Making work pay in Croatia: An ex-ante evaluation of two in-work benefits using miCROmod

The history of tax-benefit microsimulations in Croatia is short. The first microsimulations were done by Urban (2010), who developed a static arithmetic model based on the data from the Household Budget Survey (HBS) and simulated the main features of the Croatian tax-benefit system. The model was used mainly for the purpose of measuring the redistributive effects of direct taxes and social benefits in Croatia (Urban, 2010; Urban, 2014; Čok et al., 2013). Later, a new, more detailed arithmetic model was developed as a module within EUROMOD (Urban and Bezeredi, 2016). It has been available since 2016, and the tax-benefit rules for 2011 through 2015 were modelled in it, while the latest available rules are those pertaining to 2018 (Urban et al., 2017). Another recent model was developed by Inchauste and Rubil (2017) for the purpose of a detailed incidence analysis of taxes and social spending in Croatia in 2014.¹ The model uses the Income and Living Conditions Survey (ILCS) data, which is used in the present paper as well,² and the HBS data. Unlike the Urban (2010) model and EUROMOD, this model (a) does not simulate social benefits, taking their amounts as directly reported in the survey instead, and (b) includes indirect taxes (the value added tax, excises and other).³

The most comprehensive and detailed model that is currently in use is miCROmod. The model was developed as part of the research project Application of Microsimulation Models in the Analysis of Taxes and Social Benefits in Croatia (AMMATSBC) (Urban, 2016), and we use it in the present paper. miCROmod is a “family name” of two models. The first, named miCROmodA, is a standard arithmetic tax-benefit model that closely resembles EUROMOD’s module for Croatia. It is basically a EUROMOD “spin-off” model,⁴ as it has the same structure, it has been built on the EUROMOD platform, and it closely follows the EUROMOD modelling conventions.⁵ Similar to EUROMOD’s module for Croatia, miCROmodA simulates the personal income tax, social insurance contributions and social benefits paid by the central government, but it also goes beyond these by simulating the local-level benefits paid by the authorities of the four largest cities in Croatia: Zagreb, Split, Rijeka and Osijek. Another difference is that miCROmodA uses the national EU-SILC data – the abovementioned ILCS – which is more detailed than the harmonised EU-SILC data provided by Eurostat.⁶ The second model, named miCROmodB, is a behavioural microsimulation model. It is made of the following two components: miCROmodA and a structural model of the household labour supply. miCROmodA is utilised here to generate the household budget constraint implied by the tax-benefit rules and the alternatives in the household choice set. The details of the labour supply model are presented in the remainder of this section.

In the last 25 years, discrete choice models of the labour supply – those where agents make a constrained choice from a set with relatively few alternatives – have become a standard tool for obtaining labour supply responses in behavioural microsimulation (Creedy and Duncan, 2002; Aaberge and Colombino, 2018). In comparison to the traditional approach based on marginal calculus (eg, Hausman, 1985), the discrete choice approach makes it easier to handle non-linear and non-convex budget constraints, and it also allows the use of more flexible functional forms for the utility function (Creedy and Kalb, 2005). A vast majority of studies opting for the discrete choice approach use some variant of van Soest (1995) model. In this standard model, the agent chooses hours of leisure (total time endowment minus work hours) and the implied disposable income to maximise her utility subject to the budget constraint. However, despite the discrete nature of choice in Van Soest-type models, some authors argue that these models are theoretically very similar to the continuous models (Dagsvik et al., 2014; Dagsvik and Jia, 2016).

We also model the labour supply within the discrete choice framework but depart from the standard approach in that we use a theoretically sounder modelling framework where agents choose “jobs” rather than hours of leisure/work. The framework has been introduced by Dagsvik (1994), but despite its theoretical advantages, it is still relatively rarely used (eg, Aaberge et al., 1995; Aaberge et al., 1999; Aaberge et al., 2000; Dagsvik and Strøm, 2006; Kornstad and Thoresen, 2007; Di Tommaso et al., 2009; Dagsvik et al., 2009; Dagsvik et al., 2011; Aaberge and Flood, 2013; Dagsvik et al., 2014; Dagsvik and Jia, 2016; Capéau et al., 2016).

The approach seeks to acknowledge two features of the actual choice situations in the labour market. First, rather than having preferences over combinations of leisure hours and disposable income, as in the van Soest (1995) model, people have preferences over jobs, characterised as packages consisting of hours of work, a wage rate⁷ and non-wage attributes. The latter can be generally thought of as various non-pecuniary features of a job, such as the specific work environment and the tasks that the job assumes. These are observable by the agent, but not by the econometrician. Second, unlike the standard discrete choice approach, the job choice approach acknowledges the fact that there are restrictions on individual job opportunities. A given person cannot choose any job available in the economy. For example, some jobs may be unavailable to an individual due to her lack of required qualifications. Even among the jobs for which she is qualified, some may be unavailable to her due to, for example, some sort of discrimination. In addition, within a job type, she may face restrictions regarding the choice of hours of work. Her set of available jobs may also be restricted due to the type of competition in the labour market (depending on the market structure).

We first describe the model for a single-individual household and then extend it directly to households of married/cohabiting couples. Each individual has preferences over consumption C, hours of work h, and other job attributes. The job attributes are unobserved to the econometrician. Let the jobs be indexed by z, where z=0 represents “not working” (in the formal labour market). The preferences are represented by an ordinal utility function $$U(C,h,z)$$ of the following form:

where $$V(C,h)$$ is a deterministic part, and $$\epsilon \left(z\right)$$ is a stochastic term that is meant to capture the effect of unobserved job attributes on preferences. The latter is assumed random and i.i.d. across individuals and jobs according to the type I extreme value distribution with the cumulative distribution function $$F\left(\epsilon \right)=\exp (-\exp (-\epsilon \left)\right)$$.⁸

