The future of flexible labor for different economic growth scenarios.

(1)

The future of flexible labor for

different economic growth

scenarios.

Master’s Thesis Econometrics, Operations Research and

Actuarial Studies

Nynke J.J. Slagter

Supervised by: Prof. R.J.M. Alessie

(2)

(3)

The future of flexible labor for different economic

growth scenarios.

Nynke Slagter

Abstract

This paper predicts the medium-term future of flexible labor for two different definitions of flexible workers. The first one includes employees with temporary contracts without the perspective of a getting permanent contract, employees from employment agencies and freelancers. The second definition only includes the last two categories. This research uses dynamic microsimulation models as described by Knoef et al. [2013]. Households characteristics are updated and dynamic probit models based on Wooldridge [2005] are estimated to predict unemployment rates and percentages of flexible workers for different economic growth rates. The pre-dictions for flexible labor show an increase over time irrespective of definition and economic growth scenario. The predicted unemployment rates and predicted per-centages of flexible workers of the inexperienced subgroup are higher than average. For this subgroup the expected percentages of flexible workers fluctuate around a certain level for the first definition. On the contrary, the expected percent-ages of freelancers and employees from employment agencies increase over time. Analyzing four different regions in the Netherlands implies higher expected unem-ployment rates and higher expected percentages of flexible workers in the North in comparison to the rest of the Netherlands.

JEL codes: B21, B23, C33, C53, J21.

keywords: dynamic microsimulation, dynamic aging, future, flexible labor.

1 Introduction

(4)

Birsken Human Resources Innovations is interested in the future of flexible labor. The com-pany would like to respond quickly to changes in the labor market in order to facilitate and improve the link between employees and employers in case of temporary work projects. For this purpose and in light of trends in labor market sketched above, we are interested in the development of employment and flexible labor in the future. A so-called dynamic microsimula-tion model based on the methodology of the paper by Knoef et al. [2013] should be constructed for medium-term predictions in order to describe time-dependent changes. In order to predict the future of flexible workers the members present in the data set in 2012 will be dynamically aged until 2020. Subsequently, employment and being a flexible worker is predicted using the updated characteristics of the individuals until 2020.

Unemployment rates and percentages of flexible workers probably depend on economic growth. Since future economic growth rates are uncertain, different economic growth scenarios will be taken into consideration by estimating future unemployment rates and future percentages of flexible workers.

In this paper flexible workers are defined as employees with fixed-term contracts, employees from employment agencies and freelancers, all without the perspective of getting a permanent job (original definition). Additionally, this group is compared with a group of flexible work-ers, exclusively consisting of employees employed by an employment agency and freelancers (alternative definition). This paper will answer the following question. How will flexible labor develop in the future for different economic growth scenarios?

The results predict that the percentages of flexible workers increase irrespective of expected economic growth and definition. The expected percentages are higher for the northern region in comparison to the others and also people in the inexperienced subgroup are more likely to be a flexible worker than average.

The structure of this paper is as follows: the next section describes the data. Section 3 discusses the dynamic microsimulation model and the dynamic aging procedure. The following section states the estimation results. Section 5 demonstrates the simulation results and the last section concludes and discusses the paper.

2 Data

The data is obtained from the Labor Supply Panel (Arbeidsaanbodpanel, AAP) of The Dutch Institute for Social Research (Sociaal en Cultureel Planbureau). The data set contains infor-mation from surveys in the Netherlands, first carried out in 1985 and between 1986 and 2012 every other year. About 4,500 to 5,000 individuals between the age of 16 and 66 year were interviewed each period, with on average 25% from a new random sample.

(5)

dynamic models will be used and therefore only subsequent periods will be taken into consid-eration. Individuals have to participate at least two consecutive periods and this narrows the selected sample down to 24,142 observations.

The distribution of the labor force is given in Table 1. This table suggests that 774 respon-dents are unemployed, which gives an overall unemployment rate of 3.2%. The largest part of the labor force consists of employees with a permanent contract. Over 50% of the fixed-term contractors have the prospect of getting a permanent contract and are less relevant for this research in comparison to the 1,182 employees with a fixed-term contract without the perspec-tive of a permanent contract. The number of employees working for an employment agency is 693 and freelancers (1,065) are the largest group among the self-employed.

Table 1: Current labor market situation

Total Percentage Cumulative

Employee 22,062 91.4 91.4

Permanent contract 18,915 78.3 78.3

Fixed-term → permanent contract 1,272 5.3 83.6

Fixed-term 6→ permanent contract 1,182 4.9 88.5

Employment agency 693 2.9 91.4

Self-employed 1,238 5.1 96.5

Without employees 1,065 4.4 95.8

With employees 173 0.7 96.5

Other 68 0.3 96.8

Unemployed, looking for a job 774 3.2 100.0

Total 24,142 100.0

Source: AAP, own calculations

The distribution of the flexible workers, according to the original definition mentioned in the introduction, is given in Table 2. The number of employees with fixed-term contracts without the perspective of a permanent contract is 40.2% of the flexible workers. The second largest group consists of the freelancers, 36.2%. The remaining 23.6% are employees with employment agency contracts.

Table 2: Distribution flexible workers

Number Percentage Cumulative

Fixed-term contract 1,182 40.2 40.2

Employment agency contract 693 23.6 63.8

Freelancers 1,065 36.2 100

Total 2,940 100

(6)

Table 3: Distribution labor force, original definition

Item Number Percentage Cumulative

Unemployed 774 3.2 3.2

Other employment 20,428 84.6 87.8

Flex worker 2,940 12.2 100.0

Total 24,142 100.0

Using the alternative definition only 7.3% of the labor force is a flexible worker, which is al-most 4.9% less. The number of flexible workers, including employees with employment agency contracts and freelancers, is 1,758.

Table 4: Distribution labor force, alternative definition

Item Number Percentage Cumulative

Unemployed 774 3.2 3.2

Other employment 21,610 89.5 92.7

Flex worker 1,758 7.3 100.0

Total 24,142 100.0

The unemployment rates per period are given in the first column of Table 5. From Table 1 follows that the overall unemployment rate is 3.2% and the unemployment rate in 2002 is only 1.8%. The last two columns show the different percentages of flexible workers per period per definition. In 2002 and 2004, the percentages of flexible workers are low irrespective of definition, but they started to recover in 2006. However, discarding fixed-term contractors gives a small drop in 2010.

Table 5: Percentage of flexible worker per period

Year Mean

Employment Flexible worker Flexible worker

Original definition Alternative definition

1996 95.3% 13.4% 9.4% 1998 96.6% 11.8% 8.5% 2000 97.2% 11.2% 8.4% 2002 98.2% 9.7% 6.7% 2004 96.9% 9.0% 5.6% 2006 96.4% 12.5% 6.5% 2008 97.3% 14.0% 8.0% 2010 96.2% 15.3% 7.6% 2012 96.3% 16.3% 8.3% Total 96.8% 12.6% 7.5%

(7)

Some summary statistics concerning the data are given in Table 6. The average age in the data set is 42.4 years, with a minimum of 16 years and a maximum of 66 years. A percentage of 48% is women. On average the individuals in the data set have 1.2 children living at home, with a maximum of 9 children. The variable concerning experience on the labor market shows that people have 20.6 years of experience on average and at most 52 years.

