What a Skill Score Can—and Cannot—Tell Us
The subject here is the evidence behind the math–verbal divide: how a study turns school performance into measures of mathematical and verbal skill, follows those measures into adult outcomes, and how a separate survey measures skills later in life. The practical question is how to read the resulting estimates without confusing a test rank with an innate trait, an association with a guarantee, or a US adult statistic with evidence from English school cohorts.
That discipline matters because the striking contrast in the results could otherwise be dismissed in two opposite ways. A reader might assume that the numbers reveal a simple, universal truth about which kind of person succeeds. Or they might decide that the findings are too technical to interpret. The better approach is to understand the observation chain: what was measured, when it was measured, what outcome came later, and what assumptions connect the two.
Following one English cohort from school into work
The core study uses the UK Longitudinal Education Outcomes dataset. It focuses on people who completed their GCSE examinations in England between 2002 and 2004 and follows their educational records and employment outcomes into their early thirties. The descriptive sample contains almost 1.3 million people. That scale matters because the findings are not based only on a small group of university graduates or people already in work. They include people who stopped after secondary school as well as people who continued through higher education.
The study’s two main skill measures begin with ordinary school assessments. GCSE grades in mathematics are converted to a numerical scale and then placed in percentile ranks within each exam cohort. English is built from literature and language grades, then ranked in the same way. A percentile rank is a relative position: a student at the 70th percentile performed above roughly 70% of the relevant exam cohort on that measure. It does not say that the student possesses 70% of a fixed amount of mathematical or verbal ability, nor does it make scores from different tests automatically identical.
Using ranks helps the researchers compare students within the same examination period. It also gives the later coefficients an intuitive, if limited, interpretation. When the paper reports a coefficient on a skill rank, it is asking how an outcome differs as a person’s position in the cohort distribution changes, holding the listed controls constant. A coefficient of 0.1 does not mean a ten-percent increase in earnings or a ten-percent increase in intelligence. The outcome determines the unit: for a college-attendance indicator it can be read as a percentage-point relationship after the appropriate scaling; for log pay it is interpreted as an approximate percentage change in earnings.
The timing of the measurements is central. GCSEs are taken around age 16, before most of the labour-market outcomes under study. But GCSE grades can contain noise. They may reflect temporary circumstances, imperfect tests, or factors that are not the underlying subject-specific skill the researchers want to capture. There is also a reverse-causality concern in many skills-and-earnings studies: work can change what people know, and non-employment can allow skills to fade. If the skill measure is taken at the same time as earnings, it becomes difficult to know which direction the relationship runs.
To address these problems, the study uses Key Stage 2 assessments taken around age 11 as instruments for the corresponding GCSE ranks. In plain language, the earlier assessment acts as a source of predetermined variation that helps recover the part of the later GCSE measure connected to earlier subject performance. The design uses the earlier mathematics assessment for the later mathematics measure and the earlier English assessment for the later English measure. The strong relationship between the two stages supports the instrument’s relevance in the reported first-stage results.
That procedure improves the temporal ordering and reduces the influence of random measurement error, but it does not make the assumptions disappear. The interpretation requires the earlier scores to be connected to later outcomes through the relevant skill channels in the way the model assumes, once the included controls are considered. It is therefore more accurate to say that the study estimates skill relationships under an instrumental-variables design than to say that it has observed a pure, context-free effect of a school subject.
The researchers then connect the skill measures to several kinds of outcome: A-level participation, college attendance, graduation, degree classification, postgraduate study, employment, and earnings. They observe earnings annually for the relevant cohorts, with special care over which income sources are available at different ages. The main adult comparison uses people observed between ages 30 and 34, while the lifecycle analysis follows pay patterns from 21 to 34. The sample does not include people who emigrated or people who did not sit their GCSEs in an English school. Those exclusions do not invalidate the study, but they define the population to which its results most directly apply.
This sequence is the study’s main advantage over a snapshot that measures skills and wages at one moment. It lets readers see a plausible order from early performance to later outcomes. It still does not establish that raising one person’s rank by a particular amount would produce the exact average change in their future. The estimate belongs to the population, period, controls, and assumptions that generated it.
A different window: adult skills in the United States
The OECD Survey of Adult Skills answers a related but different question. In 2022–23 it assessed US adults aged 16 to 65 in three domains: literacy, numeracy, and adaptive problem solving. The assessment was delivered on digital devices and used tasks intended to reflect how skills are applied in everyday and working situations. Proficiency is reported on scales with defined levels, so a level describes the kinds of tasks an adult is likely to complete, not merely a position in a school exam cohort.
