Who Benefits from Promoting Small Enterprises? Some Empirical Evidence from Ethiopia

World Development Vol. 38, No. 4, pp. 523–540, 2010 Ó 2009 Elsevier Ltd. All rights reserved. 0305-750X/$ - see front matter www.elsevier.com/locate/worlddev

doi:10.1016/j.worlddev.2009.10.007

Who Benefits from Promoting Small Enterprises? Some Empirical Evidence from Ethiopia BOB RIJKERS University of Oxford, UK The World Bank, USA CATERINA RUGGERI LADERCHI The World Bank, USA

and FRANCIS TEAL * University of Oxford, UK Summary. — The Addis Ababa Integrated Housing Development Program (AAIHDP) aims to tackle the housing shortage and unemployment that prevail in Addis Ababa by deploying and supporting small enterprises to construct low-cost housing using technologies novel for Ethiopia. The motivation for such support is predicated on the view that small firms create more jobs per unit of investment by virtue of being more labor intensive and that the jobs so created are concentrated among the low-skilled and hence the poor. To assess whether the program has succeeded in biasing technology adoption in favor of labor and thereby contributed to poverty reduction, the impact of the program on technology usage, labor intensity, and earnings is investigated using a unique matched workers-firms dataset, the Addis Ababa Construction Enterprise Survey (AACES), collected specifically for the purpose of analyzing the impact of the program. We find that program firms do not adopt different technologies and are not more labor intensive than nonprogram firms. There is an earnings premium for program participants, who tend to be relatively well educated, which is heterogeneous and highest for those at the bottom of the earnings distribution. Ó 2009 Elsevier Ltd. All rights reserved. Key words — Africa, Active labor market programs, Ethiopia, Productivity, MSEs, Wages

This paper provides an analysis of the Addis Ababa Integrated Housing Program (AAIHDP), which is an active labor market program that attempts to tackle the severe housing shortages and high unemployment that prevail in Addis Ababa by deploying and supporting labor-intensive MSEs to construct low-cost condominium housing using technologies

1. INTRODUCTION Whether or not programs which promote micro and small enterprises (MSEs) can stimulate job creation and contribute to poverty reduction in developing countries is an important question. The motivation for MSE support is often predicated on the view that small firms create more jobs per unit of investment by virtue of being more labor intensive and that the jobs so created are concentrated among the low-skilled and hence the poor. In spite of the policy prominence of such support programs, the empirical evidence for these propositions is weak. In a recent review of the literature Betcherman, Olivas, and Dar (2004, p. 51) conclude:

* We would like to thank the staff of the Ethiopian Economic Association, in particular Ato Kibre Moges, Ato Daniel Aklilu and Ato Assefa Adamassie and the supervisors and enumerators of the Addis Ababa Construction Enterprises Survey. Discussions with Dr. Wubshet Berhanu, City Manager, Addis Ababa City Government, and Woz. Tsedale Mamo, Manager, Addis Ababa Housing Development Project Office and other staff at the Addis Ababa Housing Development Project Office have been most helpful to frame the analysis of this study. We also owe gratitude to Stefan Dercon, Paul Collier and Justin Sandefur at Oxford University, as well as Emily Kallaur, Rumana Huque, Peter Lanjouw and Louise Fox at the World Bank. Seminar participants at a CSAE seminar at Oxford University, the CSAE conference on economic development and the IZA/ World Bank conference on employment and development provided useful feedback. We also benefitted from the comments of three anonymous referees. Bob Rijkers gratefully acknowledges financial support from the Prins Bernhard Cultuurfonds and the Herbert and Ilse Frankel Memorial Fund at Oriel College, Oxford University. All errors are our own. The views expressed in this paper are those of the authors and do not necessarily reflect those of the World Bank, its Board of Executive Directors, or the countries they represent. Final revision accepted: August 24, 2009.

The evaluation literature on the labor market impacts of ALMPs [Active Labor Market Programs] is thinnest in the case of micro-enterprise development and self-employment assistance programs. There are relatively few studies and of those that do exist, many are concerned with the program’s effect on business development rather than on the future employment and earnings of participants.

The sparse available empirical evidence is entirely based on the experiences of developed and transition economies and suggests that micro-enterprise development programs typically have a positive impact on the employment prospects of participants, while their impact on earnings is mixed (see e.g., Fretwell, Benus, & O’Leary, 1999; Dar & Tzannatos, 1999; and Betcherman et al., 2004). 523

