very large sample of Norwegian males. In fact, the study included data on all males born in Norway in 1967 to 1971 who were alive at 28 years of age (N = 160,914). Due to compulsory military service at the time, these men were required to complete a test of general cognitive ability as a conscript examination, usually when they were 18 years old. Path analysis was conducted including parental education, parental income, general ability, and educational attainment. The findings revealed that general ability had by far the largest direct association with educational attainment (Figure 2):

The findings here are consistent with the studies reported above. In this case, the direct effect of cognitive ability on educational attainment was about 3 to 5 times the direct effect of parental education or parental income. The authors conclude with the following:

In this large population, comprising all male Norwegian residents born in 1967–1971, explained variances in the regression models and the path model results suggest that status attainment in the study population was dominated by general ability, parental education level and parental income level, in that order. Other parental or individual characteristics had only marginal influence.

Germany

Becker et al. (2019) analyzed the relationship between parental socioeconomic background, childhood intelligence, and socioeconomic success in a longitudinal study of a nationally representative sample of over 5,000 German children. The study began collecting information on the children in 1991 when participants were in the 7th grade (around 12 years of age). Socioeconomic success was measured in the 7th wave in 2009-2010 when participants would have been in their early 30s. Cognitive ability was measured in the first 2 waves when children were still in the 7th grade. Parental socioeconomic background was measured based on parental occupational prestige and parental education. The zero-order correlations for the variables are presented as follows:

The above table shows childhood intelligence to be a better predictor than parental socioeconomic background of educational attainment (r = 0.61 vs r = 0.43) and income (r = 0.28 vs r = 0.15). However, childhood intelligence was not a better predictor of occupational status than was parental socioeconomic background (r = 0.42 and r = 0.43). In order to compare the unique associations of childhood intelligence vs parental socioeconomic background with socioeconomic success, the authors conducted various regression analyses with childhood intelligence and parental socioeconomic background (among other variables) as the independent variables:

Model 1 is the relevant model to focus on as it only includes gender, childhood intelligence, and parental socioeconomic background as the independent variables. Thus, it does not control for mediators that might introduce overadjustment bias. This model shows that childhood intelligence has a significantly greater association with educational attainment and income than does parental socioeconomic background. In fact, parental socioeconomic background has no statistically significant association with income.

However, for adult occupational status, the coefficient for childhood intelligence is not significantly different from the coefficient for parental socioeconomic background (0.32 vs 0.24, and the confidence intervals overlap substantially). Interestingly, childhood intelligence no longer has a statistically significant association with occupational prestige or income after controlling for educational attainment and/or grades, which suggests that the influence of childhood intelligence on career success may be almost entirely mediated by academic success in Germany. That is, smarter German children go on to perform better in school, acquire higher levels of education, and consequently acquire jobs with more prestige and higher pay. But independently of academic success, intelligence does not seem to be beneficial to career success in Germany. This differs from the U.S. and Swedish data.

To represent the direct vs indirect effects of childhood intelligence and parental socioeconomic background, the authors conducted a path model to represent the pathways through which these factors influence socioeconomic success:

As you can see, the findings suggest that childhood intelligence has no statistically significant direct effect on occupational status or income in Germany. Rather, childhood intelligence has large direct effects on academic success (much larger than the direct effect of parental socioeconomic background), which in turn has large effects on occupational status and income.

In the discussion, the authors noted the following on the relative effects of different predictors of socioeconomic success:

In terms of the total effects of the continuous predictors, childhood intelligence showed the highest correlation with education, followed by adult occupational status, and then income—a pattern similar to that emerging from Strenze’s (2007) meta-analysis (see also Marks, 2015). The same pattern emerged for parental socioeconomic background, but the correlation for each indicator was lower. However, education was the predictor most closely correlated with both adult occupational status and income. The results thus support Schoon’s (2008) conclusion that “time spent in full-time education is the strongest predictor of adult social position” (p. 79) for Germany as well. The multivariate regressions revealed that childhood intelligence and parental socioeconomic background both predicted all three outcomes, even when we controlled for the other predictors. Childhood intelligence was again the most important predictor. Yet most of the effects of childhood intelligence and parental socioeconomic background on adult occupational status and income proved to be mediated by education (both quantity and quality). Only direct parental background effects on adult occupational status remained statistically significant. Income seemed to be entirely independent of family background when we controlled for intelligence and/or education. These effects persisted when we controlled for professional sector, working part time, or having children, which implies that family background is related to whether an individual aspires to a more prestigious position, but not so much to how well the position pays.

