Setting and adjusting targets for

ESF Operational Programmes

Background Paper
Client: EC, DG Employment, Social Affairs and Inclusion
Rotterdam, 26 May 2013

Setting and adjusting targets for

ESF Operational Programmes

Background Paper

Client: EC, DG Employment, Social Affairs and Inclusion

Dr Martin van der Ende

Rotterdam, 26 May 2013

About Ecorys

At Ecorys we aim to deliver real benefit to society through the work we do. We offer research,
consultancy and project management, specialising in economic, social and spatial development.
Focusing on complex market, policy and management issues we provide our clients in the public,
private and not-for-profit sectors worldwide with a unique perspective and high-value solutions.
Ecorys remarkable history spans more than 80 years. Our expertise covers economy and
competitiveness; regions, cities and real estate; energy and water; transport and mobility; social
policy, education, health and governance. We value our independence, integrity and partnerships.
Our staff are dedicated experts from academia and consultancy, who share best practices both
within our company and with our partners internationally.
Ecorys Netherlands has an active CSR policy and is ISO14001 certified (the international standard
for environmental management systems). Our sustainability goals translate into our company policy
and practical measures for people, planet and profit, such as using a 100% green electricity tariff,
purchasing carbon offsets for all our flights, incentivising staff to use public transport and printing on
FSC or PEFC certified paper. Our actions have reduced our carbon footprint by an estimated 80%
since 2007.

ECORYS Nederland BV
Watermanweg 44
3067 GG Rotterdam
P.O. Box 4175
3006 AD Rotterdam
The Netherlands
T +31 (0)10 453 88 00
F +31 (0)10 453 07 68
E [email protected]
Registration no. 24316726
W www.ecorys.nl

NL02 26598

Table of contents

Summary

Introduction

Survey on targeting methodologies

Methodologies for setting output targets

11

3.1

Introduction

11

3.2

Budget allocation as a first step

11

3.3

Calculating unit costs

22

3.4

The cost equation

26

3.5

Normative issues cost efficiency

29

3.6

Setting output targets with limited or no data

29

Methodologies for setting result targets

31

4.1

Introduction

31

4.2

Result equation

32

4.3

Application in setting targets

38

4.4

Normative issues frontier analysis

40

4.5

Setting result targets with limited or no data

42

Methodologies for adjusting targets

44

5.1

Introduction

44

5.2

Adjusting output targets

44

5.3

Application of time series models in setting output targets

47

5.4

Adjusting result targets

50

List of common indicators

53

(1) Common output indicators on participants

53

(2) Common output indicators for entities

53

(3) Common immediate result indicators on participants

53

(4) Common longer-term result indicators on participants

53

Target optimization with Excel

55

Summary

In preparation of the ESF 2014-2020 multi-annual framework, the European Commission organised
a learning seminar on setting and adjusting targets for ESF Operational Programmes. To account
for expenditures, indicators on outputs (numbers of participants, entities) and results (numbers of
participants finding employment, gaining a qualification etc.) are essential. This background paper
discusses methodologies for setting and adjusting quantified cumulative employment targets.
According to a survey of the EC, two commonly applied methods in the 2007-2013 programming
period were extrapolation from the past and a combined use of budget allocation and unit cost.
Common challenges were accounting for external factors and lack of data.
This paper first discusses the combined use of budget allocation and unit cost per participant. This
approach allows for an integrated setting of output and result targets to maximise the expected
achievement of an objective. An objective could be to maximize the number of participants in
employment (the result target). Another objective could be to maximize the total number in
employment from both participants and non-participants. This may result in different output targets
and higher employment levels overall (overall effectiveness) despite lower result targets (gross
effectiveness). Another approach is a SWOT analysis, which allows an integrated approach of topdown setting of objectives and bottom-up allocation of budgets given unit costs. Methods to allocate
budgets proportionally to target groups and by extrapolation are easier to apply but have the
drawback that they are not result oriented.
When calculating unit costs from historical data, it is also preferable to use actual historical costs
instead of historically committed costs because the actual number of participants may be different
from the initially targeted number and this could result in distorted unit costs and output targets for
the next programming period. In the case of on-going interventions, it is still preferable to use data
on actual costs incurred and to add estimates of the remaining cost to be incurred, over using
initially committed budgets for the whole period. Unit costs are also preferably calculated at a
disaggregate level to calculate more accurate output targets. When gaps in disaggregate data exist,
or perhaps when costs are influenced by factors that change in the future, a cost equation that
relates unit costs to other factors can be used to estimate unit costs. However normative issues
also play a role: are efficiency gains possible to reduce unit costs, if some projects achieve similar
results for similar target groups at lower cost? Finally, in the complete absence of any data on unit
costs, Eurostat LMP data can be used, but own judgement will still be needed.
The setting of result targets is as easy or as difficult as applying success rates to the output targets.
More than unit costs, success rates depend on a large number of factors beyond the influence of an
ESF Operational Programme, and a result equation can be used to predict future success rates. To
predict the future success rates this way, the development of the other factors need to be predicted
as well, or need to be indicated in scenarios of the future. Result equations could be tailor-made to
specific interventions, but it requires advanced econometrics to estimate such equations. A Bartik
(2009) type of model is linear and uses micro-data from individual participants and is less sensitive
to assumptions underlying the model. As for unit costs, own judgement is needed as well to assess
whether better success rates could be achieved in the future. Simply picking the project with the
highest success rate is not a sound method, because this project may have been targeted at easier
groups, where participants would have found a job without intervention as well. Randomized

experiments and matching are methods that help compare success rates for similar groups and
frontier analysis further helps choosing projects with higher result rates that should be achievable
for other projects. In the end, allocating more budget to such projects make it more likely that
(higher) result targets are achieved.
Finally, the result equation can also be used to adjust targets if at a certain moment during the
programming period, the forecasts of the factors influencing the success rates change. In this case
one needs to add the predicted results for the remainder of the programming period to the results
that could have been expected up to that moment based on the previously set targets.

Introduction

In preparation of the ESF 2014-2020 programming period, the European Commission organised a
learning seminar on setting and adjusting targets for ESF Operational Programmes. To account for
expenditures, indicators on outputs (numbers of participants, entities) and results (numbers of
participants finding employment, gaining a qualification etc.) are essential. For a limited number of
output and result indicators, quantified cumulative targets for the whole programming period need
to be set. This background paper presents and discusses methodologies for setting and adjusting
quantified cumulative employment targets.
Targets at Operational Programme level
In the ESF, Member States set national targets for outputs and results for the whole period 20142022 (up to two years beyond the ESF programming period). The results of an Operational
Programme are the sum of the results of its investment priorities and ultimately the individual
projects, as illustrated in Figure 1.1. If for example a combined total of 10,000 people participate in
projects conducted in the framework of one investment priority, and out of them 30% on average
are in employment upon, different projects under the Operational Programme may well have
varying success rates. In the example of Figure 1.1, the success rates vary from 20% to 50%.
Likewise, a target at Operational Programme level does not imply proportionate targets for each
project.
Figure 1.1

Output and results at investment priority level

Source: Ecorys, prepared for the learning seminar on targeting of 14 March 2013.

Targets at the level of investment priorities are not only different from targets at project level, they
are not targets at the macroeconomic level either. According to a survey of the EC discussed in the
next chapter, macroeconomic performance indicators such as the employment rate and the
unemployment rate are influenced by external factors beyond the control of ESF Operational
Programmes, such as the business cycle and changes in national legislation. In addition, only part
of the target population participates in ESF co-funded projects. The wider impact of ESF
Operational Programmes on for example the unemployment rate are therefore hard to estimate.

However, the output and result indicators do not relate to wider impacts, and are not discussed in
this background paper.
List of common indicators
The draft ESF regulation contains a provisional list of common indicators to be used for the period
2014-20201, which is also given in annex A in this paper.
The two types of indicators to be used are:

Output indicators (numbers of participants or entities involved);

Result indicators (mostly on the employment situation and training).

Positioning of the paper

This background paper builds on a 2011 methodological study2 on performance targets for ESF
Operational Programmes which focused on methodologies to adjust result targets. Compared to
that study, this paper focuses more on practical issues and also broadens the scope to setting
(besides adjusting) targets and to output indicators (besides result indicators).
The discussion of setting targets in this paper starts under the assumptions that historical average
data are available, and ends with discussing possible approaches without such data. This paper
discusses how to use data to set programme targets, but not how programme targets can be used
to manage projects. This paper also discusses how to use estimates of macroeconomic factors for
target setting, but does not discuss the wider impacts of the outputs and results at the macro level.
This background paper is further organized as follows:

Chapter 2 summarizes results of a EC survey on targeting methodologies;

Chapter 3 discusses methodologies for setting output targets;

Chapter 4 discusses methodologies for setting result targets;

Chapter 5 discussed methodologies for adjusting targets.

Regulation of the European Parliament and of the Council on the European Social Fund and repealing Council Regulation
EC No 1081/2006; Proposal of 14.3.2012,
http://ec.europa.eu/regional_policy/sources/docoffic/official/regulation/pdf/2014/proposals/regulation/esf/esf_proposal_en.pdf
2
Ecorys (2011), Performance Targets for ESF Operational Programmes. http://www.coprbm.eu/?q=node/494

Survey on targeting methodologies

At the start of 2013, the European Commission conducted a survey on targeting methodologies
used in the previous 2007-2013 programming period, in preparation of the learning seminar. From
this survey, it became clear that the Member States applied different methodologies. Extrapolation
of historical data was the most commonly applied method, by 10 out of 17 responding countries,
see Table 2.1. However this method was always combined with other methods. Another frequently
applied method was to calculate output targets from budgets and unit costs (7 countries). However,
new Member States had less data from the past to extrapolate from and some newer Member
States started from national strategic documents (5). Some countries explicitly indicated they
adjusted targets from the past according to the new economic situation (2) or estimated social and
labour market trends (3). Ex ante evaluations (3) and expert judgement (6) were also applied and
one country explicitly mentioned the use of benchmarking.
Table 2.1

EC Survey on targeting methods used

Method used (multiple choice)

Regression / extrapolation from the past

Number of countries
10

Output indicators from budgets and unit costs

Expert judgement

Starting from national strategic documents

Ex ante evaluation

Adjusting targets to estimated trends

Adjusting targets to new economic situation

Benchmarking

Number of countries responding to the EC survey on targeting

17

Source: EC survey on targeting, summary by Ecorys

As Member States used different methodologies, they also faced different challenges. The most
frequently mentioned challenge was the long time horizon of the ESF period, mentioned by 7
countries (Table 2.2). In the 2007-2013 programming period, external factors such as changes in
the demand for labour, demographic changes and even changes in national legislation required
adjustments of priorities and the associated targets.
Table 2.2

EC Survey on targeting challenges

Challenge (multiple choice)

Number of countries

Long time horizon; external factors

Lack of historical data

Macroeconomic impact indicators

Realistic estimations

An intervention is a mix of activities

Regional / target group disparities

Coherence with national policies

Targets in percentages or absolute numbers?

Questions on result-based management

Coherence between indicators (output, result, impact)

Few or many targets

Inclusion of other targets besides ESF indicators

Challenge (multiple choice)

Number of countries

Methodology for baseline targets

Keep records of EC methodologies

Number of countries responding to the EC survey on targeting

17

Source: EC survey on targeting, summary by Ecorys

A lack of historical data is another frequently mentioned challenge (5 countries), and indicates that
the most often applied methodology of regression / extrapolation from the past can be far from
straightforward. Lack of data is mentioned in particular by newer Member States, but also by other
countries, for new types of interventions.
Developing macroeconomic impact indicators was another challenge faced in the 2007-2013
programming period (4 countries). However, the list of common indicators for the ESF 2014-2020
period does not include macroeconomic impact indicators. Moreover, as for programme-specific
indicators, the Commission does not recommend to use impact indicators for ESF programme
monitoring. A more specific but nevertheless important question is whether to formulate output
targets in percentages (percentage of the population) or in absolute numbers (2 countries).
A reality check was another challenge mentioned specifically by a few countries (3) how realistic
will it be to deliver the targets? Quantifying targets was also a challenge for interventions that
consist of a mix of activities, in the situation of strong disparities between regions or target groups
(2 countries each).
Two countries that extrapolated data from the past, mention the challenge to ensure coherence with
national policies. Two countries question the application or even usefulness of targets in result
based management. Related issues raised by countries are whether to formulate few or many
targets, the coherence between indicators (if interventions have poor results prioritizing other
interventions to achieve better results implies that output targets for the original interventions are
not met). Also, other targets besides ESF indicators may be important for specific interventions.
The example given in the survey is getting closer to the labour market. It can be noted in this
respect that the list of common indicators for the 2014-2020 period includes participants with an
improved labour market situation 6 months after leaving.
Two specific methodological issues mentioned by countries concern the formulation of the
baselines and the harmonization of methodologies with EC guidance. In this background paper the
formulation of baselines with limited data is discussed in the concluding sections of the next two
chapters (Sections 3.6 and 4.5). Harmonization of methodologies with EC guidance is mentioned
by one country, but not further elaborated, nor in this background paper.

