[PPT] - End-of-Year Spending and the Long-Run Effects of Training Programs PowerPoint Presentation

SLIDE 1

End-of-Year Spending and the Long-Run Effects of Training Programs for the Unemployed

B. Fitzenbergera,c, M. Furdasa, C. Sajonsb

a: Humboldt Uni Berlin / b: ifm Uni Mannheim / c: ZEW Mannheim IRP - Summer Research Workshop - Madison - 20 June 2018

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 1 / 37

SLIDE 2

Introduction Motivation

Motivation

Training programs for the unemployed:

Very important part of active labor market policies (ALMPs)

Focus on providing occupation-specific skills, human capital in general,

and/or work experience

Opportunity to obtain vocational training degrees
Lock-in effect: reduced job-search intensity during participation
Intensive Use of ALMP in West Germany
Between 500-700 Tsd. entries into training p.a. in the 80s and 90s
Significant amount of resources: about 3.4 bn DM in the mid-80s

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 2 / 37

SLIDE 3

Introduction Motivation

Related Literature I

Large number of evaluation studies estimating effects on employment /

earnings accounting for selection on observables (see e.g. Card et al. 2010, 2015; Lechner et al. 2009; Biewen et al. 2014, for an overview for Germany)

Difficult to account for selection on unobservables:

⇒ Causal evidence accounting for selection on unobservables is scarce Richardson/van den Berg 2013 and Osikominu 2013 [Continuous-time duration modelling, no IVs]; Frölich/Lechner, 2010 [IV: Regional treatment intensity, complier effect]; Aakvik et al. 2005 [IV: Degree of rationing, MTE + ATT]; Caliendo et al. 2014 [No effect of additional covariates, which are typically unobservable]

Evidence for effect heterogeneity with respect to unobservables: Aakvik et
al. 2005 - Treatment effect falls with treatment probability, cream skimming
No evidence for cream skimming in Frölich/Lechner 2010

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 3 / 37

SLIDE 4

Introduction Motivation

This paper

Use IV strategy to account for unobserved selection and estimate the

long-run employment effects of training (over 10 years)

Idea: Exploit “end-of-year spending” effects caused by strict budget rules in

West Germany in the 1980s and early 1990s as source of conditional exogenous variation in training participation

Non-transferability of funds
Budget largely determined by previous years’ spending

⇒ Incentive to use all remaining funds before the end of the year and tieing up funds for the next ⇒ Use the budget surplus after the 1st half of the year as instrument for treatment in the 2nd half

Implement IV strategy in a dynamic setting using a two-step control

function approach for models with binary outcomes and binary endogenous treatment (Wooldridge, 2014)

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 4 / 37

SLIDE 5

Introduction Motivation

Related Literature II

Lots of anecdotal evidence for end-of-year spending in government agencies

and company divisions (references in paper).

Little empirical evidence due to lack of data Gao (1998).
Recently, a small literature on end-of-year spending hike associated with a

lower quality of the projects

Liebman and Mahoney (2017) [IT procurement decisions by the US

government] → precautionary motive plus decreasing returns

Later, Eichenhauer (2017) [contributions to foreign aid for a panel of

countries] → lack of planning capacity and bureaucratic effectiveness → dismisses procrastination by bureaucrats as alternative explanation

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 5 / 37

SLIDE 6

Outline

Institutional background
Data
Model of end-of-year spending
Empirical strategy - Identification
Effects of training on subsequent employment
LATE – 2SLS estimation
ATT – control function approach
Heterogeneous effects by training program
Conclusions

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 6 / 37

SLIDE 7

Institutional background

Organization Federal Employment Office West Germany

Three levels:
Central office in Nuremberg
9 regional employment offices

(REO, Landesarbeitsämter)

142 local employment offices

(LEO, Arbeitsämter)

Annual budget determined and

managed largely at the federal level

Local offices possess limited

discretion in the use of their allocated funds for training programs subject to budget plan of center/region

