Tackling disabilities in young age—Policies that work

Figure 1 shows the inflow into new spells of work impairment in our data by the month of entry. There was an increase in new work-impairment spells from 2002 to 2012, particularly in the periods 2002–2004 and 2009–2010. After 2010, the inflow into work impairment decreased. The observed pattern is mainly due to changing business cycles, with recessions in 2001–2003 and 2008–2009 causing increased inflows into nonemployment. However, changes in regulations related to the follow-up of individuals with work impairment may also have affected the inflows. First, the time-limited disability benefit was introduced in 2004. This was a temporary disability benefit targeted at young people with favorable labor market prospects, to decrease inflows into permanent disability pension. Then, a new follow-up work-impaired regime was introduced during 2009–2010. It involved a work-capability assessment, along with an expansion of the eligibility criterion for receiving temporary health-related benefits to include individuals without previous sickness or labor market history.

Figure 1

Figure 2 shows survival curves by VRP status. We see that youths spend a relatively long time in work impairment. After 36 months, 23% of the youth who do not participate in any programs is still registered as work impaired. The corresponding shares among program participants are 36% for WS, and around 60% for FU, WP, and EDU. The longer durations for VRPs partly reflects the lock-in effects of participation.

Figure 2

Figure 3 shows monthly transition rates into different VRPs. Overall, young work-impaired individuals are more likely to participate in EDU and WP and less likely to participate in FU measures and WS. We also see that EDU and WP are more frequently used early in the work-impairment spell, stabilizing around 10 months, whereas the flow into the other programs is rather stable over the 5–6-year period.

Figure 3

Figure 4 shows monthly transition rates into the outcome states work/education and social security, by VRP status. For youth who do not participate in any VRP during work impairment, the probability of having a transition to work/education decreases rapidly during the first 2 years and stabilizes at a low level thereafter. Again, we see indications of lock-in effects for the VRPs, with low probabilities and slightly increasing transition rates to work/ education during the first months of the spell. The pattern is quite different when it comes to social security, where the likelihood of experiencing a transition to social security drops the first year, but we also see a steep increase in the transition rate after around 40 months. This partly reflects the dynamic selection problem: individuals who are still work impaired after 40 months constitute a highly selected group with a high probability of entering social security and a low probability of entering employment or education.

Figure 4

Descriptive statistics of the observed characteristics of work-impaired youths are presented in Table 1. Only 46% of work-impairment spells contain VRP participation; the remaining 54% comprise the comparison group. Many young people spend considerable time in work impairment, with an average of 20 months for spells of no participation and slightly longer for VRP participation spells.

Duration is measured including time spent in VRP.

There are signs of considerable selection into different VRPs. Women participate more often in EDU, whereas they are strongly underrepresented among participants in WS. There is also a relatively low proportion of non-European youths in WS compared with participants in other programs. Participants in EDU tend to be positively selected in that they are more educated, have more recent labor market experience and also have parents with higher education and income. Participants in WP, on the other hand, seem to be negatively selected. Moreover, while nearly 44% of participants in EDU receive health-related benefits

Health-related benefits include all benefits requiring a doctor-certificated medical condition: rehabilitation benefits, vocational benefits, time-limited disability benefits (before 2010), and work assessment allowance (after 2010). The first month is chosen for practical reasons.

There are also notable differences in outcomes states, depending on VRP participation. WS and EDU seem to be more successful in terms of outcome states than WP and FU. The differences underscore the importance of modeling transitions to each VRP as well as the outcomes states separately, with controls for observable individual characteristics and previous labor market history. The large observed differences also emphasize the need to control for selection on unobserved characteristics as well.

The data do not contain information on medical diagnoses or reasons for work impairment. However, other studies provide a good indication of the types of problems that are most prevalent among youths. According to Sutterud (2017), mental health disorders are by far the most common diagnosis for recipients of temporary disability benefits; for work-impaired young people, social and psychological mental health disorders comprise between 50% and 60% of the cases of work impairment during the period of analysis (Brage and Bragstad, 2011).

Our econometric approach is a multivariate mixed proportional hazard rate model. Time is measured from the moment the individual enters work impairment (initial state) and is normalized to zero. We use spells rather than individuals as our analytical unit. Let T^p and T^o be stochastic variables denoting duration until treatment (p = {FU, WP, EDU, WS}) and duration until outcome (o = {job/education; social security}), respectively, with realizations t^p and t^o. If t ≤ t^p, there are two possible transitions: to VRP or one of the outcomes. If t > t^p, a transition to VRP has occurred, and the only transition possible is to one of the outcome states. As explained above, we only consider the effect of the first VRP of each work-impairment spell.