10

Methodologies for setting output targets

3.1

Introduction
Annex 1 of the ESF Regulation for the programming period 2014-2020 distinguishes between two
types of output indicators:

Numbers of participants in specific target groups;

Numbers of entities involved.

This chapter discusses setting output targets for participants. The main target groups are set out in
the provisional list of common indicators (see Annex A). In the previous ESF programming period,
objectives were to be set in National Action Plans (top-down approach), and unit costs were applied
to calculate what outputs were feasible given the budget (bottom-up approach).
Ideally output targets are set in such a way that the result target (gross effect of the programme), or
even better the number of persons in employment (overall effectiveness) is maximized and this
method is discussed first. A SWOT analysis that the French government applied in the previous
programming period combined both the top-down and bottom-up approach coherently. This method
determined the budget allocation. Other countries applied different methods to allocate budgets.
After determining the budget allocation, unit costs were applied to calculate output targets. All these
methods are discussed in Section 3.2.
Section 3.3 discusses the calculation of unit costs which play such an important role in the setting
of output targets. Output targets are set most accurately with disaggregate unit costs, but this
requires accurate disaggregate data. To overcome some of the data requirements, Section 3.4
discusses how a cost equation can be estimated, which relates unit costs to characteristics of the
target groups, interventions or regions. However, normative issues which are discussed in Section
3.5 also play a role: is it possible to reduce the unit costs in the future? Finally, a fall-back option for
estimates of unit costs in the complete absence of data is discussed in Section 3.6.

3.2

Budget allocation as a first step

A commonly applied method for setting output targets is to allocate budgets and use unit costs (per
participant) to calculate the number of participants. The following formula describes this method:

The calculation of the output target therefore depends on how budgets are allocated and on the
estimate of the unit cost which is discussed in the next section.
The allocation of the budget is an important issue, because the need for interventions and the
effectiveness of interventions may change over time, in different ways for different regions and
target groups. For example, when labour demand increases, job counselling is likely to be more
effective. In times of crisis, occupational mobility tends to increase3 and the same applies to
3

Moscarini, G. and F. Vella (2008), Occupational mobility and the business cycle, NBER Working Paper,
www.econ.yale.edu/~gm76/cps_nber_wp.pdf (based on USA data)

11

sectoral mobility,4 and different measures may further support this mobility such as perhaps
retraining workers.
For allocating budgets, the following methodologies are applied in Member States:
1. Ex ante analysis;
2. SWOT analysis;
3. Proportionally to the distribution of the population;
4. Extrapolation of trends.
1.

Ex ante analysis

The ideal ex ante analysis compares the situation without intervention with the situation with
interventions. This may result in a different solution compared to maximizing result targets.
Maximizing result targets optimizes the result achieved with the interventions (gross effectiveness)
whereas maximizing the total number in employment (among participants and non-participants)
optimizes the result achieved through the interventions (overall effectiveness).5
The following hypothetical example illustrates the difference between gross and overall
effectiveness. Suppose out of two interventions, intervention A has unit cost 10,000 per
participant and results in 70% of the participants in employment, and intervention B has unit cost
5,000 and results in 40% of the participants in employment.6 Without intervention, 20% of the
unemployed would find employment. The target population is 20 and the budget is 100,000. Table
3.1 illustrates the results of investing all in intervention A, intervention B or nothing:
A.

If the budget is invested in A, it allows for 10 participants, and the remaining 10 persons of
the target group do not participate. With success rates of 70% and 20% respectively, 7
participants and 2 non-participants find employment, or 9 in total.

B.

If the budget is invested in B, the budget allows for 20 participants, which equals the whole
target group. With a success rate of 40%, this means 8 participants find employment
which is also the total number in employment.

C.

If no budget is invested, 20% of the target group would be in employment or equivalently 4

persons.

Table 3.1

Simplified example of maximizing result target or numbers in employment

Number of participants

Intervention A

Intervention B

(maximizes numbers

(maximizes the

in employment)

result target)

No intervention

10

20

Number of non-participants

10

20

Target population

20

20

20

(= output target)

Participants in employment
(= gross effectiveness)
(= result target)
Non-participants in employment
Total in employment
(= overall effectiveness)
4

5
6

12

Scheele, D., R. van Gaalen en J. van Rooijen (2008), Werk en inkomsten na massaontslag: De zekerheid is niet van de
baan. Verkenning van CBS en WRR. (Work and income after mass layoff: Certainty is not in the job. Study of CBS and
WRR), Amsterdam University Press.
At the 14 March seminar the exercises maximized gross effectiveness because this can be done step by step.
Methods to obtain such data from partial data or to estimate these data without historical data, are discussed in later
paragraphs of this chapter and the next chapter.

As can be seen from the above table, intervention B maximizes the output and result targets, with 8
participants in employment compared to 7 for intervention A. Nevertheless, intervention A results in
the highest number of persons in employment: 9 compared to 8 for intervention B. In the above
example the high success rate of intervention A offsets the lower reach because of the high costs,
because a sufficiently high share of non-participants find employment, but this would not have been
the case if the success rate of intervention B had been 60% instead of 40%.
Finally, one needs to compare the benefit of the additional number in employment through the
intervention compared to no intervention, which is (9 4 = 5) for intervention A, with the cost. The
cost per additional person in employment is 100,000 / 5 = 20,000 for intervention A, and 100 / 4 =
25,000 for intervention B. For comparison to judge these costs, an order of magnitude of 50,000
can be given of an annual benefit of 7,000, discounted 10 years at 5%.7
Whether one maximizes the result target or the total number of persons in employment in an ex
ante analysis, the advantages and disadvantages are:

Advantages:
-

Method is result-based (output and results in an integrated analysis).

The choice between interventions is integrated.

Disadvantages:
-

The percentage of additional job-finders is hard to estimate;

Calculations are needed for every intervention that is considered.

Needed:
-

Preferably, administrative data per participant / person in the target group;

Second best: survey data for a sample of participants / persons in the target group and
aggregate statistics per intervention;

Data on macroeconomic factors that have an impact on interventions;

Expertise to estimate success rates with (and ideally also without) intervention, and to value
one additional person finding work.

The analysis can be done in Excel.

The example to illustrate this method is taken from the first exercise of 14 March. 10,000 workers
are threatened with unemployment, of which 50% would find a job without any intervention. The
value per job-finder is 50,000 and the budget is 10 m. The data for three interventions are
summarized in the table below. The last column reflects capacity constraints: maximum numbers of
participants that can be reached through an intervention. However, the approaches illustrated
below also work with other constraints, for example minimum numbers per region.

Alternative

Success
rate

Capacity

Job counselling

1,000

55%

5,000

Mobility scheme

4,000

60%

10,000

16,000

90%

500

Retraining

Unit cost
per
participant

If the annual benefit 7,000 and a job, if any, is found after 1 year, then the value of reduced expenditures in year 1-10
discounted at 5% per year is: 7,000/(1+5%) + 7,000/(1+5%)2 + + 7,000/(1+5%)9 = 49,755.

13

Approach a: maximize the cost effectiveness per participant.

In this approach, the interventions is identified with the best result per participant, and all the
budget is invested in that intervention until the budget or capacity is exhausted. Later it will be
shown that this approach need not result in the highest overall employment that can be achieved
with the budget.
First, the extra persons in employment through the intervention is calculated by subtracting the 50%
that find employment without intervention from the above table (see column B in the table below,
where cells shaded in green indicate already available data). The unit value per participant is found
by multiplying this difference with the value of 50,000 per person in employment (column C).
Lastly, the interventions are ranked by the difference in unit value and unit cost (column D),
compared to the unit cost. As the table below shows, each alternative has a positive value per jobfinder when the value of reduced benefit expenditure is compared to the cost (column D).

A:
Unit cost per
participant

Alternative

1. Retraining

B:
Difference in
success ratea)

C:
Unit value per
participantb)

D:
Unit value-cost
differencec)

16,000

+40%

20,000

+4,000

2. Job counselling

1,000

+5%

2,500

+1,500

3. Mobility scheme

4,000

+10%

5,000

+1,000

a)

Calculated as the success rate minus 50% without intervention

b)

Calculated as column B times 50,000

c)

Calculated as column C minus column A.

The budget of 10 million can then be allocated in two rounds as follows. First, budget is allocated
to the best intervention until the budget or capacity limit is exhausted, as illustrated in the table
below.

Alternative

A:
B:
unit cost Budget limit for
participants a)

1. Retraining

16,000

2. Job
counselling

1,000

5,000

3. Mobility
scheme

4,000

10,000

Total

14

625

C:
Capacity
limit
500

D:
Participants b)
= min[B ; C]

E:
Budget c)

500

8,000,000

500

8,000,000

a)

For the first intervention, the budget limit is 10 million divided by the unit cost (column A)

b)

The lower value of the numbers from the budget limit and capacity limit (column B and C)

c)

Column A times Column D.

In the second round, the remaining budget is allocated to the second-best intervention until
budget or capacity is exhausted, as shown in the table below.

Alternative

A:
B:
unit cost Budget limit for
participants a)

C:
Capacity
limit

D:
Participants b)

E:
Budget c)

1. Retraining

16,000

625

500

500

8,000,000

2. Job
counselling

1,000

2,000

5,000

2,000

2,000,000

3. Mobility
scheme

4,000

10,000
2,500 10,000,000

Total
a)

For the second intervention, the budget limit is the remaining budget of 2 million divided by the
unit cost (column A)

b)

The minimum value of the budget and capacity limits

c)

Column A times Column D.

The result targets are finally calculated by multiplying the number of participants (the output) with
the success rates, in the table below. Retraining results in 90% in employment for 500 participants,
or 450 persons and job counselling results in 55% in employment for 2,000 participants, or 1,100
persons. The total result target is then 1,550. This method maximizes the cost effectiveness per
participant because step for step, budget is allocated to first the alternative with the highest cost
effectiveness per participant, then the alternative with second-highest cost effectiveness per
participant and so on. In the training, the exercise was built around maximizing the cost
effectiveness per participant because it can be done step by step in Excel.

Alternative

1. Retraining

A:
Output

B:
Success rate

C:
Result target a)

500

90%

450

2. Job counselling

2,000

55%

1,100

3. Mobility scheme

60%

2,500

Total
a)

1,550

Column A times Column B.

Approach b: maximize the number of employed among participants and non-participants

As the simplified example at the start of this section showed, maximizing the result per participant
does not necessarily maximize the number of persons in employment, because this method does
not take into account that a certain number of non-participants can find employment as well, or that
a slightly lower success rate in one alternative can be compensated by a larger reach if the unit
cost per participant is lower.

15

To find the optimum number in employment, one needs to solve a so-called linear programming
problem under budget and capacity restrictions (or political restrictions of minimum numbers) and
the population restriction which specifies that the sum of participants and non-participants is equal
to the number of persons in the target group:
Maximize in employment = result%[1] * participants[1] + + result%[n] * participants[n]
Subject to:
-

Unit cost[1] * participants[1] + + unit cost[n] * participants[n] available budget

Participants[1] Maximum capacity[1]

Participants[n] Maximum capacity[n]

Participants[1] + + participants[n] + non-participants = target group.

This linear programming problem can be solved with an Excel add-in.8 One feeds the data for the
linear programming problem in an Excel sheet, and the add-in solves the optimal numbers of
participants per intervention that maximize the number of persons in employment within the
available budget. It is a black box approach which is illustrated step-for-step in Annex 2; but the
add-in is easy to apply once the user is familiar with it. The result after solving the linear
programming problem as in Annex 2 is given below.