Other findings

Other studies predicting socioeconomic outcomes

Here are other studies that I found. I didn’t cover them in detail for a variety of reasons. Either they were redundant, covered subjects born very far in the past (e.g. born before 1950), and/or had low samples.

• Erikson (2016) studied 28,000 Swedish school children and examined the relationship between parental SES, cognitive ability (age 13), and highest level of education (age 32-40). The sample included children from cohorts born during 1948, 1953, 1967, and 1972. In line with the findings in Sweden reported by Sorjonen et al. (2012), the results of this study showed that cognitive ability accounted for more variance in socioeconomic outcomes than did parental SES. I preferred Sorjonen et al. (2012) because it included a broader range of socioeconomic outcomes (i.e. it also included occupational attainment and income) and it provided standardized effect sizes which enable more natural comparisons of independent variables.
• Spengler et al. (2015) studied a sample of 745 Luxembourg children between 1968 (age 12) to 2008 (age 52). Surprisingly, the results revealed that youth IQ and parental SES had similar associations with career success. Unfortunately, the parental SES was rather limited. Parental SES was measured simply based on the children’s description of their parents’ occupations. There was no measure of parental income or parental education.
• Johnson et al. (2010) studied the relationship between participant’s father’s social class, cognitive ability (measured at age 11), social class, and participant’s offspring social class among 238 Scottish participants born in 1921. The participants were followed up at age 79 as part of the Lothian Birth Cohort 1921 Study which began in 1999. Also, see Deary et al. (2005) on the same dataset. The parental SES measure was limited, as it was based only on the father’s occupational class.
• Johnson et al. (2010) is similar to the previous study except it examines the Lothian Birth Cohort 1936.
• Strenze (2006) studied the relationship between youth cognitive ability, parental SES, and offspring SES in both the United States. The data for the United States came from the NLSY79 and findings were similar to those reported above, whereas the findings for Estonia showed parental SES to be much more influential there (comparable to that of youth cognitive ability), compared to the United States.

Miscellaneous findings

This section is for some additional findings that I came across while researching this information that don’t have a natural place in the post.

While performing research for studies for this post, I also found numerous studies showing that among individuals who ended up in a particular social class, the individuals with higher-class parents were typically more intelligent than individuals with lower-class parents (Nettle 2003, Figure 4; Deary et al. 2005, Table 1; Johnson et al. 2010, Table 1, Table 3; Johnson et al. 2010, Tables 3-4; Sorjonen et al. 2011, Table 1, Sorjonen et al. 2021). For example, Nettle 2003 reports the following:

There is no evidence of a `Robbins effect’ on class mobility. Those entering the professional class from lower social strata do not have higher GA scores than those from more privileged backgrounds. In fact the opposite is the case; within four of the five classes, those from more privileged social backgrounds had higher GA scores than those from less privileged ones. The most likely explanation for this is that within the broad occupational groups, those from more privileged backgrounds are tending to occupy the highest level and most difficult to enter positions, which would tend to be related to the higher GA levels.

These findings are likely to be due at least in part due to residual confounding, since the measure of cognitive ability is not perfect, as explained by Sorjonen et al. (2021):

It is common to adjust for potential confounding variables by including them as covariates in an analysis, in order to reduce the risk of spurious associations. However, the influence of the confounding variable may not be fully attenuated by such adjustment [12,13,14,15,16]. Residual confounding refers to confounding which remains despite adjustment. The impact of residual confounding is increased by higher true degree of confounding, larger sample size, and higher reliability in the measurements of X and Y, while it is attenuated by a high reliability in the measurement of Z [12,13,14,15,16]. With these factors in place, even if entities/individuals have the same value on observed Z they will tend to differ in their true Z and this may result in an association between observed X and observed Y even if adjusting for observed Z.

I also found one study comparing the predictive validity of parental SES and cognitive ability for job performance. Kuncel et al. (2014) recorded data on the cognitive ability and parental SES of 108 job incumbents and measured their relationship with supervisor-rated job performance. Cognitive ability was measured as a composite from scores on two tests covering various skills such as deduction, inference, numerical reasoning, as well as others. Parental SES was measured based on a composite of mother’s education, father’s education, and early family SES (a 5-point scale from “poor” to “wealthy”). Job performance was measured using a 26-item questionnaire. The ratings on 24 of the items were averaged together to create a composite measure of job performance Average Performance. On the remaining 2 items, one item provided a direct rating of overall performance and the other provided a rating of overall potential.