The agent chooses the job that maximises her utility subject to the budget constraint⁹

where $$w$$ is gross wage rate, which is assumed to be individual specific and independent of h, and I is the total exogenous non-labour income.¹⁰ Function $$T(\cdot ,\cdot )$$ is the tax-benefit function, transforming the total gross income into total disposable income.¹¹ This function takes the gross labour and non-labour incomes as arguments, and its positive (respectively, negative) value implies that the household is a net payer to (respectively, receiver from) the government budget. Similar to the papers referred to above, it is assumed that the hours of work are fixed for each given job z. In reality, for most jobs (characterised by a set of attributes), there is not much variation in the hours of work, due to specific technological (production process requiring certain work time for efficiency) or institutional constraints (eg, collective agreements with labour unions). Let H(z) be the fixed workload associated with job z. When the budget constraint and the job-specific workload are taken into account, the utility function (1) becomes the following:

where $$z\in B$$, and B is the set of jobs available to the individual. Thus, when the restrictions on job choice are taken into account, the individual’s choice becomes a discrete choice among available jobs. Let B(h) be the subset of B that contains the jobs for which H(z)= h and let m(h) be the number of jobs in B(h), a quantity unknown to the econometrician. Further, let D be the set of all possible hours of work, including zero, and let this set be finite. hen, the probability that the individual will choose a job with hours of work equal to h can then be expressed as follows (see Dagsvik et al., 2014):

where $\psi (h;w)=V\left(C\right(h,w),h))$. Following Dagsvik and Strøm (2006), it is convenient to express m(h) as a product of two factors, namely, the “opportunity measure” $$\theta ={\displaystyle {\sum }_{x\in D,x\S gt;0}m\left(x\right)}$$ and the “opportunity density” g(h): $$m\left(h\right)=\theta \cdot g\left(h\right)$$. The opportunity measure is a measure of jobs available to the individual, while the opportunity density is the faction of available jobs with the hours of work equal to h. Thus, the choice probability (4) can be rewritten as follows:

The extension to couples (married or cohabiting) is straightforward. We assume a unitary household model with a household utility function $U(C,h_F,h_M,z)$. Let s denote gender, with s=F for females and s=M for males, and let z and w be vectors instead of scalars, as follows: $$z=(z_M,z_F)$$, $$w=(w_M,w_F)$$. The utility function in this case has the following form:

Further, the set D is now the set of all possible combinations of hours of work $(h_F,h_M)$. Finally, let $m_F\left(h_F\right)$ and $m_M\left(h_M\right)$ be the number of available jobs for the female and male, respectively. Similar to the above model for single individuals, we can write $m_F\left(h_F\right)={\theta }_F\cdot g_F\left(h_F\right)$ and $m_M\left(h_M\right)={\theta }_M\cdot g_M\left(h_M\right)$. Thus, the choice probabilities for a couple are as follows:

The details of the empirical specification are here given for couples. The specification for single-individual households can be obtained straightforwardly by obvious amendments of the specification for couples.

The point of departure are the choice probabilities given in (7). To obtain an empirical model that can be estimated from microdata, we need to specify a parametric functional form of the systematic part of the utility function and to make assumptions about the opportunity distribution. For the deterministic part of the utility function of a couple, we use the quadratic specification of the following form:

where $L_s$, $s=M,F$, denotes the annual hours of leisure, defined as the total annual time endowment of 8760 hours (365 times 24) minus annual hours of work: $L_s=8760-h_s$. The parameters of the leisure terms, ${\beta }_M$ and ${\beta }_F$, are allowed to vary across the population and, thus, capture the preference heterogeneity. Precisely, they are modelled as linear functions of a set of individual characteristics and a constant: ${\beta }_s={\beta }_{0s}+{\displaystyle {\sum }_{k=1}^K{\beta }_{ks}\cdot X_{ks}}$, where the characteristics include age and its square, number of preschool children (aged 0–6), and indicators for two levels of health conditions limiting one’s activity.

Following Dagsvik and Jia (2016), we specify the opportunity measure for each gender as a linear function of observable personal characteristics and a constant. More specifically, we specify ${\displaystyle \mathrm{l}\mathrm{n}{\theta }_s}$ as ${\displaystyle \mathrm{l}\mathrm{n}{\theta }_s={\gamma }_{0s}+\sum _{\mathfrak{l}=1}^L{\gamma }_{\mathfrak{l}s}\cdot X_{\mathfrak{l}s}}$, where $$X_s={\left\{X_{\mathcal{l}s}\right\}}_{\mathcal{l}=1}^L$$ are years of age and work experience. Regarding the opportunity density, we also follow Dagsvik and Jia (2016), who model it as uniform for all positive hours of work except for a peak at hours of work observed to carry significant mass. As the only significant peak in our observed distribution of positive hours of work is the one associated with annual full-time work (2080 hours), we model ${\displaystyle \mathrm{l}\mathrm{n}{g}_s\left(h_s\right)}$ simply as a constant specific to full-time work: ${\displaystyle \mathrm{l}\mathrm{n}{g}_s\left(h_s\right)={\pi }_s}$ if $$h_s$$= 2080.

The specification for couples can be straightforwardly reworked for the case of single individuals. Instead of the choice probabilities given in (7), the points of departure are the choice probabilities in (5), with the expressions amended accordingly. Regarding the specifications of the systematic utility, the appropriate specification is obtained from (8) by dropping one of the leisure terms and all interactions in which it appears. The specifications of the opportunity measure and the opportunity density remain the same as in the previous paragraph, except that the index s can be dropped.

We now turn to the details of the estimation. Since the hourly wage rates of non-working people are not observed in the data, these need to be imputed before the estimation of the labour supply model. In the wage rate estimation, we account for possible selection into employment by using the Heckman selection model. In the wage equation, where we model the natural logarithm of the hourly wage rate,¹² we use the following covariates (along with a constant): age (in years) and its square;¹³ education (in years); work experience (in years) and its square; and a dummy for residing in an urban area (ie, an area with high population density). In addition to this set of covariates, in the selection equation, we additionally include the following ones: number of preschool children (aged 0–6); a dummy for being in a consensual union (ie, in marriage or cohabitation); two variables measuring the incomes of other household members from labour (employment and self-employment), assets (capital, property) and social insurance (unemployment, sickness, pension). The two equations are estimated simultaneously by maximum likelihood, separately for males and females. The descriptive statistics for all variables, the description of the estimation samples and the model estimates are reported in Tables A.1 and A.2 in Appendix A.