Table 6: Summary statistics

Variable Mean Std. Dev. Min. Max. N

gender 0.48 0.50 0 1 24,142

age 42.37 10.97 16 66 24,142

year of birth 1962.14 11.59 1934 1994 24,142

children living at home 1.20 1.19 0 9 24,142

years of experience 20.63 11.15 0 52 24,142

For this research we are particular interested in (future) economic growth rates. The defi-nitions for economic growth and economic growth per region are the volume mutations in gross domestic product and gross regional product, respectively. The first column of Table 7 shows the different economic growth rates in the Netherlands. The other columns show the economic growth for four different regions in the Netherlands. The first region, the North, consists of the provinces Groningen, Friesland and Drenthe. The next region, East, is Overijssel, Gelder-land and FlevoGelder-land. Region South consists of North-Brabant and Limburg and the remaining region, West, includes Zeeland, Utrecht and North- and South-Holland.

For most periods economic growth is positive in the Netherlands and in the four different regions. However, economic growth per region was negative for the regions East and South in 2002. For the other regions the growth was small, but positive. The economic growth in 2012 was negative for all regions and as a consequence for the Netherlands as a whole. The magnitude of the negative growth was the largest for the region East and the smallest for the region West. Economic growth was still negative and of a large magnitude in 2014 for the northern region, unlike the other regions. Also the expected economic growth in the North for 2016 and 2018 is negative.

In order to analyze the effect of lagged economic growth per region on employment and on flexible labor a simple linear fixed-effects model with the group variable region is estimated with clustered standard errors. The estimation results in Table 8 show that the lagged variable of economic growth per region will probably influence employment, but it will most likely not effect being a flexible worker.

3 Dynamic microsimulation model

(8)

Table 7: Economic growth

Netherlands North East South West

1996 3.6 5.4 2.9 3.8 3.2 1998 4.5 2.2 4.6 3.6 5.4 2000 4.2 2.3 4.7 5.2 4.1 2002 0.1 0.5 -0.4 -1.2 0.8 2004 2.0 0.8 1.3 2.7 1.9 2006 3.5 1.4 4.9 3.8 3.3 2008 1.7 4.2 2.1 1.4 1.5 2010 1.4 4.6 0.2 3.0 0.8 2012 -1.1 -1.3 -2.7 -0.8 -0.5 2014 1.0 -4.4 1.6 1.8 1.7 2016 0.5/2/3.5 -1.5/-2/-0.5 1/2.5/4.0 0.9/2.4/3.9 1.1/2.6/4.1 2018 0.5/2/3.5 -1.5/-2/-0.5 1/2.5/4.0 0.9/2.4/3.9 1.1/2.6/4.1

Source: Centraal Bureau voor de Statistiek

Table 8: Fixed effects model for employment, with as group variable growth with clustered standard errors

Employment Being a flex worker

lag growth per region 0.00260*** -0.00019

(0.001) (0.001)

cons 0.93998*** 0.14498***

(0.006) (0.011)

N. of observations 22,807 22,096

Source: Centraal Bureau voor de Statistiek

From the aging module the individuals move into the employment module. In this module employment is predicted using the updated characteristics of the individuals until 2020. Four different probit models are estimated for this purpose. The first two models estimate the proba-bility of being employed or not for the two different definitions of flexible labor and the last two models are random effects probit models that analyze the probability of being a flexible worker if employed per definition. These models take the effects of age and year of birth into account. Other explanatory variables are sector and region at the first time someone participates in the survey, number of children living at home and years of experience. The random effects models consider unobserved heterogeneity.

3.1 Dynamic aging

In this paper cross-sectional dynamic aging will be used, since all individuals are updated before moving to the next period.

3.1.1 Mortality.

(9)

Aging module - Mortality - Retirement

- New children being born¹ - Children leaving home¹ - Years of experience¹

Employement module

Predict employment status using a tree of (random effects) probit models, taking into account:

- age and cohort effecs - household demographics - economic growth

SCP Labor Supply Panel (wave 2012)

repeated loop until 2020

Employment status until 2020

Knoef et al. [2013] Mortality rates Marital status Estimated logit model for retirement Aging Module Employment Module

Figure 1: Design of the dynamic microsimulation models

obtained using the variables age and gender. The rates for the Netherlands are collected from www.mortality.org. The mortality rates for the year 2012 are used, since more recent ones are not available. Then Monte Carlo simulation will be applied. If the value of the random draw is lower than the mortality rate, the individual has deceased. On the other hand, if the value of the random draw is higher, age will be updated.

3.1.2 Marital status.

Then the transitions in marital status should be predicted. Those transitions are required for the prediction of children to be born and people to retire. The data considers five categories of marital status:

• Married • Cohabitation

1_{The first two variables combined are used for the explanatory variable children living at home,}

(10)

• Divorced • Widowed

• Single/unmarried

The transitions are modeled by Knoef et al. [2013]. These models are estimated from the population register in the Netherlands. Data is available from January 1, 1995 to January 1, 2008. All persons born between 1917 and 1970 in the age group 36-90 years are used for the following transitions by Knoef et al. [2013].

unmarried →married married →divorced widowed →married divorced →married

For these multinomial logit models scaled age, scaled age squared, scaled year of birth and the scaled year of birth squared are explanatory variables. The coefficients are given in Table 14 of Appendix A. The transition from being married to being widowed will require the mortality rates. For simplicity it is assumed that partners are of opposite gender and the same age as the individual in the data. We apply Monte Carlo simulation each year in order to establish whether a transition occurs.

3.1.3 Number of children.

The probability of a child leaving home depends on age and gender of the child and is modeled by a logit model and the coefficients are given in Table 15. The coefficients are obtained from Knoef et al. [2013]. They used all children in the Inkomens Panel Onderzoek (IPO) in 2006 and 2007 and modeled the probability that a child left its parental home between these years. The IPO, is a representative sample of Dutch households which consists of an administrative panel data set with income information. Most of these data are from the Dutch National Tax Administration.

Surprisingly, our data set does not contain variables about the gender of the children. Conse-quently, an average probability will be obtained and this probability will be compared with a random draw from an uniform distribution.

The probability of a newborn is also modeled by a logit model, the coefficients of this model are from Knoef et al. [2013] and are given in Table 15. For this estimation all households of the IPO data set in the years 1989 − 2006 are selected and it was checked whether a new child has entered the household during the following year, given the characteristics in t − 1. For the variable couple, we used the variable married to make it predictable. Each year also the dummy variables for having one or two children are updated, as is the dummy variable for being married using the model in section 3.1.2.

Both logit models are required to update the number of children living at home using Monte Carlo simulation.

3.1.4 Years of experience.

(11)

3.1.5 Pension

For transitions into retirement a logit model is estimated, since the model estimated by Knoef et al. [2013] did not meet the requirements. It did not estimate transition into retirement, but into occupational pension. Additionally, only the transition model for singles was given. The logit model was obtained from the AAP data set. All individuals in the labor force in 2010 are selected and a binary variable is generated whether a person retired between 2010 and 2012. Using this variable a logit model was estimated with the regressors scaled age, scaled age squared and some binary variables for marital status.