Alternative

Non-participants

A:
Budget

B:
% jobfinders

C:
Output a)

D:
Result b)

---

50%

4,688

2,344

Job counselling

5,000,000

55%

5,000

2,750

Mobility scheme

60%

5,000,000

90%

313

281

5,313

3,031

Retraining

10,000,000

Target

Total in employment

5,375

a)

The output is calculated by the black box linear programming solver add-in.

b)

The result target is also calculated by the black box linear programming solver add-in, but the results
can be verified by multiplying columns B and C.

The total number of employment of 5,375 is higher than achieved by maximizing the costeffectiveness per participant as in the first approach described above. In this approach, the number

16

http://it.usu.edu/plugins/work/sitemaps/107/files/Linear_Programming_Using_Excel.pdf

of participants is 2,500 and therefore the number of non-participants is 7,500. The total number in
employment is therefore as given below.
The black box method of linear programming finds the maximum number in employment that can
be achieved with the available budget, but does not indicate whether the value of the mix of
interventions in terms of reduced benefit expenditures outweighs the cost. To do so, one can
calculate the cost per additional person in employment compared to the situation with no
intervention. In the above example, in the case of no intervention 50% of the population of 10,000
would find a job, or equivalently 5,000 persons. The above mix of interventions therefore results in
(5,375 5,000) = 375 additional persons in employment. The cost of 10 million euro to achieve this
result is then equivalent to 26,667 per participant, which is less than the value of 50,000
associated with each job-finder in terms of reduced benefit expenditures.

Alternative

Budget

Non-participants

Output

% jobfinders

Result

---

7,500

50%

3,750

Job counselling

2,000,000

2,000

55%

1,100

Mobility scheme

60%

8,000,000

500

90%

450

10,000,000

2,500

Retraining for new jobs

in region B

Target

1,550

5,300

Total in employment

The table below compares the outcomes between the output and result targets calculated after
applying the above two methods. The first method maximizes the success rate per participant and
the second method maximizes the number of employed with the same budget.
Table 3.2

Results of the exercise in maximizing gross or overall effectiveness

a. Maximizing success rate

b. Maximizing numbers in

(gross cost effectiveness)

employment (overall effectiveness)

Output (a)

2,500

5,313

Result (b)

1,550

3,031

62%

57%

7,500

4,688

Success% per participant (a/b)

Non-participants (c)
Success% per non-participant

50%

50%

Non-participants in employment

3,750

2,344

Total in employment (b+d)

5,300

5,375

17

While the strength of the ex-ante analysis is that setting output and result targets happens in an
integrated process as illustrated above, the weakness of this approach is that it requires quite a lot
of information:

success rates per type of intervention;

success rate without intervention

the value of each job-finder in terms of reduced benefit expenditures.

In reality, the success rate can vary over the business cycle. For the next programming period, the
success rates of participants starting in 2014-2020 can be different for each year. How to set
targets for the whole programming period, was illustrated on 14 March by the third exercise. In this
exercise, only one intervention is considered. This exercise is elaborated in Chapter 5 on adjusting
targets (for the business cycle).

2.

SWOT analysis

A SWOT analysis compares strengths and weaknesses of interventions, and opportunities and
threats (at the macro level) for these interventions.
It is a logical analysis that can be used to rank interventions qualitatively rather than quantitatively
as in a full-blown ex ante analysis.

Advantages:
-

A result-based approach to allocating budgets;

Policy priorities can be taken on board;

The choice between interventions is integrated;

Avoids exact calculation of extremely hard to calculate net unit gains;

Less strict data requirements.

Disadvantages:
-

The analysis is partly subjective;

Strengths and weaknesses need to be determined for each intervention.

Needed:
-

A mix of quantitative / qualitative data per intervention to identify strengths and weaknesses;

Macroeconomic data to identify opportunities and threats;

Actual target setting can be done in Excel.

An example of a SWOT analysis can be found in a document of the French government that is
available on the internet.9 It first identified at the macroeconomic level (for France)10:

Strengths such as: a highly qualified workforce and regional business dynamics;

Weaknesses such as: the gender pay gap and difficult target groups;

Opportunities such as: jobs in personal services and jobs with durable perspectives;

Threats such as: loss of human capital and competitiveness.

After elaborating the SWOT elements, a number of policy orientations were identified (page 18-20
of the document):

Encourage access or return to work;

Improve the functioning of the labour market;

Invest in human capital;

Stimulate work in jobs with growth potential.

9
10

18

http://www.fse.gouv.fr/IMG/pdf/valuation_ex_ante_du_PO_FSE_version_definitive-2.pdf
Page 12 and 13 of the document for the full list of SWOT elements.

Chapter 2 of the document elaborated how each priority axis of the ESF Operational Programme is
expected to contribute to objectives falling under the above policy orientations. Chapter 3 of the
document described the principles behind allocating budgets to regions, which depends on a mix of
diagnostics as at the national level and budget negotiations. Another document explains how in
France historical ESF (and Cohesion Fund) budgets were adjusted according to changes in SWOT
elements in each region.11
Shift-share models can be used to adjust historical budgets while taking into account regional shifts
in underlying variables; this model is usually applied in analysis of regional growth12 but can be
applied more widely.
3.

Applying the distribution of the population

In this approach, the distribution of the population between regions and target groups is applied to
budget.

Advantages:
-

Allocation is objective and offers equal opportunities;

Less data requirements;

No comparison between interventions is required.

Disadvantages:
-

The approach is not result-based;

The choice between interventions still needs to be made.

Needed: Population data or reliable survey data (e.g. Labour Force Survey)

Can be done in Excel.

A number of countries apply this method to allocate budgets between regions, even before target
setting in countries where the Operational Programmes were already regionalised (for example,
Italy).
The method of applying the distribution of the population is however also very useful as a reality
check: can the outputs be delivered or is the targeted number of participants larger than the target
group population in a region? To apply this method, the future population and its distribution have to
be predicted.
Assuming that the ESF intends to support three different types of interventions helping school
leavers (depending on the skill levels of school leavers) Table 3.3 informs which share of money
should go to which type of intervention in the identified regions (also assuming that regions and
interventions are equally important). If the total available budget were 100 million euro for
interventions, this would imply that 4 million euro would be allocated to interventions for high
educated school leavers in the North region.
Table 3.3

Example of distribution of the population

Region

Target group

Total

North

Central

South

per education level

10%

10%

10%

30%

16%

18%

16%

50%

Low educated
school leavers
Medium educated
school leavers

11

12

The actual budgeting of ESF is integrated with the budgeting of the Cohesion Policy, see
http://www.fse.gouv.fr/IMG/pdf/Cadre_de_Reference_Strategique_National-2.pdf
E.g. Zaccomer, G. and P. Mason (2011), A new spatial shift-share decomposition for the regional growth analysis: a local
study of the employment based on Italian Business Statistical Register.

19

Region
Target group

Total

North

Central

South

per education level

4%

12%

4%

20%

30%

40%

30%

100%

High educated
school leavers
Total per region

Source: Ecorys, example prepared for the learning seminar on targeting

If the population is predicted for the whole programming period 2014-2020, a different set-up of the
table is more convenient, as illustrated in the table below.
Table 3.4

Example of distribution of the population per year

Region

2014

2015

Low

10%

8%

Medium

16%

16%

High

4%

6%

Low

10%

8%

Medium

18%

18%

High

12%

14%

Low

10%

8%

Medium

16%

16%

Target

2020

group
North

Central

South

4%

6%

Total

100%

100%

100%

100%

N school leavers

1 mln

1.02 mln

1.12 mln

High

In the above example, one may allocate proportionally higher budget (in prices of 2014) to years for
which a larger total target group population is predicted.
To set output targets, the allocated budget for each region and target group is divided by the
average unit costs, preferably the average unit cost per region and target group.

4.

Trend extrapolation

Trend extrapolation is another approach to allocate budgets. Trend extrapolation to set output
targets consists of two steps: trend extrapolation per target group / region or type of intervention,
and an adjustment to fit the allocation within the available budget.

20

Advantages:
-

Limited data requirements;

Integrates expected changes in the future.

Disadvantages:
-

The approach is not result-based;

The approach does not incorporate shifts in priorities;

Allocation of budgets is implicitly as objective or subjective as in the past.

Needed: Population data or reliable survey data (e.g. Labour Force Survey)

Can be done in Excel.

For example if the available budget remains the same and three target group populations increase
by +5%, +7.5% and +10% respectively, an adjustment will be needed as the populations increase
more than the budget. Table 3.5 illustrates this principle, where the total budget for the next
programming period remains the same. As before, the shaded cells indicate already available data.
Table 3.5

Steps of budget allocation with trend extrapolation

A: Old Budget

B: Population
increase

C: Preliminary
budget

D: Final budget

Intervention 1

20 million

+5%

21 million

19.5 million

Intervention 2

40 million

+7.5%

43 million

39.8 million

Intervention 3

40 million

+10%

44 million

40.7 million

108 million

100 million

100 million

b)

a)

a)

Old budget (column A) times (1+population increase from column B)

b)

The preliminary budget (column C) adjusted to the total budget (100 million).

In the first step, a preliminary budget is calculated per intervention by increasing the budget with the
same percentage as the increase of the population. In the above example, historically 20 million
euro was spent on the first intervention, and the target group of the first intervention increases by
5%. In a preliminary budget allocation, this results in 21 million euro (20 million + 5%). The
preliminary budgets for interventions 2 and 3 are calculated in the same way.
In the second step to allocate the budget, the preliminary budgets are adjusted to the total budget.
In Table 3.5, the preliminary budgets add up to 108 million euro, while the total available budget is
only 100 million euro. Multiplying all preliminary budgets by
suffices to adjust the preliminary
budgets to the total budget of 100 million euro.
As in all other methods discussed above, output targets are calculated as the budget divided by the
unit cost per participant. The following section discusses the calculation of the unit costs which play
such an important role in setting output targets.

3.3

Calculating unit costs

In this section, a number of aspects of calculating unit costs are discussed. The selection of
aspects is based on the survey of challenges Member States reported for the previous
programming period:

Using aggregate versus disaggregate unit costs;

Estimating costs of uncompleted interventions;

Using actual versus committed costs;

Price inflation needs to be added to historical unit costs.

Aggregate versus disaggregate unit costs

Preferably, different unit costs per participant are calculated for all regions, target groups and types
of interventions that are budgeted. The reason is that if the budget is adjusted for different regions,
target groups or types of intervention, this affects the average unit cost. This is illustrated with a
simplified example below.
Suppose that historically, 10 and 12 million euro were budgeted to interventions A and B
respectively, with respectively 1,000 and 3,000 participants. Then the unit cost per participant is
respectively 10,000 and 4,000 (budget divided by the number of participants). The average

21

historical unit cost is the total budget of 22 million euro divided by the total number of 4,000
participants, or 5,500, as shown in the table below.

Budget

Intervention A

Intervention B

Total

10,000,000

12,000,000

22,000,000

10,000

4,000

5,500

1,000

3,000

4,000

Unit cost
Participants

Now suppose that in the next programming period, the total budget remains the same but 6 million
euro is allocated to intervention A and 16 million to intervention B. If one uses for each intervention
the historical average unit cost across all interventions ( 5,500), then the output target changes
proportionally to the total budget, regardless of the budget allocation.
In the example, if the number of participants of both interventions A and B is calculated by dividing
the new budget by the historical average across both interventions, this results in 6 million / 5,500 =
1,091 participants in intervention A and 16 million / 5,500 = 2,909 in intervention B, adding up to
4,000. The number of participants is the same as the historical number, because the total budget
has remained the same in the example.
Intervention A

Intervention B

Total

6,000,000

16,000,000

22,000,000

Unit cost

5,500

5,500

Participants

1,091

2,909

Budget

4,000

However if the historical unit costs per intervention is used, this results in 6 million / 10,000 = 600
participants in intervention A and 16 million / 4,000 = 4,000 participants in intervention B, adding up
to 4,600 participants, see the table below. The reason that more participants can be reached with
the same budget, is that more budget is allocated to an intervention with lower unit costs. By
allocating more budget to lower-cost interventions, the average unit costs across interventions
decreases, from 5,500 to 22 million / 4,600 = 4,783 in this example.

