Active Labor Market Policies in a Context of High Informality: The Effect of PAE in Bolivia

Abstract

Information asymmetries and limited skills are two main factors affecting job seekers’ chances to access quality jobs in developing countries. This paper evaluates the effectiveness of a program combining job intermediation and wage subsidy in Bolivia, a country with one of the highest levels of informality in Latin-America. Using administrative and survey data, we find that the program substantially increases employment and formality. These effects are heterogeneous across different subsamples of interest. Our results suggest that Active Labor Market Policies might be an effective solution for improving access to quality jobs in the context of high informality.

Keywords:

• Job intermediation
• wage subsidy
• informality
• Bolivia

JEL classification:

• J24
• J28

1. Introduction

Labor markets in Latin American and Caribbean (LAC) and particularly in Bolivia are largely informal (62 and 81 percent of workers, respectively, do not contribute to social security) (Alaimo, Bosch, Kaplan, Pagés, & Ripani, 2015). From a policy perspective, it is important to reduce informality because it affects individuals’ welfare and productivity but also the overall economy, by imposing pressure on fiscal balance, social security, poverty, and inequality.

High labor costs, workers’ limited skills, and information asymmetries are some of the main factors affecting access to formal jobs. Individuals from vulnerable groups (e.g. youths, women, and those with lower levels of education) are more likely to be affected by these restrictions and to work informally (Attanasio, Kluger & Meghir, 2011).

Active Labor Market Policies (ALMP) represent a potentially effective policy to redress these barriers and increase individuals’ chances of accessing good quality jobs (Kluve, 2016; Pignatti, 2016). Evaluations of ALMP in LAC, and particularly of training programs, which is the most commonly implemented policy of this kind in the region, show positive results on employment and formal employment for youth and women (Card, Kluve, & Weber, 2018; Escudero, Kluve, López-Mourelo, & Pignatti, 2018; McKenzie, 2017). However, evidence about the effectiveness of other ALMP (e.g. wage subsidies, search and matching assistance programs) is still scarce in LAC (Escudero et al., 2018).

This paper evaluates the effectiveness of the Programa de Apoyo al Empleo (Program for the Support of Employment, PAE) on workers’ employment, formality (i.e. contributing to social security), and earnings in Bolivia. PAE is a public program offering job seekers registered in the Public Employment Service (PES) information about job vacancies posted by formal firms, which are also registered in PES, and three months of wage subsidy if they are selected for the vacancy. Thus, PAE provides job seekers information about the labor market, a wage subsidy to reduce firms’ hiring costs, and a job experience in a formal firm, which might also help them to signal productivity and to acquire skills for future job searches.

The fact that access to PAE is universal and that firms discretionary select candidates from the list of jobseekers sent by PES make a random assignment into the program hard to implement. To identify the effect of PAE, we combine propensity score matching and difference-in-difference techniques accounting for differences in observables and time-invariant characteristics that might affect selection into the program and labor outcomes. Additionally, we test whether our identification strategy allows us to interpret our estimates as the causal effect of PAE.

We use three sources of information: administrative records from PES and PAE and an individual survey. Data from PES provides information about jobseekers’ socioeconomic characteristics; and, characteristics of the job offers, such as the number of offers, occupational category, and offered salary. Records from PAE allow us to identify the program’s beneficiaries. These two administrative datasets are merged with an individual telephone survey applied to jobseekers registered in PES in the period between January 2016 and June 2017. The survey provides information regarding the employment characteristics of jobseekers at the time of their registration in the program and at the time of the survey (between December 2017 and February 2018).

Our results show that PAE increases the probabilities of employment in 9 percentage points (pp) and of formal employment in 4 pp. We also find a positive effect on earnings in 10 percent, although this is significant at the 15% level. Moreover, we find evidence of heterogeneous effects. The effects of PAE on the probabilities of employment and formal employment are larger for adults and for those with tertiary education.¹ Regarding gender differences, we find that relative to men, PAE has larger effect on women’s probability of having formal jobs and earnings. Moreover, we find evidence that the effects of PAE on employment and formality seem to be larger in the short term. Results from a set of robustness checks confirm these results. Finally, a cost-benefit analysis reinforces the positive returns of PAE over future individual labor outcomes.

This paper contributes to the scarce empirical literature about the effectiveness of ALMP different than job training in LAC.² In their meta-analysis, Card et al. (2018) find that, in LAC, no program estimates correspond to intermediation services, only 3 percent to employment subsidies, and 97 percent to training programs.³ Similarly, Escudero et al. (2018) find only one published evaluation of an employment subsidy program (Galasso, Ravallion, & Salvia, 2004) and one of an intermediation service (Dammert, Galdo, & Galdo, 2015) in the region. Thus, our paper is one of the first evaluations of a program combining an intermediation service, a wage subsidy, and a job experience in a formal firm in LAC. This paper is also the first evaluation of an ALMP in Bolivia (Card et al., 2018).

In a context of high informality, offering vulnerable job seekers the chance of obtaining a (scarce) formal job is expected to have positive effects on their employability. Experience in a formal job could allow them to signal better their productivity in future searches (Pallais, 2014). They could also improve their skills (by working in a more efficient environment), network, and information about the formal labor market.

The rest of the document is organized as follows. Section ALMP in LAC and PAE reviews the evidence of similar ALMP effectiveness in LAC and describes the program. Section Data sources present the data. Section Evaluation strategy and descriptive statistics describe the identification strategy and present summary statistics. Section Results presents the main results and robustness checks. Section Cost-Benefit Analysis shows a cost-benefit analysis, and Section Discussion concludes.

2. ALMP in LAC and PAE

2.1. ALMP effectiveness in LAC

Despite the interest in ALMP in developing countries, evidence about their effectiveness has begun to become available only recently (McKenzie, 2017). In LAC, this expansion has been motivated by the increasing interest in ALMP as a public policy tool aimed at improving not only labor market inefficiencies but also poverty and inequality (Escudero et al., 2018).