The estimation results are given in Table 9. The effect of scaled age on retirement is neg-ative. On the other hand, the effects of scaled age squared is positive. Being divorced or being married has a negative effect on the probability of retirement in comparison to being single or being widowed.

Table 9: Logit model for retirement

Coef SE age/10 -4.402*** 0.264 (age/10)2 0.704*** 0.043 divorced -0.812 0.428 married -1.048*** 0.250 Number of observations 3421 Log-likelihood -390.387 ∗ _{significant at α < 0.05;}∗∗ _{significant at α <} 0.01;∗∗∗ significant at α < 0.001

Monte Carlo simulation is used in order to determine if someones retires in the future. The probability of retirement is calculated using the estimated model and this probability is com-pared with a random draw from an uniform distribution.

3.2 Modeling labor market states

In order to model the probability of being a flexible worker a tree is used as given in Figure 2. The upper branch represents having a job and the lower branch considers being unemployed. The upper branch divides into two others categories, the upper one is the interesting one, being a flexible worker and the lower category demonstrates other employment.

Employment

Unemployment

Flexible worker

Other employment

Figure 2: Tree of the binary choice models

(12)

characteristics. Our model contains also the endogenous variables lagged employment and the lagged variable of being a flexible worker.

Suppose that for each individual i the observations start at time Si and ends at time Ti. The

number of observations for individual i is given by Ni. The latent variable of being employed

or being a flexible worker is given by

y_it∗ = z0_iβz+ x0itβx+ y0i,t−1ρ + ςi+ εit, (1)

where yi,t−1 = (employedi,t−1, f lexworkeri,t−1)0, zi include the time-invariant regressors and

xit all time-variant regressors. Define xi = (xi,Si+1, . . . , xi,Ti). According to Wooldridge [2001] the parameter ρ reflects true state dependence after controlling for unobserved het-erogeneity, ςi. Unobserved heterogeneity causes spurious state dependence, that is for ρ = 0,

P (yit= 1 | yi,t−1, zi, xi) 6= P (yit= 1 | zi, xi).

Wooldridge [2005] assumed a balanced data set, Si = 0 and Ti = T . Two assumptions are

made. The first one assumes that the dynamics are correctly specified, at most one lag of yit

appears in the distribution and the second one states that εit | yi,Si, zi, xi ∼ N.I.D(0, σ

2 ε),

independent over t and i. These assumptions together imply equation (3) below. The variable yit can represent being employed or being a flex worker.

P (y∗_it> 0 | yi,t−1, . . . , yi,Si, zi, xi, ςi) = P (yit= 1 | yi,t−1, . . . , yi,Si, zi, xi, ςi) (2) = Φ z_i0βz+ x0itβx+ yi,t−10 ρ + ςi , (3)

where Φ is the cumulative distribution function of a standard normal variable. Under the assumptions stated the likelihood contribution for individual i is given by

f (yi,Si+1, . . . , yi,Ti | yi,Si, zi, xi, ςi; βz, βx, ρ) =

Ti Y t=Si+1 f (yit| yi,t−1, . . . , yi,Si, zi, xit, ςi; βz, βx, ρ) (4) = Ti Y t=Si+1 Φ z0_iβz+ x0itβx+ yi,t−10 ρ + ςi yit_× 1 − Φ z0 iβz+ x0itβx+ y0i,t−1ρ + ςi 1−yit . (5) There is no log likelihood function to estimate βz, βxand ρ, because of the incidental parameter

problem. That is that the joint estimation of the unobserved effects ςi and the coefficients βz,

βx and ρ gives inconsistent estimates. This problem can be solved by integrating out ςi.

Another problem that should be solved is the initial conditions problem. Since it cannot be assumed that ςi and yi,Si are independent, Wooldridge [2005] suggested that the conditional distribution of the unobserved effect should be used in order to address this problem. We have for individual i

f (yi,Si+1, . . . , yi,Ti | yi,Si, zi, xi; βz, βx, ρ) =

∞

Z

−∞

f (yi,Si+1, . . . , yi,Ti | yi,Si, zi, xi, ςi; βz, βx, ρ)×

(6) h(ςi| yi,Si, zi, xi; αy, αz, αx)dςi.

A convenient choice is h(ς | yi,Si, zi, xi; αy, αz, αx) ∼ N.I.D.(y

0 i,Siαy + z 0 iαz + x 0 iαx, σ2η),

where ηi | yi,Si, zi, xi ∼ N.I.D(0, σ

2

η), which follows by writing

ςi= yi,S0 iαy+ z

0

(13)

However, Wooldridge [2005] assumed a balanced data set, contrary to the data set of this research. The analysis in Appendix C gives an alternative, but similar model. Define yi,Si,Ni as the interaction variables of being employed and being a flex worker at time Si with the

number of periods participating (Ni) and the interaction variables of being employed and being

a flex worker with the first period of participating (Si) and let αy includes the corresponding

coefficients. From the analysis in appendix C follows that setting ςi = yi,S0 i,Niαy+ z

0

iαz+ ¯x0iαx+ ηi (8)

gives a suitable model, where ηi| yi,Si,Ni, zi, xi ∼ N.I.D.(0, σ

2

η) and ¯xi are the means over the

different periods with the initial values included. Hence, the formula for the latent variable becomes y∗_it= z_i0βz+ x0itβx+ yi,t−10 ρ + yi,S0 i,Niαy+ z 0 iαz+ ¯x0iαx+ ηi+ εit (9) = z_i0γz+ x0itβx+ yi,t−10 ρ + y 0 i,Si,Niαy+ ¯x 0 iαx+ ηi+ εit, (10) where γz= βz+ αz.

3.3 Simulation

In order to simulate the future employment rates of 2014, 20016, 2018 and 2020 every even year, the results of the estimated probit models are required. Using the estimates given in one of the first two columns of Table 10, each one using a different definition of a flexible worker, the estimated value of the latent variable for being employed is given by the following equation:

\

employed∗_i,t+1= z0₂₀₁₂γˆz+ ˆx0i,t+1βˆx+ ˆyi,t0 ρ + yˆ 0

i,Si,Niαˆy+ ¯x

0

iαˆx+ ˆεit, (11)

where ˆx0_i,t+1 is obtained as described in section 3.1 and ˆεit is simulated from the N(0,1)

distri-bution.

The estimated latent variable for being a flex worker is defined by the equation below. The es-timates can be obtained from one of the last two columns in Table 10 each one using a different definition for a flexible worker.

\

f lexworker∗_i,t+1= z₂₀₁₂0 γˆz+ ˆx0i,t+1βˆx+ ˆy0i,tρ + yˆ 0

i,Si,Niαˆy+ ¯x

0

iαˆx+ ˆηi+ ˆεit, (12)

where ˆεit is simulated from the N(0,1) distribution for each t and each i and ˆηi is simulated

from the N (0, ση) distribution for each i.

For both models hold that if the value of the estimated latent variable for person i at time t + 1 higher is than 0, the binary variable of being employed or being a flexible worker is set to 1.