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 7 / 37

SLIDE 8

Institutional background

Budget Rules Until 1993

Stable budget rules within the Federal Employment Office (FEO) during the

1980s and early 1990s

Budget rules (1980 – 1993)
Non-transferability of funds: budget for training programs was planned and

allocated separately from other programs like job creation schemes

Allocation of funds top-down to regional and local offices was based

primarily on past levels of program participation (“head-count” calculation) and adjusted according to the expected economic development

Unused funds from one fiscal year could not be transferred to the following

year

Budget for the next year depended on the degree of utilization in the current

year and the comparison of each LEO with the other local offices

LEO has an incentive to spend the whole remaining budget before the end
f the year

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 8 / 37

SLIDE 9

Institutional background

The employment office’s budget year

⇒ Use remaining financial leeway after the first half of the year as a measure for the magnitude

f possible end-of-year spending behaviour of

local officials

Best time for readjustment was the period

immediately after the summer holidays ⇒ Consider only programs starting in the months after the summer holiday (August – November)

Figure : Monthly shares of total

year entries into training programs

4 6 8 10 12 Share in percent 1 2 3 4 5 6 7 8 9 10 11 12 Calendar month

Shares averaged over regions and time (1983–1993).

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 9 / 37

SLIDE 10

Data and training programs

Data

Data sources:
Integrated Employment Biographies (IEB) based on employees’ history and

the benefit recipients’ history

Training participation (Fortbildung und Umschulung, FuU) data
Actual and planned spending (Yearbook of the FEO, 1980-1993)
Sample: 50% sample of all program participants (1980–1993) and 3%

sample of all individuals (followed until December, 2004)

Training programs included (by length of duration):
Short-term training (STT)
Practice firms (PF)
Specific professional skills and techniques (SPST)
Retraining (RT)

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 10 / 37

SLIDE 11

Data and training programs

Sample and treatment definition

Sample restrictions:
Unemployment inflows after employment spell lasting at least three months
Individuals living in West Germany and aged 25 to 50 at the beginning of

unemployment

Treatment: First training within the first 12 months after entering

unemployment by elapsed unemployment duration (stratum)

Dynamic control group (Sianesi, 2004): Individuals with the same elapsed

unemployment duration as in the treatment group, but without or with later treatment, i.e., all individuals are replicated for each month they remain unemployed and they are eligible for treatment

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 11 / 37

SLIDE 12

Model of end-of-year spending behavior

Stylized model to rationalize end-of-year spending by LEO official in two

period model with two types of training program

Trade-off between budget absorption and maximizing returns from training
Random shocks to entries into training
End-of-year spending: Entries in period 1 lower (higher) than planned

prompts higher (lower) planned entries in period 2

Higher entries associated with composition changes towards less

effective programs → higher marginal costs to find suitable matches for more effective program

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 12 / 37

SLIDE 13

Model of end-of-year spending behavior

LEO official minimizes expected loss over two periods of the budget year L = E

(B − e1 − e2)2 + w(e1 − e∗)2 + w(e2 − e∗)2

, (1) where B: training budget (measured in entries), e1, e2 entries into training in periods 1, 2 e∗ entries maximizing net returns w relative weight of e∗, w ≥ 0 LEO official decides upon planned entries ˆ e1 and ˆ e2 at beginning of each period Actual entries: et (ˆ et) = ˆ et + ǫt where ǫt random shock with expectation zero given history and decision on ˆ et

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 13 / 37

SLIDE 14

Model of end-of-year spending behavior

End-of-year spending: Second period response to actual entries in first period ˆ e∗

2(e1) =

w 1 + w · e∗ + 1 1 + w · (B − ˆ e1 − ǫ1)

budget left

. (2) → response of planned entries ˆ e2 is less than one-for-one wrt budget left (B − e1) Surprise component ǫ1: Unanticipated → leading to a change of −

1

1 + w

· ǫ1

in planned entries in second period ˆ e∗

2(e1) and subsequently in actual entries e2.