All durations are measured in months, and thus we use a discrete version of the underlying continuous hazard rate. Let θ_kit denote the hazard rate from the initial state into state k = p, o for spell i during month t:

Equation (1) is called the selection equation and models the transition rate from work impairment into one of the VRPs. The transition rate from work impairment to one of the outcomes is explained by Eq. (2), the outcome equation. The effect of spell duration d_it is assumed to be piecewise constant. Two-month intervals are included for the first 6 months, 4-month intervals for durations of 7–18 months, 6-month intervals for durations of 19–36 months, one 12-month interval for durations of 37–48 months, and an open-ended interval for durations over 48 months.

As for observed covariates x_i, we include the individual characteristics presented in Table 1. All characteristics are measured at the start of the work-impairment spell. In addition, to capture national trends and seasonal fluctuations, we include annual and quarterly dummies as well as the local youth unemployment rate in the municipality (c_t). All variables are included as flexibly as possible, preferably using dummies for each value.

The effect of program participation is defined by the indicator function Δ_oit, taking the value of one if the treatment has been imposed before month t. This treatment effect is further divided into two effects: an on-treatment effect and an after-treatment effect. We shall provide an interpretation of the program effects in the empirical analysis when presenting the results. For the sake of simplicity, we assume that the treatment effect of one particular program is the same for all individuals; therefore, Δ_oit enters the hazard rate model just like the other explanatory variables.

The timing-of-events results of Abbring and van den Berg (2003) ensure that the abovementioned model is nonparametrically identified. With single spell data, identification hinges strongly on the proportional hazard assumption, which may be a difficult assumption to satisfy (see, e.g., van den Berg (2001) for a discussion of the proportionality assumption in a job-search setting). However, flexible modeling with a large number of time-varying calendar variables introduces exogeneity into the hazard rates and makes the proportionality assumption less important while strengthening identification (Brinch, 2007; Gaure et al., 2007; Lombardi et al., 2019).

Both the selection and the outcome equation include a set of time-invariant individual unobserved characteristics v. The unobserved characteristics enter the model as random effects and are thus assumed to be uncorrelated with the observed covariates. This may not hold in our setting. For instance, health is not observable to us and is often considered to be correlated with parental background and/or educational attainment. However, Lombardi et al. (2019) show that the ToE model is relatively robust to correlations between observed and unobserved covariates, as long as the distribution of unobserved heterogeneity is flexibly specified, the sample size is large and there is some exogenous variation in the hazard rate.

We use the modeling framework described in Gaure et al. (2007); namely, we impose a nonparametric probability distribution for v, assuming that the distribution can be characterized by an a priori unknown number of discrete points (mass points), with their associated probabilities. Further, we assume that v between different transitions may be correlated. For instance, motivated individuals are likely to profit more from program participation and are more likely to receive job offers as well. If we ignore the correlations between the unobserved heterogeneities (e.g., between job and program participation), the estimated treatment effect will be biased.

A necessary condition to interpret the treatment effects as causal is the no-anticipation assumption. This assumption states that individuals should not have private information about the exact timing of treatment ex-ante. Such information may influence their behavior; for instance, they may slow (intensify) their job search activity because they are certain that they will participate in a VRP in the future. It may be that program participation is perceived as a threat or punishment, so more effort is put into getting a job before the program starts (Maibom Pedersen et al., 2014). If this is the case, the estimated treatment effects will be biased. We do not have access to information about notification of VRP participation in our data. However, the supply of VRPs is constrained, leading to long waiting times; in about a third of cases it took more than a year from the time the user’s ability to work was assessed until a program was initiated (Lande and Selnes, 2017). Reasons for the delay were many: a program considered to be suitable was not available, or the person was too sick or negligence on the part of the public employment services (Lande and Selnes, 2017). Furthermore, around half of registered work-impaired individuals lack activity plans, and follow-up is sporadic (Riksrevisjonen, 2018). Such findings are indicative that assignment to programs is based on availability, often on short notice, and with local variations. Furthermore, the no-anticipation assumption does not rule out the possibility that some individuals know that they have a larger probability of participating in VRPs and act on this knowledge.

