Skill demand and labour market concentration: evidence from Italian vacancies

3.1 Generalised monopsonistic model

Following Manning (2006), we consider a monopsonistic model where firms compete for workers, but where, in order to set their level of employment N, they must pay both a direct and an indirect cost. The direct cost per worker is the wage W, whereas the indirect cost, I(N), can be thought of as the recruiting cost necessary to substitute the exogenously separated workers with new recruits. We assume that this recruiting cost is increasing with the share of employment working in the firm, therefore, aggregating at the market level, a higher level of concentration leads to higher recruiting costs. The rationale is that the larger the share of workers working for a firm in a market, the more difficult it becomes to find a good match amongst potential recruits. Alternatively, one can think of a framework as in Berger et al. (2022) where workers have an idiosyncratic preference or specific bundle of competencies for a workplace, therefore the larger the share of employees working in a firm, the costlier it becomes to convince the remaining workers to work in that workplace because they are those with the lowest idiosyncratic preference or the lack of necessary competencies [4]. For our purposes the key element is that it is the employment share to drive an increase in hiring costs, rather than the absolute number of employees [5].

To maintain the model simple, we are excluding factors such as the possibility for workers to move across markets as well as the entry or exit of new workers or firms. Our results will go through even incorporating these elements so long as the basic features of the monopsonistic model are preserved [6]. Empirically the evidence on workers' mobility in Italy is mixed, depending also on the geographical unit considered.

In this setting, a firm chooses N to maximise profits which are given by:

In equation (1) the level of employment N affects the cost function C(W, N) both through its direct cost (through wages) and through its indirect cost (I(N)). The latter effect operates through local labour market concentration: the larger the firm the larger its share in the local market, the more concentrated the market is.

The first order condition of equation (1) is the following

where MP_N is the marginal productivity of labour and e(N) is the inverse labour supply elasticity which depends on the employment share. As stated above we assume that the inverse labour supply elasticity is increasing with the level of employment, CN′>0 and CNN″≥0, which implies that it becomes increasingly costly to recruit workers [7].

Building on this result, we will proxy the increase in labour market concentration with an increase in the indirect cost of labour through an increase in the employment share, keeping unchanged the level of employment and thus the direct cost.

The assumption that firms can just adjust their labour force and not their wages is specific for a country with high wage rigidities and collective contracts, like Italy. Although the incentive to reduce wages from the reduction in labour market competition, the downward wage rigidity forbids them this channel, pushing them to intervene through the labour demand one [8].

3.2 Production function

We embed the approach outlined above in a canonical model of human capital with different tasks and factor-augmenting technology (Acemoglu, 2002; Autor et al., 2003; Acemoglu and Autor, 2011). Consider an economy where labour is the only input, divided in two distinct categories: “trainable” and “untrainable”. The competencies in the trainable category can be quickly learnt through on-the-job training — for example, standard technical skills. Instead, the “untrainable” category includes those competencies that are difficult to learn because they are linked to character or attitude. Some straightforward examples are competencies like leadership, problem-solving and social skills. The two groups of skills are both needed for production. Thus, they are complements and not substitutes [9].

Assume that the production function is a Cobb-Douglas function nested in a constant elasticity of substitution (CES) function:

where T is the trainable labour component, u is the untrainable one, A is the amount of training provided and θ ∈ [0, ∞) is the elasticity of substitution between trainable and untrainable labour inputs. Given that the two skill groups are complements, 0 < θ < 1. As an additional simplification, we assume that training can only improve the productivity of the “trainable” labour component [10].

3.3 Equilibrium and empirical predictions

There is a training cost τ linear in the amount of training. Both inputs belong to the same market which follows the structure described in section 3.1. Both inputs have the same direct cost W and indirect cost I(N), which depends on the total amount of labour inputs used N = T + u [11]. Thus, the profit maximisation problem can be written as:

Our primary objective is to examine the impact of employment concentration on employers' preferences for trainable and untrainable labour components. Assume that a rise in employment concentration increases the indirect cost of both inputs. Given the possibility to improve the productivity of one of the inputs through training, the optimal bundle relies not only on their relative cost and productivity but also on the ability to substitute the trainable input with training expenditures. Therefore, as input costs increase due to a rise in market concentration, measured by the Herfindahl–Hirschman Index (HHI), which represents the cumulative sum of squared employment shares of each firm within a local labour market, this alternative to substitute the training input with training becomes increasingly more financially advantageous [12].