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 14 / 37

SLIDE 15

Model of end-of-year spending behavior

Net return maximizing entries e∗ (both periods symmetric) from perspective of LEO staff Two types of training programs (j = 1, 2) Expected net returns of planned entries nj into program j are specified as Πj(nj) = ajnj − bj 2 n2

j

where aj constant treatment effect (invariant to n1, n2) (bj/2)n2

j search and management costs

Normalize b1 = 1 and set b2 = b Assumption: Program 1 less effective than program 2, a1 < a2, and involving lower marginal search costs, b1 = 1 < b2 = b. Note: Search costs are a function of planned entries reflecting the effort of the LEO staff to find suitable match; eligibility criteria are harder for program 2

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 15 / 37

SLIDE 16

Model of end-of-year spending behavior

Optimal e∗ maximizes Π1(n1) + Π2(n2) with e∗ = n∗

1 + n∗ 2 = a1 + a2/b

Actual planned entries n1, n2 set for LEO staff: Optimization subject to choice

f planned entries e by official

max

{n1,n2} Π(n1, n2) = Π1(n1) + Π2(n2)

s.t. n1 + n2 = e . Solution n1(e) = a1 − a2 1 + b + b 1 + b e and n2(e) = a2 − a1 1 + b + 1 1 + b e . Composition effect in response to increase in planned entries Entries into program 2 less responsive to an increase in total entries than entries into program 1 n′

2(e) = 1/(1 + b) < n′ 1(e) = b/(1 + b)

b/c marginal search costs are higher for program 2 than for program 1.

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 16 / 37

SLIDE 17

Model of end-of-year spending behavior

Average effectiveness of training programs (based on planned entries) a(e) = n1(e) n1(e) + n2(e)a1 + n2(e) n1(e) + n2(e)a2 = a1 + 1 1 + n1/n2 (a2 − a1) , → depends negatively on the ratio n1/n2 End-of-year spending hike reduces quality of training (and vice versa)

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 17 / 37

SLIDE 18

Empirical strategy - Identification

Operationalizing "Budget Surplus"

Problem: Budget situation of LEOs after 1st half of the year unobserved (1) Count the number of program entries for all training programs per LEO in every year and month (2) Predict the intended number of new participants at the local level for each year (“the budget”) based on entry patterns and local labor market situation

f the previous three years

(3) Adjust planned entries by the percentage change in annual federal budget for training measures (incorporate macroeconomic effects) (4) Budget leeway at the time of mid-term review for LEO l year t ⇒ Budget leewaylt =

Planned entrieslt−Actual entrieslt Eligible unemployed in Julylt [in1000]

(5) Relative surplus: budget leeway relative to all other LEOs (deviation matters most) ⇒ Relative surpluslt = Budget leewaylt −

1 L−1

L

s=l Budget leewayst

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 18 / 37

SLIDE 19

Empirical strategy - Identification

Distribution of relative surplus

Figure : Histogram (1983–1993)

5 10 15 Percent −100 −50 50 100 150 Relative surplus

µ ≃ 0, σ = 34.541, F −1(0.5) = −4.653

Figure : Spatial distribution of average

relative surplus

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 19 / 37

SLIDE 20

Empirical strategy - Identification

Validity I: Relevance condition (First Stage)

Table : Estimates of the first stage treatment probability (all training programs)

(1) (2) (3) (4) (5) (6-NEW) Standardized budget leeway 2.577∗∗∗ 2.194∗∗∗ 1.922∗∗∗ 1.883∗∗∗ 1.866∗∗∗ 2.021∗∗∗ Standard error (0.178) (0.231) (0.218) (0.223) (0.227) (.273) F-statistic 209.7 90.1 78.1 71.2 67.7 54.5 Region & time infoa) yes yes yes yes yes yes Personal characteristicsb) yes yes yes yes yes yes Relative surplus in t − 1 no yes yes yes yes no Region × year interactions no no yes yes yes yes Local UR (Jan-July)c) no no no yes yes yes Local UR (last year)d) no no no no yes yes Seasonality ur + entries ur, emp, tr no no no no no yes Unconditional treatment probability (weighted) = 65.81%; cluster size = 1 562

Note: The numbers report average partial effects (APE) per 1000 unemployed in percentage points for separate regressions of the probability to enter any program on the standardized surplus. Standard errors are clustered at local labor market level and time and reported in parentheses.