The closest conceptual matches are literacy and numeracy, but neither is identical to the English study’s measure. GCSE English is a school subject measure constructed from literature and language grades. OECD literacy concerns understanding, evaluating, and using written information in adult tasks. GCSE mathematics and OECD numeracy overlap in quantitative reasoning, but adult numeracy includes practical work with numbers, rates, ratios, tables, and graphs. Adaptive problem solving is an additional domain involving changing conditions and multiple variables; it has no direct counterpart in the two-skill school comparison.
The US results are useful for showing the distribution of adult capability. Average US performance in 2022–23 was 258 points in literacy, 249 in numeracy, and 247 in adaptive problem solving. Numeracy was below the OECD average, and 34% of US adults scored at or below Level 1, compared with 25% across the OECD. At that level, adults can handle basic whole-number or money tasks and locate single pieces of information in a chart, but may struggle with a problem requiring several steps, such as calculating a proportion. This is a population description, not a statement about a particular person’s potential.
The survey also provides a wider view of change over time. US literacy and numeracy results in 2022–23 were lower than in 2012/15, with the negative trend concentrated more strongly at the lower end while the share of high performers remained stable. The distance between the highest- and lowest-performing adults consequently widened. Adults aged 55 to 65 scored lower than adults aged 25 to 34 in all three domains; for literacy the gap was 17 points. But the OECD cautions that this difference can reflect both ageing and differences between generations in education and training. A comparison between age groups is not automatically a pure measure of what ageing does to an individual.
The survey’s repeated-cycle comparisons illustrate the same caution. A younger birth cohort scored four literacy points lower in 2022–23 than in 2012/15, a change that was not statistically significant. An older cohort scored 19 points lower when its members moved from ages 44–54 to 55–65, a significant decline. The pattern is consistent with age-related skill loss, but it is not a controlled experiment. Survey comparability also requires care because the OECD generally excludes adults who received only a doorstep interview when comparing cycles, since that option was not available in the first cycle.
The OECD note then relates adult skills to other outcomes. In the United States, after accounting for education and other relevant differences, a one-standard-deviation increase in numeracy is associated with a seven-percentage-point higher likelihood of labour-force participation. Among employed adults, the same numeracy comparison is associated with wages 11% higher. These figures show that measured skills can retain a relationship with work outcomes beyond formal qualifications. They do not show that adding a standard deviation of numeracy to an individual’s score would mechanically raise their wage by 11%.
The comparison unit also matters. The note reports a 21% wage association for a one-standard-deviation increase in education. That is not a simple contest in which education “wins” or numeracy “loses.” The two standard deviations describe different variables, and both may reflect selection, prior resources, job choices, and other factors. The useful conclusion is that formal qualifications do not exhaust the information contained in adult skill measures, not that the survey has priced two interchangeable investments.
The survey adds outcomes beyond work and pay, including self-reported life satisfaction, health, trust, and volunteering. Higher-skilled adults are more likely to report favourable outcomes, and several relationships remain after adjustments for age, gender, education, immigrant background, parental education, and family circumstances. These associations widen the meaning of skill value, but they do not establish a one-way causal chain from proficiency to health or civic engagement. Health, opportunity, confidence, and social conditions can also affect both skills and reported wellbeing.
Finally, the OECD introduces a condition that a school-to-work pathway can otherwise leave implicit: whether a person’s capabilities are actually used. About 26% of US workers are over-qualified for their jobs, 9% are under-qualified, and 38% are mismatched in field of study under the survey’s definitions. Over-qualified workers earn 19% less on average than similarly educated peers in well-matched jobs and are 12 percentage points less likely to report high life satisfaction. These figures concern employed, non-self-employed adults aged 25 to 65 and are associations, not estimates of what would happen if every mismatch were corrected.
That distinction is the evidence boundary to carry forward. The English study is the stronger source for the pathway from adolescent mathematics and English performance through educational decisions, fields, and early-adult earnings in England. The OECD note is the stronger source for contemporary US adult proficiency, distributional change, broader wellbeing, and mismatch. The sources can illuminate the same general question—how capabilities meet institutions and work—but they cannot be merged into one sample or one causal estimate.
A useful reading habit follows. For any claim about the math–verbal divide, ask four questions: Who was measured? When were they measured? What exactly was the skill measure? Which outcome is being estimated? Then ask whether education or field choice is a pathway being studied or a downstream variable that has been controlled away. With those questions in place, the next result becomes more informative: the stronger educational role of verbal skill is not a measurement mistake to be waved aside, but neither is it the whole story of what happens after students enter education.