In contrast to developed countries, ALMP in developing countries, and particularly in LAC, generally show positive, nevertheless small, effects on vulnerable groups (Card et al., 2018; Escudero et al., 2018). Recent evidence from LAC shows that ALMP is statistically more effective for women and youth (Escudero et al., 2018). Moreover, the evidence shows that effects from medium-run evaluations are not statistically significantly different from those in the short-run and that long-term evaluations are scarce in the region.

Training programs are the most commonly implemented and evaluated ALMP in LAC (Escudero et al., 2018). Despite having been implemented in many countries in the region in recent years (Vezza, 2014) and having taken recent relevance as an ALMP to smooth the consequences of the COVID-19 pandemic on the labor market (Gentilini, Almenfi, Orton, & Dale, 2020), evidence about the effectiveness of subsidized employment and search assistance programs is scarce in the region (Card et al., 2018; Escudero et al., 2018).⁴

While the evidence from developed countries shows that subsidized employment programs are generally less successful than other ALMP (Card, Kluve, & Weber, 2010; Escudero et al., 2018), developing countries show more positive results. Recent meta-analyses (Escudero et al., 2018; Levy, Montane, & Sartorio, 2019) find that subsidized employment programs have positive impacts on employment, formality, and earnings.⁵ Regarding the duration of the effects of wage subsidy programs on employment, there is mixed evidence in developing countries. On the one hand, some studies show that the effect holds only in the short-run (Groh, Krishnan, McKenzie, & Vishwanath, 2016; McKenzie, 2017). On the other hand, programs combining different ALMP or of longer duration seem to have more durable effects. Berniell and de la Mata (2017) evaluates the effectiveness of the Programa Primeros Pasos (PPP) in Argentina, a program offering youths a 12-months subsidy to acquire on-the-job training in a formal job. PPP increases the probability of formal employment, in the short-run (12 months after finishing the program) and the medium-run (4.5 years after the program started), and reduces the unemployment rate by 10 percent. In North Macedonia, a six-month wage subsidy program finds a positive impact on employment (up to 25 percent) that is still significant (15 percent) after three and a half years (Armand, Carneiro, Tagliati, & Xia, 2020). The effects are larger for individuals with lower attachment to the labor market (women and individuals with lower education). Finally, in Mexico, Abel, Carranza, and Ortega (2021) found that a program offering up to six months of subsidized wage (up to 24 percent of the minimum wage) to secondary school graduates increases formality by 12 percent, driven by an increase in permanent contracts jobs.

2.2. Background and description of PAE

Partially explained by a favorable commodity price context, Bolivia has experienced high levels of economic growth and poverty reduction since 2005.⁶ Nonetheless, the country still has the main challenge of improving productivity and job quality. In 2012, when PAE was designed, Bolivia had one of the lowest unemployment rates in the region (2 percent in 2012) but a high level of informality (81 percent). Youth, women, and people with lower levels of formal education were particularly disadvantaged in these indicators.⁷

In September 2012, the Bolivian Ministry of Work, Employment and Social Security, with the financial support of the Inter-American Development Bank, implemented PAE.⁸ The aim of the program is to facilitate the placement of jobseekers who, although accomplishing the job selection requirements, had low chances of accessing formal employment opportunities. For PAE, young workers, particularly those with lower levels of education, and workers without previous formal sector experience are the ones considered less likely to have access to formal jobs.⁹

In Bolivia, as in other developing countries, labor market conditions, such as high labor costs and restrictive labor regulations, generate disincentives for employers to formally hire workers. Unless employers know in advance the jobseeker’s contribution to the firm’s productivity or they have perfect control over his future returns (Pallais, 2014), these market conditions might incentivize the use of informal recruitment channels.¹⁰ Thus, jobseekers who are less able to signal productivity are more likely to be trapped in poor-quality jobs.

To solve this, PAE offers jobseekers a three-months subsidized job in formal firms. The wage subsidy varies between 1 and 1.5 minimum wages according to the educational level requirement and the economic sector of the vacancy.¹¹ During its lifetime, the program has benefited nearly 20 thousand individuals, 55 percent of whom were women, and 49 percent had at least some tertiary education.¹²

PAE operates through the PES.¹³ Both job seekers and firms offering vacancies need to be registered at PES to be eligible for PAE. In addition, job seekers are required to be older than 18 at the moment of registration in PES and to meet the requirements of the vacancy. In turn, firms are required to be formal (i.e. to have an active national tax identification number) and committed to offering vacancies that credibly lead to permanent hirings.¹⁴ Firms can only apply for a limited number of PAE beneficiaries depending on their size and can only reapply to the program if at least 50 percent of its previous beneficiaries were hired after graduating from the program.¹⁵

The process of matching jobseekers and vacancies is made by caseworkers. Whenever a PAE vacancy is posted, a PAE caseworker makes an initial screening of candidates, identifying those accomplishing the selection criteria and preferences, and compiles a shortlist. The intermediation process consists in matching the jobseeker’s job offer and the job vacancy’s occupation codes (Cooper, 2014). Shortlists typically include at least three jobseekers per vacancy and are provided directly to the employer.¹⁶ After receiving the shortlist, employers contact and interview jobseekers and select one of the candidates.¹⁷ Job interview and offer decisions are discretionary to the firm and the job acceptance decision is discretionary to the job seeker. Once the firm and the selected jobseeker reach a deal, the firm communicates it to PAE, which after some administrative checks, starts paying the subsidy.¹⁸

PAE combines three main components: job search support, a wage subsidy, and acquiring formal job experience. First, PAE offers a cost-free job intermediation service aiming to improve jobseekers’ (firms’) information about the quality and quantity of job vacancies and firms' (jobseekers) and to allow more efficient job matchings. Second, the wage subsidy component aims at encouraging firms to hire jobseekers that they would not hire otherwise (e.g. because they do not have experience or are unsuccessfully signaling their skills and productivity) by reducing the cost of hiring and testing a worker (McKenzie, 2017; Pallais, 2014; Vezza, 2014). Third, by offering a formal job experience, PAE improves jobseeker’s productivity signaling in future labor market searches. If deficiencies in signaling productivity affect jobseekers’ chances of obtaining formal jobs in Bolivia, then formal experience, even if it is of short duration, could act as a better signal of productivity (Pallais, 2014). In this regard, Berniell and de la Mata (2017) find that the impact of PPP may be due to the gaining in formal experience rather than improvements in human capital. A formal job experience might also help workers to acquire knowledge about and networks in formal firms, which would improve their confidence in approaching employers in further job searches. Abel et al. (2021) find that gains from formal experience are still relevant two years after the end of a subsidy program. Galasso et al. (2004) find this particularly relevant for young and female workers in Argentina. Finally, working in a formal setting could help workers to gain the skills needed in similar jobs in the future. The education system in Bolivia fails in providing individuals with the skill set demanded in the labor market, as documented by employer surveys in the country (Bagolle, Valencia, & Urquidi, 2018). Short job experiences, like the one supported by PAE, might contribute to the acquisition of these skills.