4 Estimation results

First two random effects probit models were estimated for employment with the Stata com-mand xtprobit, each one using a different definition of a flexible worker. The null hypothesis for the individual effect ηi with variance ση2 is given by H0 :

σ2 η

1+σ2

(14)

The next two random effects probit models estimate the probability of being a flexible worker if someone is employed each one using a different definition of a flexible worker. These models are estimated with xtprobit, since the null hypothesis should be rejected. The number of observations in these two models is 16, 603.

For the age and birth effects variables are created containing a linear spline. Linear splines allow estimating the relationship between the dependent variable and independent variable as a piecewise linear function. The variables are constructed such that the coefficients represent the additional change of the coefficient in comparison to the coefficient of the preceding interval. The dots for age are set at 35 and 55 years of age and for the birth cohorts at the years 1945, 1965 and 1985.

In order to observe a time trend Deaton Paxson dummy variables were included in all models at first. These dummy variables for the different periods account for the identification problems that arise because of the linear relationship between age, cohort and period. According to Deaton and Paxson [1994] the coefficients of the time dummies should sum up to 0 and should be orthogonal to the time trend in order to address the identification problem. More informa-tion about these dummy variables can be found in Appendix D. Analysis suggested that the Deaton Paxson period dummies are jointly insignificant in all models and can be removed. The interaction variables of employment and being a flex worker during the first period some-one participated times the number of periods participating and the interaction variables of employment and being a flex worker during the first period someone participated times the first period someone participated are also redundant. Next, the effect of adding xi,Si, as sug-gested in Appendix C, was also examined and as expected it was negligible. Furthermore, the variable concerning education level at the first period someone participated was evaluated and not included, just as the dummy variables about marital status.

The variable about region at the first period someone participates in the survey was jointly insignificant in the first two models and therefore not inserted in these models. On the other hand, the variable lagged economic growth per region is included in the first two models, whereas in the last two models lagged economic growth in the Netherlands and the region someone participated the first time are explanatory variables.

Next, the average partial effects are obtained using the delta method. The calculations are de-scribed in Appendix E. The average partial effects of the first two models are given in the first two columns of Table 11. The results are discussed below and the results for the alternative definition are given in brackets.

The average partial effects imply that being employed two years ago increases the expected probability of being employed with 26.3 (23.5) percentage points, holding all other factors con-stant. However, for a flexible worker this probability is predicted to be 2.5 (0.8) percentage points lower, but for the alternative definition this effect is insignificant.

Having children living at home has a significant negative effect on the expected probability of being employed, ceteris paribus, the effect increases with the number of children. Finally, the predicted effect of an one percent increase in economic growth per region gives an increase in the expected probability of being employed of 0.2 percentage points.

(15)

Table 10: Estimation results

model being employed2 _{being employed}2 _{being a} _{being a}

flex worker flex worker

definition original alternative original alternative

lag work=1 1.67919*** 1.56975*** -1.12625*** -0.61753***

(0.109) (0.106) (0.132) (0.169)

lag flex worker=1 -0.43894*** -0.16390 1.11783*** 1.27222***

(0.091) (0.118) (0.078) (0.103) children=1 -0.20797* -0.20405* -0.09661 -0.08153 (0.093) (0.092) (0.103) (0.128) children=2 -0.35830** -0.35853** -0.13843 -0.03903 (0.130) (0.128) (0.100) (0.127) children=3 -0.57476** -0.58276** 0.04327 0.28585 (0.190) (0.188) (0.136) (0.169) children≥4 -0.94306*** -0.94566*** 0.13649 0.28077 (0.278) (0.276) (0.213) (0.265) mean children 0.22524*** 0.22299*** -0.01519 -0.02856 (0.066) (0.066) (0.022) (0.034) age3_{: (.,34)} _0.02663 _0.03266 _0.00041 _0.00067 (0.037) (0.037) (0.027) (0.035) age3_{: (35,54)} _-0.08735* _-0.09401* _0.04647 _0.04386 (0.040) (0.039) (0.030) (0.038) age3: (55,.) -0.09198** -0.09217** 0.03167 0.05610 (0.032) (0.032) (0.029) (0.034) mean age3_{: (.,34)} _-0.01370 _-0.01296 _0.03191 _0.01692 (0.037) (0.036) (0.031) (0.040) mean age3_{: (35,54)} _0.03886 _0.03561 _-0.08004* _-0.07463 (0.041) (0.040) (0.035) (0.045) mean age3: (55,.) 0.08638 0.08419 0.03702 0.02197 (0.047) (0.047) (0.045) (0.054) years of experience 0.03305*** 0.03460*** (0.009) (0.009)

mean years of experience -0.00473 -0.00321

(0.010) (0.010)

lag growth per region 0.04702** 0.04377*

per region (0.017) (0.017)

mean lag growth -0.00170 0.00036

per region (0.035) (0.035) workS=1 0.15448 0.17139 0.07629 0.35755 (0.114) (0.111) (0.209) (0.253) flex workerS=1 0.01659 0.03597 1.19614*** 1.71622*** (0.092) (0.120) (0.125) (0.196) 2 _{probit model}

(16)

Table 10 continued: Estimation results

sectorS (agriculture) industry -0.85429*** -0.85094*** (0.176) (0.210) construction -0.54733** -0.48981* (0.192) (0.227) trade, catering -0.83540*** -0.83267*** (0.169) (0.201) transport -0.96046*** -1.34514*** (0.189) (0.246) business services -0.76969*** -0.69952*** (0.169) (0.198) well-being -1.04397*** -0.97455*** (0.172) (0.203) other services -0.74265*** -0.73037*** (0.187) (0.221) government -1.40690*** -1.25426*** (0.199) (0.234) education -1.30682*** -1.40760*** (0.189) (0.236)

regionS (reference North)

East -0.24655** -0.23760* (0.092) (0.115) South -0.16930 -0.16991 (0.090) (0.112) West -0.16823* -0.16866 (0.084) (0.104) female=1 0.00431 -0.06486 (0.076) (0.098) female=1 × children=1 0.26642* 0.31308 (0.134) (0.169) female=1 × children=2 0.32545** 0.47046** (0.120) (0.150) female=1 × children=3 0.34757* 0.33800 (0.165) (0.196) female=1 × children≥4 0.33779 0.33034 (0.282) (0.337)

∗ _{significant at α < 0.05;}∗∗ _{significant at α < 0.01;} ∗∗∗ _{significant at α < 0.001} 2 _{probit model}

(17)

Table 10 continued: Estimation results

year of birth3: (.,1944) -0.02541 0.00314 (0.040) (0.048) year of birth3_{: (1945,1964)} _0.01651 _-0.02806 (0.042) (0.050) year of birth3_{: (1965,1984)} _0.04802** _0.05117** (0.015) (0.019) year of birth3: (1985,.) 0.17105*** 0.08006* (0.028) (0.036) lag growth 0.04414* 0.07081** (0.018) (0.022) constant -0.31734 -0.54416* 47.87682 -8.14162 (0.218) (0.214) (77.295) (92.598) ln(σ2) -0.05260 0.19667 (0.147) (0.174) N. of observations 17046 17046 16603 16603 Log-likelihood -1494.436 -1513.724 -3662.120 -2556.707

Likelihood ratio test

H0 : ρ = 0 0.796 1.941 154.815 127.416

p-value 0.186 0.082 0.000 0.000

∗ _{significant at α < 0.05;}∗∗ _{significant at α < 0.01;} ∗∗∗ _{significant at α < 0.001} 2 _{probit model}

3 _{contains linear spline}

the probability of being a flexible worker with 15.6 percentage points. However, if someone was employed as a flex worker the expected probability increases with 15.5 percentage points. Furthermore, for each increase of one percent in economic growth, the probability of being a flexible worker is expected to increase with 0.4 percentage points.