Budget
Unit cost
Participants

Intervention A

Intervention B

Total

6,000,000

16,000,000

22,000,000

10,000

4,000

600

4000

4,600

The advantage of using disaggregate unit costs is that the output targets can be set more
accurately, but the data requirements are large and certain disaggregate data may be missing. How
to deal with a situation of partially missing disaggregate data, is discussed in the Section 3.4.
Estimating costs of still running interventions
If an intervention is fully completed, one can simply divide the actual expenditures by the total
number of participants. However, if an intervention is still running, one may assume that all
expenditures have been incurred for those participants who have left the programme, but not for
those who are still in the programme. Figure 3.1 illustrates this situation, where 9 euros have been
spent, and the number of participants is 5. Naively dividing the spent budget of 9 by the number of
participants would result in an estimated unit cost of 1.8 per participant. However, recognizing that
one euro still needs to be spent for those participants who are still in the program, the unit cost is
calculated by dividing 10 by 5, resulting in a unit cost of 2.

22

Figure 3.1

Unit costs for running interventions

The unit cost in the situation of running intervention can be calculated by assuming that for those
participants still in the program, only a percentage p of the unit cost need to be incurred. This can
be done by solving the equation:

In the above equation, one fills in the assumed expected budget:

The resulting formula for the calculation of unit cost of still running interventions is then:

) (

This generalization of the unit cost calculation reduces to the standard formula when the
intervention is completed and 100% of the costs are incurred, because then p=1.
Actual versus committed cost
A pragmatic solution for still running interventions might seem to calculate the unit cost from
historically committed budgets and historically targeted numbers of participants. However, this is
not preferred because the actual number of participants could be quite different from the targeted
number whereas the actual cost is likely to be closer to the committed costs because some costs
are fixed, such as start-up costs and overhead.
Figure 3.2 illustrates this situation, where 8 million euro is committed for a targeted number of 4,000
participants. Based on this committed budget and output the unit cost would be 2,000. However if
actually 6 million euro is actually spent on an actual number of 2,000 participants, the unit cost is in
reality 3,000.

23

Figure 3.2

Unit costs based on committed versus actual costs

Whether the historically committed or the historical actual unit costs are used for setting output
targets, matters substantially in the above example. In the first case, the budget allocated for the
next programming period would be divided by 2,000 and in the second case it would be divided
by 3,000. The method to estimate the costs that still need to be incurred for those participants
who are still in the program results in more reliable estimates of the unit cost because it uses more
actual data.
Price inflation
The average unit costs of the 2007-2013 ESF programme need to be inflated to prices of 2014 and
beyond for the 2014-2020 ESF programme. In formula, if expenditure[t] is the expenditure of year t
and pi[t] is the price inflation in year t, then the average unit costs in prices of 2014 is calculated as
follows:
[

Unit cost[2014] =

]
[

For example if prices increased by 2% each year in 2008-2012 except in 2009 when prices
remained constant, and for 2013 and 2014 an inflation of 2% is expected, the unit cost in prices can
be calculated as illustrated in Table 3.6.
Table 3.6

Illustration of the use of price inflation

In prices of
2007

Expenditures

2007
2008
2009
2010
2011
2012
2013

2007-13

24

100

2008

2009

2010

2011

2012

2013

2014

102.0

102.0

104.0

106.1

108.2

110.4

112.6

102

102.0

104.0

106.1

108.2

110.4

112.6

104

106.1

108.2

110.4

112.6

114.8

106

108.1

110.3

112.5

114.7

108

110.2

112.4

114.6

110

112.2

114.4

112

114.2

799.1

In Table 3.6, the shaded cells indicate the actually incurred expenditures, in prices of the going
year. Expenditures were only 100 in 2007 compared to 112 in 2013. However when the inflation of
2008-2014 is accumulated for the expenditures of 2007, the expenditures of 100 in 2007 amount to
112.6 in expected prices of 2014. The expected expenditures of 112 in 2013 needs to be inflated by
only the expected price index of 2014 (1.02) which results in 114.2 in prices of 2014. In prices of
2014, the total expenditures of 2007-2013 is therefore not
100 + + 112 = 742 (sum of shaded cells in Table 3.6) but
112.6 + + 114,2 = 799.1 (sum of last column in Table 3.6).
If the total number of participants were 100 over the period 2007-2013, then the unit cost in prices
of 2014 should in the above example be calculated as 7.991 instead of 7.420.
For the period 2014-2020, future price inflations should be forecasted and applied to both receipts
of ESF funds and expenditures, but this matter of financing goes beyond the scope of this paper.

3.4

The cost equation

A large number of projects will be conducted under a certain investment priority, differing between
regions, target groups, scales of magnitude, types of intervention and years. This raises the
question how valid historical unit costs are when different circumstances have developed in any of
these aspects. In addition, projects can run in multiple regions, et cetera, and data may be lacking.
Together, these difficulties create a multi-dimensional problem, and the question is how to
effectively calculate appropriate unit costs. One approach is that of a cost equation, which relates
the cost of a project to a number of factors. A cost equation is a tool that can be used to greatly
reduce the dimensionality of calculating unit costs, and to estimate costs of projects for which no
historical data are available.13
Attention should be paid to ensure that the estimated relation applies to a whole range of projects.
To this end, it should be avoided that one large project dominates the estimation. One solution is to
relate unit cost rather than total cost to a number of factors. But even then, a particular project can
have extremely high unit costs, for example business start-up support, if a business loan is
included. To further scale unit costs of different projects, experts calculate the logarithm of costs
and after their analysis transform the logarithm of costs back into normal values. To illustrate the
principle of using a cost equation to differentiate unit costs in an efficient way, the following example
illustrates a cost equation without such transformations.
The following formula gives an example of a cost equation for the unit cost per participant:

In this formula, the value a0 reflects fixed costs such as start-up costs and overhead that are
incurred even when there are no participants. The unit cost per participant can be related to the
13

However, when the type of future target groups and interventions are not known, one still needs to make assumptions
about future target groups and interventions, perhaps for different scenarios.

25

number of participants if economies of scale exist: projects with more participants may have lower
unit costs, although unit costs tend to increase again for very large projects. In the above formula,
the unit cost is assumed proportional to the number of participant, with value a1.
The formula further explained:

The variable: Region2 takes on the value 1 if the project is active in Region2, and otherwise
the value 0

The values b2, b3 indicate euro mark-ups for regions compared to a reference region
(Region1 in this example) which must not be included in the equation.

The values c2, c3, are euro mark-ups for target groups targeted by the project, compared to
the reference target group

The values d2, d3, are euro mark-ups for the activities of the project.

The last term is the prediction error of the formula, where 2 indicates the variance of the difference
between the actual unit cost (left hand side of the equation) and the predicted unit cost (right hand
side of the equation).
If there are 30 regions and 10 target groups, this means that 40 values must be estimated to relate
unit costs to these factors. However, often many values need not to be used to calculate unit costs:
1. Values that are virtually zero;
2. Values that are statistically not reliable;
3. Values in a close range.
The principle of Ockhams razor in econometrics is to remove variables with zero, unreliable or
similar values, to greatly reduce the number of variables. For example, consider a cost equation for
unit costs per project, with the following estimated values, with the error margin given below the
estimate (the example abstracts from different activities)14:

Then the value for the second region may be neglected because the mark-up of 3 is very small.
The value for the third region may be neglected because the error margin ( 100) is larger than
the value of the percentage mark-up ( 50). This implies that regions 1, 2 and 3 can be grouped
together as the reference region. The euro mark-ups for regions 4, 5 and 6 may all be assumed to
be 300, because for all three regions, 300 is within the error margin of the estimated values.
The unit cost in the last region is 500 lower than for equally large projects for the same target
groups in the reference region. After applying Ockhams razor the equation becomes (ignoring the
error margins of the estimated values):

14

26

Usually the error margin is determined as the 95% range of values that would be estimated for random selections of the
population which can be calculated as 2 times the standard deviation of the estimate.

Table 3.7 illustrates how with the previous formula, the unit cost of one region (R3) with missing
data for some projects can be calculated. The example is simplified by assuming that all projects in
one region have the same target group, but similar calculations apply for real life data:

Table 3.7
Region

Estimating unit cost

Target group

Number

Historical

Historical unit

Estimated unit

projects

participants

cost

cost

per project
R1

20

50

7500

5500

R2

20

100

4000

5100

R3

20

200

4200

R4

20

50

8000

5800

R5

2&3

20

100

5000

5600

R6

20

200

3500

4500

R7

20

400

2000

2000

120

18,000

63,500,000

66,700,000

20

4,000

+4000 x ?

16,800,000

140

22,000

83,500,000

Total excl. R3
R3
Total incl. R3

Source: Ecorys, example prepared for the learning seminar on targeting

The cost equation provides an objective estimate for projects with missing data, but can also be
used for projects that have a mix of target groups, as in region 5 in the above example, or for new
target groups in new regions (not illustrated in the above example). Without estimating the cost
equation, arbitrary choices would have to be made. For example:

The average cost of all other projects = 63,500,000 / 18,000 = 3,528

The average cost of other projects for target group 3 = 2,750.

However, the power of the cost equation lies not in correctly predicting the unit cost of each and
every project. In the above example, the estimated unit cost is too low for projects in region 1 and
too high for projects in region 2. The power of the cost equation lies in the ability to correctly
estimate the average unit cost over all projects, taking account of differences between regions,
target groups and activities, economies of scale and whatever else is relevant. In the above
example, the total cost exclusive region 3 are estimated at 66.7 million, which is 5% higher than the
actual total cost of 63.5 million; a difference that is generally considered acceptable.
What makes the above model easy to apply, is that a linear cost function is assumed. One
drawback of a linear cost function is that projects with very high unit costs (outliers) have a
disproportionate impact on the estimated cost of projects with no data (historical or future): they
elevate the average. To overcome the disproportionate effect of outliers, one option is to identify
them and remove them from the data, and another is to scale the unit cost, for example by taking
logarithms. However if the cost equation is specified in logarithms:
Log(unit cost) = a + b*x + N(0,2)
One needs to be aware that then the expected cost is not:
exp(a + b*x) but
exp(a + b*x + 2).

27

Non-linear functions for unit costs have in turn the drawback that estimates are very sensitive to the
exact function that is used, and the expected cost can be complex to calculate. For practical
reasons, the linear model, with possibly removing outliers from the data, and applying Ockhams
razor to keep the model transparent and accessible, should be considered first.

3.5

Normative issues cost efficiency

An alternative to using historical unit costs is to set normative unit costs. Projects that have
exceptionally high unit costs compared to similar projects are then excluded from the calculation.
Again, the cost equation helps to take into account multiple factors influencing the unit costs. To set
the normative unit costs, the historical unit cost is compared with the estimated unit cost. If the ratio
is very high, it can be considered to require a lower unit cost for future projects. To continue from
the previous example, the historical unit costs are highest compared to the estimated unit costs for
projects in regions R1 and R4.
Table 3.8
Region

Estimate of cost efficiency

Historical unit cost

Estimated unit cost

Ratio historical to
estimated unit cost

R1

7500

5500

1.36

R2

4000

5100

0.78

R3

4200

R4

8000

5800

1.38

R5

5000

5600

0.89

R6

3500

4500

0.78

R7

2000

2000

1.00

Source: Ecorys, example prepared for the learning seminar on targeting

Both regions have in common that projects are relatively small: only 50 participants per project
compared to 100 and above in other regions. One might then consider procuring projects with 100
or more participants only. According to the cost equation, the unit cost should then be 500 euros
less, because the cost decreases by 10 euro for every extra participant from 50 to 100.
However, whether to actually require lower unit costs remains a normative choice, because lean
and mean projects are not necessarily the best; comparing historical and estimated unit costs is a
tool to identify potential cost inefficiencies, but not more than that.

3.6

Setting output targets with limited or no data

All the methodologies described above assume more or less accurate knowledge of unit costs, and
methods to complete partially known unit costs.
If no historical data are available at all, for example because a new type of intervention is
programmed, aggregate historical unit cost for that type of intervention can be used. Such data
exist for all 27 Member States, for all types of interventions. All Member States contribute the
relevant data to the Eurostat Labour Market Policy database, which is available from the following
web-link:
http://epp.eurostat.ec.europa.eu/portal/page/portal/labour_market/labour_market_policy/database

28

From this database, the following tables can be downloaded:

Expenditures per Member State per intervention;

Number of participants per Member State per intervention: entrants and average stock per year.