Therefore employers facing a concentrated labour market are more likely to demand relatively more untrainable competencies. As the concentration rises, the inverse labour supply becomes steeper, increasing the marginal cost of the labour inputs; given that the two inputs are complements, an employer will find it more profitable to divert part of the investment from the trainable input to the untrainable one, substituting the former with an increase in the training investment. Indeed, as a corollary, it can be shown that this simple model also predicts an increase in training spending following an increase in employment concentration.

4.1 Sources

The source of the vacancy data is Wollybi, [13] a project that collects online vacancies in Italy from job portals since February 2013. For internal data consistency, we concentrate on the years from 2015 to 2018, and we select only primary sources, neglecting secondary sources such as aggregators (e.g. websites that re-post vacancies retrieved from other websites) [14]. Each vacancy includes detailed information such as location, industry, education and skill requirements [15].

To measure the level of concentration across local labour markets, we exploit the Italian ORBIS dataset, AIDA, by Bureau van Dijk, from 2015 to 2018. This dataset contains the full balance sheets and income statements of Italian firms. Similar data have been used in recent research, see for example Gopinath et al. (2017).

One potential drawback of online vacancies is that they capture only vacancies posted on the Internet and may not be representative of the universe of vacancies. Online vacancies have been used by other papers and have been found fairly representative of the universe of job openings (Hershbein and Kahn, 2018; Modestino, 2020). In the appendix we provide a detailed assessment of the representativeness of online data; however, it is worth mentioning that our paper focuses on the skill distribution within occupation across markets characterised by different degrees of concentration, therefore any bias that online vacancies may have is likely to be greatly weakened.

We restrict the analysis only to those vacancies that report both the province and 2-digit NACE industry code, this leads to a final sample of 553,132 vacancies, distributed over 4 years, 106 provinces, 380 occupation codes and 73 industry codes. In addition to the skill content, OJA contains information about the level of education (ISCED) and the level of experience required (expressed in years). Tables A9 and A10 show the summary statistics of this final sample.

4.2 Skill classification

To allow comparability with other papers of the literature we have used the skill taxonomy of O*NET, developed by the Bureau of Labour Statistics [16]. Skills extracted from OJA are classified into the finest level of the O*NET taxonomy which is organised into three hierarchical levels. We used the finest level as the building block to construct two classifications. The first is the broadest O*NET level composed of the following categories: Knowledge, Skills, Abilities, Work Activities and Work Styles [17]. The broad classification available in O*NET however does not lend itself to a clear interpretation as there are subtle differences between what is classified as skill and what is classified as, say, ability or work activity. For example “mathematical reasoning” is classified as an ability under the category of “cognitive abilities”; on the other hand “use of mathematics to solve problems” is classified as a skill under the category of “basic skills”. Moreover “developing and building teams” is considered as a work activity under the category of “interacting with others” whilst “persuasion” and “coordination” are considered as skills (social skills). Starting from the finest level we have therefore constructed a different skill classification composed of the following groups Cognitive, Social, Digital, Hard (technical), Organisational skills. We did not regroup items of the Knowledge category leaving them separate as we believe that these refer to a set of principles and facts applying in general domains which can be easily linked to the educational system [18]. Table 1 lists the competency classifications and the corresponding description of each category.

4.3 Measuring skill intensity

Once extracted the information from vacancies and mapped into a skill taxonomy, the final challenge pertains to the creation of measures of intensity of a given skill (or category of skills). Measuring the intensity with which a job vacancy demands each skill is challenging. To address this issue, we define two different measures: a binary and a continuous measure, the term frequency–inverse document frequency (TF-IDF), which is similar to the local-quotient measure used by Alabdulkareem et al. (2018). On the one hand, the binary measure provides a more straightforward interpretation of the results. However, on the other hand, it fails to measure how much a particular skill is essential for that specific vacancy compared to the other vacancies in the same occupation.