With the predicted wage rates at hand, we are able to construct the likelihood and log-likelihood functions for the labour supply model. For those in the sample who work, we use the observed wage rate, whereas for those who do not work, we predict the wage rate using the estimated wage equation including the error terms. For notational simplicity, write the wage equation for person i as ${\displaystyle \mathrm{l}\mathrm{n}{w}_i={\overline{w}}_i+\sigma \cdot {\eta }_i}$, where ${\overline{w}}_i$ represents the systematic part, $\sigma$ is the standard deviation of the error term in the wage equation of the Heckman model, and ${\eta }_i$ is a standard normally distributed random term. Let us first give the log-likelihood function for the case of a sample of single-individual households. This function will take account of the fact that the missing wage rates of non-workers are predicted with an error. Ignoring these errors in the estimation of the labour supply model may lead to inconsistent estimates of the model’s parameters. For a sample of N single-individual households indexed by i, the log-likelihood function can be expressed as follows:

where $$({\tilde{h}}_i;{\tilde{w}}_i)$$ are the observed hours and wage rate of individual i, $$\Omega =(\beta ,\gamma ,\pi )$$ are the parameters of the model and E is the expectation operator with respect to ${\eta }_i$. The first term on the right-hand side pertains to workers, and the second to non-workers. In practice, the evaluation of the expectation in the second term is approximated by simulation. To this end, let ${\eta }_i^r$, r=1, 2, …, R, be independent draws from the standard normal distribution (for each individual). Then, if R is large, the following approximation will hold:

In our case, R=50, and the draws are generated as follows. For each individual whose wage rate is predicted rather than observed, we randomly draw 50 errors from the distribution of errors in the wage equation of the Heckman selection model.

For couples, the log-likelihood function is analogous to (9), with straightforward amendments. First, the following substitutions are to be made: $${\tilde{h}}_i=({\tilde{h}}_{Mi},{\tilde{h}}_{Fi})$$, $$h=(h_M,h_F)$$; ${\displaystyle \mathrm{l}\mathrm{n}{\theta }_{Mi}+\mathrm{l}\mathrm{n}{\theta }_{Fi}}$ instead of ${\displaystyle \mathrm{l}\mathrm{n}{\theta }_i}$; ${\displaystyle \mathrm{l}\mathrm{n}{g}_M\left(h_{Mi}\right)+\mathrm{l}\mathrm{n}{g}_F\left(h_{Fi}\right)}$ instead of ${\displaystyle \mathrm{l}\mathrm{n}g\left(h\right)}$; ${\displaystyle \mathrm{l}\mathrm{n}{g}_M\left({\tilde{h}}_{Mi}\right)+\mathrm{l}\mathrm{n}{g}_F\left({\tilde{h}}_{Fi}\right)}$ instead of ${\displaystyle \mathrm{l}\mathrm{n}g\left({\tilde{h}}_i\right)}$. Second, the expectation with respect to ${\eta }_i$ is to be taken over all households where the wage rate is predicted rather than observed for at least one partner.

The estimates of the labour supply model parameters for couples are displayed in Table 2.^[18] The deterministic part of utility is increasing in consumption as follows: the parameters of the variables involving consumption imply a positive marginal utility of consumption, and this holds for all values of consumption and for all male-female combinations of leisure hours observed in the data. Utility is increasing as well in each spouse’s leisure, and again this is true for all observed levels of consumption and leisure hours. We find evidence of a decreasing marginal utility of income and each spouse’s leisure hours, although in the case of male leisure hours, the result is not significant. The parameters of the interaction between consumption and leisure hours are not significant for either gender, suggesting no solid evidence that consumption and leisure are either substitutes or complements. Nevertheless, if anything, the positive signs of the estimates suggest that the relationship is closer to complementarity than to substitutability. On the other hand, there is significant evidence that the leisure hours of spouses are complements to one another.

The parameters of the interactions of leisure hours with observable characteristics point to a significant preference heterogeneity with respect to age, health conditions and parenting. For both spouses, the marginal utility of leisure is U-shaped with respect to age, being higher for younger and older people than for those in-between. Moreover, an additional hour of leisure is more preferred by people with health conditions that limit activity, especially if the limitation is severe. Finally, parenting more preschool children is associated with a higher marginal utility of leisure, but only for women. This result is in accordance with women’s behaviour and attitudes conforming to the rather firmly established tradition of considering child rearing as “women’s business”.

For both genders, there is a significant heterogeneity observed in the market opportunity measure: the job opportunities differ with respect to both age and work experience. Being younger and having more work experience expands one’s job opportunities. The negative impact of age might be due to employers’ preference for younger workers, conditional on work experience, a phenomenon that may be partly due to a type of discrimination and partly due to older worker’s relatively low readiness to adapt to the rapid technological changes characterising the modern economy. Concerning the positive effect of work experience, this is an intuitively expected result, given how employers value experience when hiring. The effects of both age and experience are stronger for men, perhaps because other factors affecting market opportunities might be more important for women. For example, whether a person has (or plans to have) children is likely more important for women’s market opportunities, especially given the recent unfortunate tendency of some employers to prefer childless female workers over those with small children. Finally, the parameters of the full-time peaks, which are meant to capture the intensity of full-time job offers for males and females compared to job offers with other work time regimes, are both positive and highly significant, indicating that for both males and females full-time jobs are offered significantly more frequently than other working time arrangements.

We now assess the in-sample and out-of-sample prediction performance of the model. If the model will be used for simulations of the effects of tax-benefit reforms, it should perform well in this regard.

We first assess the in-sample prediction performance by checking if the model is able to reproduce the observed discretised distribution of annual hours of work in the samples of couples and singles used for the estimation of the model. Precisely, we compare the shares of individuals observed to choose each of the discrete alternatives with the corresponding model-predicted choice probabilities. The fit is assessed for the marginal distribution of each gender, as well as for the joint male-female distribution. In addition, the fit is assessed with respect to household disposable income as well: the observed distribution is compared to the predicted.

The comparisons of the observed and predicted distributions of work hours for males and females in couples are displayed in Figure 2. The fit for the marginal distributions appears to be very good for both males (panel a) and females (panel b). For both genders, the average predicted choice probability of non-participation is slightly smaller, ie, by approximately one percentage point, than the share of people observed not to participate. The opposite holds for the full-time hours (2080 hours). The fit is good for the remaining alternatives as well. It is almost perfect for men, with differences of at most 0.2 percentage points and is somewhat worse for women, but still reasonably good. A good fit for the marginal distributions does not, however, necessarily guarantee a good fit of the joint distribution. Thus, we check the fit for the latter as well. As shown on panel c, the predicted distribution is very well in line with the observed one, not only for the alternatives with relatively large masses but also for those scarcely represented in the sample.

Figure 2

Download asset Open asset

In Figure 3, we compare the observed distributions of total household disposable income compared to the predicted ones. This comparison also confirms the model’s good prediction performance. Although there are some differences between the observed and predicted densities, the two are aligned reasonably well, though not perfectly.