524

WORLD DEVELOPMENT

novel for Ethiopia. To assess whether the program has succeeded in biasing technology adoption in favor of labor, and thereby contributed to poverty reduction, the impact of the program on technology usage, labor intensity, and earnings of participants is investigated using a unique matched workers-firms dataset, the Addis Ababa Construction Enterprise Survey (AACES), which is representative of all registered housing construction firms in Addis, and was designed specifically for the purpose of analyzing the impact of the program. 1 The paper is organized as follows. The next section describes the program and its rationale. The data collected to analyze it are presented in Section 3, which also presents descriptive statistics on the characteristics of program and nonprogram firms. Section 4 assesses the impact of the program on technology adoption. Section 5 describes the differences between program participants and nonparticipants and assesses the program’s impact on earnings. Section 6 discusses our results. A final section concludes. 2. THE ADDIS ABABA INTEGRATED HOUSING DEVELOPMENT PROGRAM (AAIHDP) Severe housing shortages and high unemployment constitute critical development challenges in Addis Ababa. The unemployment rate is 42% and estimates of the housing shortage in Addis vary between 250,000 and 300,000 housing units (HDPO, 2004). In addition, at least a third of the existing 640,000 housing units are dilapidated and the shortage is increasing by approximately 40,000 units each year (Construction Ahead, 2005). The housing shortage has arisen because housing supply has not kept up with increasing housing demand, primarily due to the shortages of key construction inputs. The demand for cement, for example, is almost 60% higher than the available supply (World Bank, 2007). 2 By creating and supporting MSEs the AAIHDP aims to simultaneously stimulate housing construction and job creation. The specific objectives of the programme include ‘promotion of micro and small-scale enterprises, which can absorb more labor force and operate at a lower overhead cost’ as well as ‘promotion of cost efficient housing construction technology’ (HDPO, 2004, p. 1). To achieve these objectives, the AAIHDP aims to construct 192,500 houses, generate 80,000 job opportunities, support 1,300 existing MSEs and create another 1,000 new ones, over a 5-year period. To construct housing affordable to low-income dwellers, the AAIHDP has designed a production process that deviates from the one conventionally used in the construction sector. The lowcost aspect of the program consists in building a less luxurious, homogeneous type of housing using novel low-cost construction technologies designed to minimize input requirements, such as pre-cast beams and ribslabs, fixed-price contracts and standardized production procedures permitting greater specialization. To implement this production process, the AAIHDP intervenes in the construction sector by both creating new SMEs and providing support for firms in the program. Participation in the program is conditional on passing a test. Anybody who has either graduated from a Technical and Vocational Education and Training (TVET) college or can show proof of having experience in the construction sector may take a test to participate in the program. After attending an orientation session, successful candidates can establish an enterprise, either by themselves or together with other successful applicants. They can choose to form a single proprietorship, cooperative, or a trade association (share-company, plc or partnership). Individuals who fail the test are allowed to

resit the test at a later date and may attempt to upgrade their skills by joining successful candidates in the project as apprentices. Once firms are formed, they can register their interest in AAIHDP construction work with the Program Office, which is in charge of implementing the program. Only micro and small firms, that is firms employing fewer than 50 people and with a capital stock worth less than 5,00,000 Birr (approximately 55,000 USD), were allowed to compete for AAIHDP jobs. Consequently, all firms formed by the program were small. Existing firms were not allowed to compete for AAIHDP contracts, unless the capacity of the newly formed program MSEs did not suffice. In theory, incumbent firms could attempt to join the program by having their employees take the test and regroup as a “new” firm. Qualitative evidence suggests that this is not very common (World Bank, 2007). In addition, when the managers of nonparticipating firms surveyed by the AACES were asked why they had not attempted to participate in the AAIHDP, 54% of the firm managers claimed they were not aware of the program, 21% claimed participation would not have been beneficial on balance and the rest either claimed they did not fulfill the criteria or had no opportunity to register. The AAIHDP does not create firms which can execute foundational and structural works. Instead, it hires the existing firms for these tasks. For the purposes of this paper, firms which can execute such tasks are referred to as contractors, while firms which cannot execute are referred to as constructors. The latter category consists predominantly of MSEs and is consequently of focal interest. 3 The AAIHDP provides wide-ranging support to firms participating in the program by (i) providing and, in certain cases, subsidizing a place to work, (ii) facilitating access to credit, (iii) providing training and access to inputs (on credit), (iv) subsidizing machinery for firms producing rebars (reinforcement bars) or hollow blocks, (v) providing training to firms engaging in pre-cast beam and hollow block production, and (vi) awarding contracts to program firms. Not all newly created SMEs are awarded contracts, but AAIHDP contracts are almost exclusively awarded to firms created by the AAIHDP (except for contracts assigned to contractors). By providing these different types of support the program alters the factor prices different firms face, which in turn affect technology adoption and factor choices. The program does not affect technology adoption or factor usage directly. The modeling challenge is to identify the impact of the AAIHDP interventions on technology and factor choices, as well as earnings, and separate the effect of the program from those differences between participants and nonparticipants which are not due to participation. Given the magnitude of the housing shortage and the existence of very severe input constraints, it is almost certain that the inputs now taken up by the program would have been used by other firms to construct housing had the program not been introduced (World Bank, 2007); the question is not whether to build housing, but how. Housing construction by the “conventional” private sector, the existing modus operandi, is likely to reflect the production process that would have prevailed in the absence of the AAIHDP intervention. To assess the impact of the AAIHDP on technology and factor choices we therefore use the choices of nonprogram firms as the counterfactual, drawing a distinction between constructors and contractors. To assess program impact on earnings, we compare the earnings of workers in program firms with those of workers in nonprogram firms. Since the AAIHDP explicitly targets those with skills or experience in the construction sector, workers in nonprogram firms constitute a relevant comparison group.