The binary measure describes whether a vacancy demands at least one skill of that category. In contrast, the TF-IDF documents how important a particular skill is for a vacancy relative to the importance of that skill in the vacancy's occupation.

For a skill category j in vacancy i for occupation o, the term frequency (tf_ijo) is the share of skills of category j demanded. The inverse document frequency (idf_ijo) is the log of the share of vacancies in occupation o demanding at least a competency of the category j. Formally, the TF-IDF is computed as:

The TF-IDF is a standard measure in the literature of information retrieval, [19] and is our preferred measure as it gives more importance to occupation-specific skills rather than to general skills. Indeed, skill categories that are unimportant for a vacancy will have a low TF-IDF score because the tf_ijo will be low. On the other hand, very common skill categories will instead have a low TF-IDF score because that category will be demanded in most of the vacancies in that occupation; thus, the idf_ijo will be very low. On the opposite, specific skills in high demand for a given occupation will be characterised by a high TF-IDF score.

Figure 1 shows the distribution of the number of skills demanded for each job ad. We can see that almost half of the vacancies demand less than 5 skills. Figure 2 displays the average number of skills demanded by each group. The categories Skill, Activity and Knowledge are the most requested with an average of more than two competencies belonging to these categories per job ad. Tables A13 and A14 report the correlation matrices between the different categories for the two different intensity measures.

4.4 Measuring labour market concentration

Following the literature, we define a local labour market as the combination between a province, [20] an industry/sector and a year.

As a measure of concentration, we use the HHI, defined as the sum of squares of each firm's employment shares in a local labour market. Figure 3 shows the logarithmic distribution of the HHI at the local labour market level, unweighted or weighted for the number of vacancies posted. The average local labour market is moderately concentrated, with a mode around log(HHI) = 7, equivalent to an HHI of 0.11 or an Inverse Herfindahl–Hirschman Index (IHHI) of 9.2 [21]. The IHHI can be interpreted as the number of equal-sized firms that will induce the same observed HHI [22]. However, the average job vacancy is posted in a low-concentrated market with a mode around log(HHI) = 6. As also observable in Figure 4, which plots the maps of the average HHI and the number of vacancies posted at the province level, employers in less concentrated markets post more job vacancies [23].

4.5 Empirical strategy

For our empirical specification, we regress the two measures of skill demand at the vacancy level on the log-HHI index of the local labour market where the vacancy was posted, formally

This empirical specification aims to understand whether differences in the concentration level are associated with differences in the demand for competencies. Controlling for the fixed effects, we want to pick up the difference in skill demand associated with the difference in the concentration level and not pick up province- or occupation-wide differences.

The specification outlined above has been enriched with a number of possible controls and augmented with an IV estimator. Section 5.5 describes them in detail.

5.1 Effect of labour market concentration on experience and education

In a well-known paper Modestino (2020) show that, in the USA, following an increase in the supply of workers, employers' requirements in terms of education and experience increase, denoting some form of opportunistic upskilling. This effect should be similar considering firms with stronger monopsony power.

Therefore we start by analysing the effect of labour market concentration on experience and education. Table 2 reports the estimates of labour market concentration on whether a vacancy requires less than 1 year of experience (No Exp. required), the years of experience demanded (Experience), [25] whether it requires a university degree (Graduate) and the total number of skills demanded. Overall labour market concentration is negatively correlated with experience and positively with the level of education. Specifically, one standard deviation increase in the labour market concentration increases the probability that the vacancy does not require any experience by 5.6%, i.e. an increase of 1% point [26]. The same change in HHI decreases the amount of experience required by 6% points, or 2.2%, which amounts to almost 25 days less of experience required. Furthermore, labour market concentration is positively correlated with the probability that the job ad requires a university degree.