Figure 3

Download asset Open asset

If the model’s parameters are truly structural, the model should also perform well in out-of-sample predictions. However, a good out-of-sample prediction performance is not guaranteed by a good in-sample prediction performance. Thus, we use the model to test the out-of-sample prediction performance. We take the parameters estimated for the data for the year 2015 (reported and discussed above) and the data for the year 2014 to assess how well the model can reproduce the distributions of hours of work and household disposable income for the latter year. The data source is the same, namely, the ILCS, and miCROmod is used to model the details of the tax-benefit system in 2014. As shown in Figure 4, the observed and predicted distributions of the hours of work are aligned in the 2014 data as well as they were in the 2015 data. This holds equally for the gender-specific marginal distributions (panels a and b) and the joint male-female distribution (panel c). The good out-of-sample prediction performance is corroborated through the comparison of the observed and predicted densities of household disposable income (Figure 5).

Figure 4

Download asset Open asset

Figure 5

Download asset Open asset

A word of caution is in order here. A more credible assessment of the model’s out-of-sample performance would be to check if the model fits the data well from a year more distant from 2015 than 2014 is. Precisely, we would ideally use a year when the tax-benefit system and the distributions of hours of work and disposable income were more different from those in 2015. With respect to the tax-benefit system and the distributions of hours of work and disposable income, 2014 is very similar to 2015, thus, good out-of-sample performance is expected. This, however, makes the assessment of the out-of-sample prediction performance less credible. Unfortunately, only the data for 2015 (ILCS 2016) and 2014 (ILCS 2015) were available to us.

We compute the following two elasticities of labour supply: the elasticity of participation and the elasticity of work hours. More precisely, the elasticity of participation is the elasticity of the probability of participation (ie, the probability of working non-zero hours), as predicted by the model. It is defined as the proportional change in the probability of participation induced by a proportional increase in the gross wage rate, divided by that proportional increase in the gross wage rate. The elasticity of work hours is the elasticity of the expected work hours, where the expectation is taken over the discrete alternatives (36 for people from couples, 6 for singles), with the model-predicted choice probabilities as weights.

Both elasticities are the so-called “aggregate” elasticities. Precisely, in the case of elasticity of participation, the average individual probability of participation is first obtained by averaging over the sample. The aggregate elasticity of the probability of participation is then calculated as the proportional change in the average probability of participation divided by the proportional rise in the gross wage. Similarly, in the case of elasticity of work hours, the average individual expected work hours are calculated by averaging over the sample. The aggregate elasticity of work hours is then obtained as the proportional change in the average individual expected hours divided by the proportional increase in the gross wage.^[19]

For the calculation of both elasticities, a 10% gross wage increase is assumed. For males and females in couples, the elasticities are calculated with respect to both their own gross wage (own elasticity) and their partner’s/spouse’s gross wage (cross elasticity).^[20] The elasticities are displayed in Table 4 for couples and in Table 5 for singles. In the first row of each table (labelled “All”) are the elasticities computed for the whole sample, without conditioning on any attribute apart from gender and relationship status. The remaining elasticities pertain to narrower samples, defined according to such attributes as the household’s income quintile, age, education, having a preschool child and the number of minor children.

All the own elasticities are positive, as expected, indicating that higher wages, other things equal, make work more attractive, thus increasing both participation and work hours. In contrast, most cross elasticities are negative but very small: a higher spouse’s/partner’s wage acts as a disincentive to work (more), but the disincentive effect is practically negligible, especially for males. For that reason, from here on, we focus exclusively on the elasticities with respect to the individual’s own wage. In couples, females’ elasticities are, in general, considerably larger than males’, which is a consequence of generally lower participation rates and average work hours for females. Similarly, the elasticities for single females tend to be higher than those for single males, but the tendency is not as strong as in the case of couples. For example, single females’ elasticities are higher than those of single males in the bottom three income quintiles, in age groups younger than 50, in groups with at most a middle educational level and irrespective of the number of minor children; in the remaining groups, the opposite holds. The elasticities for single males are mostly larger than for males from couples, as the former have a lower labour supply, in terms of both participation and average annual hours. The opposite holds for females, as females from couples participate less and work fewer hours on average than single females.

There is quite some heterogeneity in the elasticities by the income quintile, age, education, whether one has a preschool child, and the number of minor children. The elasticities are highest in the first income quintile, declining monotonically towards the fifth quintile. This pattern, which is commonly observed in the literature, is due primarily to the fact that the participation rate and the average number of work hours are lowest at the bottom of the income distribution. Indeed, given the high importance, on average, of labour income for total family income, the low labour supply is a major reason for being at the bottom of the income distribution. In most cases, the same monotonic pattern is observed with respect to education: elasticities are declining with the level of education. In regard to age, the elasticities for the youngest (30 or less) and oldest (50 or more) tend to be higher than for those in-between, mainly due to higher labour supply of the latter. Females’ elasticities are much more affected than those of males by whether one parents a preschool child and how many minor children one has: those with preschool children and those with more minor children have higher elasticities.

In terms of labour market conditions, Croatia is among the worst performers within the EU. The inactivity rate for the population aged 20–64 is as high as 28%, which is well above the EU-28 average of 23% and the third highest in the EU. Inactivity is especially high for females (33%) and is 10 percentage points higher than for males.^[21] In addition to high inactivity, as much as 16% of the active population is unemployed, which is seven percentage points above the EU-28 average and is the third highest; little difference is found between males and females. Long-term unemployment is another problem: the share of long-term unemployment is two thirds (again the third highest) compared to the EU-28 average of approximately one half.

Although the above figures are telling enough about the issues of the Croatian labour market, we also consider a labour market aggregate that better corresponds to our classification of flexible people into participants (h>0) and non-participants (h=0). Recall from Section three that, besides the employed and the unemployed, the flexible persons include all inactive persons who are not in education or training, are not retired and are not unable to work. Then, the employed are considered to be participating, and the rest are considered to be not participating. In a recent paper, Nestić and Tomić (2018) extend the NEET (not in employment, education or training) concept beyond the youth population. In particular, they consider the whole population aged 20–64, and define the extended NEETs^[22] as those not in employment, education, training, retirement or disability. Thus, conceptually this aggregate corresponds largely, though not perfectly, to our non-participating flexibles.^[23] As shown in Figure 6, Croatia has one of the highest extended-NEET rates in the EU; only five countries have a higher rate.