Conclusions

Cognitive ability is a much better predictor than parental SES

The purpose of this post was to analyze the relative predictive validity of youth cognitive ability vs parental SES on socioeconomic outcomes. When analyzing zero-order correlations without controlling for any other variables, some of the studies show that cognitive ability and parental SES have roughly similar associations with socioeconomic outcomes. However, when analyzing the unique associations between each predictor and socioeconomic outcomes (e.g., with regression models), the studies converge on the conclusion that cognitive ability is a far better predictor of socioeconomic outcomes than is parental SES. In this section, I will summarize the standardized coefficients from the main studies considered in this post.

Let us begin with the standardized coefficients for educational attainment:

• Note: the coefficients in the United Kingdom are likely relatively low because these indicate the associations of cognitive ability and parental SES on occupational attainment while controlling for educational attainment. These coefficients are the direct path coefficients from cognitive ability and parental SES to occupational attainment, which means they represent the effects that are not mediated through educational attainment. We should expect these coefficients to be low, given that much of the effects of cognitive ability and parental SES on occupational attainment are in fact mediated through educational attainment.

Aside from Germany, the findings here are similar to the findings for educational attainment. That is, in the United States, United Kingdom, and Sweden, the standardized coefficient for cognitive ability on occupational attainment was about 2 to 3 times the corresponding coefficient for parental SES. In Germany, the coefficient for cognitive ability is only about 33% greater.

Income

The findings in the United States for income are similar to the prior findings on educational and occupational attainment. That is, the standardized coefficients for cognitive ability were about 2 to 3 times the coefficients for parental SES. In Sweden and Germany, cognitive ability had a modest association with income whereas parental SES had no statistically significant association with income.

In short, if you want to predict a child’s future educational attainment, occupational attainment, or income, it is much more important to know the child’s cognitive ability than their parents’ education, occupation, and income. A child’s future success is much more predicted by their own intelligence than the success of their parents.

Cognitive ability increases in importance over time

A common pattern that emerged from different datasets is that youth cognitive ability is a better predictor of one’s long-term socioeconomic outcomes than outcomes that emerged shortly after secondary school. For example, several studies showed that while cognitive ability is a better predictor than parental SES of income in late adulthood, it is not a better predictor of income in early adulthood:

Education is the great mediator

Much of the effect of cognitive ability and parental SES on occupational attainment and income was mediated through educational attainment. That is, cognitive ability (and parental ability) improve occupational attainment and income in part indirectly. For example, more intelligent individuals acquire more education which results in more prestigious jobs with higher pay. However, the degree to which education mediates these effects varies slightly by country.

To summarize the degree to which education mediates the effect of cognitive ability and parental SES, I’ve gathered effect sizes before and after including controls for education. The percent reduction in the effect size indicates the percentage of the association that is mediated through educational attainment. The effect size before controlling for education is often said to measure the total effect whereas the effect size after controlling for education is said to measure the direct effect.

For example, if cognitive ability has a standardized coefficient of 0.30 in a model without educational attainment as an independent variable, and this coefficient drops to 0.10 after including educational attainment in the model, this suggest that about two thirds of the association between cognitive ability and the dependent variable of the model is explained by education. To calculate this, it is not necessary to standardize the effect sizes, since I’m mainly concerned with the percentage reduction, which will be the same regardless of standardization.

Here are the coefficients for cognitive ability predicting occupational attainment before and after introducing controls for educational attainment:

Occupational attainment regressed on cognitive ability

In all of the datasets, Most of the effect of cognitive ability on occupational attainment (50 to 70%) was mediated through educational attainment.

Now, consider the coefficients for parental SES predicting occupational attainment:

Occupational attainment regressed on parental SES

• Note: for the Marks (2022) data, I included the effect sizes for parental education and family income separate by |.

Roughly half of the effect of parental SES (as a composite) on occupational attainment was mediated through education. Nearly all of the effect of parental education was mediated through offspring education. And 30 to 50% of the effect of parental income was so mediated.