WHO BENEFITS FROM PROMOTING SMALL ENTERPRISES? SOME EMPIRICAL EVIDENCE FROM ETHIOPIA

Of course, had the program not been implemented, those currently employed in program firms could instead have been operating as casual construction workers rather than being employed in nonprogram firms. 4 Regrettably, reliable data on such casual construction workers are not available and it is thus not possible to convincingly compare program participants with casual workers. At this juncture, it is also important to note that our unit of analysis is the firm and not the construction sector as a whole. While we can assess the impact of the program on technology and factor choices of individual firms, the factor demands and productivity of the housing construction sector as a whole may also be affected by the mix between constructors and contractors in the construction process, if these firms differ in their labor intensity and/or adopt different technologies. Unfortunately, our data do not enable us to rigorously assess whether, and, if so, to what extent, such substitution effects are occurring. These effects and potential spillover effects are discussed in Section 6. 3. THE ADDIS ABABA CONSTRUCTION ENTERPRISE SURVEY (AACES) The AACES is a survey of matched firms and workers in the construction sector in Addis Ababa, designed specifically for the purpose of analyzing the impact of the AAIHDP on labor market outcomes and firms’ technology and factor choices. For a detailed description of the data and the sampling design, the reader is referred to World Bank (2007). Here, we focus on the most important features of the survey. The AACES was conducted by the Ethiopian Economic Association in collaboration with the World Bank in December 2006 and early January 2007, and covered 240 firms and 971 workers, 241 of whom were casual workers. The workers’ data contain detailed information on workers’ earnings, their employment history, experience, skills, educational background, program participation, job satisfaction, motivation for choosing their current activity, and on a number of socio-demographic char-

525

acteristics including household characteristics and parental background. 5 Program firms were sampled from the AAIHDP Program Office List of Program Participants. For nonprogram firms, the registry of firms of the Addis Ababa City Administration, which contains information on all construction firms in Addis, constituted the sample frame. The AACES is thus representative of all housing construction firms in Addis. The sample proportions of firms engaging in different activities approximate the corresponding population shares. Contractors, however, were oversampled since there are relatively fewer of them. Up to four workers per firm were interviewed, stratified by occupational category in order to ensure occupational heterogeneity in our sample. The sample of workers consequently contains disproportionately many workers from higher occupational echelons. In addition, women, particularly those not employed as casual workers, were oversampled in order to document their disadvantaged position in the labor market. 6 The worker data will be further considered in Section 5; here we focus on the firm data. The sample of firms contains data on 103 nonprogram constructors, 92 program constructors, 17 nonprogram contractors, and 28 program contractors. A program firm is defined as a firm which fulfills at least one of three criteria; (i) having been created by the program, (ii) having received support from the program, or (iii) having worked for the program. 7 The firm-level data provide information on a rich set of firm characteristics, including their activities, age, size, capital stock, inputs, outputs, expenditures, revenues, organizational and occupational structure, program participation and receipt of program support, access to finance and inputs, skilled personnel, constraints, expectations, their labor force and wages. In addition, the data contain information on the volume and total costs of inputs and outputs, enabling us to compute firm-specific input and output prices, which provide natural instruments for factor choices. Table 1 presents descriptive statistics on program firms and nonprogram firms. On average, program firms employ more workers, have more capital, use more inputs, and have a better educated workforce. In Graph 1, the amount of capital per

Table 1. Descriptive statistics firms. Nonprogram constructor

Program constructor

Nonprogram contractor

Program contractor

Mean

Sd

Mean

Sd

Mean

Sd

Mean

Sd

Firm age

3.78

4.33

2.27

1.80

6.95

4.90

6.65

6.00

Revenue Revenue (log)

11.56

1.39

11.94

1.26

13.92

1.59

14.30

1.59

Factors Workers (log) Capital (log) Inputs(log) Capital per worker (log)

1.86 9.97 10.94 8.11

0.68 2.26 1.64 2.14

2.56 10.71 11.37 8.15

0.62 1.51 1.65 1.46

3.09 12.39 13.30 9.30

1.48 2.06 1.44 1.99

3.33 12.97 13.95 9.65

1.37 1.80 1.48 1.87

Labor intensity Inputs per worker(log) Education of the workforce (years)

9.09 7.65

1.62 2.86

8.81 9.51

1.50 2.76

10.20 8.77

1.33 2.92

10.62 9.27

1.25 2.75

Characteristics of the workforce Age of the workforce Labor force at start-up (log)

29.68 1.49

7.70 0.87

28.52 2.40

6.49 0.84

27.37 1.92

6.68 1.60

26.96 2.83

6.36 1.25

Price deflators Output price deflator Input price deflator Observations

1.11 1.01 78

0.46 0.45

0.97 1.03 70

0.36 0.40

1.06 0.93 9

0.41 0.41

1.08 0.91 13

0.34 0.33

Note: All monetary measures in Ethiopian Birr (ETB). 1 USD is worth approximately 9.1 Ethiopian Birr (ETB).

WORLD DEVELOPMENT

10

12

Table 2. Explaining capital intensity – OLS. Dependent variable: log of capital per worker. (1) Coef./sd

(2) Coef./sd

(3) Coef./sd

(4) Coef./sd

0.141 (0.12)

0.140 (0.12) 0.006 (0.27)

0.213 (0.14) 0.122 (0.26) 1.766*** (0.35)

0.215 (0.138)

1.862*** (0.373)

8.546*** (0.31) 222 0.108 0.096

0.082 (0.379) 0.277 (0.418) 0.330 (0.437) 0.210 (0.417) 0.233 (0.342) 0.463 (0.956) 8.537*** (0.319) 222 0.118 0.085

8

Workers (log)

6

Program dummy Contractor 4

Capital per worker (log)

526

0

2

4

6

8

Workers (log) Non-Program Constructor Non-Program Contractor

Program Constructor Program Contractor

Graph 1. Capital intensity vs. firm-size.