Figure 6

Download asset Open asset

This unfavourable labour market situation has arisen due partly to the economic crisis that hit Croatia particularly severely. However, the underlying reasons for the high inactivity and unemployment seem to be mainly structural, rather than cyclical, and Croatia is, in that respect, similar to other south and south-east European countries. Since the adverse labour market impact of the economic downturn was stronger (and the subsequent recovery weaker) than in other EU countries, it also seems that the structural weaknesses amplified the cyclical factors.

As many theoretical and empirical contributions suggest, one structural factor that may be at play is the design of the tax-benefit system, insofar as it impacts work incentives.^[24] Various redistributive instruments – both taxes and benefits – and their disincentive effects with respect to the labour supply have been particularly emphasised in both scholarly work and policy circles. Policy makers across countries have realised that redistribution towards the needy, in the form of income support programmes for those out of work, can create significant work disincentives for the beneficiaries or their family members. In other words, pursuing equity objectives may have not only direct costs (social protection expenditures) but also indirect, efficiency costs (lower labour supply). Thus, some countries decided to try to reconcile the equity and efficiency objectives by implementing social protection instruments that “make work pay” (for a review, see Matsaganis and Figari, 2016).

One of these types of instruments are the so-called in-work benefits (henceforth, IWBs).^[25] The idea is to incentivise the needy to find employment by providing higher transfers to those who work but earn a low income than to those who do not work (although they are able to work). Unlike the traditional social benefits that are granted to people out of work, IWBs are conditional on employment, which is their most important distinguishing feature. An IWB is usually designed to have the following features: (i) a phase-in region, starting at some relatively low number of work hours (or some low amount of income earned), in which the amount of the benefit is increasing with work hours (or income), implying a negative marginal effective tax rate; (ii) a maximum amount of benefit paid to those who reach a certain number of hours (amount of income); (iii) a phase-out region in which the amount of the benefit is falling with the work hours (or income), up to a maximum number of work hours (amount of income), where the benefit drops to zero. In an optimal taxation framework, Saez (2002) has shown that if people choose their labour supply on both the extensive margin (whether to work or not) and the intensive margin (how much to work conditional on working), and if the labour supply responses of people from low-income families are more concentrated at the extensive than at the intensive margin, then social benefits with the above three features are likely to be optimal (that is, social welfare-maximising).

Whether an IWB will have the intended effects of making it more attractive for people to work and receive higher transfers than to stay out of work and receive lower transfers, depends on its design. A design feature that has proven crucial so far concerns the level at which the income test of eligibility is done. In that respect, the following two types of IWBs are distinguished: (a) family-based and (b) individual-based. The pioneering IWBs, introduced in the US and the UK in the early 1970s, known today as the Earned Income Tax Credit (EITC) and the Working Tax Credit (WTC), respectively, are family-based. The income test is done at the family level, by adding up the market incomes of all family members and comparing it to a threshold. Family based IWBs have the quality of being well targeted towards needy families, rather than to low-income individuals from relatively well-off families. However, this, in principle, comes at the price of significant work disincentives (high participation tax rate) for potential second earners if their potential earnings (to be realised upon eventual employment) are high enough to make the family ineligible for the IWB, according to the family income test. For that reason, some countries (for example, the Netherlands, Belgium, Sweden and Finland) decided to implement individual-based IWBs, where the income test is done at the level of the individual. The work disincentives for potential second earners are avoided in this way, but again, there is a price to be paid: individual-based IWBs may not be well targeted, as a beneficiary can be a low-earning person from a relatively well-off family. Thus, in general, there is a trade-off between how well targeted an IWB is and how much of a work incentive it provides for potential second earners.^[26]

A growing empirical literature using quasi-experimental identification methods finds, in general, that IWBs increase the labour supply of the beneficiaries at the extensive margin, thus, raising their families’ disposable income. On the other hand, there is little evidence of substantial labour supply responses at the intensive margin. Most quasi-experimental evaluations of the labour supply effects of the EITC in the US show that it has a significant positive impact on employment (for example, Eissa and Liebman, 1996; Eissa and Hoynes, 2004; Eissa and Hoynes, 2006; Meyer and Rosenbaum, 2001; Chetty et al., 2013).^[27] The results for other countries – such as the UK (Blundell et al., 2005; Leigh, 2007; Francesconi and van der Klaauw, 2007), Canada (Card and Hyslop, 2005; Milligan and Stabile, 2007), Sweden (Laun, 2017) and the Netherlands (Bettendorf et al., 2014) – are similar in nature, if not in magnitude, all showing positive employment impacts of IWBs.^[28]

Another growing strand of literature, where the present paper belongs to, evaluates the labour supply effects of IWBs ex-ante, using behavioural microsimulations. These studies, mostly done for European countries, often focus on the dependence of the labour supply effects of IWBs on their design, especially with respect to the family/individual-based distinction. An exemplary study is that of Bargain and Orsini (2006), who simulate the effects of two IWBs – one family-based, one individual-based – on social inclusion (employment) and poverty in Finland, France and Germany. They conclude, as the theory predicts, that the family-based IWB is more suitable for poverty reduction, while the individual-based IWS is better for increasing employment. Thus, the choice between the two crucially depends on the policy maker’s objectives. Broadly similar conclusions are reached by Vandelannoote and Verbist (2016), based on simulations for Belgium. Perhaps of greater importance for the present paper are the analyses for south and southeast European countries. Figari (2010) simulates family- and individual-based IWBs for Greece, Italy, Portugal and Spain. He finds that the expected trade-off between the redistributive and incentive effects is present but is rather weak if the employment rate of secondary earners (mostly women) is low. This points to the importance of the current labour market circumstances on the effects of IWBs. Focusing exclusively on married women and single mothers in Italy, Figari (2015) finds that both individual- and family-based IWB schemes can increase female participation and that it is possible to achieve more redistribution and efficiency at the same time. In another Italian study, Luca et al. (2014) show that the trade-off between efficiency and redistribution can be mitigated to some extent through a double-earner premium, with the remaining trade-off depending on the size of the premium. Ayala and Paniagua (2018) simulations suggest that IWBs would increase female participation in Spain as well. In the Western Balkan context, the studies of Mojsoska Blazevski et al. (2015) and Ranđelović et al. (2013) indicate potentially favourable employment effects of IWBs in Macedonia and Serbia, respectively. Both studies show that individual-based IWBs are more effective in incentivising potential second earners.

