Details Matter: How Skills Influence Economic Recovery

For a long time, the labour market has been studied in isolation from other macroeconomic indicators. This has led to a paradox in contemporary economic literature. On the one hand, we know quite a lot about processes within the labour market: how employees are hired, how salaries are agreed upon, what happens to employee skills over the course of unemployment, and how new hires undergo onboarding, to name but a few. On the other hand, we have very limited knowledge of how specific processes in this market affect other macroeconomic indicators such as GDP or inflation.

In the large macroeconomic models (e.g. DSGE models), which are commonly used to support decision-making at central banks and ministries around the world, the labour market is assumed to consist of the average employee and the average employer. The employee performs certain work for the employer and receives a salary for it. However, experience suggests that the labour market is in fact far more complicated.

My recent study shows that the realities of the labour market, including industry requirements for skills, professional mobility, or the loss of competencies during unemployment, can have a long-term negative impact on the economy. If ignored, these factors lead to the systematic underestimation of the consequences of economic crises. We may have a different outlook for economic recovery if industry competencies are taken into account. Furthermore, targeted retraining programmes may add momentum to an economic recovery.

Key role of skills in labour market

DSGE models assume that workers are interchangeable in the labour market. But in real life, companies are very discerning about who they hire. Consider HeadHunter or any other job aggregator. Each vacancy there comes with a list of skills the employer expects from the future employee. It is intuitively clear that the skills required are related to the industry. For example, digital skills are central to information technology, while personal interaction skills are indispensable in education.

This intuitive conclusion is supported by employee surveys, such as the European Skills and Jobs Survey (ESJS), which uses 21 skills, from teamwork ability to heavy lifting. Workers in the 27 EU member states were asked: ‘How often do you use this skill in your work?’ This is how each industry’s need for a specific skill is quantified based on the percentage of answers of ‘Often’ or ‘Very often’.

We can now discard the unrealistic idea that all employees are the same to an employer. In fact, each firm has certain needs for certain skills, depending on the industry, and aims to hire employees who correspond to them. For employees intending to switch industries (this is called ‘occupational mobility’), this means that the more their skills align with the target industry, the greater the chances of success. This brings us a matrix of possibilities for transition from industry to industry, which will identify opportunities for occupational mobility.

Current research finds that skills are generally ‘tied’ to labour market processes (1, 2). Consider the fundamental work by Jacob Mincer, one of the founders of modern labour economics and human capital theory, written in 1982 in collaboration with Haim Ofek. The authors show that real wages after a work interruption are lower than at the time of dismissal, and the longer the interruption, the greater the decline. They link this effect to the loss of skills during unemployment.

This link has been confirmed by an overwhelming number of empirical studies (for example, 1, 2). Admittedly, one study on German and US data fails to find a decline in the skills of the unemployed, but that is probably due to the fairly basic skills the study is based upon, such as mathematical skills, speech fluency and conscientiousness.

As the ESJS survey suggests, many industries require more specific skills, such as implementing new approaches or using digital instruments, and job applicants who lose such skills may struggle to become re-employed. Thus one can conclude that not only skills matter, but it is essential to have data on a sufficiently broad range of skills.

Other research focuses less on the loss of skills than on their acquisition. A large body of studies document the success of on-the-job training, internships and onboarding programmes (1, 2, 3, 4). Perhaps the most striking experiment took place in Uganda, where the authors of the study helped to organise training for 1,700 workers and 1,500 firms. The workers were randomly assigned to one of three groups: 1) the control; 2) on-the-job trainees; and 3) trainees under government programmes. Occupational skills were assessed using a bespoke test, and data were collected over four years. The authors found that, relative to the control group, the workers who received training, regardless of the type of programme, showed significant and broadly similar gains in industry skills and markedly better employment outcomes on average over the three years after the intervention.