3. Data sources

Data for the present evaluation comes from three sources. First, administrative records from PES contain socioeconomic information and information about the job offer (i.e. interest in a vacancy) of the jobseekers registered in the PES between January 2016 and June 2017. Second, administrative data from PAE allow us to identify the program’s beneficiaries. Finally, we use information from a telephone survey, collected between December 2017 and February 2018, to the jobseekers registered in PES between January 2016 and June 2017.¹⁹

Out of 25 220 jobseekers registered in PES in this period, 74 percent were reached by interviewers (18 713), and 46 percent (11 541) completed the survey. To do this, a team of interviewers received a list of individuals registered in PES (PAE beneficiaries and non-beneficiaries) divided by years. Interviewers were asked to first contact all individuals registered in 2017 and then, those registered in 2016.²⁰ Table SM1 in the Supplementary Materials section (SM) shows that the proportion of individuals registered who were contacted and completed the survey in 2017 and 2016 are similar (74 percent were contacted for each year and 48 and 45 percent completed, respectively).²¹

Collecting information through telephone surveys presents advantages over face-to-face (e.g. lower costs and speed) and online surveys (e.g. accessibility) (Szolnoki & Hoffmann, 2013). However, it also introduces some challenges. By conducting a telephone survey, one might introduce a sample selection bias by excluding job seekers who do not have a telephone. Also, collecting data by phone, in contrast to other alternatives, might affect the quality of the information reported or affect the decision to participate in the survey (i.e. mode effect). The overall response rate might also be lower, or it could change across population groups, making some of them overrepresented (Holbrook, Green, & Krosnick, 2003; Nandi & Platt, 2011; Szolnoki & Hoffmann, 2013).

We explore the presence of such problems in our data. First, the chances of introducing sample selection bias for not having a telephone are neglectable in our sample. Despite that 2020, Household Survey in Bolivia indicates that 13 percent of households in urban areas do not have access to mobile phones, and only 0.6 percent of jobseekers registered in our PES dataset do not show a valid phone number. Second, we test whether the quality of the data reported in the telephone survey differs from the administrative records from the PES. As mentioned above, given that the PES registry includes only a few variables, we restrict the analysis to only three variables of interest: age, sex, and education. Correlation coefficients between the information reported in both datasets for these three variables are high and statistically significant (Table SM2 in the SM). Finally, using the information available in the PES database, we test for systematic differences between those who answer or do not the telephone survey (Table SM3 in the SM). Although we find that the probability of completing the telephone survey is associated with individual socioeconomic characteristics, we find that there is not selective attrition bias (i.e. the conditional probability of completing the survey is not affected by being a PAE beneficiary or not).²²

4. Evaluation strategy and descriptive statistics

Considering that assignment into PAE is not random, we combine kernel propensity score matching (PSM) and difference-in-difference (DID) methods to estimate the effect of PAE on labor outcomes. PSM allows us to create a comparison group that is similar in terms of observable characteristics to PAE beneficiaries. In addition, DID allows us to control for time-invariant unobservable characteristics that might affect both participation in PAE and labor outcomes. The key identifying assumption is that, in the absence of the treatment, the labor outcomes of individuals in the control and treatment groups would have followed a parallel trend.

We start estimating the propensity scores by running a probit model of the treatment variable T on a vector of covariates X corresponding to a period before the treatment (i.e. at the time of registration in PES). We include a rich set of individual and household characteristics, including a dummy variable for whether the beneficiary had previous working experience; a dummy for whether the beneficiary was working at the moment of registration at PES; a dummy variable for whether the jobseeker was ever promoted in a previous job; a dummy if the jobseeker defines himself as indigenous; sex; age; age squared; years of education; a dummy for having a disability; civil status; a dummy for being the head of household; the number of children; a dummy if the jobseeker holds a tertiary education diploma; household income; dummies for the year, month and city of registration at PES; the id of the caseworker matching jobseekers and vacancies; and the logarithm of the expected salary (i.e. the wage the jobseeker expected to obtain when manifested interest in a job vacancy at registration in the PES). (1) $$P_i^{\mathrm{*}}=\gamma +\delta X_i+{ϵ}_i$$(1) where P is a latent variable that determines the value of T under the following scheme: $$T_i=\{\begin{array}{c}0\mathit{\text{ if }}{P}_i^{\mathrm{*}}\le \overline{p}\\ 1\mathit{\text{ if }}{P}_i^{\mathrm{*}}>\overline{p}\end{array}$$

After estimating the propensity score, we restrict the sample to the common support.²³ Then, we implement a kernel PSM, where each individual in the treatment group is matched with a weighted average of individuals in the control group. For calculating these weights, we use the Epanechnikov kernel function, where weights are proportional to the proximity of the propensity scores of the treatment and control individuals in a determined neighborhood.

Finally, we estimate a DID regression on the weighted outcomes generated previously: (2) $${\widehat{y}}_{it}={\beta }_0+{\beta }_1{\mathit{time}}_t+{\beta }_2{T}_i+{\beta }_3{\left(\mathit{time}·T\right)}_{it}+{ϵ}_{it}$$(2) where ${\widehat{y}}_{it}$ is a labor market outcome (i.e. having a job, having a formal job, or the logarithm of earnings); time is a dummy variable indicating the time of registration at the PES (pre-treatment) and the time of interview (post-treatment); and $T_i$ is a dummy variable indicating the treatment status.