The alternative definition gives different average partial effects according to the last column of Table 11. The predicted effect of being employed is only −4.9 percentage points, holding all other factors constant, whereas the additional effect of being a flexworker in the previous period is assumed to be 13.3 percentage points. Hence, the expected effect of being employed is substantially smaller for the alternative definition.

The effect of an one percent increase in economic growth in the previous period probably gives an increase of 0.4 percentage points in the probability of being a flexible worker according to the alternative definition, ceteris paribus.

(18)

Table 11: Average partial effects

model being employed being employed being a being a

lag work=1 0.26269*** 0.23541*** -0.15624*** -0.04895**

(0.033) (0.031) (0.026) (0.017)

lag flexible worker=1 -0.02513*** -0.00827 0.15526*** 0.13329***

(0.007) (0.007) (0.019) (0.022) children=1 -0.00841* -0.00827* 0.00402 0.00447 (0.004) (0.004) (0.007) (0.005) children=2 -0.01641* -0.01656* 0.00316 0.01258* (0.007) (0.007) (0.007) (0.006) children=3 -0.03154* -0.03258* 0.02262* 0.03095** (0.013) (0.014) (0.011) (0.010) children≥4 -0.07013* -0.07168* 0.03232 0.03021 (0.032) (0.033) (0.018) (0.016) age: (.,34) 0.00119 0.00147 0.00004 0.00004 (0.002) (0.002) (0.003) (0.002) age: (35,54) -0.00390* -0.00423* 0.00442 0.00272 (0.002) (0.002) (0.003) (0.002) age: (55,.) -0.00411** -0.00415** 0.00301 0.00348 (0.001) (0.001) (0.003) (0.002) years of experience 0.00148*** 0.00156*** (0.000) (0.000)

lag growth per region 0.00210** 0.00197*

(0.001) (0.001)

lag growth 0.00420* 0.00439**

(0.002) (0.001)

∗ _{significant at α < 0.05;}∗∗ _{significant at α < 0.01;}∗∗∗ _{significant at α < 0.001}

younger than 35, the expected probability of being employed increases with 0.12 (0.15) per-centage points, holding all other factors constant. For people in the labor force between 35 and 55 years of age, the expected probability decreases with 0.27 (0.28) percentage points for an increase in age of one year and for the oldest group the effect is even higher, -0.68 (-0.69) percentage points.

(19)

Table 12: Average partial effects of age

model being employed being employed being a being a

age: -,34 0.00119 0.00147 0.00004 0.00004 (0.002) (0.002) (0.003) (0.002) age: 35,54 -0.00271*** -0.00276*** 0.00446*** 0.00276*** (0.001) (0.001) (0.001) (0.001) age: 55,- -0.00682*** -0.00691*** 0.00747** 0.00624*** (0.001) (0.001) (0.002) (0.002)

∗ _{significant at α < 0.05;}∗∗_{significant at α < 0.01;} ∗∗∗ _{significant at α < 0.001}

−.01

−.005

0

.005

Average partial effects

age: −,34 age: 35,54 age: 55,− Effect with respect to

−.01

−.005

0

.005

Average partial effect of age with 95% confidence intervals.

(a) Original definition (b) Alternative definition

Figure 3: Average partial effects per age cohort for the dynamic probit models of being employed

5 Simulation results

For estimating the labor market states 50 simulations are performed per economic growth sce-nario and average unemployment rates and percentage of flexible workers are calculated. The same individual effect is used, such that the effect of economic growth can be obtained. The results of the simulation are analyzed by a decomposition of the effects of the differ-ent factors. Each factor was changed separately in each period. The results of this analysis are given in Appendix G. It should be noticed that the order in which the factors are changed was chosen at random. Besides the factors described below, the percentages will be influenced by people leaving the data set due to death or retirement.

(20)

eco-−.01 −.005 0 .005 .01 .015

−.01 −.005 0 .005 .01 .015

Average partial effect of age with 95% confidence intervals.

Figure 4: Average partial effects per age cohort for the dynamic probit models of being a flexible worker if being employed

nomic growth scenarios per region are given in Table 7. The regions are used as given in 2012. This figure predicts a large rise in unemployment in 2014. This can be explained by the low economic growth per region in the previous period, 2012. Additionally, an expected increase in the number of children living at home, which follows from the aging module, and increasing age raise the predicted unemployment rate.

As a result of the higher economic growth per region, unemployment is expected to decrease by 1.3% in 2016 in comparison to 2014. The drop is limited as a consequence of the high unemployment rate in 2014. A high unemployment rate in the previous period implies a higher unemployment rate in the current period, holding all other factors constant according to the average partial effects given in Table 11. Furthermore, the increasing expected number of chil-dren will most likely influence employment negatively, as will the increasing age.

Different scenarios for economic growth per region in 2016 are required in order to predict the unemployment rate of 2018, since the exact figures for economic growth are unknown at the moment. A range of 3 percent is chosen. As expected, low economic growth in the previ-ous period implies a higher unemployment rate. For the analysis below we used the average economic growth scenario for 2018 and 2020.

(21)

Again, increasing age seems to have a negative effect on expected employment in 2020, like the increasing unemployment rate of the previous period. The expected number of children has no clear effect in this period.

In conclusion, the most relevant factors are age and the lagged variables of economic growth per region and employment. The expected increase in the number of children living at home in the first three periods is also of importance.

4 5 6 7 8 Percentage 2010 2012 2014 2016 2018 2020 Year

Pessimistic growth rate Average growth rate

Optimistic growth rate

Original definition for flexible workers

Unemployment rates for different economic growth rates

Figure 5: Unemployment rates using the original definition

Figure 6 demonstrates the different percentages of flexible workers for different economic sce-narios. Three different scenarios will be considered for 2018 and 2020; a pessimistic one (0.5% economic growth), an optimistic one (3.5% economic growth) and an average one (2% economic growth). In Figure 6 the original definition for a flexible worker is assumed.

Most of the time the expected percentages of flex workers will increase with time, except for the year 2014. This can be explained by the negative economic growth in the previous period (2012). Factors which are most likely responsible for the rising line are increasing age and the positive average partial effect of the lagged variable of being a flexible worker. During the first two periods, the effect of children living at home on the expected percentage of flex-ible workers is also positive. Furthermore, the high unemployment rate of 2014 restrains the growth of the expected percentage of flexible workers in 2016. In 2018 and 2020, the expected percentages of flex workers will increase for the chosen range of economic growth, although the slope depends on it.

