What the Numbers Actually Measure
This chapter examines how digital-skill wage premiums are measured, because the same percentage can describe very different things. The central task is to distinguish an advertised salary in a job posting from a worker’s realised hourly earnings, and to distinguish a robust association from a causal personal return. Once those distinctions are clear, the figures become useful evidence about jobs and markets rather than promises about what any one learner will earn.
Two datasets, two links in the chain
The World Bank study starts with vacancies. It uses more than 67 million online job postings from 29 countries between 2021 and 2024. Each posting can contain a wage offer, occupation, sector, location, education and experience requirements, work arrangement, and a list of requested skills. This makes it possible to ask a hiring-market question: do postings that request digital skills offer different advertised wages from otherwise comparable postings?
That is a valuable question, but it is not the same as asking what a worker is paid after being hired. An advertised wage is an offer attached to a role. It may be negotiated, changed, or never accepted. The posting may describe an ideal candidate rather than the tasks that occupy most of the eventual employee’s time. The data also cover online vacancies, not every job. The paper notes that online postings overrepresent cities, high-skill industries, and white-collar work; they omit much informal, freelance, local, and unadvertised employment, with especially uneven coverage in lower- and middle-income countries.
The OECD report starts from workers rather than vacancies. Its 2023 Survey of Adult Skills asks people how often they perform tasks involving ICT, reading, writing, numeracy, problem solving, discretion, learning, influencing, co-operation, and other skills at work. It also measures proficiency in important information-processing skills and collects hourly earnings. This allows a different question: among workers with observed individual and job characteristics, is more frequent use of a skill associated with higher wages or better work outcomes?
The OECD measure of use is based mainly on frequency. Respondents report whether they perform relevant tasks never, less than monthly, less than weekly, at least weekly, or every day, and the responses are combined into skill-use indicators. Frequency is a practical way to compare workplaces, but it does not fully capture how difficult, consequential, or valuable a task is. Using specialised software once a week might matter more than opening a routine spreadsheet every day. The measure can reveal whether a capability is activated, not the full economic value of every activation.
The two studies therefore observe different stages. The vacancy data describe what employers request and what they offer. The worker survey describes what people can do, what they report doing, and what they earn. One is closer to demand and recruitment; the other is closer to realised work and worker experience. If their findings point in the same direction, that strengthens the case that digital capability matters in labour markets. It does not make the studies interchangeable or turn either one into a controlled test of training.
The World Bank study organises digital demand along three margins. The extensive margin asks whether a posting requires any digital skill at all. The intensive margin asks how many digital skills it lists. The qualitative margin asks which type of skill is named—basic, intermediate, advanced, traditional AI, or a GenAI category. These distinctions answer progressively sharper questions. A job may need one basic digital skill as a threshold. Another may need several advanced capabilities that change the scope of the role. Treating both situations as simply “digital” would hide the difference.
The diagram's negative arrow matters. Better controls can make a comparison more credible, but they do not by themselves show what would happen if the same worker were randomly given a digital skill and followed over time. The result remains a conditional association: a relationship measured after accounting for selected differences.
Read a percentage with its comparator attached
The World Bank’s first result is easy to misread. In the preferred specification, a job requiring at least one digital skill is associated with a 1.6% higher advertised wage on average. The estimate is about 1.3% in high-income countries and approximately 7.5% in low- and middle-income countries. Those are not three versions of a universal “digital raise.” They are estimates for different comparisons and labour markets.
The size changes as the model becomes more selective. With country, year, sector, and occupation effects, the reported premium is about 3.9%. Adding education, experience, employment type, and work arrangement reduces it to about 2.1%. Adding location-by-year controls produces the preferred 1.6% estimate. The decline does not mean the final estimate is the only “true” number and the earlier ones are useless. It shows that digital vacancies differ from other vacancies in several ways that are also related to pay. High-paying roles may be more likely to use technology, demand more education, offer more autonomy, or appear in expensive cities. A raw wage gap mixes all of those differences together.
The preferred estimate tries to compare postings more narrowly: within the same broad country-location-time environment, sector, and occupation, while also accounting for listed education, experience, employment type, and work arrangement. The remaining relationship is more informative than a simple average, because it asks whether digital requirements still matter among more comparable jobs. It is still not a randomised comparison. Employers choose whether to adopt tools, how to write postings, and what wage to offer. Workers choose whether to apply and accept. Unobserved differences can remain.
The paper tests one important alternative explanation by adding employer fixed effects. In plain language, this compares vacancies from the same employer where possible, reducing the chance that the premium is only a signature of high-paying firms. The reported relationship remains. That is evidence against the narrow claim that firm identity explains everything. It does not prove that a digital requirement itself caused the wage offer. Within-firm vacancies can still differ in responsibility, urgency, team, budget, or unmeasured difficulty, and the estimate applies to the observed posting sample.
The second margin changes the question. Each additional digital skill listed in a posting is associated with about a 0.5% higher advertised wage globally and in high-income countries, compared with about 2.6% in low- and middle-income countries. This does not mean that a worker can accumulate five certificates and claim five increments. A listed skill may be redundant with another, may signal a broader role, or may appear because an employer writes unusually detailed adverts.