Overall, our results suggest a different interpretation than Modestino (2020). We observe in fact that an increase in local labour market concentration reduces the experience required, but, at the same time, it increases the demand for graduate workers. These results are in line with the training hypothesis. If employers in a more concentrated labour market are more prone to training new workers, they do not demand that workers already possess job experience; instead, they look for workers who can acquire and learn new competencies fast and efficiently, as signalled by their education level [27].

Finally, considering the skill variable, we do find evidence of an “upskilling” effect but it is somewhat different from the standard interpretation, in our case an increase in labour market concentration leads to an increase in the number of skills required. However, so far we have not analysed the type and nature of the skills required. The next section deals with these issues.

5.2 Market concentration and skill demand

For reasons of space we report in the text the results of the TF-IDF measure whilst we leave the tables related to the binary measure in the Appendix. We start with the broad O*NET classification of the competencies set, Tables 3 and A11 report the results for the ordinary least squares estimations of the TF-IDF and binary intensity measure, respectively. Each table includes five different categories of skill/competency as dependent variable: Skill, Knowledge, Ability, (Work) Activity and (Work) Style [28]. Overall Knowledge is negatively correlated with labour market concentration whilst work styles activities and skills are positively correlated although the results are sharper when considering the TF-IDF measure.

Note that the Knowledge pillar consists in the “organised set of principles and facts”, whereas the Skills pillar defines “developed capacities to facilitate learning” and Work Activities are “general types of job behaviours occurring on multiple jobs” [29]. These results are in line with the training rationale. Employers in more concentrated markets are more willing to provide on-the-job training, so they are more interested in workers that are able to learn faster rather than workers who already possess knowledge. Knowledge pertains to competencies that are strongly connected with formal training and can be taught by the firm internally. On the contrary, working attitudes such as being a quick learner or being good at interacting with others are less easily trainable and are acquired by the firm on the market through hiring.

To give a clearer sense of these results, Figures 5 and A5 plot the estimated coefficient for all the different competencies and for the two different intensity measures. Regarding the magnitude of the coefficients, a standard deviation rise in the labour market concentration decreases the Knowledge TF-IDF score by around 1.5%. The same increase in local labour market concentration, ceteris paribus, leads to an increase of Skill TF-IDF score by around 1.1%. Therefore, a rise in the HHI decreases the importance of Knowledge competencies and increases that of Skill competencies compared to their usual relevance for that occupation.

To better explore these issues we regrouped skills and competencies into a classification that allows us to shed more light on what type of skills are requested in concentrated markets. Tables 4 and A12 report the results [30]. Social and hard skills are positively related to labour market concentration using both intensity measures. On the other hand, cognitive, digital and organisational skills are negatively correlated with labour market concentration using the TF-IDF measure whilst the results are less sharp with the binary intensity measure.

Results presented so far fit with the training rationale. Higher concentration in the labour market lead firms to demand fewer trainable competencies (e.g. the knowledge pillar) and more that are un-trainable (e.g. social skills) [31]. However there some results are not completely in line. For example, following the training hypothesis, we would expect cognitive skills to be positively correlated with market concentration whilst our results show that the relationship is negative for the TF-IDF measure and null for the binary measure.

5.3 Skill demand heterogeneity across occupations

To shed light on these issues we have analysed separately different classes of occupation. Following the ISCO classification, we divided occupations into high- and medium/low-skill occupations. Specifically, those occupations with the 1-digit ISCO code between 1 and 3 are high-skill occupations, whereas the occupations with codes between 4 and 9 are low/medium occupations [32]. We, therefore, estimate separately the effect on high and medium/low occupations [33]. Figures 6 and A6 show the heterogeneous effect of local labour market concentration on the demand for skills [34]. Two results emerge. First, the binary measure delivers sharper differences, this is expected as it does not weigh skills by their relative importance. Second, it emerges a clear difference between high and medium-low-skill occupations.

Compared with medium-low skill occupations, employers posting vacancies in high-skill occupations in highly concentrated labour markets increase the relative demands for Cognitive skills and reduce the demand for Knowledge. On the contrary, for low-skill occupations Hard and Social skills become relatively more important. Given the different nature of the tasks performed in each occupation class, employees in high-skill occupations are required to perform more complex tasks and duties than employees in low occupations. Also, organisational skills are positively correlated with market concentration for high skills occupations whilst they are negatively correlated with low-skill ones.

