El presente Reglamento será obligatorio en todos sus elementos y di rectamente aplicable en cada Estado miembro.
CATEGORÍA 1 — MATERIALES ESPECIALES Y EQUIPOS CONEXOS 1A Sistemas, equipos y componentes
N. B.: VÉANSE TAMBIÉN LA RELACIÓN DE MATERIAL DE DE FENSA Y EL ARTÍCULO 1C450.
DiD Methodology
The goal of this chapter is to investigate the effect of a decrease in theUIbenefit level on the self-employment probability in Spain. Such an UI benefit level change was implemented in the context of the labor market reform agenda 2012. The new law lowered the replacement rate after the first six months of an individual’sUIbenefit receipt by about 16.67% (seeSection 2.4.2). As this reform affected only individuals entitled to more than six months of benefits who entered theirUIbenefit spell after 15 July 2012, one can differentiate between treated and untreated individuals who entered theirUI benefit spell either during the pre- or during the post-reform period.
We exploit this quasi-experimental setup to identify the causal reform effect of thisUIbenefit level reduction using a parametric Difference-in-Differences (DiD) ap- proach. WhileRebollo-Sanz & Rodríguez-Planas(2020) estimate the reform effect on the job-finding probability in employment, we apply theDiD strategy additionally to self-employment and total employment (= employment + self-employment). In other words, we decompose the total employment effect of the reform into the effect on self- employment and employment (vs. staying unemployed). Consequently, we are able to identify the potential bias that emerges through ignoring self-employment and only focusing on employment. We implement this strategy estimating a linear probability model separately for each month23 afterUIspell entry as illustrated below:
P(Yit = 1|Ti, POSTi, Xi j) =α + β · Ti+ γ · POSTi+ δ · Ti× POSTi + J
∑
j=1 λj· Xi j+ εit (2.1)Individuals who entered into unemployment prior to 15 July (1 January 2011 - 14 July 2012) are assigned to the pre-reform period, whereas those who entered into unemployment on 15 July or later (15 July 2012 - 31 December 2013) are assigned to the post-reform period. This circumstance is indicated by the binary variable POSTiwhich takes either the value 0 when an individual belongs to the pre-reform period (POSTi= 0), or the value 1 when belonging to the post-reform period group (POSTi= 1). Moreover, individuals are assigned to either treatment or control group. Those entitled to more than six months ofUIbenefits constitute the treatment group which is represented by the dummy variable, Ti, taking the value 1 (Ti= 1). Conversely, individuals entitled toUI benefits of not more than six months form the control group (Ti= 0).
Our estimation sample includes individuals who received UI benefits within the period between 1 January 2011 and 31 December 2013 (large sample). We follow them over their unemployment24 spell until they exit into a new job, become self-employed or their observations get censored. Furthermore, the sample is restricted to individuals of age 35-52 with anUIentitlement length of at least 120 days to align treatment and control groups. In addition, we apply the same strategy to a sample with an age restriction that is in line withRebollo-Sanz & Rodríguez-Planas(2020), i.e. with individuals who are 35 to 52 years old.
Three different sets of outcome variables are used. In the first set, our dependent variable Yit represents a binary outcome that takes the value 1 if individual i exits from unemployment into the state of interest (either self-employment, employment or total employment - seeFigure 2·3) in month t, given that we still observe this individual as being unemployed at the beginning of month t. The outcome variable takes the value 0 if the individual stays unemployed in month t. Consequently, the effects on three different outcome variables per month t can be estimated. We follow each individual’sUIspell over 26 months and identify the dynamic development of theAverage Treatment Effect (ATE). The second set of outcome variables measures whether the individual became self-employed or employed within a certain amount of months t. We chose intervals of 6, 12, 18, and 24 months. Finally, a set of unemployment duration variables is used as outcome variable to estimate different types of duration elasticities. In particular, we distinguish between the total unemployment spell duration (considering UI, UAand unregistered periods as job seeker to constitute unemployment) and theUIspell duration.
In our basic setting, only models including the group dummy Ti, POSTi and their interaction are estimated. In further steps, different sets of control variables are added (represented by vector Xi j). All of them are measured at an individual’sUIspell entry. Socioeconomic variables include a female dummy, log of age, educational level dummies (lower, secondary, and university education), an indicator whether the individual has Figure 2·3:Illustration of Binary Outcome Variables
Notes:BesidesUIspells, unemployment also includesUAspells and unregistered spells which we count as being unemployed without receiving any kind of benefits. For more details, see also AppendixII.3.3.
Source:Authors’ own illustration.
24The unemployment spell includes bothUIandUAreceipt and counts unregistered periods as unemployment spells without receiving any kind of benefits.
children, and an immigrant dummy. Macroeconomic control variables include quar- terly real GDPgrowth rate, month indicators, and dummy variables for all Spanish Autonomous Regions. Ultimately, the vector of controls entails a set of pre-displacement job characteristics: log of employment experience, log of self-employment experience, and occupational skill level (high, medium, low skilled). The summary statistics of all variables used in this chapter are presented inTables II.4andII.5of the Appendix. The variables’ exact definitions can be inferred from AppendixII.3.3.
DiD Identification
The identification strategy of the reform effect can be summarized in the following steps: first, estimating the (self-)employment probabilities of treated relative to non- treated individuals; and second, comparing both groups across time, i.e. those who were displaced in the post-reform period with those displaced in the pre-reform period. The additional comparison with workers displaced at the same time but assigned to the control group (CG) is used to cancel out other factors that may have systematically affected both groups.