The probability that spells i has a transition to state k during month t can be expressed as follows:

K_it denotes the set of feasible transitions for spell i during month t. Before entering treatment, K_it = {p, o}. During and after treatment, K_it = {o}. θ_kit is the hazard rate defined in Eqs (1) and (2) above. Let y_kit be an outcome indicator variable equal to 1 if spell i has a transition to state k in month t and 0 otherwise, and let Y_i denote the complete set of outcome indicators for spell i. The conditional likelihood contribution by spell i can then be formulated as follows:

The distribution of the unobserved heterogeneity v_i is approximated in a nonparametric way by means of a discrete distribution (Heckman and Singer, 1984). As the unobserved heterogeneity terms are unknown by the researcher, they must be integrated from the likelihood function. We follow Gaure et al. (2007) and let the number of mass points be determined endogenously in the estimation process together with the other parameters. The estimation procedure starts with one mass point (no unobserved heterogeneity), and then more points are added until the likelihood can no longer be improved.

This is considered to be the case when the log-likelihood increases by <0.01.

In this section, we show the main results from the estimated multivariate mixed proportional hazard rate model with unobserved heterogeneity outlined above. The preferred model has eight mass points in the heterogeneity distribution. The number of mass points is selected using the AIC (Lombardi et al., 2019). We start by showing the effect of participating in the different VRPs on transitions to job/education and social security.

As mentioned above, the model consists of six transitions that are estimated simultaneously: transitions from reduced working capacity to one of the four labor market programs and transitions to one of the two outcome states: job/education or social security. The first is definitely a measure of success of VRPs. A transition to social security may indicate that the program did not have the intended effect—but not necessarily. If participation in a VRP helps to realize that the person is incapable of taking up work at all, then a transition to social security may be interpreted as a positive outcome.

Table 2 presents the main results.

Complete estimation results are presented in Table A1 and Table A2, both in the Appendix.

In order to facilitate the interpretation of the estimates, we have calculated the probability that a reference person who has been registered with reduced working capacity for a period of 6–10 months exits to one of the outcomes states. The reference person is a male native Norwegian, 22–25 years old, who lives in the east of Norway and has only completed compulsory school. His parents have low education and average income. The dashed line in Figure 5 (6) shows the probability of having a transition to social security (job/education), provided that such a transition has not yet occurred, for a reference person who has not participated in any program. The bars show how participation in a VRP changes this probability, during participation in the VRP (left) and after completion of VRP (right), ceteris paribus. The gray bars show significant effects, while the white bars show effects that are not significantly different from zero.

Figure 5

Figure 5 (left) shows that both EDU and WS are associated with lock-in effects; that is, during program participation the likelihood of having a transition to social security is about 50% lower for a reference participant in training and 45% lower for reference participants in WS relative to not participating in any program. However, the likelihood of experiencing a transition to social security is very small: slightly over 0.3% for a reference person and nearly 0.2 for a participant in WS or EDU.

It is important to note that this is a conditional probability, conditional on not yet having experienced a transition.

Figure 6 illustrates the effect of VRP participation on the transition to job/education. Somewhat surprising, as the figure on the left shows, we do not find any lock-in effects related to program participation. On the contrary, all programs show positive on-program effects on the transition to job/education of roughly between 0.5 and 4% points relative to the nonparticipation alternative. This could reflect that caseworkers more actively use VRPs as a springboard to working life than is the case with ordinary ALMP.

Figure 6

The figure on the right shows the positive effects of all programs after participation. WS is associated with the largest positive effects, in line with most studies of both ordinary ALMPs and VRPs, also in an international context. The impact is an increase of about 4% points, from 2% to 6%. Meanwhile, EDU shows an increase of nearly 2% points compared with the nonparticipation alternative. The likelihood of getting an ordinary job or starting an education is nearly three times as high after participation in WS relative to not participating in any program. The objective of WS is that participants continue to work for the firm that receives the subsidy after the subsidy is removed, which can partly explain the positive effect. However, for those participants who return to work impairment after the subsidy period, there is still a significantly increased likelihood of a transition to work or education. This indicates that the impact of wage subsidies is not exclusively a deadweight effect (i.e., it is not the case that employers only hire people they would have employed anyway).

The literature often argues that WS has greater deadweight and displacement effects than the other types of measures. Caliendo et al. (2017) point out that the selection to WS might be more complex than for other programs because it involves more active participation on the part of the employed during the hiring. Attempts to fully control for positive selection might not be altogether successful.