The unit cost can be calculated by dividing the expenditures by the number of participants. For the
number of participants, the larger value of the number of entrants and the average stock per year
can be used. Figure 3.3 illustrates the costs of training calculated with the above Eurostat
database: typically, the costs of training are in the 3,000 - 8,000 range.
Figure 3.3

Costs of training per participant

Source: Ecorys and IZA (2012), in Euro prices of 2009.15

The database can help to estimate unit costs in the absence of any data on ESF projects. However
if a Member States experiments with more effective and expensive interventions, or a lighter than
average intervention, one still needs judgement in the estimate of the cost for the new type of
intervention.

15

Ecorys and IZA (2012), Analysis of costs and benefits of active compared to passive measures, Final report, Rotterdam.

29

Methodologies for setting result targets

4.1

Introduction
Setting result targets is as easy or as difficult as applying success rates to the output targets. The
previous Chapter 3 discussed methods for setting the output targets, ideally integrated with setting
result targets.
The success rate can in principle be calculated as the percentage of participants leaving
successfully, in particular the participants leaving and finding employment. However, the success
rates depend on many factors which are likely to change in the future, including:

The business cycle;

The characteristics (composition) of the target group;

Wages and their role in the matching of supply and demand;

Labour market regulation;

Changes in skills needs;

Other structural changes in the demand for labour;

Because all of the factors which influence the success rates are likely to change in the future,
applying historical success rates to output targets does not necessarily result in achievable result
targets. A result equation relates the success rates to such underlying factors. Once these factors
are forecasted, the result equation can be used to predict future success rates, as discussed in
Section 4.2.
Another aspect is whether higher success rates would be achievable if best practices were
adopted in all projects, where the best practice should be defined as the practice that results in
the highest success rates, highest cost efficiency or ideally the highest overall effectiveness.
Benchmarking is the general method to identify best practices. However, a project aimed at a more
difficult target group may have a larger overall effect than a project aimed at persons who would
have found a job without intervention as well. Benchmarking is only useful after controlling for
differences in the composition of the target group and other relevant effects. Techniques for
controlling for these effects include, with references to empirical literature:

Randomized experiments16;

Matching17.

Randomized experiments have to be set up before the intervention, and consist of randomly
(blindly) assigning some persons to the one or the other intervention, or no intervention at all: the
random assignment ensures that effects of other factors average out, and the intervention with the
highest success rate can arguably be attributed to the intervention itself. Matching consists of
comparing the success rates of one group of participants in the one intervention with a similar
group in the other intervention (or with a similar group of non-participants). The project with the
highest success rate for that group, is arguably the best practice for that group.

16

17

Gautier, P., P. Muller, B. van der Klaauw, M. Rosholm and M. Svarer (2012), Estimating equilibrium effects of job search
assistance, working paper, http://www.oecd.org/els/emp/DanishExperiment.pdf.
Ecorys and IZA (2012), Analysis of costs and benefits of active compared to passive measures, Annex F provides
references to studies based on matching techniques in Austria, Germany, the Netherlands, Poland and the UK.

31

The first method requires large efforts in setting up interventions, whereas the second method
requires detailed data on individual participants. Still, both instruments are strongest in providing ex
post information on the net-effect of an intervention in the past.
Instead, Section 4.4 discusses the frontier analysis, which can be applied after controlling for other
effects, whether through the result equation, randomized experiments or matching. Section 4.5
concludes the chapter with discussing practical methods in the absence of data on historical
success rates for particular groups.

4.2

Result equation
The result equation relates observed results to relevant variables such as target groups, regions
and the phase of the business cycle (impact variables) and a variable that captures the effect of the
participation. In its most abstract form, the result equation reads:

The value a would be the average result if all other values b and c were zero, and in general is
needed to ensure the average result for average values of the other factors. The value b controls
for the impact of relevant factors such as the business cycle, composition of the target group
etcetera. The value of real interest in this equation is the third value c. If the value c is greater
than zero, than the programme is effective towards the result.

Advantages:
-

Method is direct;

Observations from all programmes are used to estimate likely results in specific situations;

Knowing what the result would have been in specific situations in the past enables setting
attainable result targets for such a situation in the future.

Disadvantages:
-

No wider impacts beyond the results are taken into account;

The target setter still needs to judge whether better results in the future compared to the
past are realistic.

Needed:
-

Preferably, administrative data per participant / person in the target group;

Second best: survey data for a sample of participants / persons in the target group and
aggregate statistics per intervention;

Data on macroeconomic factors that have an impact on interventions;

Statistical software such as Stata, Eviews or SAS18;

Expertise to estimate unit results.

With regard to the disadvantage that the result equation does not take into account wider impacts
beyond the results, it should be noted that the estimates of wider impacts are not part of the target
setting procedures for the ESF 2014-2020 period. Macroeconomic models could in theory be used
to explore wider impacts, including:

Substitution effect (employers replace non-participants with participants);

Displacement effect (employers without participants lose jobs to those with participants).

However, macroeconomic models have shortcomings as well, in particular:

18

32

Gauss and TSP cannot handle large datasets well, the Excel add-in only serves for basic regression analysis and SPSS
has no convenient programming options to use estimated values in further analysis.

Modelling macroeconomic effects in the labour market is difficult;

Macro-models tend to treat all target groups equal.

In the micro-based approach, the result is typically an improved likelihood of employment. Two
standard forms of the result equation are:

A model of binary outcomes (found a job: yes or no);

A duration model (time till job is found, if any).

The model of binary outcomes is relatively safe, because relatively few modelling assumptions are
required, compared to for example a duration model. In the standard model of Bartik,19 the
probability Pit to find a job is defined for every individual i in a given period t, and is related to
personal characteristics Xi, the business cycle Xt and a participation variable Iit.

Historical values of Pit are defined as 0 for all periods t in which the individual i did not find a job,
until the moment at which the individual did find a job in which case Pit is defined as 1. The values
a0, a1 and a2 are estimated by choosing the values that minimize the difference between the lefthand side and the right-hand side of the above equation.
For example, consider a minimalist dataset below with data for six participants (indicated with 1)
and six nonparticipants (indicated with 0, in column C in the print-out below). The impact variable is
the regional unemployment rate, which is 6%, 7% or 8% depending on the region (column B in the
print-out below). Finally, the first column indicates whether the person found employment (indicated
with 1) or not (value), in column A of the print-out below. In reality, other factors should be included
and a far larger number of observations is needed, time series effects may need to be taken
account of etcetera; the example only serves to illustrate the principle of estimating a regression
equation.

19

Bartik, T. J., R. W. Eberts and W.-J. Huang (2009), Methodology for adjusting GPRA workforce development program
performance targets for the effects of business cycles. Upjohn Institute Working Paper No. 09-154. Kalamazoo, MI: W.E.
Upjohn Institute for Employment Research.

33

In the above example, half of the participants find employment compared to two-third of the
nonparticipants. This might suggest that the intervention is counterproductive, however the
participants are in regions with a slightly higher unemployment rate (7.0% on average) compared to
non-participants (6.8% on average), and the question is whether the higher unemployment rates in
regions of participants can account for the negative effect of participation in the intervention.
After installing the Data Analysis add-in, in a similar way as the Solver add-in as discussed in
Annex 2, one can go to the tab Data and click on Data Analysis in the ribbon and choose the
option Regression

And select the Y range (the first column including the header), the X range (the second and third
column including the header), indicate the labels (the headers) are included and that the constant
(the value a) needs to be included (is not zero).

34

The result of the regression appears in a new sheet as shown in the print-out below.

The adjusted R square is a measure of the goodness of fit, and is an adjustment of the R square for
small numbers of observations as is the case in the above example. In reality relations are not so
strong and an (adjusted) R square of 15% can be considered as very good.

35

The intercept in the above example is 4.0. With an average unemployment rate of 6.9%, this
implies for non-participants an average result percentage of 4.0 0.49 x 6.9 = 61.9%, which is
somewhat less than the actual 66.7% (four non-participants find work and two do not in the
example). The P-values of 0.3% and 0.7% indicate that the value a and the regional
unemployment rate are different from 0 with high certainty (more than 99%).
However, the estimated effect of participation in the intervention is negligible (between -0.57 and
+0.40 with an average value of -0.09), after controlling for the unemployment rate. Hence the
intervention is not counterproductive as seemed at first sight, but does not improve the success rate
either.
Standard individual characteristics that are generally found to have an impact on the probability to
find a job are for example educational level and age, and standard time-variables are for example
the unemployment rate and the activation/benefit expenditure ratio.
The result equation estimated above is a minimalist model, which does not take account of time
series effects, which are discussed in the next Chapter 5. Also, the appropriate model to estimate
depends on the type of intervention, because different types of interventions influence the likelihood
of employment in different ways. Two examples are given below.
Modelling employment incentives
The intervention variable depends on the type of intervention. For employment incentives, this
could be a tax credit for job-finders. According to the econometric literature, the impact of financial
incentives on the probability to find a job is not always significant. 20 However, if gain in household
income is used instead of the wage, the impact of financial incentives is much more likely to be
significant, as is evident from the empirical literature. The example below illustrates the difference
between wages and changes in net household income:

Where indicates the difference between the new situation (participant finds a job) and the old
situation (participant does not have a job). For example if without the job the husband will have a
monthly unemployment benefit of 1,000 and the wife has a part-time job earning another 1,000
per month which is taxed at 20%, and the husband finds a job earning 2,000 per month causing
the wage of both husband and wife to be taxed at 30%, then the gain in net household income is
not 2,000 but only 300:
Wage
Tax of worker
Lost benefit
Extra tax of spouse
Net gain

+ 2000
600 -/1000 -/100 -/= 300

On an annual basis, this worker stands to gain 3,600 (12 times 300) and a one-time tax credit of
1,200 would increase this gain by one third, which is far more substantial than the gain of 5%
compared to the annual wage of 24,000.

20

36

This discussion is based on Ecorys (2004), Benefit systems and their interaction with active labour market policies,
ec.europa.eu/social/BlobServlet?docId=3195&langId=en. This report describes estimates of micro-models for employment
incentives for six EU15 countries.

Since the net gain is roughly 1/7th of the wage in the above example, the estimate of the value a3
should be roughly 7 times as large if net gain is used as the financial incentive variable instead of
the wage, and the estimate of the impact of the tax credit should also be roughly 7 times as large.
To model employment incentives, it is important to compare future alternatives instead of benefits
and wages from the past. Future benefits can be calculated given the national tax benefit system.
For those unemployed that find a job, the wage of the future job is also known. However, for those
that do not find a job, the wage of potential future jobs is unobserved and needs to be estimated.
These unobserved wages are estimated through a wage equation, relating observed wages to
relevant characteristics, in a similar way as the cost equation described in the previous chapter.
Modelling training
For modelling training, the intervention variable has value 1 for participants and value 0 for nonparticipants. While modelling training, the one complication is to take account of the so-called lockin effect. This lock-in effect implies that during the training, participants have a lower probability of
finding employment. It is only after the training that the chance of finding employment should be
improved. Table 4.1 illustrates that if training lasts six months (including waiting time till the training
starts), a comparison of the results of participants and non-participants would be extremely
negative for the intervention, while positive result of the training could be evident in the longer run.
Table 4.1
Region

Illustration of the lock-in effect on results of training

Employed on 1 January

Employed on 1 July

Employed on 31
December

Non-participants

0%

50%

60%

Participants

0%

10%

65%

Source: Ecorys, example prepared for the learning seminar on targeting

If one estimates the impact of training with a binary probability model (found employment, yes or
no), it suffices to compare the employment of participants and non-participants in the medium term.
If a duration model is estimated, it is important to include both an intervention variable for the
chance of employment during the training, and an intervention variable for the chance of
employment after the training. If training has various durations, even assuming the average
duration for all participants improves the quality of the estimates.
If only one training variable instead of both is used, for example has attended training, the effect of
training tends to be overestimated. It is not intuitive to see why, but the following factors play a role:

The probability of employment diminishes over time;

The long period after training contributes more to the fit of the model than the short period of
training;

After the period of training, the probability of participants just starting to look for a job is
compared with the already diminished probability of non-participants who started looking for job
earlier. Apparently, a high value of one single training variable to predict employment correctly
after training, contributes more to the overall fit than a low value to predict employment correctly
during training. The solution is, to use two training variables, for during and after training.