Types of program support Land Building Machinery Credit Training Other support Constant N R2 Adjusted R2

8.084*** (0.31) 222 0.007 0.002

8.083*** (0.31) 222 0.007 0.002

12 10 8 6

Inputs per worker (log)

14

Note: F-Test of joint significance of land, buildings machinery, credit, training, and other support in column 4: F(6,213) = 0.13 (Prob > F) = 0.87. *** Indicate significance at the 1% level.

4

worker is plotted against the size of the firm measured in terms of the number of employees. Capital intensity does not seem to differ between program and nonprogram firms and also does not vary systematically with firm-size, though contractors are significantly more capital-intensive than constructors. These findings are backed up by regressions of capital intensity on firm-size (column 1), firm-size and program participation (column 2), and firm-size, program participation, and being a contractor (column 3), presented in Table 2. In column 4, capital intensity is regressed on firm-size, being a contractor and dummies indicating receipt of different types of support. Neither the program dummy, nor the interactions of the program dummy with firm-size or being a contractor are significant, confirming that program firms are not more or less capital-intensive than nonprogram firms. In addition, capital intensity is not correlated with different types of program-support, even though the coefficients on the support dummies vary. At first glance, it thus seems unlikely that the absence of a correlation between program participation and capital per worker masks countervailing impacts of different types of program support on capital intensity. Yet, these support dummies are potentially endogenous. Moreover, the results do not suggest a positive relationship between firm-size and capital intensity. Contractors are significantly more capital-intensive than constructors, but the capital intensity of contractors also does not vary with firm-size. Input intensity does not seem to depend on program participation or on firm-size either, as illustrated by Graph 2, which plots input usage per worker against the number of workers. The graph does show that contractors typically use more inputs per worker. Regressions presented in Table 3 reveal that input intensity is strongly correlated with capital intensity and confirm that contractors use more input per worker, even after controlling for firm-size and capital intensity (see columns 3 and 4). Indeed, size variables and program participation have no impact, corroborating the intuition from the graphs. In addition, input intensity is not correlated with different types of program support (column 5). Although these support dummies are potentially endogenous, prima facie, these results suggest that the absence of a significant correlation between inputs usage per worker and program participation does not obscure offsetting impacts of different types of program support on inputs usage. The absence of positive relationships between firm-size and labor intensity contrasts with findings from manufacturing surveys in Sub-Saharan Africa, which reveal that capital intensity and input intensity rise with firm-size (see e.g., Teal, 2009, or Rijkers, 2008). However, Little (1987) and Little, Mazumdar, and Page (1987) find for India that this pattern does not hold.

0

2

4

6

8

Workers (log) Non-Program Constructor Non-Program Contractor

Program Constructor Program Contractor

Graph 2. Input intensity vs. firm-size.

4. ASSESSING PROGRAM IMPACT ON TECHNOLOGY ADOPTION (a) A Cobb–Douglas technology To assess the impact of the AAIHDP on technology adoption and productivity, a human capital augmented Cobb–Douglas production function is estimated; output, Y, is modeled as a function of physical capital, K, human capital, H, material inputs, M, and a technology parameter, A, using

WHO BENEFITS FROM PROMOTING SMALL ENTERPRISES? SOME EMPIRICAL EVIDENCE FROM ETHIOPIA

527

Table 3. Explaining input intensity – OLS. Dependent variable: log of input per worker.

Workers (log)

(1) Coef./sd

(2) Coef./sd

(3) Coef./sd

(4) Coef./sd

(5) Coef./sd

0.135 (0.128)

0.102 (0.142) 0.148 (0.264)

0.359** (0.140) 0.049 (0.247) 1.901*** (0.355)

0.283* (0.145) 0.097 (0.247) 1.586*** (0.375) 0.201*** (0.062)

0.407*** (0.144)

Program dummy Contractor Capital per worker (log) Types of program support Land Buildings Machinery Credit Training Other support Constant N R2 Adjusted R2

9.459*** (0.318) 185 0.006 0.001

9.454*** (0.319) 185 0.008 0.003

9.750*** (0.302) 185 0.143 0.129

7.955*** (0.616) 175 0.186 0.166

1.976*** (0.372)

0.079 (0.345) 0.262 (0.385) 0.059 (0.396) 0.109 (0.400) 0.411 (0.319) 0.287 (0.967) 9.762*** (0.313) 185 0.154 0.116

Note: F-Test of joint significance of land, buildings, machinery, credit, training, and other support in column 4: F(6,167) = 0.38 (Prob > F) = 0.89. * Indicate significance at the 10% level. ** Indicate significance at the 5% level. *** Indicate significance at the 1% level.