(22)

14 16 18 20 Percentage 2010 2012 2014 2016 2018 2020 Year

Percentage of flexible workers for different economic growth rates

Figure 6: Percentages of flexible workers using the original definition

the alternative definition and the effect of children living at home is larger. Furthermore, the high unemployment rate of 2014 has almost no effect using the alternative definition, which also follows from the smaller average partial effect.

4 5 6 7 8 Percentage 2010 2012 2014 2016 2018 2020 Year

Alternative definition for flexible workers

Figure 7: Unemployment rates using the alternative definition

(23)

sub-6 8 10 12 14 Percentage 2010 2012 2014 2016 2018 2020 Year

Percentage of flexible workers for different economic growth rates

Figure 8: Percentages of flexible workers using the alternative definition

group considers people who entered the labor force recently, the maximum age of this group in 2012 depends on education attainment and can be found in Table 13.

Table 13: Subgroup inexperienced

Education age2012 ≤

Primary school 20

Lower vocational education (LBO)/Lower general secondary education (VMBO) 23

Intermediate vocational education (MBO)/Higher general secondary 23

education (HAVO)/ Pre-university education (VWO)

Higher Vocational Education (HBO) 26

University 27

(24)

5 10 15 Percentage 2010 2012 2014 2016 2018 2020 Year

Average economic growth Original definition Unemployment rates 4 6 8 10 12 14 Percentage 2010 2012 2014 2016 2018 2020 Year

Average economic growth Alternative definition Unemployment rates

Labor force Low−educated

Inexperienced

Figure 9: Unemployment rates for average economic growth per subgroup

first year the expected percentages of flexible worker are about 3% higher for the inexperienced people, but after 2014 there is a substantial increase.

The low-educated subgroup follows the same pattern as the reference group again, but the percentages are slightly higher.

Subsequently, four different regions in the Netherlands are compared. These regions are the same as mentioned in section 2, except for Zeeland, this province belongs to the region South here. The predicted unemployment rates per region for the average economic growth scenario are given in Figure 11. The estimated unemployment rates in the three northern provinces are higher than in the rest of the Netherlands. The same can be concluded for the estimated percentages of flexible workers using one of the two definitions according to Figure 12. These observations can be related to the lower predicted economic growth rates for the North in com-parison to the other regions. Furthermore, the given percentages of flexible workers of 2010 and 2012 are lower in the West, but for the simulated percentages this is not the case.

6 Conclusions

For both definitions the percentages of flexible workers is expected to increase with time, except for the year 2014, probably due the negative economic growth in 2012. For the last two periods we predict an increasing line of which the slope depends on the economic growth scenario. High economic growth implies high employment and a high percentage of flexible workers.

(25)

10 20 30 40 50 Percentage 2010 2012 2014 2016 2018 2020 Year

Average economic growth Original definition Percentages of flexible workers

5 10 15 20 25 Percentage 2010 2012 2014 2016 2018 2020 Year

Average economic growth Alternative definition Percentages of flexible workers

Inexperienced

Figure 10: Percentages of flexible workers for average economic growth per subgroup

2 4 6 8 10 Percentage 2010 2012 2014 2016 2018 2020 Year

Average economic growth Original definition Unemployment rates 2 4 6 8 10 Percentage 2010 2012 2014 2016 2018 2020 Year

Average economic growth Alternative definition

Unemployment rates

North East

South West

(26)

14 16 18 20 22 Percentage 2010 2012 2014 2016 2018 2020 Year

Average economic growth Original definition

Percentages of flexible workers

6 8 10 12 14 Percentage 2010 2012 2014 2016 2018 2020 Year

Average economic growth Alternative definition

North East

South West

Figure 12: Percentages of flexible workers per region for average economic growth

higher than average. However, unemployment is expected to be larger under people who just entered the labor market. The expected percentages of flexible workers are about 25% higher if we consider the original definition. If employees with fixed-term contracts are disregarded, the expected percentages of flexible workers are almost 5% higher until 2014 and in 2020 the difference will even rise to more than 10%. From this can be concluded that for young people the percentages of freelancers and employes with employment agency contracts are expected to increase from 2014, whereby the percentages of employees with fixed-term contracts will only decrease.

Analyzing four different regions in the Netherlands implies that the expected unemployment rates and expected percentages of flexible workers will probably be higher in the North than in any of the other regions. One explanation could be that for most of the years the economic growth in the North is lower than in the other regions.

The following remarks should be taken into consideration. The final data set seems biased, since the overall unemployment rate is higher under the individuals with missing values. The overall unemployment rate of the raw sample was 4.8% and after removing the observations with missing values or which are nonconsecutive, the overall unemployment rate decreases to 3.2%.

(27)

References

A.S. Deaton and C.H. Paxson. Saving, growth, and aging in Taiwan, pages 331–357. University of Chicago Press, 1994. doi: 10.3386/w4330.

W. Greene. Econometric Analysis. Prentice Hall, 7 edition, 2012. ISBN 9780131395381. M. Knoef, R. J. M. Alessie, and A. Kalwij. Changes in the Income Distribution of the Dutch

Elderly between 1989 and 2020: A Dynamic Microsimulation. Review of Income and Wealth, Series(3):460–485, September 2013.

Averill M. Law and David M. Kelton. Simulation Modeling and Analysis. McGraw-Hill Higher Education, 3rd edition, 1999. ISBN 0070592926.

S. Rabe-Hesketh and A. Skrondal. Avoiding biased versions of Wooldridge’s simple solution to the initial conditions problem. Economics Letters, 120(2):346–349, 2013.

A. Skrondal and S. Rabe-Hesketh. Handling initial conditions and endogenous covariates in dynamic/transition models for binary data with unobserved heterogeneity. Journal of the Royal Statistical Society: Series C (Applied Statistics), 63(2):211–237, February 2014. J. Wooldridge. Simple solutions to the initial conditions problem in dynamic, nonlinear panel

data models with unobserved heterogeneity. Journal of Applied Econometrics, 20(1):39–54, 2005.

(28)

Appendix A

Marital status

Table 14: Multinomial logit model for marital status

Men Women Coef SE Coef SE Married → Divorced age/10 1.461 0.0414 1.611 0.0482 (age/10)2 -0.190 0.0043 -0.224 0.0052 (year of birth-1900)/10 0.179 0.0404 -0.097 0.0477 ((year of birth-1900)/10)2 _0.013 _0.0037 _0.038 _0.0042 constant -8.556 0.0894 -7.888 0.1054 Divorced → Married age/10 -0.388 0.0560 0.167 0.0704 (age/10)2 _-0.032 _0.0057 _-0.098 _0.0074 (year of birth-1900)/10 -0.068 0.0539 0.290 0.0708 ((year of birth-1900)/10)2 _-0.010 _0.0050 _-0.031 _0.0064 constant -0.069 0.1197 -2.975 0.1533 Unmarried → Married age/10 -2.732 0.0751 -1.954 0.1066 (age/10)2 0.203 0.0086 0.124 0.0123 (year of birth-1900)/10 1.654 0.0919 1.927 0.1305 ((year of birth-1900)/10)2 -0.138 0.0077 -0.151 0.0108 constant -0.918 0.1987 -4.154 0.2755 Widowed→ Married age/10 1.254 0.1353 1.234 0.1467 (age/10)2 _-0.175 _0.0114 _-0.188 _0.0130 (year of birth-1900)/10 0.204 0.0919 0.419 0.1101 ((year of birth-1900)/10)2 -0.026 0.0106 -0.018 0.0118 constant -5.406 0.3319 -7.481 0.3376