Results are robust when

(6-NEW) does not subtract leave-one-out average
using 2 (4) years for out-of-sample prediction of planned entries (partial effects are a bit smaller)
including more lags in the local unemployment rate
mitting the correction of budget surplus w.r.t. changes in intended spending

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 20 / 37

SLIDE 21

Empirical strategy - Identification

Validity II: Exclusion condition

Threat: Third variables influencing both surplus and employment conditions after treatment start (during and after the end of the program) But: ...

Labor market shocks: Likely to affect treated and non-treated in the same

way

Management style of LEO regarding seasonality of training:

Accounted for by the lagged value of relative surplus

Almost no LEO consistently above or below average
No significant correlation b/w surplus now and local employment conditions

later on (after program completion) → programs are small!

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 21 / 37

SLIDE 22

Effects of Training on Subsequent Employment

OLS ATT, 2SLS Estimation LATE

Estimation of the Effect of Treatment on the Treated (ATT): Benchmark

based on selection on observables assumption: Dynamic OLS [sufficiently flexible] to replicate semiparametric matching estimates

Align treated and not-yet-treated-but-eligible unemployed during month of

treatment start; report averages over months 1 to 12 of elapsed U-duration

Average across months of treatment starts, i.e. by months of elapsed

unemployment durations (1 to 12 months)

Outcome variable: Employment rates during years 1 to 10 after treatment

start

Results are robust to Inverse Probability Weighting and to Fractional Probit
Specifications (4) and (5): W/o or w/ lagged lagged value of relative

surplus as control in outcome equation and first stage, resp.

2SLS: Same specification as OLS

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 22 / 37

SLIDE 23

Effects of Training on Subsequent Employment

LATE – 2SLS Estimation

Table : Effect of training on subsequent employment of compliers (LATE)

Specification (4) Specification (5) Specification (6-NEW) Year Desc OLS-ATT IV-Red IV-LATE IV-Red IV-LATE IV-Red IV-LATE 1

.192∗∗∗
.198∗∗∗
.003∗∗
.171∗∗
.003∗
.138∗
.141
.070

(.002) (.002) (.002) (.078) (.002) (.080) (.172) (.084) 2–3

.054∗∗∗
.055∗∗∗
.002
.132
.002
.103
.017
.008

(.002) (.002) (.002) (.104) (.002) (.106) (.214) (.107) 4–10 .023∗∗∗ .021∗∗∗

.002
.083
0.001
0.030
.170
.084

(.002) (.002) (.002) (.096) (.002) (.098) (.198) (.099)

Note: ∗∗∗, ∗∗, and ∗ indicate statistical significance at 1%, 5%, and 10% level, respectively. Standard errors (in parentheses) are clustered at time and local labor market level and obtained through weighted bootstrapping based on 200 replications. Specification (5) includes the local unemployment rate from previous year in addition to the controls from specification (4).

LATE effects negative but quite imprecisely estimated,
Treated: Negative selection during lock-in period / positive selection in the long run
2SLS results sensitive (in the magnitude) wrt selected sets of control variables, but

always negative in the long run.