37

4.3

Application in setting targets

To illustrate the use of a Bartik-type result equation, consider an intervention for minority group
unemployed school leavers, for which the value b of the youth unemployment rate is -2, indicating
that for each percentage point of higher youth unemployment rate, the probability of finding a job
decreases by 2 percentage points. In the example below, it is estimated that 40% of the participants
would find employment in 2014, at a youth unemployment rate of 20%. Furthermore, the youth
unemployment rate is forecasted to decrease in the period 2014-2020 as indicated below.
The budget allows for 10,000 participants for the whole programming period, which is 10% of the
target group population of 100,000 minority group unemployed school leavers over the period 20142020. The number of participants is targeted at 10% of the target group in each year.
All the above data are summarized in Table 4.2 below. The question is how to use the relationship
between the success rate and the youth unemployment rate to set result targets for the period
2014-2020?
Table 4.2

Basic data for setting result targets

2014

Youth unemployment rate

20%

% in employment 6 months after leaving

40%

Number minority group unemployed

school leavers
N participants

2015

2016

2017

2018

2019

Sum

20%

16%

14%

12%

10%

8%

20,000

20,000

16,000

14,000

12,000

10,000

8,000

100,000

2,000

2,000

1,600

1,400

1,200

1,000

800

10,000

N in employment 6 months after leaving

The answer is to first add 2% to the success rate for each percentage lower youth unemployment
rate, to predict the success rates for 2015-2020. Secondly, these success rates are multiplied with
the numbers of participants in the respective years to calculate the number in employment (6
months after leaving) in each year. The sum of the participants in employment 6 months after
leaving over all years gives the final result target.
First step: predicting future success rates
The formula to use in this example is:
Success%[t] = success%[2014] 2*(youth unemployment rate[t] youth unemployment rate[2014])
Success%[t] = 40% - 2*(youth unemployment rate[t] 20%).
Table 4.3 shows the results. As before, the shaded cells indicate values that already are provided
or have been calculated. For example, the success rate of 2020 is
64% = 40% - 2*(8% - 20%) = 40% + 2 * 12%.

38

2020

Table 4.3

Predicting future success rates with Bartik-type equations

2014

2015

2016

2017

2018

2019

2020

Sum

Youth unemployment rate

20%

20%

16%

14%

12%

10%

8%

% in employment 6 months after leaving

40%

40%

48%

52%

56%

60%

64%

20,000

20,000

16,000

14,000

12,000

10,000

8,000

100,000

2,000

2,000

1,600

1,400

1,200

1,000

800

10,000

Number minority group unemployed

school leavers
N participants

N in employment 6 months after leaving

Calculating participants in employment per year

In the second step, the predicted success rates are multiplied with the numbers of participants in
each year. In the above example, this means multiplying the success rates in the second row with
the respective numbers of participants in the fourth row. Table 4.4 shows the results of this
calculation in the last row. For example, the number of 512 for 2020 is 64% of the 800 participants
of 2020.
Table 4.4

Predicting participants in employment with Bartik-type equations

2014

2015

2016

2017

2018

2019

2020

Sum

Youth unemployment rate

20%

20%

16%

14%

12%

10%

8%

% in employment 6 months after leaving

40%

40%

48%

52%

56%

60%

64%

20,000

20,000

16,000

14,000

12,000

10,000

8,000

100,000

2,000

2,000

1,600

1,400

1,200

1,000

800

10,000

800

800

768

728

672

600

512

Number minority group unemployed

school leavers
N participants

N in employment 6 months after leaving

Lastly, the result target for 2014-2020 is the sum of the results of each years, which add up to 4,880
as indicated in Table 4.5: 4,880 = 800 + 800 + + 512.

39

Table 4.5

Setting the result target with Bartik-type equations

2014

2015

2016

2017

2018

2019

Sum

Youth unemployment rate

20%

20%

16%

14%

12%

10%

8%

% in employment 6 months after leaving

40%

40%

48%

52%

56%

60%

64%

20,000

20,000

16,000

14,000

12,000

10,000

8,000

100,000

2,000

2,000

1,600

1,400

1,200

1,000

800

10,000

800

800

768

728

672

600

512

4,880

Number minority group unemployed

school leavers
N participants

N in employment 6 months after leaving

4.4

Normative issues frontier analysis

To set target results, it helps to use historical data from various projects to assess which results are
attainable. With a result equation, success rates can be calculated to adjust for changes in
underlying factors that influence the success rates. Another aspect which needs to be judged is
whether better results could be achieved in the future, compared to the past.
In this section, we illustrate the issues of assessing attainable results and setting result targets for
different situations in the future with the method of frontier analysis, also called data envelopment
analysis. First, issues of assessing attainable results are illustrated in the (unlikely) case of many
observations in a comparable situation. Then, issues of applying estimates to new situations are
illustrated.
Even for comparable target groups, in the same region and the same phase of the business cycle,
different projects can achieve different results, for example due to differences in the services
provided or the scale of the project. If there is a large group of projects with similar results, the
result of those projects could be used as an estimate of attainable results in the future. Figure 4.1
illustrates that with an example where the result is plotted against the output (number of training
participants). The majority of the results in this example lie on one line, and the solid line provides
an estimate of attainable results.
The solid line connects the best results at given output levels and reflects the frontier of attainable
results. Projects with poorer results such as bottom left in Figure 4.1 are not taken into account.
Single projects with results far above or below the frontier should be considered as outliers
unique projects that may be difficult to replicate in the future.21
A normative issue even when comparing projects in similar situations (with respect to impact
variables such as target group, region, business cycle), is how to treat isolated clusters of projects
with very good results should these determine the target for the future? In Figure 4.1 this situation
is illustrated with the dashed line. The dashed lines connects higher result levels (around 4.5) given
a number of training participants around 10, ignoring other projects with lower result levels (around
3.2) given the same number of training participants around 10. If the dashed line is applied to target
results in the future, the target levels are in fact set higher than for most projects in the past.
21

40

2020

Stochastic frontier analysis is a method that identifies such outliers and leaves them out of the calculations, but this is
beyond this background paper.

Whether the dashed or the solid line should be used, is a normative issue, based on judgement.
However, the dashed line has the advantage of being based on evidential information and in that
sense helps in setting higher targets.
Figure 4.1

Use of frontier analysis as an aide in the decision whether to set higher targets

In the above example it was assumed that the situation was comparable for all projects. In reality
impact variables such target groups, regions or the phase of the business cycle will have been
different between projects. The result equation discussed in the previous section helps to control for
differences in impact variables. Figure 4.2 illustrates how the result equation can be used to set
targets for new groups (actually, a group with very few observations).
In Figure 4.2, the lower red dashed line connects the three observed results for a new group y.
The results are poorer than for the old group x of Figure 4.1, but the reason might be that the new
group has a lower likelihood of employment with or without the programme. The result equation can
be used to predict the results for the new group y, by replacing the relevant impact variables with
the characteristics of group y. In the example of Figure 4.2 the predicted result levels for group y
are below the observed levels for group x but above the observed level for group y.
Again, it is a normative issue based on judgement whether to set higher result targets for the new
group based on higher predicted values than the observed results and again the frontier analysis
has the advantage of being based on evidential information (in combination with the result
equation).

41

Figure 4.2

4.5

Use of frontier analysis to set targets for a new group y

Setting result targets with limited or no data

In the case there are limited or no data available, one needs to apply practical solutions. In this
section, two practical solutions are discussed:

One solution assumes that historical results are available for the relevant target group in the
situation without the programme (the baseline) although not for the situation with the
programme;

The other solution assumes that no historical results are available for the relevant target group
at all, whether with or without the programme.

Limited availability: historical results for baseline and unit costs but not for the programme
In the first case the result for the target group is known for the situation without intervention, for
example 25% of the older unemployed find employment within one year of becoming unemployed.
Then this no policy result can serve as a starting point (first step) for target setting.
In a second step, one estimates what minimum result is needed to justify the cost of the programme
and this would be the target level. The cost of the programme would be justified if the unit gain is
greater than or at least equal to the unit cost. The unit cost is assumed to be known (see Section
3.6 for the case no data on unit costs are available), for example 5,000 per participant. To
calculate the minimum result needed to justify the cost, one also needs to assign a value to a unit
result, for example 50,000 per job-finder.
The requirement that the unit gain is at least the unit cost, can be expressed in a formula:
[Percentage extra job-finders] * Unit value Unit cost
Or equivalently:
Percentage extra job-finders [Unit cost] / [Unit value].
In the above example, the minimum percentage of extra job-finders to justify the cost of the
programme is therefore 5,000 / 50,000 = 10%.

42

Lastly, the minimum percentage of extra job-finders is added to the baseline result to set a
minimum target percentage, in the above example 25% + 10% = 35%. If the output target would be
10,000 older unemployed participants, then the result target would be 3,500 participants in
employment upon leaving in this example

No historical results available at all, nor unit costs

The previous method uses unit costs. If unit costs are missing, as well as historical results, only
rules of the thumb remain. A rule of the thumb from the author of this background paper is:

75% of the new unemployed find a job the first year;

25% of the long-term unemployed ( 1 year) find a job after their first year.

The above rule of the thumb is based on long experience with Dutch unemployment data and
international meta-studies. However the author has no such rule of the thumb for other target
groups, except for his home country. This is because a reasonable target percentage for other
groups depends on national institutional settings including for example:

Wage cost of older workers;

Equal pay enforcement for minorities;

As an alternative for other target groups besides the long-term unemployed, the unemployment
distribution of the target group can be combined with the above rule of the thumb percentage for
short-term and long-term unemployed. For example if 20% of the older unemployed is short-term
unemployed and 80% is long-term unemployed, then:

The 75% of the target percentage for short-term unemployed is weighted by 20%.

The 25% of the target percentage for long-term unemployed is weighted by 80%.

As a result, the target percentage for the whole group of unemployed older workers is in this
example: target% = 20% * 75% + 80% * 25% = 35%. As in the previous example, if the output
target is 10,000 older unemployed participants, then the result target would be 3,500 participants in
employment upon leaving.
Table 4.6

Application of rule of the thumb to a new target group.

Short-term

Long-term

Total

Distribution

20%

80%

100%

Result %

75%

25%

35%

43

Methodologies for adjusting targets

5.1

Introduction
This chapter briefly discusses methodologies for adjusting output targets and result targets. When
outputs are adjusted by reallocating a given year budget between different target groups, one can
simply adjust the output targets to the most recent developments, in a similar way as the trend
extrapolation discussed in Section 3.2. Only when budgets are reallocated between years
according to the situation of the labour market, a methodology needs to be developed in advance.
For adjusting result targets, the micro-model of Bartik described in Section 4.2 can be used again,
but then focusing on the variables capturing the impact of external factors on the results.

5.2

Adjusting output targets

To reallocate budgets between years, according to the needs and the size of the target group
population, a model needs to be developed that predicts the target group population. Similar to the
unit cost and wage equations discussed in previous chapters, a target group equation can be
formulated to relate the target group population to external factors.
While the author has advised against estimating a macro-economic model for results with annual
data, it is quite different for estimating the size of target groups based on monthly data. Estimating
such a relationship with macro-economic data quickly involves time series aspects.
Macro-economic time series models

Advantages:
-

The required data are usually available, for example from national accounts and PES data;

The model is very suitable for predicting developments over time;

The model can be flexibly specified.

Disadvantages:
-

The limited number of observations allows inclusion of only a limited number of explanatory
variables (relevant factors);

Time series estimations come with a number of pitfalls (which are discussed below);

Basic software such as the Excel add-in is not sufficient to estimate these models.

Needed:
-

Econometric expertise;

Statistical software such as Stata, Eviews, TSP, Gauss.

A set of equations for the size of target group (say, the unemployed in group i) in month m can
be formulated as follows:

[
[ ]
[

[ ]
]

]
[

]
[ ]

[ ]
(
)
(
)

Here, the symbol represents a change compared to the previous period, the values a0, a1, a2, b0
and b1 are coefficients to be estimated by running a regression, and N(0,2) indicates Normally
distributed white noise with variance 2.