Now, consider the coefficients for cognitive ability predicting income:

Independent of human capital, parental SES weakly predicts career success

Another important finding to note is that parental SES has very small associations with career success (occupational attainment and income) after controlling for human capital (offspring cognitive ability and education). The standardized coefficients of the career success measures regressed on parental SES were consistently small when including controls for human capital (e.g., β ~ 0.10 or lower). Note that these are equivalent to the standardized coefficients reported in the previous tables for parental SES after introducing controls for educational attainment. Consider the following:

Career success regressed on parental SES, controlling for human capital

Notes:

• For the Marks (2022) rows, I only included the coefficients for parental income, because parental education had no statistically significant positive association with either occupational attainment or income.
• For K&W (2000), the standardized coefficient for parental SES was calculated by dividing the coefficient for zSES on annual earnings (910, Table 7.2) by the standard deviation of annual earnings (16,083, Table 7.1), which is 910/16,083 = 0.06.

The findings here are clear. When controlling for offspring cognitive ability and educational attainment, parental SES has a very weak association with offspring occupational attainment or income. In fact, only 2 of 13 coefficients considered above surpassed 0.10 which is considered a small effect size.

The finding that parental SES has small effects on economic success has been replicated in other studies as well.

• In a nationally representative sample of 8,901 8th graders in the National Education Longitudinal Study, Rumberger et al. (2009) [archived] reports a standardized coefficient of parental SES on adulthood earnings of just β = 0.11 after accounting for the effects of educational attainment, 8th-grade grades, and 8th-grade test scores (Table 1).
• Kingston (2006) also conducted a review of studies comparing the predictive validity of parental SES, educational attainment, and cognitive ability on socioeconomic outcomes in the United States. He reviews additional studies showing that the net effect of family SES on occupational status and earnings is “minor” after controlling for education and cognitive ability (see page 121).

The causal effect of parental SES on career success is likely very weak

What I mean by this is that, in developed populations, variation of parental SES is likely causally responsible for very little of the variation of career success.

The previous section shows that parental SES has a very weak association with career success when controlling for human capital. Since much of even this association is likely not causal (due to likely confounding, that I mention below), this implies that the causal effect of parental SES on career success is actually very weak (i.e. even weaker than the small associations imply).

• Most of the variation in cognitive ability is caused by variation in genetics, rather than variation in environmental factors. For example, a review of the literature on genetics and intelligence differences reports that up to 80% of the variation in intelligence is due to genetic differences in late adulthood (Plomin and Deary 2015). Thus, only a small proportion of the variation in cognitive ability can be explained by environmental factors. Furthermore, most of the effect of environmental factors comes from unshared environmental factors rather than shared environmental factors. Because parental SES belongs to the shared environmental factors, that leaves very little room for parental SES to explain variation in cognitive ability.
• Much of the association between cognitive ability and socioeconomic outcomes is due to genetic factors. For example, one study finds that 59% of the association between cognitive ability and educational attainment is actually due to shared genetic factors rather than shared environmental factors (Marioni et al. 2014). Thus, only a minority of the association between cognitive ability and educational attainment is explained by shared environmental factors, which limits the degree to which parental SES can have an effect on educational attainment mediated through cognitive ability.
• Much of the association between parental SES and offspring cognitive ability is due to genetic factors, because high-SES parents are likely to have higher levels of intelligence and individuals with high intelligence have higher intelligence largely due to genetics (see Plomin and Deary 2015). Thus, high-SES parents will have offspring with higher cognitive ability, not just because they have offered superior environments, but also because they pass on better genes for cognitive ability. In fact, in a DNA analysis of thousands of twins in the United Kingdom, Trzaskowski et al. (2014) have reported that “the same genes are largely responsible for genetic effects on family SES and children’s IQ”.
• In addition to genetic confounding, the association between parental SES and offspring cognitive ability is probably also due to other forms of confounding. For example, as already noted, high-SES parents tend to be smarter than average. If there are environmental benefits to having smarter parents (independent of socioeconomic background), then part of the association between parental SES and academic achievement will be due to environmental confounding with parental cognitive ability. There is some evidence that such confounding actually exists. For example, Marks and O’Connell (2021) find that the effect of parental SES on test scores decreased by 50-60% after controlling for a measure of maternal cognitive ability. Similar points apply to any parental traits that may cause both high SES and smart offspring (e.g., parental intelligence, parenting styles, personality, etc.).