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 23 / 37

SLIDE 24

ATT – Control Function Approach

Control function approach for nonlinear models with discrete endogenous covariates (Wooldridge, 2014)

Selection on unobservables: Control function approach for nonlinear models with discrete endogenous covariates (Wooldridge, 2014), applied to Fractional Probit Random Coefficients Model: bd

t random coefficient on d

yt = 1 [y ∗

t ≥ 0] = 1

at0 + z1b0 + (bt0 + z1(b1 − b0))d + bd

t d + ut ≥ 0

(3a)

d = 1 [d∗ ≥ 0] = 1 [γ0 + z1γ1 + z2γ2 + ν ≥ 0] = 1 [zγ + ν ≥ 0] , (3b) where yt employment, d treatment dummy, y ∗

t and d∗ are latent indices,

z = (z1, z2) exogenous, z2 instruments Selection/Endogeneity: ν and (ut, bd

t ) are not independent

Specification of z1 analogous to OLS

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 24 / 37

SLIDE 25

ATT – Control Function Approach

Average effect of treatment on the treated (ATT)

τATT,t ≡ E

y 1

t − y 0 t | d = 1

= Pr(y 1

t = 1 | d = 1) − Pr(y 0 t = 1 | d = 1)

Average Potential Outcomes y ˜

d t with ˜

d = 0, 1, integrating out the distribution

f (z1, bd

t , ut) among the treated d = 1:

Pr(y

˜ d t = 1 | d = 1) =

Ez1,bd

t ,ut|d=1

1
at0 + z1b0 + (bt0 + z1(b1 − b0))˜

d + bd

t ˜

d + ut ≥ 0

.
Builds on Average Structural Function (Blundell and Powell, 2003, 2004)

ASF(˜ d, z1, d = 1) = Pr(y ˜

d t = 1 | z1, d = 1)

Note that Pr(y ˜

d t = 1 | d = 1) does not depend upon z2

(exclusion restrictions).

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 25 / 37

SLIDE 26

ATT – Control Function Approach

Estimating Counterfactuals accounting for Selection on Unobservables

Wooldridge (2014, section 6.3): Set of control functions e2 = k2(d, zi, θ)

redundant in outcome equation for employment (3a)
sufficient statistic for capturing the endogeneity of d, i.e. the distribution of

(bd

t , ut) conditional upon d and z depends only upon e2

ASF by integrating out e2:

ASF(˜ d, z1, d = 1) = Ee2|d=1,z1

Prob

at0 + z1b0 + (bt0 + z1(b1 − b0))˜ d + bd

t ˜

d + ut > 0 | d = 1, z1, e2

.

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 26 / 37

SLIDE 27

ATT – Control Function Approach

Implementation

For estimation purposes: Assume bd

t ˜

d + ut is normally distributed conditional on e2 and expectation is a linear function of e2. Wooldridge (2014, section 6): e2 involves

generalized residual: gr = dλ(zγ) − (1 − d)λ(−zγ)

≡ standard Heckman (1978) selection correction

interactions of gr with z1
interaction of gr with the treatment dummy d

Note: square gr 2 can not be used due to multicollinearity Note: Effect of gr is not non-parametrically identified

> flexible, smooth functional form in gr as parametric approximation

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 27 / 37

SLIDE 28

ATT – Control Function Approach

Two-stage Procedure

Calculate the generalized residuals,

gr i, based on the estimated coefficients from the first-stage probit regression for the probability of treatment: ˆ gr i = diφi/Φi − (1 − di)φi/(1 − Φi)

Use the estimated generalized residuals in a flexible way in the second-stage

fractional probit model for employment:

Pr(yit = 1 | di, zi1, ei2) =

(4) Φ z1ˆ b0 + z1ˆ δ1di + ˆ ω0 gr i + ˆ ω1 gr idi + zi1 gr i ˆ ψ ,

Estimate ATT by integrating out discrete differences over distribution of
bservables and gri in the treatment sample
τATT,t = 1

N1

di =1
Φ

z1ˆ b0 + z1ˆ δ1 + ˆ ω0 gr i + ˆ ω1 gr i + zi1 gr i ˆ ψ (5) −Φ z1ˆ b0 + ˆ ω0 gr i + zi1 gr i ˆ ψ .