the familiar formula: Y ¼ AK a H b M c . 8 Human capital, H, is assumed to be a function of the educational attainment of the workforce, E, and the number of workers L: H ¼ L expðhðEÞÞ. This formulation enables us to control for both the quantity of labor, proxied by the number of workers, L, and the “quality” of labor, proxied by some function h(E) of the average educational attainment of workers. Controlling for the skill level of workers is important, since it will be shown that workers in program firms are on average better educated. Taking logs and imposing a linear functional form, i.e., hðEÞ ¼ @b E, yields an estimable equation of the human capital augmented production function: ln Y ¼ ln A þ a ln K þ b ln L þ dE þ c ln M þ u: The “quality” of the labor force matters if d–0. To test whether program firms use a technology that differs from nonprogram firms, all the parameters of the production function are interacted with program participation, P ln Y ¼ ln A þ a ln K þ b ln L þ dE þ c ln M þ qA P  ln A þ qK P  ln K þ qL P  ln L þ qM P  ln M þ e: Under the null hypothesis of no differences in technology the interaction terms should be zero. Table 4 presents the results of OLS estimation of the production function. The dependent variable is the log of annual revenue. 9 The baseline explanatory variables are the factors of production, firm age, and activity dummies. Information on one or more explanatory variables are missing for 71 firms, forcing us to reduce our sample to 169 observations. 10 The specification presented in column 1 tests whether the technol-

ogy used by contractors is different from that used by constructors. This seems to be the case; the dummy on being a contractor and the interaction terms of this dummy with labor, capital, and inputs usage are jointly significant at the 10% level. 11 Despite the fact that these interactions are imprecisely estimated due to the limited number of contractors in our sample. Since the group of constructors is of focal interest, and because the sample of contractors for whom production functions can be estimated is very small, we focus on constructors for the rest of this subsection. The specification in column 2 tests for differences in technology between program constructors and nonprogram constructors. The null hypothesis that no such differences exist cannot be rejected since neither the program dummy nor the interaction terms of program participation with labor, capital, and inputs usage are individually or jointly significant. This result is robust to exclusion of activity dummies (column 3) and entering the different types of program support separately (column 4). However, our sample is small and the production function parameters are consequently imprecisely estimated. Though they are always statistically insignificant, estimates of program impact vary across specifications. The specification in column 5 attempts to assess the impact of the heterogeneity of labor by including measures of the educational attainment and age of the workforce as additional explanatory variables. Neither enters significantly; the higher educational achievement of workers in program firms does not translate into higher productivity. In addition, the absence of a statistically significant program impact is not driven by differences in firm age since the coefficient on firm age is very low and statistically insignificant in all specifications,

528

WORLD DEVELOPMENT Table 4. Production functions – OLS dependent variable: log of annual revenue.

Factors Workers (log) Capital (log) Inputs (log) Firm age Firm age Interaction terms and firm-type Contractor Contractor * workers Contractor * capital Contractor * inputs

(1) Coef./sd

(2) Coef./sd

(3) Coef./sd

(4) Coef./sd

(5) Coef./sd

0.102 (0.124) 0.048 (0.039) 0.592*** (0.051)

0.196 (0.173) 0.051 (0.046) 0.598*** (0.070)

0.277* (0.150) 0.057 (0.046) 0.606*** (0.061)

0.170 (0.118) 0.067* (0.038) 0.580*** (0.046)

0.208 (0.180) 0.046 (0.047) 0.597*** (0.070)

0.004 (0.019)

0.021 (0.023)

0.013 (0.021)

0.008 (0.021)

0.020 (0.023)

0.845 (1.177) 0.198 (0.268) 0.010 (0.098) 0.034 (0.103)

0.409 (1.105) 0.028 (0.234) 0.026 (0.085) 0.067 (0.093)

0.212 (0.147) 0.032 (0.240) 0.119 (0.214) 2.794 (2.130)

Program dummy Program * workers Program * capital Program * inputs Types of program support Land

0.670 (1.201) 0.185 (0.272) 0.003 (0.099) 0.033 (0.103) 0.012 (0.220) 0.356 (0.236) 0.154 (0.239) 0.167 (0.245) 0.123 (0.195) 0.507 (0.887)

Buildings Machinery Credit Training Other support Human capital Education of the workforce Age of the workforce Activity dummies? Constant N R2 Adjusted R2 * ***

Yes 4.295*** (0.557) 169 0.777 0.745

Yes 3.989*** (0.745) 147 0.673 0.619

No 3.887*** (0.692) 147 0.638 0.617

Yes 4.192*** (0.535) 147 0.646 0.620

0.025 (0.027) 0.000 (0.015) Yes 3.815*** (0.916) 147 0.676 0.615

Indicate significance at the 10% level. Indicate significance at the 1% level.

indicating that young firms are not less productive than older firms. Overall, the basic Cobb–Douglas specifications fit the data rather well, as judged by the adjusted R2 consistently exceeding 0.6. Parameter estimates are furthermore consistent with Constant Returns to Scale for both program and nonprogram firms.