Considering that our sample is composed of jobseekers registered in the PES (i.e. receive information about vacancies), the coefficient of interest, ${\beta }_3,$ should be interpreted as the marginal effect of the other two PAE components combined (i.e. receiving a wage subsidy and the chance of having a formal job experience).

Because of data restrictions (i.e. we only have pre-treatment information at the time of registration and post-treatment information at the time of the telephone interview) we are not able to test whether the parallel trend assumption of the DID model holds. Although we are not able to test for this because pre-application data is not available, results from a robustness test (Oster, 2019) validate that selection bias on unobservables does not threaten our identification of causal effects.

Table A1, in the Appendix, presents the mean values of the covariates included in the PSM estimation at the baseline (i.e. at registration in PES), for individuals in the control and treatment groups.²⁴ The first three columns (unweighted variables) show that jobseekers in our sample were 31 years old, mostly women (58 and 57 percent in the control and treatment groups, respectively) and not indigenous (11 and 7 percent in the control and treatment groups, respectively). They also have 14 years of education, no disabilities (6 and 2 percent in the control and treatment groups reported having a disability, respectively), and have reduced chances of having a tertiary education diploma (32 and 30 percent). Finally, individuals in the treatment group have a higher household monthly income (Bs$2,276 or US$327 and Bs$2,089 or US$300, respectively)²⁵ and higher expected wages than those in the control group. Implementing the PSM allows us to have a balanced sample in the pre-treatment covariates, as shown in the last three columns of Table A1.

5. Results

Table 1 presents the estimated effect of PAE on the probabilities of being employed and having a formal job, and on the logarithm of monthly earnings.²⁶ PAE increases the probability of being employed by 8 pp and the probability of having a formal job by 4 pp. Moreover, PAE increases earnings, conditional on working, by 9 percent, although this is only statistically significant at the 15 percent level.²⁷ While the probabilities of employment and having a formal job and earnings of the treatment and control groups were statistically similar before the intervention, beneficiaries’ labor outcomes were improved after treatment.

Table 1. PAE effects on employment, formal employment and (log) earnings

	Employment				Formal employment				(Log) Monthly earnings conditional on working before and after PAE
	Coef.	Std. Error	t		Coef.	Std. Error	t		Coef.	Std. Error	t
Baseline
Control	0.303				0.024				7.265
Treated	0.306				0.018				7.327
Diff (T-C)	0.004	0.011	0.330		−0.006	0.003	−1.760	*	0.063	0.036	1.740	*
Follow up
Control	0.461				0.082				7.422
Treated	0.546				0.116				7.573
Diff (T-C)	0.085	0.012	7.150	***	0.034	0.007	4.670	***	0.152	0.044	3.450	***
Impact
Diff-in-diff	0.082	0.016	5.030	***	0.040	0.008	4.980	***	0.089	0.057	1.560
Observations	22 322				22 322				5 186
R^2	0.04				0.03				0.02

Note: Robust Standard errors are reported. *Significant at 10%; **significant at 5%; *** significant at 1%.

5.1. Heterogeneous effects

We explore the heterogeneous effects of PAE by gender, age, and level of education (Table 2). First, while the effect of PAE for males and females is positive, the effects on formal employment and earnings for women are larger than for men. Similarly, PAE has positive effects on employment and formal employment of adults (older than 28) and youths (between 18 and 28), but the effects are larger for adults in the three outcomes analyzed. Finally, we find that the effect of PAE on employment and formality is larger for those with a higher level of education (i.e. having some tertiary education) than for those with a lower level of education (i.e. those having completed high school at most). These results are aligned with similar evaluations in LAC countries (i.e. positive impact on employment outcomes and higher impact for women) and also add new evidence in terms of heterogeneous effects by education level (Escudero et al., 2018).

5.2. Effects over time

To explore whether PAE has differentiated effects over time, Table 3 shows the effects of PAE for the cohorts registered in 2016 and 2017. PAE has positive effects on employment and formality for the two cohorts of applicants. However, aligned with McKenzie (2017), impacts reduce over time. The effect on employment reduces to 43 percent after the first year and the one on formality even more, to 29 percent.²⁸ The fact that PAE effects reduce over time suggests that the program requires additional interventions (e.g. training) to have a sustained effect on workers’ employability.

Table 2. PAE heterogeneous effects

	Employment					Formal employment					(Log) Monthly earnings conditional on working before and after PAE
	Diff-in-diff	Std. Error	t		N	Diff-in-diff	Std. Error	t		N	Diff-in-diff	Std. Error	t			N
Men	0.080	0.025	3.180	***	9370	0.033	0.014	2.470	**	9370	0.029	0.078	0.370		2442
Women	0.087	0.022	4.050	***	12902	0.047	0.010	4.650	***	12902	0.146	0.083	1.760		2661
Difference (Men-Women)	−0.008	0.011	−0.689			−0.013	0.007	−1.775	*		−0.117	0.027	−4.345	***
Adults (>28 years)	0.102	0.024	4.320	***	11016	0.049	0.012	4.140	***	11016	0.116	0.081	1.430		2995
Youths (18–28 years)	0.063	0.023	2.710	***	11150	0.028	0.011	2.460	**	11150	0.026	0.084	0.310		2116
Difference (Adults-Youth)	0.038	0.004	10.833	***		0.021	0.002	8.525	***		0.091	0.024	3.781	***
High education (HE)	0.103	0.019	5.490	***	16490	0.045	0.010	4.450	***	16490	0.088	0.072	1.210		3606
Low education (LE)	0.029	0.033	0.860		5778	0.022	0.013	1.650	*	5778	0.035	0.101	0.350		1345
Difference (HE-LE)	0.075	0.021	3.554	***		0.023	0.007	3.510	***		0.052	0.063	0.836

Note: Robust Standard errors are reported. *Significant at 10%; **significant at 5%; *** significant at 1%.