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 28 / 37

SLIDE 29

ATT – Control Function Approach

Specification Tests Generalized Residuals (Spec. 5)

Year 1 2–3 4–10 Joint sig. [p-value] CF-ATT 1: Treatment dummy di, gri, gri · di ω0 .039 (.091)

.030 (.090)
.251 (.077)∗∗∗

[.003]∗∗∗ ω1 .143 (.040)∗∗∗

.009 (.034)

.021 (.031) [.000]∗∗∗ ω0 = ω1 = 0 [.000]∗∗∗ [.893] [.000]∗∗∗ [.000]∗∗∗ CF-ATT 2: As CF-ATT 1 plus interactions di · z1 ω0 .030 (.088)

.066 (.091)
.239 (.077)∗∗∗

[.010]∗∗ ω1 .011 (.058) .107 (.041)∗∗∗ .132 (.041)∗∗∗ [.005]∗∗∗ ω0 = ω1 = 0 [.943] [.002]∗∗∗ [.000]∗∗∗ [.000]∗∗∗ CF-ATT 3: As CF-ATT 2 plus interactions gri · z1 ω0 .274 (.152)∗ .051 (.175) .049 (.163) [.300] ω1

.040 (.056)

.099 (.040)∗∗ .148 (.040)∗∗∗ [.001]∗∗∗ ω0 = ω1 = 0 [.062]∗ [.045]∗∗ [.001]∗∗∗ [.001]∗∗∗ ψ = 0 [.000]∗∗∗ [.000]∗∗∗ [.000]∗∗∗ Note: Standard errors (in parentheses), p-values [in brackets] of χ2-statistics for Wald test of joint significance of parameters based on clustering at time and local labor market level and

btained through weighted bootstrapping based on 200 replications.

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 29 / 37

SLIDE 30

ATT – Control Function Approach

ATT – Effect of training on Employment (Spec. 5)

Year OLS-ATT FP-ATT IV-LATE CF-ATT 1 CF-ATT 2 CF-ATT 3 1

.198∗∗∗
.199∗∗∗
.138∗
.225∗∗∗
.218∗∗∗
.165∗∗∗

(.002) (.002) (.080) (.058) (.057) (.051) 2–3

.055∗∗∗
.055∗∗∗
.103
.037
.014
.016

(.002) (.002) (.106) (.058) (.059) (.056) 4–10 .021∗∗∗ .021∗∗∗

.030

.180∗∗∗ .172∗∗∗ .124∗∗∗ (.002) (.002) (.098) (.048) (.048) (.045)

Note: Standard errors (in parentheses) are clustered by time and local labor market

level. Specification controls for regional and time variation, individual characteristics,

lagged value of relative surplus, time path (January-July) of local unemployment rate in the year of treatment start, and local unemployment rate from the previous year.

Strong negative lock-in effects for program participants
Long-run treatment effects significantly positive and higher than OLS estimates
In total, negative selection into training
Differences between compliers and treated on average → higher budget induced

entries lower effectiveness of programs

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 30 / 37

SLIDE 31

ATT – Control Function Approach

ATT – Effect of training on Employment (Spec. 6-NEW)

Year OLS-ATT FP-ATT IV-LATE CF-ATT 1 CF-ATT 2 CF-ATT 3 1

.198∗∗∗
.199∗∗∗
.070
.219∗∗∗
.208∗∗∗
.149∗∗∗

(.002) (.002) (.084) (.055) (.054) (.048) 2–3

.055∗∗∗
.055∗∗∗
.008
.007

.015 .016 (.002) (.002) (.107) (.058) (.059) (.055) 4–10 .021∗∗∗ .021∗∗∗

.084

.162∗∗∗ .150∗∗∗ .101∗∗ (.002) (.002) (.099) (.048) (.048) (.044)

Note: Standard errors (in parentheses) are clustered by time and local labor market

level. Specification controls for regional and time variation, individual characteristics,

lagged value of relative surplus, time path (January-July) of local unemployment rate in the year of treatment start, and local unemployment rate from the previous year.