44

For the number of unemployed, monthly data is typically available. For GDP, typically quarterly data
is available. To predict the development of a target group population, the above set of equations
needs to be estimated, and the development of GDP needs to be predicted.
The second equation of the above set is a time series equation, because the development of total
unemployment is partly related to the development of total unemployment in the previous month,
which in econometrics is called a lagged variable.
The main pitfalls of time series equations are in general:

A correlation does not indicate causality;

The unit root problem: the coefficient of the lagged variable is roughly 1;

Autocorrelation: the prediction error of one month is correlated with the prediction error of the
previous month;

Multi-collinearity: the estimated values of external variables are mutually correlated;

End-point bias.

These problems and the solutions are briefly discussed below.

Correlation versus causality
In the second equation, the value a2 does not imply that a given change of GDP causes
unemployment to change by a2 times the GDP change. The causality might equally well be the
other way round: if there are more unemployed, employment is lower, and fewer employed produce
lower output (GDP). The idea behind using predictions of GDP to predict the development of
unemployment is that perhaps GDP is easier to predict based on the demand of consumers,
imports, exports, and investments. The impact of all these factors on unemployment is implicit
through their impact on GDP, if GDP and unemployment are strongly correlated. In fact, the target
group equation can be used to predict unemployment regardless of the actual causality. The
causality is more of a concern when predicting the results of an intervention, because then naturally
the causal relation from the intervention to the result is of primary interest.
Unit root and autocorrelation
The practical problem of a unit root is that in that case the last observation is the best predictor. The
standard assumption of independent prediction errors is not applicable in the case of a unit root,
and even the standard assumption of normally distributed prediction errors is no longer applicable.
The practical problem of autocorrelation is that the prediction error of the last observation carries
over to the prediction for the subsequent period; autocorrelation is a milder version of the unit root
problem.
For both problems, the solution is to formulate an equation for the change of the variable that is to
be explained, instead of the variable to be explained itself (on the left-hand side of the equation). In
the above set of equations, the equation is already formulated for the change of unemployment,
exactly to solve the potential unit root and autocorrelation problems.
Multi-collinearity
The practical problem with multi-collinearity is that a strong mutual correlation between explanatory
variables (on the right-hand side of the equation) causes the weights a1, a2, to be extremely
large. While this may result in a perfect fit of the past observations, it also is likely to result in
extremely large (and implausible) predictions. The solution to this problem is to use fewer
explanatory variables. The above set of equations already has few explanatory variables.

45

Multi-collinearity can be identified through the correlation matrix of the estimates of a1, a2, , but
practically speaking, one can already conclude multi-collinearity if the estimates are extremely
large.
Endpoint bias
The endpoint bias refers to the fact that in a single time series, the first and the last values have a
disproportionate weight in the estimated of time series coefficients. To illustrate the endpoint bias,
Figure 5.1 shows a series with two extreme values, at the 25th in the middle of the time series and
the 50th data point at the end of the time series. If a linear line is estimated based on all 50 data
points, then the estimated 51st data point would be 7.6 instead of 8.1 if both extreme values were
ignored, an error of 6%. If the 50th data point would be ignored but the extreme value in the middle
of the time series is included in the estimation, then the estimated 51st data point would be 8.0, an
error of only 1% compared to 8.1.
To correct for the endpoint bias, one judges by sight and leaves out the first or the last observation
if necessary. In statistical software, the so-called Hodrick-Prescott filter automatically corrects for an
endpoint bias. Correcting for the endpoint bias can be important in short time series. However if the
time series is jointly estimated for a large number of regions, it is unlikely that the endpoint bias is
the same in all regions, and the issue of the endpoint bias can be ignored.
Figure 5.1

46

Illustration of the endpoint bias

5.3

Application of time series models in setting output targets

Examples of estimating macroeconomic time series models are not provided as this paper is about
target setting. However, the example below illustrates how the results of a macroeconomic time
series can be used in setting output targets.
Suppose that the following equations have been estimated:

[ ]

[ ]

]
[ ]

]
[

[ ]

Unemployment tends to decrease if the economy (GDP) grows and its estimated coefficient of -0.5
reflects this. But unemployment also has its own dynamics, perhaps because it takes time to find a
new job. The estimate of 0.3 for the effect of a change in the number of unemployed in the previous
month reflects this. The unemployment rate tends to decrease with educational level, but also tends
to be more sensitive to the business cycle as reflected by the coefficient of +0.5 in the second
equation. Among the low educated unemployed a relatively large core may exist that remains
unemployed throughout the business cycle, stabilizing the unemployment rate among this group
at a relatively high minimum level. Thus if the share of high educated unemployed increases, the
share of low educated unemployed decreases. The coefficient for medium educated unemployed
may be insignificant and is set at 0 in this example, implying that the share of medium educated
unemployed is constant.
Table 5.1 provides the base data for setting output targets using the above set of equations. The
four quarterly GDP growth rates for 2014 add up to 1.15%, a realistic GDP growth for 2014 as a
whole. The initial unemployment forecast for January 2014 is also provided.
Table 5.1

Base data for setting output targets with macroeconomic time series data
2014

gdp
u

M1

M2

M3

M4

M5

M6

M7

M8

M9

M10

M11

M12

0.25%

0.25%

0.25%

0.50%

0.50%

0.50%

0.30%

0.30%

0.30%

0.10%

0.10%

0.10%

-1.00%

%[high]

10%

%[medium]

50%

%[low]
Unemployed

40%
100,000

High

10,000

Medium

50,000

Low

40,000

To apply the above model, first the change in the number of unemployed needs to be forecasted,
based on the change in the preceding month and the quarterly GDP growth. For example for
February 2014 this gives:
u[February 2014]

= -0.5* gdp[First quarter 2014] + 0.3* u[January 2014]

= -0.5*0.25% + 0.3*(-1%) = -0.43%.

u[March 2014]

= -0.5* gdp[First quarter 2014] + 0.3* u[February 2014]

= -0.5*0.25% + 0.3*(-0.43%) = -0.25%.

47

The number of unemployed in February 2014 is therefore the 100,000 of January minus the
predicted 0.43% decline or equivalently 99,575 persons. The number of unemployed in March 2014
is the 99,575 of February minus the predicted 0.25% decline or equivalently 99,324 persons.
Continuing in this way for the other months of 2014 gives the result in Table 5.2 below.
Table 5.2

Base data for setting output targets with predicted changes in unemployment rates
2014

gdp
u

M1

M2

M3

M4

M5

M6

M7

M8

M9

M10

M11

M12

0.25%

0.25%

0.25%

0.50%

0.50%

0.50%

0.30%

0.30%

0.30%

0.10%

0.10%

0.10%

-1.00%

-0.43%

-0.25%

-0.33%

-0.35%

-0.35%

-0.26%

-0.23%

-0.22%

-0.12%

-0.08%

-0.08%

99,575

99,324

99,000

98,656

98,306

98,054

97,832

97,618

97,506

97,423

97,350

M11

M12

%[high]

10%

%[medium]

50%

%[low]

40%

Unemployed

100,000

High

10,000

Medium

50,000

Low

40,000

The forecasts of u in Table 5.2 are in turn used to forecast the shares of high, medium and low
educated unemployed. For February 2014, the predicted change in the share of high educated
unemployed is:
Dshare[high] = +0.5* -0.43% = -0.21%
This results in a share of high educated unemployed of 10% - 0.21% = 9.79% (in Table 5.3 the
rounded figure 9.8% is shown), and applying this percentage to the total number of unemployed in
February 2014 (which is 99,575) gives 9,746 high educated unemployed. Continuing in this way for
the other educational levels and the other months of 2014 gives the table below.
Table 5.3

Base data for setting targets broken down by educational level

2014
M1

gdp

M2

M3

M4

M5

M6

M7

M8

M9

M10

0.25%

0.25%

0.25%

0.50%

0.50%

0.50%

0.30%

0.30%

0.30%

0.10%

0.10%

0.10%

-1.00%

-0.43%

-0.25%

-0.33%

-0.35%

-0.35%

-0.26%

-0.23%

-0.22%

-0.12%

-0.08%

-0.08%

%[high]

10%

9.8%

9.7%

9.5%

9.3%

9.1%

9.0%

8.9%

8.8%

8.7%

8.7%

8.7%

%[medium]

50%

50%

50%

50%

50%

50%

50%

50%

50%

50%

50%

50%

%[low]

40%

40.2%

40.3%

40.5%

40.7%

40.9%

41.0%

41.1%

41.2%

41.3%

41.3%

41.3%

100,000

99,575

99,324

99,000

98,656

98,306

98,054

97,832

97,618

97,506

97,423

97,350

Unemployed
High

10,000

9,746

9,596

9,403

9,199

8,992

8,844

8,713

8,587

8,521

8,473

8,430

Medium

50,000

49,788

49,662

49,500

49,328

49,153

49,027

48,916

48,809

48,753

48,712

48,675

Low

40,000

40,042

40,066

40,097

40,129

40,161

40,183

40,203

40,222

40,232

40,239

40,245

The average numbers of unemployed in 2014 can then be used to set output targets for the
subgroups of unemployed, for example assume:

a total budget is available for 10,000 participants,

costs per participant are equal for all groups and

subgroups are targeted proportionally to the population.

Table 5.4 gives the results. In the second column, the average numbers of unemployed are
calculated by averaging the numbers of January up to December. In the third column, the shares of
high, medium and low unemployed are calculated from the second column. In the fourth and last

48

column, the number of participants is targeted by applying the shares to the total number of
participants for which budget is available.
Table 5.4

Initial target setting

Educational level

Average number

Share

Participants

9,042

9.2%

919

Medium

49,194

50.0%

5,000

Low

40,152

40.8%

4,081

Total

98,387

100%

10,000

(=output target)
High

Now suppose that in June, the forecasts for the second half have been adjusted significantly
downward. The new situation is then given as in Table 5.5.

Table 5.5

Base data for adjusting output targets with macroeconomic time series data
2014
M1

M2

M3

M4

M5

M6

M7

M8

M9

M10

M11

M12

0.25%

0.25%

0.25%

0.50%

0.50%

0.50%

-0.70%

-0.70%

-0.70%

-0.90%

-0.90%

-0.90%

-1.00%

-0.43%

-0.25%

-0.33%

-0.35%

-0.35%

%[high]

10%

9.8%

9.7%

9.5%

9.3%

9.1%

%[medium]

50%

50%

50%

50%

50%

50%

%[low]

40%

40.2%

40.3%

40.5%

40.7%

40.9%

gdp
u

Unemployed

100,000

99,575

99,324

99,000

98,656

98,306

High

10,000

9,746

9,596

9,403

9,199

8,992

Medium

50,000

49,788

49,662

49,500

49,328

49,153

Low

40,000

40,042

40,066

40,097

40,129

40,161

Repeating the same calculations as for Table 5.2 and Table 5.3 results in Table 5.6 where due to
the downward adjusted GDP forecasts, the unemployment level and the share of high educated
unemployed people increase.
Table 5.6

Data for adjusting targets with unemployment forecasts and educational breakdown
2014
M1

gdp

M2

M3

M4

M5

M6

M7

M8

M9

M10

M11

M12

0.25%

0.25%

0.25%

0.50%

0.50%

0.50%

-0.70%

-0.70%

-0.70%

-0.90%

-0.90%

-0.90%

-1.00%

-0.43%

-0.25%

-0.33%

-0.35%

-0.35%

0.25%

0.42%

0.48%

0.59%

0.63%

0.64%

%[high]

10%

9.8%

9.7%

9.5%

9.3%

9.1%

9.3%

9.5%

9.7%

10.0%

10.3%

10.6%

%[medium]

50%

50%

50%

50%

50%

50%

50.0%

50.0%

50.0%

50.0%

50.0%

50.0%

%[low]

40%

40.2%

40.3%

40.5%

40.7%

40.9%

40.7%

40.5%

40.3%

40.0%

39.7%

39.4%

Unemployed

100,000

99,575

99,324

99,000

98,656

98,306

98,547

98,964

99,436

100,026

100,654

101,297

High

10,000

9,746

9,596

9,403

9,199

8,992

9,133

9,381

9,663

10,017

10,396

10,786

Medium

50,000

49,788

49,662

49,500

49,328

49,153

49,273

49,482

49,718

50,013

50,327

50,648

Low

40,000

40,042

40,066

40,097

40,129

40,161

40,141

40,101

40,055

39,996

39,931

39,863

Suppose that in the first 6 months 5,000 unemployed have participated in the intervention, of which
460 were high educated; 2,500 were medium educated and 2,040 were low educated. Then the

49

analysis for M7-M12 in the new situation can be repeated as was done above for M1-M12 in the old
situation. This gives the result of Table 5.7.
In June 2014, the participants of the first six months of 2014 are given. The remaining 5,000
participants for which budget is left for the second half of 2014, are allocated proportionally to the
new forecasts of the unemployed in the second half of 2014. Lastly, the adjusted target for the
whole of 2014 is the sum of the participants who already have entered the programme in the first
half of 2014, plus the adjusted target for the remainder of 2014.
Table 5.7

Adjusted targets based on new mid-2014 GDP forecasts

M1-M6

M7-M12

M1-M12

Participants

Unemployed

Shares

Participants

Adjusted target

460

9,896

9.9%

496

956

Medium

2,500

49,910

50.0%

2,500

5,000

Low

2,040

40,015

40.1%

2,004

4,044

Total

5,000

99,821

100%

5,000

10,000

High

Comparing the targets of Table 5.4 and Table 5.7 shows that in this case, the adjustment of output
targets would be small. In the course of years instead of months, adjustments may be larger, but
the underlying calculations are the same as illustrated above.