5.3. Robustness check

The fact that our evaluation sample does come from a random selection of jobseekers registered in PES in the period of interest might raise concerns about the external validity of our estimated effects. Another potential concern is related to the absence of data for testing the parallel trend assumption and the capacity of our identification strategy to provide causal estimates of the effects of PAE. In this section, we perform three robustness checks and analyze whether our results hold. First, we recalculate our estimates using an ex-post randomized sample. Second, we estimate the impacts of PAE using inverse probability weighting (IPW).²⁹ Finally, we test to what extent our identification strategy is capable of reducing the potential selection bias among PAE beneficiaries (Oster, 2019).

For the ex-post randomization, we generate an ex-post probabilistic sampling from the total population of beneficiaries and controls registered in PES in 2016–2017, in which a 50 percent chance of ending in the sample is assigned to everyone. Given that in our case answering the telephone survey is independent of the chosen sampling selection procedure, we assume that individuals who did not reply to the survey would have done it neither if a random sample was originally implemented. Therefore, individuals who ended up being selected for the ex-post random sample and for whom we do not have information from the telephone survey are treated as regular refusal. Table 4 shows the results of estimating the previous effects for individuals randomly assigned to the ex-post sample who answered the telephone survey.

Table 3. Impact of PAE disaggregated by year

	Employment					Formal employment					(Log) Monthly earnings conditional on working before and after PAE
Year	Diff-in-diff	Std. Error	t		N	Diff-in-diff	Std. Error	t		N	Diff-in-diff	Std. Error	t		N
2016	0.063	0.018	3.490	***	15654	0.027	0.009	3.100	***	15654	0.044	0.068	0.650		3363
2017	0.147	0.041	3.560	***	6042	0.092	0.024	3.810	***	6042	0.163	0.127	1.280		1761

Note: Robust Standard errors are reported. *Significant at 10%; **significant at 5%; *** significant at 1%.

Results are consistent with the effects presented above. The effects on employment and on formal employment are 9 pp and 5 pp, respectively, and similar to the ones for the original sample (Table 1). We also find similar heterogeneous effects, qualitatively favoring adults, those with a higher level of education, and women.³⁰

The second robustness check deals with the potential sample selection due to survey non-response. For this, we use IPW to estimate the effects of PAE. First, we estimate a logistic model for the probability of completing the survey, conditional on being reached by interviewers and using information from the PES administrative records. Then, we proceed to calculate the probability of completing the survey and calculate the inverse of the probability to estimate the impact of PAE. Table 5 shows that the IPW results are similar in magnitude and significance to the ones presented in Tables 2, 3, and 4.

Table 4. Impact of PAE on labor variables for the ex-post randomization sample

	Employment					Formal employment					(Log) Monthly earnings conditional on working before and after PAE
	Coef.	Std. Error	t		N	Coef.	Std. Error	t		N	Coef.	Std. Error	t		N
Full sample	0.095	0.023	4.180	***	11152	0.051	0.011	4.410	***	11152	0.100	0.081	1.220		2358
2016	0.081	0.025	3.220	***	7896	0.033	0.012	2.680	***	7896	0.061	0.096	0.630		1621
2017	0.124	0.059	2.110	**	2914	0.121	0.033	3.610	***	2914	0.258	0.197	1.310		709
Men	0.102	0.035	2.930	***	4586	0.039	0.019	1.990	**	4586	0.007	0.111	0.070		1203
Women	0.087	0.030	2.890	***	6380	0.056	0.013	4.220	***	6380	0.206	0.121	1.710	*	1145
Adults (>28 years)	0.102	0.033	3.120	***	5470	0.054	0.016	3.320	***	5470	0.152	0.117	1.300		1456
Youths (18–28 years)	0.088	0.033	2.690	***	5540	0.048	0.016	2.910	***	5540	0.022	0.130	0.170		1033
High education	0.116	0.026	4.390	***	8178	0.055	0.014	3.780	***	8178	0.108	0.105	1.030		1632
Low education	0.033	0.046	0.710		2796	0.033	0.018	1.830	*	2796	0.129	0.135	0.960		640

Note: Robust Standard errors are reported. *Significant at 10%; **significant at 5%; ***significant at 1%.

Table 5. Impact of PAE on labor variables correcting sample selection with Inverse Probability Weighted Estimators

	Employment					Formal employment					(Log) Monthly earnings conditional on working before and after PAE
	Coef.	Std. Error	t		N	Coef.	Std. Error	t		N	Coef.	Std. Error	t		N
Full Sample	0.084	0.016	5.16	***	22336	0.04	0.008	4.89	***	22336	0.092	0.057	1.62		5256
2016	0.065	0.018	3.59	***	15694	0.026	0.009	3.05	***	15694	0.059	0.069	0.85		3332
2017	0.151	0.041	3.64	***	6202	0.092	0.024	3.81	***	6202	0.173	0.125	1.38		1798
Men	0.09	0.025	3.58	***	9404	0.033	0.014	2.44	**	9404	0.032	0.076	0.42		2418
Women	0.084	0.022	3.89	***	12906	0.046	0.01	4.62	***	12906	0.158	0.083	1.92	*	2673
Adults (>28 years)	0.107	0.024	4.53	***	11014	0.049	0.012	4.18	***	11014	0.126	0.08	1.57		3037
Youths (18–28 years)	0.064	0.023	2.73	***	11200	0.029	0.012	2.56	**	11200	0.025	0.083	0.3		2098
High education	0.105	0.019	5.57	***	16482	0.045	0.01	4.47	***	16482	0.097	0.075	1.29		3608
Low education	0.035	0.034	1.02		5798	0.021	0.013	1.59		5798	0.044	0.099	0.45		1393

Note: Robust Standard errors are reported. *Significant at 10%; **significant at 5%; *** significant at 1%. A logistic regression is estimating using available information at the PES such as age; gender; a dummy if the jobseeker complete tertiary education; a dummy for being single; the logarithm of the expected salary; dummies for main cities (La Paz, El Alto, Cochabamba, and Santa Cruz); dummies for the year of registration; the number of offers registered in the SPE; and a dummy that takes the value of 1 if the job seeker ended up being a beneficiary of PAE.