We briefly describe the Croatian system of direct taxes and cash social benefits in 2015, focusing on the most important instruments.^[29] All the amounts that we mention here are monthly amounts. For comparison, the average gross monthly employment income, calculated from the ILCS data for 2015, is approximately HRK 6,800. Employer and employee social insurance contributions (SIC) amount to 17.2% and 20% of gross earnings. The personal income tax (PIT) base for an employee equals gross earnings minus employee SIC, basic personal allowance (HRK 2,600) and allowances for dependent members (HRK 1,300; 1,820; 2,600 for the first, the second, the third child; and HRK 1,300 for adult dependants). The PIT base is divided into three parts, taxed by the marginal rates of 12% (HRK 0–2,200), 25% (HRK 2,200–13,200) and 40% (beyond HRK 13,200). Cities and municipalities impose a surtax, whose base is the PIT amount, and the rates are 0–10% in municipalities and 0–18% in cities.

A social insurance system provides replacement incomes during periods of unemployment, sickness and parental leave. The unemployment benefit lasts between 3 and 15 months, depending on the time spent in employment before the job loss. During the first three months of unemployment, the benefit amounts to 70% of the net monthly earnings, while in the remaining period, this amount drops to 35% of the net monthly earnings. The maximum benefit amount equals 70% of the national average net monthly wage for the first three months and 35% of the national average net monthly wage for the remaining period.^[30]

Two most important social benefits are the child benefit and the guaranteed minimum benefit. The child benefit is an income-tested benefit for households with children aged up to 15, or up to 19 if in secondary education. The net monthly household income per member may not be higher than HRK 1,663, and the benefit amount per child equals HRK 300, 250 or 200, depending on the income. There is also a supplement of HRK 500 (1000) for households with three (four or more) children, which was introduced as a pronatalist policy measure. The guaranteed minimum benefit is an asset- and income-tested benefit scheme for households with incomes below the “minimum subsistence amount”. The benefit amount equals the minimum subsistence amount minus the average net household income in the preceding 3 months. The minimum subsistence amounts for some typical households are as follows: HRK 800 for a single person; HRK 960 for a couple; HRK 1,600 for a couple with two children; HRK 1,660 for a lone parent with two children. In addition, the guaranteed minimum benefit recipients are eligible to receive compensation for electricity costs of HRK 200 per month, as well as compensation for housing costs, which are provided by cities and municipalities, with the ceiling amount set to 50% of the guaranteed minimum benefit amount.

We simulate the effects of two in-work benefits. One is inspired by the Working Tax Credit in the United Kingdom (hereafter: WTC), and the other is inspired by the Employee Tax Credit in Slovakia (hereafter: ETC). EUROMOD (Sutherland and Figari, 2013) is used to “export” the policies from the UK and Slovakian tax-benefit systems and “import” them into the 2015 Croatian tax-benefit system, as modelled in miCROmodA. We do not import them in the exact form in which they exist in their home countries, however. While preserving the basic mechanism of the original policies, we simplify and adapt the original WTC and ETC.^[31] We do these amendments for the sake of comparability between the WTC and ETC and to better fit the Croatian context. For example, some elements, such as the eligibility rules and definitions of the relevant income concepts, are aligned among the two benefits. No other changes to the Croatian tax-benefit system are introduced besides implementing the WTC or ETC, and both are excluded from the income used for income-testing the child benefit and the guaranteed minimum benefit.

The eligibility rules, income tests and benefit formulas for WTC and ETC are as follows:

Eligibility: Every person aged 18–65 who worked at least 520 hours in a given year is eligible. It does not matter how the minimum required hours are distributed over the year, eg, one may work 40 hours a week for three months, or 20 hours a week for six months, in continuity or with interruptions. It only matters whether a person worked at least 520 hours during the year.

Income test: Both benefits are income tested. The test income comprises the sum of the post-tax original and replacement incomes. In the case of ETC, only the income of the eligible person is taken into account, irrespective of whether the person is single or married. In contrast, the WTC also considers the income of the spouse if the person is married.

Benefit formula: The benefit amount (B) is given by the following formula:

where M is the maximum benefit amount, I is the test income, T is the income threshold, and r is the benefit withdrawal rate.

Table 6 shows the parameters for the two benefits. For the WTC, the maximum benefit amount is the sum of the following four components: (a) Basic component – for all eligible workers, (b) Lone parent component – for lone parents, (c) Couple component – for persons who have partners, and (d) 1560 hours component – for persons who have worked at least 1560 hours per year, which is the equivalent of 9 months of work at a full-time job. In the case of ETC, all eligible workers are entitled to the same maximum benefit amount. Note that the maximum benefit amount is much larger in the WTC, but the withdrawal rate is also higher than in the ETC.

To illustrate the functioning of the two in-work benefits, we use examples with two hypothetical couples. An “actor” is a hypothetical non-employed person who considers what to do regarding his or her activity in the next 12 months and has two options in this “pre-decision situation”.^[32] One option is to remain non-employed for the whole considered period. Alternatively, he or she can become employed for a certain number of months, with certain work hours per week and a certain monthly gross wage. We have two couples, A and B, in which the actors are spouses A2 and B2, respectively. In couple A, the other spouse, A1, is non-employed. In couple B, the spouse B1 is full-time employed (12 months during the year, 40 hours per week), earning 50% of the average monthly gross wage (AGW).

Figure 7 depicts the situations for couple A (panel a) and couple B (panel b). Spouses A2 and B2 consider whether to remain non-employed for the whole year, or to get a full-time job (40 hours per week) for the whole year. If they choose non-employment, the annual gross wage of A2 and B2 is zero. Otherwise, A2 and B2 can earn a monthly gross wage in the range from 40% to 100% of the AGW. The panels in Figure 7 show the benefit amount at the household level (left graph), the household disposable income (middle graph), and the participation tax rate faced by the actor (right graph), with each plotted against the household’s total gross employment income.

Figure 7

Download asset Open asset

Both spouses of couple A are non-employed in the pre-decision situation. If spouse A2 decides to remain non-employed, the benefit amount is zero for both WTC and ETC. If A2 becomes employed, the WTC provides much more generous support than the ETC (at least for gross wages of up to 80% of the AWG). For example, if A2 becomes employed at a gross wage of 40% of the AGW, the amount of the WTC is 13.4% of the AGW, compared to only 3.5% of the AGW in the case of the ETC. Both benefits provide incentive for participation, as both decrease the participation tax rate, but the WTC is much more effective than the ETC in this respect.