Main results robust, especially for CF approach

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 31 / 37

SLIDE 32

Heterogeneous Effects by Training Program

Heterogeneous Effects: Size of Treatment Group by Training Program and Elapsed Unemployment Duration

Table : Treatment group size

Elapsed duration STTa) PF SPST RT All Control group 1–3 3 813 3 070 19 446 10 453 36 782 771 139 4–6 3 186 2 327 13 017 5 731 24 261 426 544 7–9 2 698 2 040 9 535 4 941 19 214 320 852 10–12 2 358 1 670 7 057 3 594 14 679 240 373 1–12 12 055 9 070 49 055 24 719 94 936 1 758 908

Note: a) Since STT programs existed only until the end of 1992, the control group for this training is a bit smaller and amounts to 1 470 480 non-treated observations. Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 32 / 37

SLIDE 33

Heterogeneous Effects by Training Program

Having only one instrument?

Heckman, Urzua, Vytlacil (2008): One instrument per alternative for

comparisons among multiple treatments, where treatment-specific instrument excluded in value for alternative treatments Only one instrument needed in pairwise comparison for one program versus next best alternative

Here: Pairwise comparison of one treatment versus no treatment
no joint equation with multiple endogenous treatments
probit differs by training program: specific approximation of selection
ATT: selection of treated wrt to their nontreatment outcome
nontreatment group very large, i.e. selection of this group small

→ innocuous for treatment in one specific training program against the alternative of no training at all → refrain from estimating pairwise effects of one training program against another training program

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 33 / 37

SLIDE 34

Heterogeneous Effects by Training Program

Effect of Relative Surplus on Treatment Probability

Table : First stage results (benchmark specification)

(1) (2) (3) (4) STT PF SPST RT Relative surplus 2.171∗∗∗ 2.381∗∗∗ 4.716∗∗∗ 1.080∗∗ Standard error (0.576) (0.361) (0.918) (0.349) F-statistic 14.201 43.672 26.414 9.583

Note: The numbers report average partial effects (APE) per 1000 unemployed in percentage points for separate regressions of the probability to enter the respective training program on the standardized surplus. STT: Short-term training; PF: Practice firms; SPST: Specific professional skills and techniques; RT: Retraining; SSPST: Short SPST (planned duration <= 6 months); LSPST: Long SPST (planned duration > 6 months). Standard errors are clustered at local labor market level and time and reported in parentheses. The corresponding F-statistics are reported in square

brackets. Benchmark specification: controls for regional and time variation, individual characteristics, lagged value
f relative surplus, time path (January-July) of local unemployment rate in the year of treatment start, and local

unemployment rate from the previous year. ∗∗∗, ∗∗, and ∗ indicate statistical significance at 1%, 5%, and 10% level, respectively.

All four programs positively affected by the instrument ⇒ LEOs increase/reduce

participation in all four programs in case of surplus/deficit

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 34 / 37

SLIDE 35

Heterogeneous Effects by Training Program

Effect of training on Subsequent Employment for STT and PF

(1) (2) (3) (4) (5) (6) (7) (8) Year Des OLS-ATT FP-ATT IV-RED IV-LATE CF-ATT 1 CF-ATT 2 CF-ATT 3 Short-term training (STT) 1

0.168∗∗∗
0.151∗∗∗
0.151∗∗∗
0.001
0.052
0.182∗∗∗
0.182∗∗∗
0.147∗∗∗

(0.003) (0.003) (0.002) (0.002) (0.058) (0.044) (0.045) (0.057) 2–3

0.120∗∗∗
0.098∗∗∗
0.098∗∗∗
0.001
0.041
0.182∗∗∗
0.189∗∗∗
0.119∗

(0.005) (0.004) (0.002) (0.003) (0.058) (0.052) (0.052) (0.064) 4–10

0.057∗∗∗
0.045∗∗∗
0.045∗∗∗
0.002
0.063
0.068
0.074

0.017 (0.005) (0.004) (0.002) (0.002) (0.048) (0.052) (0.052) (0.067) Practice firms (PF) 1