Finally, since we are not able to test for the parallel trends assumption due to a lack of data, there might be concerns about the presence of selection bias. To test to what extent our identification reduces this, we proceed with a robustness check based on the validation test developed by Oster (2019). The objective of the method is to estimate an approximation of the real treatment effect (i.e. without bias). The author proposes a method to validate whether adding control variables to linear regression for treatment estimation reduces the selection bias on unobservables. To do that, the method considers the existence of a proportional relation between the selection bias due to unobservables and observables characteristics, which allows recovering the selection bias related to unobservables characteristics from an available set of controls.

The method firstly estimates the treatment effect with controls (controlled estimation) and without them (un-controlled estimation). Then, based on a specific proportional relation of bias, the controlled estimation is adjusted by the change of the treatment effect laying between the un-controlled and controlled estimations. Finally, this adjustment is weighted by the difference between the controlled R-square and the assumed R-square (i.e. if unobservable variables were accessible). Intuitively, this aims at identifying the change that the controlled estimation would suffer if we would have access to unobservable variables. If the adjusted estimation is extremely different from the controlled estimation, selection bias plays an important role in the controlled estimation.

From the adjusted estimation, it is possible to implement two robustness tests. The first consists on constructing a confidence interval based on the adjusted estimation and the controlled estimation. To get the adjusted estimation we need to set the proportional relation of bias and the maximum R-square. Based on the confidence interval, we could test the null hypothesis of the unbiased estimator being equal to cero. The second test consists of estimating the proportional relation of bias needed to get an adjusted estimator equal to cero, giving a maximum R-square. If the estimated proportional relation of bias is greater than one, we reject the unbiased treatment estimation equal to cero. This is because in all cases it is assumed that observable and unobservable characteristics are, at most, equally important in determining the bias.

To implement these robustness tests in our setting, we define a set of controls expected to reduce the bias and that the test will validate. This will be the controlled estimation in the adjustment procedure. Also, following the empirical suggestions by Oster (2019), the proportional relation of bias is set to one (i.e. the equal importance of observable and unobservable) and the maximum R-square is set to 1.3 times the controlled R-square. The robustness tests shown in the last two columns of Table 6 confirm the results of our matching with the difference in difference estimation. First, the confidence interval for estimated coefficients does not include zero for the employment and formal employment outcome, so we can reject the null hypothesis of unbiased treatment effect equal to cero. Second, the proportional relation of the bias test estimates a value greater than one for the employment and formal employment outcomes. This means that to get an unbiased treatment effect equal to cero, the importance of unobservable variables has to be greater than the observables in the selection bias. This is inconsistent with the assumption of equal importance, revealing that the unbiased treatment effect is different from cero. Both tests validate the results of our empirical approach. The results of the tests for earnings are also coherent with the fact the estimated effect of PAE on the variable is only statistically significant at 15%. Based on this, it is possible to argue that our empirical procedure successfully reduced the selection bias and that the estimated effects correspond to the causal effect of PAE on labor market outcomes.

Table 6. Robustness test estimation

Outcome	Diff in Diff estimation	Un-Controlled estimation	Controlled estimation	Bias adjusted estimation	Confidence interval test	Proportional relation of bias
Employment	0.082 [0.044]	0.189 [0.028]	0.082 [0.311]	0.012	[0.012–0.082]	1.133
Formal employment	0.041 [0.029]	0.074 [0.018]	0.041 [0.051]	0.010	[0.010–0.041]	1.220
(Log) Monthly earnings conditional on working before and after PAE	0.089 [0.020]	0.250 [0.016]	0.085 [0.138]	−0.003	[−0.003–0.086]	0.966

Notes: The table shows the estimated coefficients and R-square in brackets. Controlled estimation is the difference in difference estimation with additional controls.

6. Cost-Benefit analysis

We present a simple and conservative calculation of a cost-benefit analysis for PAE. For this, we consider the effects of PAE on earnings and the probability of employment and use a discount rate of 5 percent per year to estimate the Net Present Value. Regarding earnings, as reported in Table 1, the estimated effect of PAE on earnings is 8.9 percent. Using earnings at baseline reported for individuals in the control group (Bs$1,672 or US$240), we estimate a benefit attributable to PAE of US$21,4 (Bs$148) per month, which implies an annual benefit of US$257 (Bs$1,786).

Following Attanasio, Guarín, Medina, and Meghir (2015), we estimate the impact on earnings for 5 years. The calculation considers two scenarios: one, where gains are permanent; and the other, where gains depreciate at an annual rate of 10 percent. We also assume that the program does not affect the growth rate of earnings.

Regarding the impact on employment, we use the average unemployment duration at the time of registration in PES reported by jobseekers. On average, jobseekers were unemployed for 13.5 months. To estimate the benefits of the PAE on employment, we use the earnings obtained by the treatment group after PAE and only for 13.5 months, which implies a monthly benefit of US$17.9 (Bs$125) and a total benefit for the 13.5 months period of US$249 (Bs$1,734).

Costs associated with the program operation are calculated based on the 2017 minimum salary and the average stipend payment that can be granted, which is 1.25 minimum salaries for the 3 months. Thus, the cost of subsidizing employment reaches US$1,077 per beneficiary. The opportunity cost incurred by beneficiaries for participating in the program and not receiving another income is not considered.

In the first scenario, the net lifecycle gain of PAE is US338, while in the most conservative scenario the gain is US$247. The internal rate of return (IRR) is 12 and 9 percent, respectively, a figure slightly lower than the impact that Attanasio, Kugler, and Meghir (2011) found for Jóvenes en Acción, an ALMP for vulnerable youth in Colombia (they found an IRR between 35 and 21 percent). Regarding that the impact of PAE is only evident (through this evaluation) for two years and considering that the impact is reduced from the second year, the program could pay between 77 and 75 percent of the cost of the program. In a sensitivity analysis, we considered the observed reduction in the impact on employment and earnings, and fix the lowest impact for years 3, 4 and 5. In this case, we find that IRR is negative at 9 and 10 percent. In this scenario, ten years of impact is required to have a positive cost-benefit analysis.