(b) Tackling endogeneity It is well known that OLS estimates of the production function may be misleading when one or more of the explanatory variables are correlated with the error term, for example because of measurement error, simultaneous determination of inputs and outputs, or omitted variable bias (see e.g.,

WHO BENEFITS FROM PROMOTING SMALL ENTERPRISES? SOME EMPIRICAL EVIDENCE FROM ETHIOPIA

Ackerberg, Caves, & Frazer, 2005). In addition to these conventional endogeneity concerns, we have to be on guard for selection bias, which arises when program participation is correlated with the error term, for example because there are unobserved factors not controlled for in the production function which are correlated with both productivity and selection into the program. As pointed out by Aivazian and Santor (2008), the selection problem is two-sided since firms self-select and because the program selects firms. Instrumental variables methods are used to deal with these endogeneity problems. 12 To alleviate instrumenting requirements Constant Returns to Scale restrictions are imposed, enabling us to estimate output per worker as a function of the capital labor ratio and inputs per worker. To further facilitate identification, capital per worker is assumed to be exogenous. The motivation for the latter assumption is that the capital stock is unlikely to change very much in the short run given high adjustment costs, while input usage is typically much more variable. Given the severity of identification requirements, our strategy is to first assess the impact of the two types of endogeneity in isolation, before attempting to tackle their joint effect. An inputs price index is used as an instrument for inputs usage. Prices of inputs vary across firms since input markets are far from perfect; the rationing of key construction inputs such as cement price had led to the development of a large black market and associated price dispersion (Global Development Solutions, 2006). While price is a theoretically appealing instrument since it is likely to affect output only through its impact on factor usage. Ackerberg, Lanier Benkard, Berry, and Pakes (2007) warn us that orthogonality may nevertheless be violated when price differences are correlated with unobserved quality differences in both inputs and outputs. Alternatively, price differences may be correlated with entrepreneurial ability and skills. 13 Program participation is instrumented using the size of the firm at start-up. The AAIHDP orientation provided a forum for aspiring entrepreneurs to team up. Credit constraints in combination with the legal requirement that cooperatives should employ at least 10 employees might help explain why program firms are larger at start-up than nonprogram firms. Most IHDP participants lack the capital required to set up a profitable sole proprietorship or trade association by themselves. Pooling resources through the formation of cooperatives might enable them to nevertheless improve their incomes. At start-up, the firm has not yet benefited from AAIHDP support; size at start-up will be exogenous if there is no direct link between initial size and productivity once current size is accounted for. The results of our instrumenting regressions are presented in Table 5 and the corresponding first stage regressions are presented in Tables A1.1 and A1.2 in Appendix. In column 1 endogeneity of inputs is the sole concern; program firms and nonprogram firms are pooled together and input usage is instrumented using the input price deflator, which is a good instrument as judged by the Cragg–Donald F statistic of 12.93. The resulting parameter estimate of the contribution of inputs to revenue is somewhat higher than the corresponding OLS estimate (presented in the bottom part of the table) yet the Hausman test does not reject the null hypothesis of no endogeneity. In columns 2–4, inputs and capital per worker are assumed exogenous and selection bias is the key concern. To facilitate identification, column 2 models the impact of the program as a technology shifter. The results do not indicate that the program succeeds in increasing productivity; if anything, pro-

529

gram firms are less productive. Moreover, IV estimates are lower than OLS estimates, suggesting upward rather than downward selection bias. A possible explanation for such upward bias is that the program is capable of selecting the most able workers (amongst applicants). As a robustness check, we drop activity dummies in column 3. The results do not change markedly, though the improved performance of our instrumenting strategy is noteworthy. The specification in column 4 is more general as it also allows for an impact of the program on capital and inputs usage. The null hypothesis of no difference in technology cannot be rejected, though the parameter estimates are imprecisely estimated as our sample is small and the instruments are weak, as is indicated by the low Cragg–Donald F statistic of 1.41. Hausman tests do not reject the null of no exogeneity for all three specifications, suggesting that the impact of selection bias is not large. Columns 5 and 6 allow for the coexistence of selection bias and endogeneity of inputs usage. In column 5, the program impact is again modeled using a dummy, which is negative and insignificant. Column 6 adopts a more general specification. Parameter estimates are poorly behaved, probably because the predicted program participation and instrumented interactions are highly correlated. In addition, instruments remain weak in both specifications, as evidenced by the Cragg– Donald F statistics of 3.79 (column 5) and 0.79 (column 6). While our point estimates are imprecise, the null hypothesis that the program has no impact on technology adoption is not rejected in either specification, and Hausman tests do not reject the null hypothesis of exogeneity. 14 In conclusion, the finding that program firms do not use a different technology is unlikely to be driven by selection bias or endogeneity of factor inputs. 5. ASSESSING PROGRAM IMPACT ON EARNINGS Descriptive statistics, presented in Table 6, confirm that workers in program firms are on average better educated than workers in nonprogram firms. Table 7 presents probit models of the probability of being hired by a program firm. The first column is the baseline specification with age, gender, education and a dummy for having completed an apprenticeship as explanatory variables. Column 2 adds dummies for prior activities to this baseline specification to test whether program firms draw disproportionately on workers with an underprivileged labor market status. Column 3 adds dummies for workers’ entire employment history, and column 4 includes both employment history and prior activity dummies. As expected, having a TVET degree is strongly positively associated with the probability of being employed in a program firm. The importance of training is further evidenced by the positive and significant coefficient on having completed an apprenticeship in the past, although this may also reflect the fact that those who do not pass the test can join the firm as apprentices. The insignificance of prior activity dummies reveals that program firms do not draw disproportionately on unemployed workers, casual workers and workers in otherwise marginal jobs. Focusing on the entire employment history of individuals instead demonstrates that workers who have experienced a significant unemployment spell (e.g., longer than 3 months) in the past, workers who have experience working in a cooperative, as well as workers who have experience as a domestic employee are significantly more likely to be employed in program firms, whereas workers who have experience as casual workers are less likely to be employed in program firms.