0.165∗∗∗
0.146∗∗∗
0.146∗∗∗
0.005∗∗∗
0.187∗∗∗
0.217∗∗∗
0.228∗∗∗
0.192∗∗∗

(0.004) (0.002) (0.003) (0.002) (0.063) (0.045) (0.046) (0.058) 2–3

0.074∗∗∗
0.048∗∗∗
0.048∗∗∗
0.004∗
0.130∗
0.143∗∗∗
0.149∗∗∗
0.098∗

(0.005) (0.004) (0.004) (0.002) (0.075) (0.039) (0.059) (0.054) 4–10

0.037∗∗∗
0.011∗∗∗
0.011∗∗∗
0.004∗∗
0.143∗∗
0.276∗∗∗
0.285∗∗∗
0.215∗∗∗

(0.005) (0.004) (0.004) (0.002) (0.073) (0.048) (0.048) (0.047)

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 35 / 37

SLIDE 36

Heterogeneous Effects by Training Program

Effect of training on Subsequent Employment for SPST and RT

(1) (2) (3) (4) (5) (6) (7) (8) Year Des OLS-ATT FP-ATT IV-RED IV-LATE CF-ATT 1 CF-ATT 2 CF-ATT 3 Specific professional skills and techniques (SPST) 1

0.146∗∗∗
0.159∗∗∗
0.160∗∗∗
0.003∗
0.185∗
0.065
0.071
0.107∗∗

(0.002) (0.002) (0.002) (0.002) (0.101) (0.044) (0.044) (0.048) 2–3

0.006∗
0.008∗∗∗
0.008∗∗∗
0.004∗
0.234∗
0.010
0.008
0.008

(0.003) (0.003) (0.003) (0.002) (0.141) (0.055) (0.054) (0.056) 4–10 0.026∗∗∗ 0.013∗∗∗ 0.013∗∗∗

0.002
0.076

0.050 0.014 0.025 (0.003) (0.002) (0.002) (0.002) (0.117) (0.045) (0.043) (0.043) Retraining (RT) 1

0.308∗∗∗
0.316∗∗∗
0.316∗∗∗
0.004∗∗
0.400∗∗∗
0.127∗∗∗
0.115∗∗∗
0.122∗∗

(0.002) (0.002) (0.002) (0.002) (0.058) (0.033) (0.057) (0.050) 2–3

0.136∗∗∗
0.129∗∗∗
0.129∗∗∗
0.002
0.217
0.177∗∗∗
0.156∗∗∗
0.032

(0.003) (0.003) (0.003) (0.002) (0.058) (0.044) (0.059) (0.061) 4–10 0.076∗∗∗ 0.076∗∗∗ 0.076∗∗∗

0.001

0.017 0.096∗∗ 0.110∗∗ 0.140∗∗ (0.003) (0.003) (0.003) (0.002) (0.048) (0.043) (0.048) (0.066)

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 36 / 37

SLIDE 37

Conclusions

Budget surplus in first 6 months increases training during remainder of year
Effects of training on the long-run (4-10 years after) employment chances of

unemployed individuals differ strongly

No effect (or even negative), if unemployed individuals were assigned to

spend down remaining funds (compliers)

Significant ATT of ≥ 10 percentage points
Findings are consistent with model predicting lower quality of training

matches when planned entries into training increase b/c of budget surplus

Different training programs respond positively to budget surplus with

suggestive evidence that less effective programs respond more strongly

Strong heterogeneity of treatment effects by the type of training (treatment

vs no training at all):

Significantly negative effects for short-term training and practice firms
Significantly positive effects for longer training (HC accumulation)

Fitzenberger, Furdas, Sajons End-of-year Spending 20 June 2018 37 / 37