The cost-benefit analysis shows the effectiveness of PAE even under conservative assumptions and without considering the benefits associated to access to formal employment (e.g. retirement savings, health insurance, vacations). Also, the analysis does not consider that beneficiaries can improve their earnings and job prospects due to the program and the screening provided by it. The analysis also assumes that individuals would permanently earn labor income for 5 years, which could not be the case if they stop participating in the labor market for personal (e.g. maternity) or professional reasons.

A relevant fact that we do not consider in the present cost-benefit analysis is that the benefits among PAE beneficiaries could be explained at the expense of other job seekers in the market, which would diminish the impact through a general equilibrium effect. However, evidence in the literature in this regard is not conclusive yet. For instance, Crepon, Duflo, Gurgand, Rathelot, and Zamora (2013) found that a labor intermediation program in France obtains benefits at the expense of a decrease in the employment rate of non-beneficiaries. In contrast, Berniell and de la Mata (2017) found no evidence of displacement effect for PPP in Argentina.

7. Discussion

This paper offers evidence about the effectiveness of PAE, an ALMP combining labor intermediation, a wage subsidy, and the chance of having a formal job experience in a high-informality context. This evidence is particularly relevant given that such ALMP, in contrast to training programs, have received almost no attention in the empirical literature in LAC. From a policy perspective, we contribute to the literature showing evidence about a cost-effective policy that contributes to improving individuals’ labor market outcomes.

Our results show that PAE substantially improves the probabilities of employment and formality. These results are particularly larger among women, adults, and those with a higher level of education. The magnitude of the effects of PAE on employment, formality, and earnings (although imprecisely estimated) is larger than the ones for wage subsidies (Galasso et al., 2004), labor intermediation (Dammert, Galdo, & Galdo, 2015) and well-known training programs (Attanasio et al., 2011; Card, Ibarrarán, Regalia, Rosas-Shady, & Soares, 2011) in LAC.

It is important to remind that the restriction to identify individuals in the control group who benefit from the other programs offered by PES provokes that our results potentially represent a lower bound of the real effect of PAE. Additionally, the non-experimental setting of this evaluation requires that the assumptions discussed above hold. Results from the robustness tests performed to confirm that selection bias on unobservables does not represent a threat to our estimation of causal effects. Additionally, recent meta-analyses of the effectiveness of ALMP show that average program effects from randomized experiments are not very different from the average effects from non-experimental designs (Card et al., 2018; Escudero et al., 2018).

PAE proves to be effective in the context of high informality as the one in Bolivia. Providing job seekers information about job vacancies and a subsidy to work in a formal firm improves their chances of employment, formality, and earnings. These results are particularly important for groups that are usually marginalized in formal labor markets, such as women. However, the program has the challenge of improving the access of other marginalized workers and of making their effects sustainable over time. The program could benefit from introducing a component of skills training, where jobseekers gain the skills required in the classroom and on-the-job training.

Acknowledgment

We thank Mauricio Chumacero, Soraya Roman, Nelson Chacon, Graciana Rucci, and Manuel Urquidi for support and helpful comments. We also thank Ricardo Avalos for his research assistance. The data that support the findings of this study are available from the Inter-American Development Bank and PAE.

Disclosure statement

The authors report there are no competing interests to declare.

Notes

1 Following the national regulation of Bolivia (Law 342), we define youth as individuals younger than 29.

2 Key differences between the context of develop and developing countries (e.g., informality, access to information, quality and relevance of education and skills formation) constrains the chances of extrapolating the results from the former. Generating evidence from developing countries is particularly important now that many countries in LAC have started implementing or reinforced their wage subsidy programmes as a response to COVID-19 (United Nations International Children's Emergency Fund [UNICEF], 2020).

3 Evidence about intermediation services are more frequent in Nordic and Anglo countries, while evidence about subsidies are more frequent in Germanic and Nordic countries.

4 Most of the programmes and evidence about effectiveness comes from developed countries, where wage subsidies have a long-standing tradition (O’Higgins, 2017). For instance, Gelber, Isen, and Kessler (2016) find that an internship programme in New York increases earnings and employment in the short-term.

5 Galasso et al. (2004) analyse the impact of a wage subsidy programme in Argentina, finding that it improves the probability of employment and does not affect earnings. The effects are larger among women and youth.

6 Between 2005 and 2017 the poverty rate in Bolivia reduced from 59.6 to 36.4 percent and the yearly average GDP was 4.9 percent.

7 In 2012, the unemployment rate for youth, women, and people with low level of education were 3, 3, and 1 percent, respectively. In turn, informality rates were 87, 83, and 96 percent, respectively. Informality rate is defined as percentage of employed workers contributing to social security. Inter-American Development Bank, Labour Market and Social Security Information System (SIMS).

8 In 2017, a second version of PAE was implemented. The new version maintains the same logic and logistic of the original programme but includes three specific pilots aiming at targeting three vulnerable groups (youths, women, and people with disabilities). Our evaluation corresponds to the original programme design.

9 Although PAE aims at targeting those with lower levels of education, the programme does not include any eligibility criteria in this regard. Beneficiaries tend to have higher education levels than the general population (Table SM7).

10 According to Mazza (2017), nearly 80 percent of workers in Bolivia finds jobs through informal channels.

11 Wage subsidies increases with educational level and by economic sector. For each educational level, subsidies in manufacture are higher than in services, which in turn are higher than in commerce.

12 Jobseekers could be beneficiaries of PAE only once in their lifetime.

13 PES works as a single window for different labour programmes, such as: Mi Primer Empleo Digno, for disadvantaged youths; Intermediación Directa, which is a search and intermediation assistance programme; and Plan de Empleo, which is a platform combining several policies aiming at increasing employability. Unfortunately, due to data restrictions, it is not possible to identify individuals who are beneficiaries of these interventions. Therefore, our estimates are likely to be a lower bound of the real effect of PAE.

14 At registration with the programme, firms sign an affidavit stating the intention that the vacancy leads to a permanent hiring.

15 The maximum number of beneficiaries of a firm could not exceed 50 percent of its current stock of workers. Once a firm achieved its quota of beneficiaries, it could only reapply a year after the first beneficiaries graduated from PAE.