As stressed above, results for the binary measure are sharper than those of the TF-IDF measure, especially for some skills such as cognitive and digital. This can be partly explained by the increased generalised diffusion of these competencies. In a more concentrated labour market is more common to require such competencies: it is in fact more frequent that advertisements contain at least one of these skills (binary measure). At the same time, these competencies are becoming increasingly required in all vacancies. This reduces their relative weight in the TF-IDF measure. We know that labour market concentration is also associated with an increase in the number of skills (Table 2). Splitting these estimates by occupation (Table 5) shows that this effect is driven by high-skill occupations. Thus labour market concentration is associated with an increase in the number of competencies in high-skill occupations. In addition, cognitive and digital skills are increasing but also becoming more general. This explains why the effect on the TF-IDF measure becomes zero or even negative.

5.4 Alternative measure of skill intensity: effective use

The previous paragraph shows that there is not an ideal approach to measure skill intensity as there is always a tension between skills that are in high demand in general and skills that are in high demand because are occupation specific. A possible way to reconcile the different approaches is to rely on the effective-use measure described in Alabdulkareem et al. (2018). The intuition is simple starting from the TF-IDF measure, or more precisely a variant of it, it is possible to identify a threshold that defines the average demand for skill within an occupation. The effective use is a dichotomous measure that takes the value of 1 if that particular skill is demanded more than the average and 0 otherwise.

Figures 7 and 8 displays the estimates following the same methodology described in Section 4.5. The effective-use measure shows clearer differences between high and low-skill occupations which reinforce the interpretation provided so far. Compared with low-skill occupations, employers posting vacancies in high-skill occupations in highly concentrated labour markets increase the relative demands for Cognitive skills and reduce the demand for Knowledge. Moreover, digital skills are positively correlated with labour market concentration for high-skill occupations, whilst they are negatively correlated for medium-low skill occupations.

Overall these results support the training rationale as they can be explained in terms of different training requirements of high and low-skill occupations. Presumably, new hires in high-skill occupations need to learn in-depth and complex competencies, which are difficult to be taught through mentoring or assistance from senior colleagues, instead, these competencies require formal teaching as in-class courses. For example, the type of organisational skills that are needed for high-skill occupations are generally managerial skills that can be acquired in graduate studies. Hence firms in more concentrated markets tend to demand more of these skills alongside a higher level of education. Table 5 shows that labour market concentration increases the level of education and the share of graduates and this effect is higher for high-skill occupations. On the other hand, new hires in low-skill occupations, performing simpler duties and tasks, can learn by being assisted and guided by an experienced colleague. Thus, having good social and hard skills can particularly help recruits in low-skill occupations to learn the required competencies. Whilst in high-skill occupations, where workers are more likely to be trained through more formal training activities, cognitive abilities become particularly important to enable the workers to learn and acquire new knowledge. Finally, it is important to stress that several cognitive skills, being often general, are less likely to be explicitly mentioned for high-skill occupations as they are subsumed by the higher level of education. This can explain the low significance of the coefficient for high-skill occupations.

5.5 Robustness

Although our results are robust to different specifications and include several controls for sectors/occupations etc. we acknowledge that they do not settle the issue of possible unobserved heterogeneity/endogeneity. In order to address this issue we have performed several robustness checks. First, we have controlled for the average firm size in the local labour market (larger firms may demand different skills); second we have controlled for the level of unemployment in the local labour market (unemployment may be correlated with concentration). Third, given the significant correlation between the different skills, we have controlled for the other skills in each specification. Fourth, we have estimated a more rigourous model using year times province, year times industry and year times occupation-fixed effects to control for potential province, industry and occupation-specific shocks. Fifth, we have implemented an IV estimator using the Hausman–Nevo instrument. Our findings are robust to all these different specifications/controls. The detailed results are available in the appendix.