Quite a different picture is revealed for couple B. Because spouse B1 is already employed in B2’s pre-decision situation, the couple is receiving the WTC. If actor B2 becomes employed, the amount of the WTC falls due to the income test, which takes into account the couple’s total income. Even if B2 becomes employed at a low gross wage, for example, 40% of the AGW, the benefit amount falls to zero. This mechanism reduces incentives for participation by increasing the participation tax rate, which is substantially higher than in the baseline scenario. In contrast, in the case of the ETC, the benefit amount of a household increases if B2 obtains a job, and this increase occurs because the income test considers the incomes of the two spouses independently. The ETC decreases the participation tax rate if B2 becomes employed at a low gross wage, ie, up to 60% of the AGW.

The Baseline scenario simulates the actual tax-benefit system in 2015. Two counterfactual scenarios are created as follows: the WTC reform (ETC reform) scenario implements the WTC (ETC) into the tax-benefit system. We first simulate the effects on the labour supply, and then, taking the behavioural responses into account, we simulate the effects on the state budget, poverty and inequality. Simulations are performed for couples only, and the sample is the same as that for which the labour supply model for couples is estimated.

The two in-work benefits have different effects on the labour supply (Table 7). The WTC decreases the average probability of being employed at 2080 and 2600 hours for both males and females. In other words, individuals are more likely to choose less time intensive employment or to become non-employed. The probability of non-participation for males increases by 0.37 percentage points. For females, the increase is approximately twice as large, 0.79 percentage points. In contrast, the ETC decreases the probability of non-participation for both males and females by 0.35 and 1.17 percentage points, respectively. It also slightly decreases the probability of engagement in over-time work (2600 hours).

The labour supply effects reported in Table 7 are aggregate effects, pertaining to the entire sample of couples. The aggregate effects may, however, come as a net outcome of heterogeneous effects across different groups. We are particularly interested in the participation effects for males and females depending on the employment status of their spouses. Thus, for each gender, we compute the change in the average participation probability with the WTC and ETC reforms for those whose spouse is employed and for those whose spouse is not employed. As discussed before, the employment status of one’s spouse is likely to matter for one’s work incentives, and the importance of the spouse’s employment status is likely to depend on the type of in-work benefit (ie, whether it is individual- or family-based).

The results are shown in the upper part of Table 8. In the first row, we report the aggregate effect; these figures are equal to those reported in the first column of Table 7 (ie, for zero hours), multiplied by −one for convenience, so that they represent changes in the probability of participation, rather than non-participation. There are notable differences between the effects for persons whose spouses work and those whose spouses do not work. This is especially so in the case of the WTC, where one effect is positive and the other is negative. Precisely, upon the introduction of the WTC, the participation probability raises for males whose spouses do not work, and declines for those whose spouses work. The same holds for females, yet with notably stronger effects since the introduction of the WTC presumably leads to a greater income effect among females. For both genders, the negative aggregate effect is driven by the negative effect – decline in incentives for participation – for persons whose spouses work. In contrast, the ETC increases the participation probability for both genders irrespective of whether their spouses do or do not work, with an expectedly larger increase for those with non-working spouses. In summary, while the ETC improves the incentives for work independent of the spouse’s employment status, the WTC does so only if the spouse is not employed.

These results are in line with the structure of work incentives incorporated into the WTC, as we explained in the example with the hypothetical couple B (see Section 5.4), where only one partner participates at the time the other is considering whether or not to obtain a job. We saw that when one partner is already working, the other partner’s incentive to participate diminishes when the WTC is introduced and improves when the ETC is introduced. Moreover, note that for females whose spouse is not employed, the WTC increases the probability of participation more than the ETC does. This could also have been predicted based on differences in the structures of incentives to participate between the WTC and ETC for a couple where neither spouse works and one is considering whether to obtain employment. As we saw in the example with the hypothetical couple A, the higher generosity of the WTC makes it substantially more effective than the ETC in improving incentives for participation (up to a gross wage of 80% of the AWG). For males whose spouse is not employed, however, the ETC brings a larger increase in the participation probability. Nevertheless, the differences between the effects of the WTC and ETC are quite small both for males and females whose spouses do not work; thus, we should perhaps not draw any firm conclusions from these results.

The budgetary effects of the WTC and ETC are shown in Table 9. We report the aggregate amounts of income from employment (measuring labour supply), employer and employee SIC, personal income tax and surtax (PITS), benefits other than WTC/ETC (means-tested and other^[33]), and WTC/ETC. The amounts are reported for the baseline situation and two post-reform situations. The first post-reform situation does not take into account the labour supply responses to the introduction of the WTC/ETC, but only the mechanical effects. The second post-reform situation is the final situation which, in addition to the mechanical effects, takes into account the behavioural effects in the form of labour supply responses to the WTC/ETC. Given our calibration of the WTC and ETC, in both cases, the only mechanical effect is a 300-million increase in expenditures on social benefits (see rows 7 and 9), leading to a 300-million decrease in net taxes (rows 10 and 11).

Regarding the behavioural effects, note first that the changes in aggregate employment income have the same direction as the changes in labour supply: the ETC reform increases the aggregate employment income, while the WTC reform reduces it. The changes are small, however: a 0.76% decline in the case of the WTC reform and a 0.29% increase in the case of the ETC reform are observed. Obviously, the SICs and PITS change in the same direction as the employment income. In both cases, the means-tested and other benefits decrease slightly, due to the net increase of labour supply among benefit recipients. The largest difference is with respect to the change in the amount of the WTC/ETC. The behavioural responses to the WTC reform make the expenditures for the WTC 93 million or 31% higher relative to the amount that would be spent if there was no decline in the labour supply. This might seem puzzling given that the WTC reform reduces the probability of participation, and participation is among the eligibility criteria. Recall, however, that the WTC reform reduces the probability of full-time work (2080 hours) and increases the probabilities of lower positive hours for both genders (see Table 7). Thus, the WTC reform induces a fraction of those working full-time to reduce their labour supply to become eligible for the benefit. On the other hand, the introduction of the ETC induces labour supply responses that make its aggregate amount 16 million or 5.3% larger than it would be without the behavioural responses. Unlike the WTC reform, the ETC reform increases the participation probability and does not reduce the probability of full-time work; indeed, the latter increases slightly (see Table 7). These differences between the WTC and ETC reforms regarding their total (mechanical plus behavioural) effects stem from three crucial differences in their respective designs (see Table 6). First, the WTC is a family-based in-work benefit; thus, the sum of both spouses’ incomes is relevant for income testing. In contrast, the ETC is an individual-based benefit; thus, the spouses’ individual incomes are considered for income testing. This is one factor contributing to the decline in the probability of full-time work in the case of the WTC reform, but not in the case of the ETC reform. Second, the WTC is more generous than the ETC, even after allowing for the fact that the former is family-based and the latter individual-based. Third, the WTC is phased out at a much higher rate than the ETC. As a result, the total budgetary effects, expressed as changes in net taxes, amount to HRK –513 million for the WTC and HRK –252 million for the ETC. Therefore, the ETC reform increases the labour supply – unlike the WTC reform, which reduces it – and is 50% cheaper.