Unlike in a laboratory experiment, we only observe individuals in one of the four states (in the pre-reform or post-reform period, and belonging to the control or treatment group). So first of all, our identification strategy requires that the composition of treatment and control group is not affected by the reform itself. Additionally, our identification strategy requires that the treatment and control group behave similarly regarding our outcome variable in the pre-reform period. In other words, theDiDestimator can only be unbiased if the parallel-trend assumption holds. Then, the average of the control group captures the counterfactual development of the treatment group and we can identify the causal reform effect. As long as their composition stays constant and time shocks are common to both groups, the parallel trend assumption holds. In this case, treatment and control group are allowed to start at different levels of the outcome variable.
Figure 2·4illustrates the quarterly number25 ofUIinflows by group. The quarterly inflow level is constantly higher with regards to the treatment group compared to the control group. However, the composition of both groups’ inflows seems to develop fairly parallel, which could serve as evidence in favor of a fixed group composition. Moreover, there is no evidence ofUIentry date manipulation, i.e. there are no suspicious spikes right before the reform was implemented (red line). This finding speaks in favor of a fixed group composition and emphasizes the statement ofRebollo-Sanz & Rodríguez-Planas
(2020) that strategic lay-offs to avoid the replacement rate reduction are rather unlikely since the reform was implemented already two days after its announcement.
25In the Appendix,Figure II·18shows the corresponding figure in percentage terms orFigure II·19in a monthly dimension.
Figure 2·4:UI Transitions (Total Numbers) 0 1000 2000 3000 4000 5000 6000 7000
Total number of individuals in each group
2008q12008q32009q12009q32010q12010q32011q12011q32012q12012q32013q12013q32014q12014q3
UI transition quarters
Treatment Control
Notes:This figure illustrates the quarterly transitions intoUI, i.e. the total number of individuals in both the treatment and the control group who switch intoUIin each quarter. The sample is restricted to individuals who are 20 to 52 years old, with anUIbenefit entitlement length of at least 120 days, whose transition takes place between the first quarter of 2008 and the last quarter of 2014. The reform quarter is highlighted with a red dashed line. In the Appendix, Figure II·18shows the corresponding figure in percentage terms orFigure II·19in a monthly dimension.
Source:Authors’ calculations based onMCVL2005-2017 data.
Figure 2·5illustrates the validity of the parallel trend assumption. It plots quarterly average probabilities of exiting into self-employment26 in the period between Q1/2010 and Q4/2014. The sample is restricted to individuals of age 20-52, with anUIentitlement length of at least 120 days; and the reform quarter is highlighted in red. This figure shows that the common trend assumption holds, i.e. that treatment and control group seem to have parallel trends with respect to their outcome variables before the reform took place.
Even though our evidence speaks in favor of both a fixed group composition and a valid common trend assumption, some risk remains that theDiDestimator is biased due to inherent differences between the groups. Tables II.6andII.7show mean comparison tests of some interesting covariates between the two groups for different age restrictions. Indeed, most of the variables are significantly different between treatment and control groups, and some of them are included into our model (through Xi j) to control for group differences to make parallel trends more plausible. Moreover, the conditional independence assumption27 requires a full common support of both treated and non- treated individuals’ characteristics. Thus, we decided to pre-select our sample based on the propensity score, under the awareness that implementing such trimming may come at the expense of some external validity, since the focus is set on a subset of the original sample.
26The parallel trend checks for employment and total employment as outcome variables can be inferred from AppendixFigure II·20. They are in line with the findings with regards to self-employment.
Figure 2·5:Parallel Trends Check for Self-Employment 0 .005 .01 .015 .02 .025 .03
SE exit indicator (mean)
2010q1 2010q3 2011q1 2011q3 2012q1 2012q3 2013q1 2013q3 2014q1 2014q3
Quarterly date of individual's UI spell
Treatment Control
Notes:This figure illustrates quarterly average probabilities of exiting into self-employment in the period between the first quarter of 2010 and the last quarter of 2014. The sample is restricted to individuals who are 20 to 52 years old, with anUIentitlement length of at least 120 days. The reform quarter is highlighted with a red dashed line.
Source:Authors’ calculations based onMCVL2005-2017 data.
In line with Crump et al. (2009), we only include individuals with a treatment propensity score between 0.1 and 0.9. Figures 2·4and2·5andTables II.6and II.7are already based on the pre-selected sample. However note, that without pscore trimming these figures look almost identical in shape.28 The mean comparison tests improved slightly through pscore trimming and some of the differences turned insignificant or smaller in its magnitude.29 TheDiDanalysis which follows inSection 2.6.1is based on the pscore trimmed sample. Since the new law from 2012 also changed labor market rules for workers older than 52, this seems to be a reasonable maximum age restriction to avoid that other sections of the reform bias our results. In our main settings, we restrict our estimation sample even further to individuals who are between 35 and 52 years old. The reason for this is the Royal Decree-Law 4/2013, which was adopted on 22 February with the goal of promoting self-employment among young workers (defined as men younger than 30 and women younger than 35), and which could affect our results as well. To sum up, theDiDapproach allows us to estimate the average treatment effects of the reform in our quasi-experimental scenario. However, we can also take advantage of the discontinuity in the replacement rate which was the object of the 2012 labor market reform. In fact, focusing only on treated individuals, we would be able to identify local average treatment effects only for the respective group, providing us with a more complete picture for this impact evaluation.
28Figures without pscore trimming can be provided upon request. Note, that the levels inFigure 2·4decrease as individuals with the lowest and highest pscore percentiles are excluded.