(29)

Appendix B

Children

Table 15: Logit models for the number of children living at home

Children not leaving their parental home Coef SE

age child -0.008 0.0147

age child2 _-0.005 _0.0004

male 0.474 0.0445

constant 4.072 0.1198

New children being born Coef SE

age/10 -5.983 0.1449 age/102 0.459 0.0154 (year of birth-1900)/10 -0.432 0.1666 ((year of birth-1900)/10)2 _0.073 _0.0144 man 0.679 0.0232 couple 1.514 0.0473 one child 0.753 0.0287 two children -0.637 0.0306 constant 11.220 0.4473

(30)

Appendix C

Model

The method suggested by Wooldridge [2005] requires a balanced dataset. Wooldridge solution with interaction variables

Since the data set is unbalanced, the possibility of including interaction variables, where yi,Si is interacted with the number of periods participating (Ni) and where yi,Si is interacted with the first period participating (Si) are considered.

Define yi,Si,Ni as the interaction variables described above. The vector αy includes all the corresponding coefficients. A choice is h(ς | yi,Si,Ni, zi, xi; αy, αz, αx) ∼ N.I.D.(y

0

i,Si,Niαy + z0_iαz+ x0iαx, ση2), where ηi| yi,Si,Ni, zi, xi ∼ N.I.D.(0, σ

2

η). This follows from setting

ςi= yi,S0 i,Niαy+ z

0

iαz+ x0iαx+ ηi.

Only time-variant variables can be estimated consistently. Next, the article of Skrondal and Rabe-Hesketh [2014] was considered.

Constrained Wooldridge solution

Using the solution of Wooldridge [2005] implies including all observations of the time-variant regressors. As a result the model becomes really large. Because of this and of the unbalanced data set the means ¯x are used with the initial values included. Define the following

ςi= yi,S0 i,Niαy+ z

0

iαz+ ¯x0iαx+ ηi.

Rabe-Hesketh and Skrondal [2013] show that the initial values of the time-varying covariates have an additional effect on the individual effect. It is therefore unreasonable to restrict the coefficient corresponding to xi,Si to be the same as the coefficients of xit, t > Si. This can lead to severe sample bias for β, unless T is large.

Alternative restricted model

The recommended model suggested by Rabe-Hesketh and Skrondal [2013] defines ςi= yi,S0 i,Niαy+ z

0

iαz+ ¯x0iαx+ x0i,SiαxS+ ηi, where ηi | yi,Si, zi, xi ∼ N.I.D.(0, σ

2

η). This model takes into account the dependence of the

distribution on covariates and can be used on an unbalanced panel. Rabe-Hesketh and Skrondal [2013] showed that the bias is negligible if xi,Si was included in the model or if the panel is long enough. The relative bias is significant at the 5% level for panels up to T = 5 and substantial for panels of length T = 3 and T = 4. Most likely including xi,Si is not necessary, since the number of periods in our data panel is larger than 5.

Appendix D

Deaton Paxson dummy variables

Since age+cohort=period, an identification problem should be solved. The method proposed by Deaton and Paxson [1994] can be used to identify the different effects; assume that all time dummy coefficients add up to zero and are orthogonal to a linear time trend,

(31)

From this follows θ1 = − T X t=2 θt ! θ2 = − T X t=3 (t − 1)θt ! and θ1= T X t=3 (t − 1)θt ! − T X t=3 θt ! θ1= T X t=3 (t − 2)θt.

Then, we can write

T X t=1 θtdt= T X t=3 (t − 2)θt ! d1− T X t=3 (t − 1)θt ! d2+ T X t=3 θtdt = T X t=3 θt(dt− (t − 1)d2+ (t − 2)d1).

Hence, these are the dummy variables introduced by Deaton and Paxson [1994].

Appendix E

Average partial effects

The average partial effects of the random effects probit model are discussed below. For the probit model set σ_η2= 0. Define xall,it= (zi, xit, yi,t−1, yi,Si,Ni, ¯xi) and β = (γz, βx, ρ, αy, αx).

Continuous variables

In the random effects probit model the marginal effect of an unit chance in the j-th regressor on the probability P (yit = 1 | xall,it, ηi, β) is equal to

P (yit= 1 | xall,it, ηi; β)

xall,it,j

= βjφ(x0all,itβ + ηi),

since the value of the individual effect ηi is unknown the above equation cannot be obtained.

We have

E(yit | xall,it) = E(E(yit | xall,it, ηi)) = P (yit = 1 | xall,it; β, ση) = Φ(xall,itβ),˜

where ˜β = √β

1+σ2 η .

Define ˆγ˜it = φ(x0_all,itβ)ˆ˜ β. The average partial effects (APE) for the random effects probitˆ˜

model are given by

AP E = E ∂E(yit| xall,it, ηi) ∂xall,it

(32)

According to Greene [2012] this can be estimated by [ AP E = 1 n 1 T n X i=1 T X t=1 ˆ ˜ γit = 1 n 1 T n X i=1 T X t=1 φ(x0_all,itβ)ˆ˜ βˆ˜ =γ¯ˆ˜it.

The delta method can be used to estimate the asymptotic variance of the average partial ef-fects. Defineθˆ˜0 = (βˆ˜0, ln(ˆσ2_η)) andγ¯ˆ˜it= _n1_T1

n P i=1 T P t=1 ˆ ˜

γit. For the probit model set θˆ˜0 =βˆ˜0.

The asymptotic variance is given by

Est.Asy.V ar(γ¯ˆ˜it) = 1 n2 1 T2Est.Asy.V ar n X i=1 T X t=1 ˆ ˜ γit ! = 1 n2 1 T2 n X i=1 T X t=1 n X j=1 T X s=1 Est.Asy.Cov(ˆγ˜it, ˆγ˜js) = 1 n2 1 T2 n X i=1 T X t=1 n X j=1 T X s=1 Git(θ) ˆˆ˜ V G0js(θ)ˆ˜ = 1 n 1 T n X i=1 T X t=1 Git(θ)ˆ˜ ! ˆ V   1 n 1 T n X j=1 T X s=1 G0_js(θ)ˆ˜  

where Git is the matrix of partial derivatives

Git(θ) =ˆ˜

∂φ(x0_itβ)ˆ˜ βˆ˜ ∂θˆ˜0 .

and ˆV is the estimator of the asymptotic variance ofθ.ˆ˜ The k-th row of Gi(θ) ifˆ˜ β has K elements is given byˆ˜

Git,k(θ) =ˆ˜ ∂φ(x0_itβ)ˆ˜ βˆ˜k ∂βˆ˜1 , . . . ,∂φ(x 0 it ˆ ˜ β)βˆ˜k ∂βˆ˜K ,∂φ(x 0 it ˆ ˜ β)βˆ˜k ∂ ln ˆσ2 η ! . Discrete variables

The marginal effects of discrete variables are given by 1 n 1 T n X i=1 T X t=1 Φ(x0_l,itβ) − Φ(xˆ˜ 0_base,itβ)ˆ˜ .