Finally, we assess the distributive impacts of the WTC and ETC. In Table 10, we report the changes in inequality (measured by the Gini coefficient) and poverty (measured by the poverty headcount, gap and squared gap indices). Both the inequality and poverty measures are based on the equivalised (OECD scale) disposable household incomes. Again, we distinguish between the mechanical and total (mechanical plus behavioural) effects. Considering only the mechanical effects on inequality, the WTC reduces the Gini coefficient twice as much as the ETC. However, since the WTC reform decreases employment for people with a working spouse, the total decrease (0.60 Gini points) is smaller than the mechanical one (0.80 Gini points). The ETC reform reduces the Gini coefficient by virtually the same amount (0.62 Gini points). We have a similar situation with poverty incidence, as evidenced by approximately the same total effects of the WTC and ETC on the poverty headcount measure. The poverty depth and severity – measured by the poverty gap and squared gap, respectively – are reduced more by the WTC reform than the ETC reform. However, taking account of their respective costs discussed above, the ETC reform is more cost-effective in reducing inequality and poverty.

The model has not been updated for the years after 2014.

See section 3.

The amounts of indirect taxes are imputed through a survey-to-survey imputation from the HBS as the “donor” survey into the ILCS as the “recipient” survey.

SRMOD for Serbia and MAKMOD for Macedonia are other examples of EUROMOD spin-off models.

See EUROMOD (2018).

For more details on miCROmodA, see Urban et al. (2018).

The wage rate is a job characteristic in a general formulation of the model. In our application, however, wages differ across individuals, rather than across jobs. Thus, in this paper, jobs are not characterised by the wage rate, but only by the hours of work and a set of unobserved features.

This distribution is sometimes alternatively called the Gumbel distribution.

In this static setting, saving is part of consumption (i.e., future consumption) and therefore disposable income equals consumption.

For example, income from financial assets or from renting property.

It embodies the personal income tax rules, including those pertaining to mandatory social security contributions, and the rules determining the eligibility for, and the amounts of, various social benefits, be they social insurance (e.g., unemployment benefit) or social assistance (e.g., minimum income support).

For details on how the hourly wages are obtained, see the next section.

We can include both experience and age, as experience is not defined as potential experience (age minus years of schooling minus six), but rather taken as actual experience, reported by survey respondents.

The original name of the survey in Croatian is Anketa o dohotku stanovništva (ADS).

For example, someone who was in school in some months and then employed for the rest of the year, or who was employed for a number of months and retired for the rest of the year. Such people are excluded because we model the annual labour supply, and they cannot be considered as flexible (in the above explained sense) for the whole year. The self-employed are also excluded since it is not clear what exactly their labour supply is, or at least how to measure it in hours.

There may be other persons in the household, but none of these may be flexible. For example, a flexible woman living with her inflexible son is included in the sample of single-individual households, or, more precisely, in the sample of single females.

Or monthly, or weekly, depending on how the hourly wage is obtained.

The descriptive statistics for all variables used in the estimation are reported in Table B.1 in the Appendix.

The alternative to the aggregate elasticities is to compute individual elasticities for all persons in the sample and average them. The two elasticities need not be the same. A sufficient condition for equality between them is that all individual elasticities are the same.

For simplicity, from here on, we will use the term “elasticity of participation” instead of the precise term “aggregate elasticity of the average probability of participation” and the term “elasticity of work hours” instead of the precise term “aggregate elasticity of the average expected work hours”.

However, males fare worse than females in comparison to their counterparts in other EU countries: Croatia has the highest male inactivity rate, while that for females is the sixth highest.

Nestić and Tomić (2018) call them the “NEET adults”, but we find the term “extended NEET” more appropriate.

Using the extended-NEET concept instead of the traditional concept of inactivity is especially relevant in the Croatian context, given that approximately a third of the working-age inactive people are retired, a consequence of a relatively large number of war veterans and early retirees.

For reviews, see, for example, Moffitt (2002), Krueger and Meyer (2002), Kleven and Thrustrup (2005) and Keane (2011).

Other instruments are minimum wages and wage subsidies (see Immervoll and Pearson, 2009).

Some researchers have proposed a “premium” for double-earner families as a design feature aimed at dealing with the trade-off (Luca et al., 2014; Kurowska et al., 2017).

See Hotz and Scholz (2003) for a review of studies focusing on the EITC.

Lundberg and Norell (2018) provide an extensive review of quasi-experimental evidence.

For more information, see the EUROMOD Country report for Croatia (Urban et al., 2018).

For the description of parental benefits and sickness benefits, see the EUROMOD Country report for Croatia (Urban et al., 2018).

For an explanation of original policies, see the following EUROMOD Country reports: De Agostini (2017) for the UK and Gabik and Paur (2017) for Slovakia.

The terminology used here (“actor”, “pre-decision situation”) does not mean that our empirical results can say anything about the dynamics of employment. This terminology is used for illustrative purposes only.

For example, unemployment benefits.

The estimation sample consists of all flexible persons. For the criteria that a person must satisfy to be considered flexible, see Section 3 (Data).

The hourly wage refers to the gross hourly wage. For a flexible person, it is calculated as follows. The reported monthly amounts of cash and in-kind employment incomes (both taxable and non-taxable) are added together, multiplied by 12 to obtain the annual amount, and then divided by the annual number of work hours to obtain the hourly amount. The annual work hours are obtained as described in Section 3 (Data). The actual, rather than discretised, number of hours is used.

The descriptive statistics for all variables used for the estimation of the Heckman selection model are displayed in Table A.1 (quadratics omitted). The estimates of the parameters from the wage and selection equations are given in Table A.2.

This work was fully supported by the Croatian Science Foundation, grant number UIP-2014-09-4096.

The authors thank the two anonymous referees for their valuable comments. The authors also thank John K. Dagsvik for the fruitful discussions and useful suggestions related to an earlier version of this paper. All the remaining errors and omissions are the sole responsibility of the authors.

1. Version of Record published: December 31, 2019 (version 1)

© 2019, Bezeredi, Ledić, Rubil, Urban

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.