5.4

Adjusting result targets

To adjust result targets, the same model as the Bartik model of Section 4.2 can be used, but now
the values of the macroeconomic variables are used to adjust result targets. Table 5.8 illustrates
the adjustment of a result target (the percentage of unemployed participants finding employment
within 6 months after leaving) to the business cycle.22 In this example, the chance to find
employment is estimated to decrease by 1.1 percentage point for every per cent point of
unemployment rate. In this example, every per cent point of unemployed corresponds to 20,000
persons. Over the period 2014-2020 the implied numbers of unemployed in Table 5.8 add up to 1
million. If the ESF budget covers 200,000 participants over this seven-year period, per year one fifth
of the unemployed can be covered by the ESF Operational Programme. For 2014, this means the
output would be targeted at 32,000 participants (one fifth of the population of 160,000 unemployed).
Table 5.8

Illustration of adjusting result targets

2014

2015

2016

2017

2018

2019

2020

Unemployment rate

8%

10%

9%

8%

6%

5%

4%

Share of job finders

76.7%

74.5%

75.6%

76.7%

78.9%

80.0%

81.1%

Number of unemployed

160,000

200,000

180,000

160,000

120,000

100,000

80,000

1,000,000

Number of participants

32,000

40,000

36,000

32,000

24,000

20,000

16,000

200,000

Targeted job-finders

24,544

29,800

27,216

24,544

18,936

16,000

12,976

154,016

The question is, how to adjust the result targets if the chance to find employment decreases by 1.1
percentage point for every per cent point of unemployment rate? One first needs to set a target
norm, for example at a 5% unemployment rate, 80% of the unemployed participants should find a

22

50

This example is slightly modified from Ecorys (2011), Performance targets for ESF Operational Programmes.
http://www.coprbm.eu/?q=node/494

Sum

job within a six months after leaving an ESF-supported project. If that is the norm, then the adjusted
share of job-finders for higher and lower unemployment rates can be calculated as follows:

[ ]

For example for 2020 the predicted unemployment rate is 4% and the above formula gives a share
of job-finders of 81.1% and in 2018 the predicted unemployment rate is 6% and the above formula
gives a share of job-finders of 78.9%.
Once the share of job-finders is calculated, the result target is easy to calculate:

In 2019, this formula gives that 80% of the 20,000 participants should find a job, or equivalently
16,000 persons. Repeating this for each year and taking the sum over the period 2014-2020 finally
gives the result target of 154,016 participants who should find a job, which is naturally rounded to
154,000. Of course, since the ESF programming period ends in 2020 and the result target is that
the unemployed find a job within a year, the result target actually applies for the period 2014-2021,
since the last unemployed of 2020 that find a job, will likely find that job in 2021.

51

List of common indicators

In this annex a list of common indicators is presented. The list of common indicators can be found
in the annex to the draft ESF regulation.23 Further definitions can be found in the 2013 draft ESF
monitoring and evaluation guidance.24
(1) Common output indicators on participants
Participants refer to persons benefiting directly from an ESF investment and who can be identified
and asked for their characteristics, and for whom specific expenditure is earmarked. Other
beneficiaries should not be counted as participants.

unemployed, including long-term unemployed

long-term unemployed

inactive

inactive, not in education or training

employed, including self-employed

below 25 years

above 54 years

with primary (ISCED 1) or lower secondary education (ISCED 2)

with upper secondary (ISCED 3) or post-secondary education (ISCED 4)

with tertiary education (ISCED 5 to 8)

migrants, people with a foreign background, minorities (including marginalised communities

such as the Roma)

disabled

other disadvantaged

The total number of participants is calculated automatically on the basis of the output indicators.
These data on participants entering an ESF supported operation are to be provided in the annual
implementation reports. All data are to be broken down by gender.
(2) Common output indicators for entities

number of projects fully or partially implemented by social partners or non-governmental

organisations

number of projects targeting public administrations or public services

number of micro, small and medium-sized enterprises supported

These data are to be provided in the annual implementation reports.

(3) Common immediate result indicators on participants

inactive participants newly engaged in job searching upon leaving

participants in education/training upon leaving

participants gaining a qualification upon leaving

participants in employment upon leaving

These data are to be provided in the annual implementation reports. All data are to be broken down
by gender.

23
24

http://ec.europa.eu/regional_policy/sources/docoffic/official/regulation/pdf/2014/proposals/regulation/esf/esf_proposal_en.pdf
European Commission, Programming Period 2014-2020, Monitoring and Evaluation of European Cohesion Policy,
European Social Fund, Draft Guidance document, 9 January 2013,
http://ec.europa.eu/social/main.jsp?catId=701&langId=en

53

(4) Common longer-term result indicators on participants

participants in employment 6 months after leaving

participants in self-employment 6 months after leaving

participants with an improved labour market situation 6 months after leaving

These data are to be provided in the annual implementation reports. They are to be collected based
on a representative sample of participants within each priority axis. Internal validity of the sample
should be ensured in such a way that the data can be generalised at the level of priority axis. All
data are to be broken down by gender.

54

Target optimization with Excel

In Microsoft Excel, a standard built-in add-in can be of great help in optimizing the result targets
subject to economic (and perhaps political) constraints. This annex illustrates how to install this
add-in in Excel 2010 and how to use it for the first exercise. This add-in is called Solver and solves
linear programming problems. This is very useful for maximizing result targets under budget,
capacity and minimum investment constraints, because budgeting is essentially a linear
programming problem.
The Excel add-in can be found behind Options under the tab File, as shown in the screen-print
below.

In the Excel Options screen, go to the option Add-Ins. Another screen will then pop up as
illustrated below. In this popup-screen, choose Solver Add-in and then click on Go at the
bottom, and then OK.

55

Finally, under the tab Data the Solver add-in is now available.

56

To use the Solver add-in, the unit costs, maximum budget and other constraints need to be
specified, as well as the contribution of each alternative to the target. For Exercise 1, the screen
print below shows how this information can be prepared. Besides the policy alternatives
(counselling, mobility scheme, re-training in this example), the contribution of the no-policy
alternative to the result needs to be specified (50% of the non-participants are in employment in this
example). The cells for the numbers of participants are left blank for the moment, these will be filled
by the Solver add-in in such a way that the result (persons in employment) is maximized.
The budget constraint states that the sum of participants weighted with their unit costs must be less
than 10m euro in the example of exercise 1:
1,000 * (counselling) + 4,000 * (mobility scheme) + 16,000 * (re-training) 10,000,000.
As well as the unit costs for the alternative, the zero unit costs for nonparticipants need to be filled
in.

The capacity constraint for counselling is that one times the number of participants in the
counselling alternative must be less than 5,000, and so on for the other alternatives. Finally, it
needs to be specified that the numbers of participants and non-participants add up to the total,
10,000 in the example of exercise 1:
1*(nonparticipants) + 1*(counselling) + 1*(mobility scheme) + 1*(re-training) = 10,000.

57

After the ground work is done, the result needs to be specified. The number of persons in
employment is:
50%*(nonparticipants) + 55%*(counselling) + 60%*(mobility scheme) + 90%*(re-training).
This formula needs to be filled in in cell B6 in the sheet of the screen print below, where the cells
B3:E3 refer to the expected result per participant, and the cells B4:E4 to the numbers of
participants that will be solved by the add-in.

The formula for the budget constraint needs to be filled in in cell B16 in the sheet of the screen print
below. Here, the cells B9:E9 refer to the unit costs and the cells B4:E4 to the numbers of
participants.

58

The same is done for the capacity restraint of counselling as in the screen print below. Here, cell
B10:E10 refer to the specification that the capacity limit applies to the numbers of participants in
counselling (value 1) and not to the other alternatives (value 0).

59

After filling the similar formulas for the capacity constraints of the mobility scheme and re-training in
cells B18 and B19, the formula for the count restriction is filled in in cell B20 in the screen-print
below.

In cells C16:C20 the right hand side values of cells G9:G13 are copied, but this is purely cosmetic.
After all the formulas have been specified as above, the Solver add-in can be applied to find the
numbers of participants that maximize the result of persons in employment. For this, one needs to
go to the tab Data and then click on the Solver add in at the top right side of the screen, as
illustrated in the screen below.

60

In the pop-up screen of the Solver add-in, the user can specify which cells contain the required
information:

Set objective: cell B6 for the formulas for the result target (persons in employment);

Indicate that the result target needs to be maximized (choose option Max);

By changing variable cells: B4:E4 (numbers of participants).

61

To add a constraint, one needs to open another popup screen by clicking on the Add button. The
budget and capacity constraints are all subject to maximum limits and can be grouped together as
illustrated in the pop-up below. Reference needs to be made to the cells with the formulas
(B16:B20), not the cells with the numbers that are used in the formulas (B9:E12).

The count restriction is an equality and this constraint needs to be added as well.

62

After specifying the constraints, the Add Constraint window can be exited by clicking on Cancel.

The Solver window then should look as illustrated below.

63

Lastly, a Solving Method needs to be specified; in this case the Simplex Linear Programming
method is most appropriate, because the number of persons in employment is a linear function of
the numbers of participants.

Clicking on the Solve button makes the Solver add-in calculate the numbers of participants
(=output target) and the numbers of persons in employment (= result target).
After clicking on Solve, a message should pop up to indicate that a solution is found which
satisfies all the constraints and one can exit this message by clicking on OK.

64

The results should look as in the screen print below.

According to the linear programming solver, it is optimal to target 5,000 participants in counselling
(exhausting the capacity constraint for counselling) and 313 participants in re-training. Together
with the 50% of nonparticipants in training who are in employment, this results in 5,375 persons in
employment.

65

The solver identifies the optimal allocation of participants, but still one final check is needed: is
benefit per additional job-finder sufficient to justify the cost per additional participant? Without any
intervention, 50% of all 10,000 persons threatened with job loss would have found a job in this
example, resulting in 5,000. Hence the additional number of job-finders attributable to the mix of
alternatives in this example is (5,375 5,000) = 375 persons.
The cost per additional job-finders is therefore the budget of 10m euro divided by 375, which is
26,667. In this exercise, each additional job-finder is valued at 50,000, which justifies the cost of
26,667.
Finally, the output target is calculated as the sum of participants over all alternatives, namely:

Output target = 5,000 + 313 = 5,313.

The result target is calculated as the sum of participants weighted with the result% per participant:

66

Result target = 5,000*55% + 313*90% = 3,032.

P.O. Box 4175

3006 AD Rotterdam
The Netherlands

Watermanweg 44
3067 GG Rotterdam
The Netherlands

T +31 (0)10 453 88 00

F +31 (0)10 453 07 68
E [email protected]

W www.ecorys.nl

Sound analysis, inspiring ideas

BELGIUM BULGARIA HUNGARY INDIA THE NETHERLANDS POLAND RUSSIAN FEDERATION SOUTH AFRICA SPAIN TURKEY UNITED KINGDOM