(33)

(34)

Appendix F

Simulations

F.1 Low education level

4 5 6 7 8 9 Percentage 2010 2012 2014 2016 2018 2020 Year

Low eduction level

Figure 13: Unemployment rates of the low-educated subgroup using the original defini-tion 16 18 20 22 Percentage 2010 2012 2014 2016 2018 2020 Year

Low eduction level

Percentages of flexible workers for different economic growth rates

(35)

4 5 6 7 8 Percentage 2010 2012 2014 2016 2018 2020 Year

Low eduction level

Figure 15: Unemployment rates of the low-educated subgroup using the alternative definition 8 10 12 14 16 Percentage 2010 2012 2014 2016 2018 2020 Year

Low eduction level

(36)

F.2 Inexperienced

8 10 12 14 16 Percentage 2010 2012 2014 2016 2018 2020 Year

Inexperienced labor force Original definition for flexible workers

Figure 17: Unemployment rates of the inexperienced subgroup using the original defini-tion 46 47 48 49 50 51 Percentage 2010 2012 2014 2016 2018 2020 Year

Inexperienced labor force Original definition for flexible workers

(37)

8 10 12 14 16 Percentage 2010 2012 2014 2016 2018 2020 Year

Inexperienced labor force

Figure 19: Unemployment rates of the inexperienced subgroup using the alternative definition 10 15 20 25 Percentage 2010 2012 2014 2016 2018 2020 Year

Inexperienced labor force

(38)

F.3 Low economic growth rates

5 10 15 20 Percentage 2010 2012 2014 2016 2018 2020 Year

Low economic growth Original definition Unemployment rates 4 6 8 10 12 14 Percentage 2010 2012 2014 2016 2018 2020 Year

Low economic growth Alternative definition

Unemployment rates

Labor force Low−educted Inexperienced

Figure 21: Unemployment rates for different subgroups for low economic growth

10 20 30 40 50 Percentage 2010 2012 2014 2016 2018 2020 Year

Low economic growth Original definition

5 10 15 20 25 Percentage 2010 2012 2014 2016 2018 2020 Year

Inexperienced

(39)

4 6 8 10 12 Percentage 2010 2012 2014 2016 2018 2020 Year

Low economic growth Original definition Unemployment rates 4 6 8 10 Percentage 2010 2012 2014 2016 2018 2020 Year

Unemployment rates

North East

South West

Figure 23: Unemployment rates per region for low economic growth

14 16 18 20 22 Percentage 2010 2012 2014 2016 2018 2020 Year

Low economic growth Original definition

6 8 10 12 14 Percentage 2010 2012 2014 2016 2018 2020 Year

North East

South West

(40)

F.4 High economic growth rates

5 10 15 Percentage 2010 2012 2014 2016 2018 2020 Year

High economic growth Original definition Unemployment rates 4 6 8 10 12 14 Percentage 2010 2012 2014 2016 2018 2020 Year

High economic growth Alternative definition

Unemployment rates

Labor force Low−educated Inexperienced

Figure 25: Unemployment rates for different subgroups for high economic growth

10 20 30 40 50 Percentage 2010 2012 2014 2016 2018 2020 Year

High economic growth Original definition

5 10 15 20 25 Percentage 2010 2012 2014 2016 2018 2020 Year

Labor force Low−educated Inexperienced

(41)

2 4 6 8 10 Percentage 2010 2012 2014 2016 2018 2020 Year

High economic growth Original definition Unemployment rates 3 4 5 6 7 8 Percentage 2010 2012 2014 2016 2018 2020 Year

Unemployment rates

North East

South West

Figure 27: Unemployment rates per region for high economic growth

14 16 18 20 22 24 Percentage 2010 2012 2014 2016 2018 2020 Year

High economic growth Original definition

6 8 10 12 14 Percentage 2010 2012 2014 2016 2018 2020 Year

North East

South West

(42)

Appendix G

Simulation analysis

This section gives an impression of the effects of the different factors in the simulation section. For each model and each period the factors are changed separately. One fixed seed was used in order to compare the factors. Note that for the last three simulated periods changing the num-ber of flex workers in the previous period influence the numnum-ber of observations for employment. Take for example unemployment in 2018. For the basis all time-variant explanatory vari-ables are set to the values of the previous period. Next, the values for employment in 2014 are changed into the values of 2016. As a consequence the predicted number of unemployed people increases. The same procedure is performed for the lagged variable of being a flexible worker. However, the number of observations for being a flexible worker is lower for 2016 than for 2014 and the number of obervations decreases. Therefore, we also consider this change for the same number of observations. The following steps are changing the age of people, the number of children living at home in 2016 to the value of 2018, etc.

Table 16: Unemployment 2014 (a) Original definition

basis 79 3.36 2,352

lag employment 83 3.53 2,352

lag flexible worker 87 3.70 2,352

age 106 4.51 2,352

children living at home 121 5.14 2,352

years of experience 120 5.10 2,352

lag growth (per region) 151 6.42 2,352

Source: Own calculations

(b) Alternative definition

basis 74 3.15 2,352

age 101 4.29 2,352

lag growth (per region) 138 5.87 2,352

Table 17: Flexible worker 2014 (a) Original definition

basis 351 15.95 2,201

age 377 17.13 2,201

lag growth 351 15.95 2,201

basis 170 7.68 2,201

age 189 8.54 2,201

lag growth 169 7.63 2,201

(43)

basis 129 5.80 2,223

lag employment 159 7.15 2,223 79 3.79 2,087

age 108 5.17 2,087

lag growth 106 5.08 2,087

basis 126 5.67 2,223

lag employment 143 6.43 2,223 73 3.48 2,098

age 110 5.24 2,098

lag growth 96 4.58 2,098

basis 306 15.45 1,981

age 340 17.16 1,981

lag growth 367 18.53 1,981

basis 133 6.64 2,002

age 167 8.34 2,002

lag growth 194 9.69 2,002

(44)

basis 115 5.92 1,944

lag employment 164 8.44 1,944 101 5.46 1,850

age 137 7.41 1,850

lag growth 130 7.03 1,850

basis 108 5.52 1,955

lag employment 147 7.52 1,955 98 5.24 1,871

age 136 7.27 1,871

lag growth 129 6.89 1,871

basis 301 17.50 1,720

age 343 19.94 1,720

lag growth 356 20.70 1,720

basis 168 9.64 1,742

age 193 11.08 1,742

lag growth 199 11.42 1,742

(45)

basis 98 5.74 1,707

lag employment 150 8.79 1,707 83 5.19 1,600

age 122 7.63 1,600

lag growth 119 7.44 1,600

basis 98 5.67 1,729

lag employment 139 8.04 1,729 86 5.30 1,622

age 113 6.97 1,622

lag growth 112 6.91 1,622

basis 291 19.65 1,481

age 339 22.89 1,481

lag growth 337 22.75 1,481

basis 178 11.79 1,510

age 195 12.91 1,510

lag growth 192 12.72 1,510