16 For each jobseeker, shortlists provide national identity number, name and surname, date of birth, address, phone number, and occupation sought. These shortlists do not rank jobseekers but listed them by the date of registration into the PES.

17 Employers received the information that was available at the time of PES registration only. During the intermediation process, caseworkers only matched vacancies with jobseekers and did not offer more services (e.g., vocational information, training) (Cooper, 2014).

18 Table SM10 in the SM show the characteristics of PAE beneficiaries.

19 Data collection about jobseekers was performed for those registered between January 2015 and June 2017. Nevertheless, we restrict our analysis to 2016 and 2017, due a high level of attrition for 2015 jobseekers. The difficulty in finding jobseekers prior to 2016 is explained by the high percentage of people who changed their contact telephone numbers, a common occurrence in Bolivia.

20 To consider an interview as “refused” interviewers were required to make at least five communication attempts, at different times.

21 Similar rates hold by beneficiaries and no-beneficiaries status (Table SM1).

22 Table SM3 in the SM shows the differences between those who completed or not (including the ones who were not contacted) the survey. Those who completed the survey are older, more likely to be women and married, have a higher level of education, higher expected wage and manifested interest in fewer vacancies than those who did not complete the survey. As expected by the data collection protocol, those registered in 2017 are slightly more likely to complete the survey.

23 Graph SM1 in the SM shows the distribution of propensity scores for the treatment and control groups. Less than 1 percent of the sample falls outside the common support.

24 In Table SM7 we compare some baseline characteristics of PES beneficiaries with household survey data. This allow us to get a sense about how representative is our sample. We could conclude that, on average, those in SPE are younger, more educated and less indigenous.

25 Through the paper, income is deflated to 2017 prices and converted to US $ using an exchange rate of 6.96 bolivianos per US$.

26 Table SM4 in the SM shows that the effect of PAE on employment, formal employment, and log earnings presented in this Section hold when different kernel bandwidths (0.09, 0.03 and 0.01, rather than the 0.06 default) are used. Similarly, results hold when standard errors are calculated by a bootstrap with 500 replications (Table SM5 in the SM).

27 Table SM6 in the SM shows the effect of PAE on earnings, unconditional on working. For this, we calculate an inverse hyperbolic sine transformation (IHS).

28 These results need to be taken carefully because differences in the composition of applicants to the different cohorts might confound with the effects of PAE over time.

29 Tables SM8 and SM9 show that baseline characteristics are balanced under the two methods (IPW and ex-post randomization).

30 To strengthen our analysis, we additionally performed the ex-post randomization exercise 100 times and plotted the distribution of the estimated coefficients for each outcome. Graph SM2 in the SM show that the distributions are also consistent with the results found with the original sample.

Appendix

Table A1. Descriptive statistics at baseline

	Unweighted variables				Weighted variables
	Mean control	Mean treated	t		Mean control	Mean treated	t
Employment status at the time of registration	0.311	0.306	0.463		0.306	0.306	0.010
Ever received a promotion	0.333	0.324	0.811		0.327	0.324	0.310
Has work experience	0.419	0.349	6.379	***	0.345	0.350	0.540
Indigenous	0.106	0.069	5.576	***	0.072	0.069	0.460
Woman	0.580	0.572	0.724		0.568	0.571	0.350
Age	31.371	30.730	3.238	***	30.819	30.717	0.650
Age squared	1064.689	1013.140	3.441	***	1019.320	1012.432	0.590
Education	14.280	14.277	0.029		14.291	14.285	0.100
Has a disability	0.060	0.019	8.424	***	0.021	0.019	0.790
Married	0.256	0.265	−0.927		0.264	0.264	0.030
Head of household	0.353	0.357	−0.363		0.354	0.356	0.150
# of children	0.827	0.897	−2.364	**	0.897	0.892	0.190
Tertiary education diploma	0.322	0.298	2.339	**	0.302	0.299	0.350
Household monthly income (Bs)	2089.066	2276.116	−6.040	***	2274.320	2265.362	0.320
Year of registration in the PES
Year 2016	0.664	0.854	−18.921	***	0.845	0.853	1.280
Year 2017	0.336	0.146	18.921	***	0.155	0.147	1.280
Month of registration in the PES
January	0.085	0.052	5.525	***	0.061	0.052	2.020	**
February	0.110	0.056	8.175	***	0.058	0.056	0.420
March	0.111	0.122	−1.631		0.124	0.120	0.630
April	0.105	0.075	4.440	***	0.073	0.076	0.590
May	0.141	0.079	8.352	***	0.075	0.079	0.770
June	0.086	0.064	3.559	***	0.061	0.064	0.730
July	0.062	0.081	−3.344	***	0.080	0.081	0.240
August	0.066	0.109	−7.341	***	0.107	0.110	0.480
September	0.058	0.106	−8.390	***	0.105	0.105	0.000
October	0.067	0.123	−9.295	***	0.127	0.123	0.540
November	0.066	0.087	−3.754	***	0.086	0.088	0.400
December	0.044	0.045	−0.409		0.044	0.046	0.420
PES office of registration
El Alto	0.220	0.192	3.027	***	0.199	0.193	0.780
Sucre	0.030	0.078	−10.908	***	0.078	0.078	0.010
La Paz	0.377	0.260	11.027	***	0.273	0.260	1.490
Cochabamba	0.084	0.082	0.294		0.077	0.083	1.170
Oruro	0.064	0.080	−2.805	***	0.081	0.080	0.220
Potosi	0.040	0.048	−1.817	*	0.047	0.048	0.210
Tarija	0.040	0.059	−3.931	***	0.056	0.059	0.550
Santa Cruz	0.114	0.123	−1.220		0.118	0.124	0.850
Trinidad	0.013	0.019	−2.337	**	0.021	0.020	0.550
Cobija	0.018	0.060	−11.405	***	0.049	0.056	1.540
(Log) wage offer (PES)	7.674	7.636	6.106		7.639	7.637	0.480
PES caseworker id	2200.459	2271.836	−2.686	***	2256.118	2268.407	0.590
Number of job applications	1.490	1.471	0.638		1.500	1.472	1.200

Note: *Significant at 10%; **significant at 5%; ***significant at 1%.