As mentioned earlier, the ToE model makes some assumptions that are difficult to test. In order to investigate the sensitivity of our results, we run three robustness tests. The first two concern unobserved heterogeneity, while the last test concerns the sample composition.

The ToE model assumes that unobserved heterogeneity is time invariant. However, during the spell of work impairment, health may change. If changes in health status influence the probability of treatment as well as the probability of having a transition to one of the outcomes, our effect estimates may be biased. For instance, youths with deteriorating health statuses may be less likely to participate in VRPs because they need to get better in order to benefit from participation. At the same time, deteriorating health may be associated with an increased probability of having a transition to social security and a decreased probability of having a transition to job/education. Although our data do not contain any direct information about individual health status, we observe whether the youths receive health-related benefits during the spell of work impairment. A medical-certified reduced work capacity of at least 50% is required in order to be eligible for a health-related benefit. Receipt of a health-related benefit may thus serve as a signal of the gravity of the health condition. We include a time-varying indicator equal to 1 with the receipt of health-related benefits and 0 otherwise as a proxy to changes in health status.

Holm et al. (2017) conduct a similar robustness test in their evaluation of active labor market programs for sick-listed workers in Denmark.

The second robustness test investigates the sensitivity of our results to the choice of information criterion used to select the number of mass points in the distribution of unobserved heterogeneity. Lombardi et al. (2019) show that selecting too few or too many mass points for the distribution of unobserved heterogeneity may seriously bias the treatment effects. They compare the performance of the Maximum Likelihood (ML) criterion (i.e., choose the number of mass points where there is no further improvement in the log likelihood) to information criteria penalizing parameter abundance: the AIC, the Bayesian Information Criterion (BIC), and the Hannan–Quinn Information Criterion (HQIC) in a ToE framework. They conclude that all information criteria perform better than the ML criterion, but no single criterion performs better in all settings. They thus recommend using all three criteria and report the results from the different criteria as a robustness check. However, they also show that the risk of overcorrection is larger in small samples than in large samples, which implies using a less restrictive criterion, such as the AIC, in our case.

The last robustness test concerns changes to the composition of our sample related to the introduction of the new work-impairment regime in 2009–2010. Importantly, the regime introduced obligatory work capability assessments and expanded the target group to include individuals without previous labor market or sickness histories. This change is particularly relevant for our target group, as young people often lack labor market experience. In addition, the regime emphasized intensified follow-up and early activation. As shown in Figure 7, the reform led to a steep rise in VRP participation. We investigate whether the implementation of the WAA reform affects our results by estimating the model solely on work-impairment spells starting before 2009.

Ideally, we would also like to estimate the model on spells starting after 2009. However, the time span is too short for many to experience a permanent transition. Hence, a large portion of the observations is censored.

Figure 7

Results from the robustness tests are presented in Table 3. The first column shows results from the original model of Table 2, where health is assumed to be time invariant, the number of mass points is chosen according to the AIC and the whole sample is used. The next column then introduces time-varying health, as explained above. The third and fourth column presents results using the two other information criteria, whereas the last column shows results for spells starting before the work-impairment reform in 2009. As shown in Table 3, the results seem to be largely robust to the inclusion of time-varying health. However, some of the positive impacts become more prevalent, and EDU and WS now significantly reduce the likelihood of a transition to social security. We also see that the results are robust to the information criterion used; both the BIC and the HQIC give very similar results as the original model using the AIC.

When estimating the model using only spells starting before 2009, two interesting differences appear regarding FU and WP. The results indicate that participants in FU and WP are more likely to have a transition to social security while on the program when we disregard the period after 2009. This may be associated with the use of VRPs as a screening device. Rules are such that activation (work-oriented measures) is compulsory before permanent disability benefits can be granted. As the eligibility criteria for receiving health-related benefits became less strict after the WAA reform, it is likely that the sample was comprised of relatively more youths with serious health conditions before the implementation of WAA than after it. FU and WP are more likely than EDU and WS to have been more frequently offered to youth with unclear work capability prospects in the period before 2010. The other noticeable change is that the positive job/education effect of having participated in WP and FU disappears. This suggests that the positive effects of the period after 2009 are driving the positive results in the original model. Recall that inflow into VRPs increased dramatically with WAA, with VRP participants likely to have better employment prospects since the rules for entitlement became less strict. The observed pattern seems to reflect this change in the composition of the target group.