All models are wrong, but some are useful

Do we lose important information by neglecting this knowledge in large macroeconomic models? Undoubtedly we do. But the more important and less obvious question is whether it matters for the answers we obtain from DSGE models.

In the 1970s, UK statistician George Box set out one of the most important principles for building models: “The statistician knows, for example, that in nature there never was a normal distribution, there never was a straight line, yet with normal and linear assumptions, known to be false, he can often derive results which match, to a useful approximation, those found in the real world”. What he meant was that every model is a simplification, but some nonetheless improve our understanding of a case and help answer specific questions.

At first glance, it might seem that the more mechanisms and data a model uses, the better. In practice, when ever more components are added, it is increasingly harder to interpret the results and to understand the mechanisms behind them and the proportions in which they drive the results. So mechanisms need to be added selectively and deliberately.

My research sets out to establish whether all these specific labour market processes matter for questions that may at first glance appear unrelated to them. Suppose, for instance, there is a temporary large-scale shutdown of factories (as a result of a natural disaster or epidemic): what level of GDP is expected several years later? Will the conclusions differ depending on how the labour market is set up in the model used for these calculations?

This question is difficult to answer using the gold standard of macroeconomic modelling – DSGE models. Their underlying ‘average’ economic agents, rational expectations and necessary return to equilibrium after a shock make it hard or even impossible in certain cases to build behavioural patterns into them.

At the same time, agent-based models are gaining traction in macroeconomics, and they can incorporate microdata and simulate 'imperfect', heterogeneous agents (1, 2, 3). Drawing on survey and experimental data on how real people behave in the economy, these models dispense with the representative ‘average’ agent in favour of synthetic populations made up of many heterogeneous agents with the characteristics of real people. This makes it possible to model a country's population at a scale of 1:1.

The origins of two approaches

The fundamental difference between DSGE and agent-based models lies in their development history.

DSGE models grew out of the idea of extending the formal methods used in physics to economics. On the one hand, the 1874 publication of Elements of Pure Economics by Léon Walras, the creator of general equilibrium theory, in many ways marked the transformation of economics into a genuine science built on rigorous formal mathematics – without which it is hard to imagine any good research today. On the other, as economics has evolved, economists have come to understand that mechanisms suited to describing physical phenomena and processes are far from always fit for describing unique human beings endowed with emotion and judgement.

In many physical models, for example, the mechanism is in a state of rest and requires an external shock to 'set it going'. But in biological systems, the catalyst for change is often born within the system itself.

Let us imagine a hypothetical experiment. Assume we have a room completely isolated from any external influence. What happens if we put five machine tools in it? Opening the door a week later, we will find them in the same place with nothing in the room changed. But assume we leave five people in the room for a week – the outcome is far less obvious. In a week, there will be changes visible both in the people and in the room. What is more, the changes will depend on which people are picked for the experiment. Moreover, the result is unlikely to be the same should we repeat this experiment even with the same people.

Another important consideration: the outcome (such as the state of the room) will be a consequence of all the people’s actions, and it will not be the same as if they were left in the room one at a time. They might, for instance, unite their knowledge and efforts to rearrange the room to suit themselves, or even find a way out of the room altogether. This property of a system is called 'emergentness' – when the agents in the system (the people in our experiment) may jointly create something new, something they could not create individually.

Essentially, the economy is such a system: very many people acting in their own interests, periodically joining forces and striking deals, with the aggregate indicators (GDP, for one) forming out of the actions of the individual agents. This knowledge has led economists to look for models capable of describing such systems appropriately. This is why agent-based models, designed precisely to describe complex systems, are being more widely used in economics.

The idea of the agent-oriented approach dates back to the 1940s, to the works on cellular automaton of John von Neumann (link in Russian) – a mathematician, physicist and one of the creators of game theory. A cellular automaton is essentially a space (such as a chessboard) divided into identical cells. Each cell is assigned a value of 0 or 1. At each step, the algorithm ‘looks’ at the ‘neighbours’ of a cell and on that basis sets a new value for it. This is repeated for every cell. Though the concept may appear a little divorced from reality, it does in fact help describe many practical problems.