The paper finds a concave relationship: the marginal association becomes smaller as more digital skills are added. It remains positive across almost the whole observed distribution and approaches zero only around 37 listed skills, far above the average of about two digital skills per posting. The useful lesson is not the boundary number itself. It is that requirements are a bundle. The first few capabilities may signal entry into a more complex role; later additions may overlap, describe the same workflow, or add less expected productivity than the earlier ones.
The qualitative margin provides another warning against headline reading. A basic digital requirement and an advanced analytical or cloud skill do not carry the same signal, even if both count as one digital skill. The category matters because the comparison is about the role that calls for it, not only the word appearing in the advert.
What robustness supports—and what it leaves open
A robust association is worth taking seriously, but “robust” has a specific meaning here. In the World Bank analysis, digital-skill coefficients remain positive in specifications that add job characteristics, detailed fixed effects, other non-digital skills, and in some cases employer fixed effects. That pattern makes it less plausible that the entire result is caused by one obvious difference, such as digital jobs being concentrated in a particular occupation. It supports the conclusion that digital requirements are systematically related to higher advertised wages in the studied data.
It does not identify the worker-level process behind the relationship. A firm may list a digital skill because it is building a more productive team. The premium may then reflect the whole team, better management, more valuable products, or greater responsibility, with the skill serving as a marker for the package. The paper attempts to account for broader skill bundles, but no set of observed controls captures every difference in technology, management quality, or job difficulty.
Nor does the study tell us that the person who already holds a job will receive the coefficient as a raise. The estimate compares vacancies. It says that postings with a particular requirement tend to offer more, conditional on the model. A worker asking for higher pay would still need to show that their capability changes the work they perform, improves a decision or process, is scarce relative to available candidates, and belongs to a role whose pay-setting rules can recognise the contribution.
The OECD evidence has its own boundary. Its hourly wage measure is closer to realised pay, but skill use is reported by workers and measured mainly by frequency. The relationship between use and wages may also reflect sorting: people in better-paid occupations may have more opportunities to use certain skills, or workers who are already more capable may be selected into jobs that use them. The report controls for occupation, industry, education, proficiency, firm size, contract and other characteristics, but the remaining coefficients are still associations.
This difference is why the two sources should be combined carefully. If the World Bank finds that employers advertise more for digital requirements and the OECD finds that workers who use information-processing skills more often tend to earn more, the combination suggests a coherent chain from demand to use to pay. It does not prove every link for every person. The stronger interpretation is comparative: digital capability is more likely to be rewarded when it is attached to work that values and deploys it.
The country comparison also requires a denominator. A 7.5% association in an LMIC and a 1.3% association in a high-income country are percentages applied to different wage structures, currencies, labour institutions, and samples. A percentage is not an absolute amount. Nor does it directly measure purchasing power or living standards. We need to see both the relative premium and the underlying salary level before judging who is economically better off.
That point becomes sharper in the reported averages. Advertised annual salaries across the sample average roughly $58,811 in high-income countries and $7,590 in low- and middle-income countries. Jobs requiring at least one digital skill average about $65,736 versus $9,389. A larger proportional premium in a lower-income market can therefore coexist with a much smaller dollar addition. Scarcity can make a capability more valuable relative to the local baseline without erasing the international wage gap.
A careful reader should therefore keep five questions beside every premium claim:
- What is the outcome? Is it an advertised salary, accepted compensation, hourly earnings, promotion, productivity, or something else?
- What is the comparison? Is the estimate against jobs without any digital requirement, against similar jobs in the same occupation, or against another skill category?
- What is the skill object? Is it proficiency, a requirement in text, or reported use? Is it basic, advanced, traditional AI, or GenAI development or literacy?
- What context is held constant? Are occupation, sector, place, time, education, experience, work arrangement, employer, or other skills included?
- What remains outside the design? Does the sample omit informal work? Could the skill be a proxy for firm quality or role scope? Is the result correlational, self-reported, or based on a new and sparse category?
These questions do not make the evidence evasive. They make it portable. A number stripped from its outcome and comparator can support almost any story. A number kept with its design can guide a more specific judgment: the market appears to value this requirement in these kinds of roles, under these conditions, in this observed sample.
The next question is the mechanism
Measurement can tell us where a premium appears. It cannot by itself explain why. The remaining question is what happens inside the work. Does the digital capability reduce a bottleneck, improve a decision, coordinate people, or allow a firm to produce something it could not produce as effectively before? Does the worker have the domain knowledge and authority to act on the result? Does the organisation provide enough discretion and learning for the capability to be used repeatedly?
Those questions matter because the same requirement can be a threshold in one job and a differentiator in another. A spreadsheet skill may be routine in a finance team but transformative in a small organisation that previously had no reliable reporting. An AI literacy requirement may describe superficial tool familiarity in one vacancy and the ability to redesign a high-value professional workflow in another. The label alone cannot resolve the difference.
The evidence has now narrowed the claim. Digital-skill requirements are associated with higher advertised wages, and more frequent use of several information-processing skills is associated with higher hourly wages. The estimates survive meaningful controls, but they are not personal raise schedules. To understand the premium, we must next follow the capability into the task, production system, and role where its value is either realised or left unused.