In the 1970s, Thomas Schelling, subsequently a Nobel prize winner, used this simple idea to explain segregation in US cities. He showed that the gradual emergence of ‘black’ and ‘white’ neighbourhoods is due to people’s desire to have a share (typically 30–50%) of their neighbours be the same ethnic group as themselves and their willingness to move house to achieve this. The emergence of African-American ghettos, then, may be explained not by decisions taken ‘from above’ but by the sum of individuals’ actions taken purely in their own interests.

In economics, these models increasingly gained traction in the 1990s and have been in intense use in this decade for macroeconomic purposes. The Bank of Canada is already using agent-based models to support its rate decision-making along with DSGE models.

Putting it into practice

I use an agent-based model in my study. The model contains around 13,000 agents, including individuals, firms, a central bank, a government, a commercial bank, and importers and exporters. Each individual has a particular set of skills that allow them to work in a particular sector. Naturally, the model is calibrated based on microdata: how many people and firms work in each sector, what skills a sector needs and what skills its employees have, and so on.

I compare the simulation results of this model with a benchmark model, in which the concept of skills is absent and individuals can work only in the sector of their original employment. My results show that the way the labour market is modelled qualitatively changes the outcomes.

In the benchmark model, a supply shock (that is, a sudden contraction in economic output) is followed by a full GDP recovery after five years. In simulations of the skills model, by contrast, GDP fails to recover after the same shock, and the gap between the simulations with the shock and the simulations of the same model without it persists even five years on.

The mechanism behind this result is as follows. During the shock, certain employees are made redundant. In the benchmark model, once the shock is over, firms rehire the employees who were dismissed, and this ensures the swift recovery of GDP to its pre-crisis level.

However, if the hiring process is centred on skills, then, first, skills may be lost over the course of the crisis and workers may not be re-employed in their previous positions. Second, the skills model includes occupational mobility. That means that some redundant employees may find work in other sectors. This leads to the emergence of a new structure in the labour market after the crisis, which prevents GDP from returning to its pre-crisis path.

What is more, a detailed analysis shows that the impact of the shock is uneven across sectors. In sectors where higher-level skills are needed, firms struggle to hire, since, first, the pool of sufficiently qualified employees is limited, and, second, these workers are also being sought by sectors with lower skill requirements, creating additional competition for them.

It is clear then that there are labour market mechanisms that prevent GDP recovery after a crisis and that these mechanisms are at least partially determined by skills.

This raises a question: can workers’ skills be upgraded through training programmes so that GDP can return to its pre-crisis path?

I test how well two different training programmes can do this: a universal programme, i.e. one not related to any particular sector, and a targeted programme, tailored for specific sectors.

The first involves training randomly selected unemployed people in random skills. In practice, there are various government retraining courses for the unemployed to attend and acquire skills. My experiments show that such programmes, if designed in a particular way, can be effective in returning GDP to its pre-crisis level. That said, while I do not assess the cost and organisational difficulty of such training, it probably requires considerable effort and expense.

As I note above, however, only a number of sectors face serious hiring difficulties: those with very high skill requirements. So the second programme considered is targeted training of students for two sectors. The assumption is that graduates from universities enter the labour market with skills that match those sectors exactly and that firms in those sectors are their priority employers. If some of these students are not hired in the first round, they can try to find work as anyone else at a firm in any other sector.

The simulations suggest that the result of this programme is a full return of GDP to its pre-crisis path, just as with the previous, larger-scale, training programme.

The main conclusion of my study is thus that the simplifications of labour market processes which are common in large macroeconomic models can lead to an underestimation of the negative consequences of shocks for the economy. Modelling a more realistic labour market can not only predict the negative effects of a shock on the economy, but also make it possible to design training programmes to offset these effects.