530

WORLD DEVELOPMENT Table 5. Production functions – constructors only. Dependent variable: log of annual revenue per worker.

Instrumental variable estimation Inputs per worker (log) Capital per worker (log)

(1) Coef./sd

(2) Coef./sd

(3) Coef./sd

(4) Coef./sd

(5) Coef./sd

(6) Coef./sd

0.719*** (0.173) 0.044 (0.046)

0.582*** (0.052) 0.061 (0.038) 0.597 (0.566)

0.578*** (0.045) 0.067* (0.038) 0.265 (0.307)

0.019 (0.022) 2.742** (1.309) Yes

0.024 (0.022) 4.016*** (0.548) Yes

0.014 (0.022) 3.994*** (0.489) No

0.492*** (0.106) 0.096 (0.070) 1.542 (1.818) 0.224 (0.198) 0.154 (0.249) 0.025 (0.023) 4.655*** (0.912) Yes

0.492*** (0.106) 0.096 (0.070) 1.542 (1.818) 0.224 (0.198) 0.154 (0.249) 0.025 (0.023) 4.655*** (0.912) Yes

0.917*** (0.301) 0.050 (0.085) 0.026 (0.025) 4.430 (3.162) 0.034 (0.332) 0.620 (0.456) 1.229 (2.481) Yes

Yes

Yes

Yes Yes Yes

Yes Yes

Yes Yes Yes Yes

Yes

Yes

Yes Yes Yes

Yes Yes

13.3 12.93 147 0.77(16) 1.00

12.41 11.89 147 0.59 (17) 1.00

29.54 35.71 147 0.07(4) 0.99

4.73 1.41 147 3.93(19) 0.99

8.16 3.79 147 1.13(17) 1.00

Yes Yes Yes Yes Yes 3.63 0.56 147 4.16(19) 0.99

0.575*** (0.054) 0.065 (0.040)

0.577*** (0.054) 0.064 (0.040) 0.182 (0.172)

0.579*** (0.045) 0.066* (0.038) 0.191 (0.140)

0.022 (0.022) 3.796*** (0.529) Yes 147 0.604 0.556

0.023 (0.022) 3.863*** (0.533) Yes 147 0.608 0.556

0.012 (0.021) 3.945*** (0.461) No 147 0.578 0.566

0.605*** (0.070) 0.052 (0.046) 0.016 (0.981) 0.068 (0.103) 0.050 (0.095) 0.023 (0.023) 3.699*** (0.667) Yes 147 0.609 0.551

0.605*** (0.070) 0.052 (0.046) 0.016 (0.981) 0.068 (0.103) 0.050 (0.095) 0.023 (0.023) 3.699*** (0.667) Yes 147 0.609 0.551

0.605*** (0.070) 0.052 (0.046) 0.016 (0.981) 0.068 (0.103) 0.050 (0.095) 0.023 (0.023) 3.699*** (0.667) Yes 147 0.609 0.551

Program dummy Program * inputs per worker (log) Program * capital per worker Firm age Constant Activity dummies Endogenous Inputs per worker Program dummy Program * inputs per worker (log) Program * capital per worker (log) Instruments included Input price Labor force at start-up (log) Labor force at start * inputs per worker (log) Labor force at start * capital per worker (log) Labor force at start * input price (log) Anderson LR Cragg–Donald F Number of observations Hausman Chi(2) (df) Prob Chi(2) For comparison: ordinary least squares Inputs per worker (log) Capital per worker (log)

Yes

Yes

Program dummy Program * inputs per worker (log) Program * capital per worker Firm age Constant Activity dummies Number of observations R2 Adjusted R2 * ** ***

Indicate significance at the 10% level. Indicate significance at the 5% level. Indicate significance at the 1% level.

To model the impact of the program on wages, Ln(W), a standard earnings function approach is used. Wages are modeled as a function of observable characteristics, X, such as age, education, and firm characteristics, and the type of firm the worker is employed in, Dj. The potential endogeneity of participation and of schooling are the two major obstacles to suc-

cessfully estimating this earnings function. Selection bias arises when participants differ from nonparticipants in terms of unobserved characteristics that affect both earning potential and the probability of being employed in a program firm, thus creating a correlation between the error term and the regressors and consequently causing the OLS estimator to be biased.

WHO BENEFITS FROM PROMOTING SMALL ENTERPRISES? SOME EMPIRICAL EVIDENCE FROM ETHIOPIA

To tackle this selection problem, the Lee selection correction is used. 15 The potential endogeneity of schooling, arising for example because of a correlation between schooling and ability, is also addressed by means of a control function approach. As pointed out by So¨derbom, Teal, Wambugu, and Kahyarara (2006), when the return to education is nonlinear, control function estimators of the returns to schooling are likely to result in more precise estimates of schooling than 2SLS estimators. In addition, control function estimates of the returns to schooling are more robust than 2SLS estimators when slope parameters vary with unobserved factors (see also Card, 2001). A related modeling problem is that the benefits from participation might well be heterogeneous; those who lack alternative opportunities and have low earning potential are likely to benefit more from program participation than individuals with high earning potential. In particular, it is anticipated that the benefits from participating in the program are inversely related to the level of educational attainment. To allow for this possibility, the gains from participation are allowed to depend on observables following Blundell and Costa Dias (2002). Assuming a log-linear Mincerian relationship between earnings, our estimable equation takes the form:

531

LnðW i Þ ¼ a þ bX X þ bDj þ bX X  Dj þ hðX ; ZÞ þ ; where hðX ; ZÞ is a vector of terms that correct for endogeneity bias due to selection and the endogeneity of schooling and the subscript j denotes firm-type. 16 These selection correction terms are a function of observables that affect both participation and earnings, X, and variables that affect participation only, Z, the exclusion–restrictions. Table 8 presents the results from OLS estimation of earnings functions. Column 1 begins by regressing monthly income on days worked, a dummy for being a casual worker, gender, age and its square, years of schooling and its square, employment history dummies and firm-type dummies, forcing the gains from program participation to be homogenous. The results are consistent with the literature; returns to schooling are convex, wages for casual workers are significantly lower than those for permanent workers, the age-earnings profile is concave, and the coefficient on the gender dummy is negative. Turning to the results of central interest, workers in program constructors earn 29% more than workers in nonprogram constructing firms, though there is a 103% premium associated with working for a contractor. In column 2 the

Table 6. Descriptive statistics workers. Nonprogram constructor

Program constructor

Nonprogram contractor

Program contractor

Mean

Sd

Mean

Sd

Mean

Sd

Mean

Sd

Age Sex Monthly pay (log) Days worked per month (log) Casual dummy

28.62 0.81 5.77 3.05 0.23

8.63 0.40 0.76 0.40 0.42

28.65 0.85 6.12 3.09 0.22

6.85 0.36 0.84 0.34 0.42

30.51 0.86 6.52 3.01 0.27

7.61 0.35 1.00 0.49 0.45

27.62 0.77 6.42 3.13 0.29

6.41 0.42 1.01 0.27 0.46

(Highest level of) educational attainment Years of schooling Primary Secondary College TVETcomplete

8.64 0.43 0.47 0.01 0.15

4.31 0.50 0.50 0.08 0.35

10.33 0.31 0.65 0.01 0.35

3.79 0.46 0.48 0.09 0.48

10.88 0.24 0.67 0.02 0.27

3.36 0.43 0.47 0.14 0.45

10.26 0.33 0.59 0.03 0.39

4.36 0.48 0.50 0.17 0.49

Entire employment history Apprenticeship completed in the past Government employee Employee in a firm Domestic employee Family worker Self-employed Employed in a cooperative Casual workers

0.25 0.10 0.42 0.06 0.03 0.07 0.02 0.12

0.44 0.30 0.49 0.23 0.17 0.26 0.14 0.33

0.46 0.18 0.44 0.09 0.03 0.06 0.14 0.09

0.50 0.38 0.50 0.29 0.17 0.24 0.34 0.29

0.35 0.22 0.51 0.12 0.00 0.08 0.04 0.08

0.48 0.42 0.51 0.33 0.00 0.28 0.20 0.28

0.39 0.20 0.44 0.15 0.05 0.05 0.08 0.03

0.49 0.40 0.50 0.36 0.21 0.21 0.27 0.17

Prior activity Unemployed Employee in a firm Self-employed Casual workers Student Government employee Inactive

0.20 0.25 0.06 0.16 0.22 0.04 0.01

0.40 0.43 0.24 0.37 0.41 0.21 0.08

0.17 0.31 0.07 0.22 0.16 0.04 0.00

0.38 0.46 0.25 0.42 0.37 0.19 0.00

0.14 0.41 0.10 0.06 0.24 0.02 0.02

0.35 0.50 0.31 0.24 0.43 0.14 0.14

0.21 0.36 0.01 0.18 0.15 0.07 0.00

0.41 0.48 0.12 0.39 0.36 0.26 0.00

Instruments for education Distance to school (log) Father was a farmer Father was a civil servant Mother was a farmer Mother was a housewife Observations

3.34 0.37 0.21 0.15 0.64 310

0.75 0.48 0.41 0.36 0.48

3.19 0.24 0.32 0.11 0.62 224

0.71 0.43 0.47 0.31 0.49

3.20 0.20 0.31 0.10 0.61 48

0.65 0.41 0.47 0.31 0.49

3.12 0.32 0.30 0.06 0.68 64

0.66 0.47 0.46 0.24 0.47

Note: All monetary measures in Ethiopian Birr (ETB). 1 USD is worth approximately 9.1 Ethiopian Birr (ETB).

532

WORLD DEVELOPMENT Table 7. Being hired by a program firm: probit selection models. Dependent variable: working in a program firm.

Age Primary Secondary College TVET Sex Apprenticeship completed

(1) Coef./sd

(2) Coef./sd

(3) Coef./sd

(4) Coef./sd

0.00 (0.00) 0.19 (0.09) 0.19* (0.09) 0.25 (0.19) 0.19*** (0.05) 0.00 (0.05) 0.17 (0.04)

0.00 (0.00) 0.22 (0.09) 0.23* (0.09) 0.26 (0.19) 0.22** (0.05) 0.00 (0.06) 0.15 (0.04)

0.00 (0.00) 0.18 (0.10) 0.18 (0.10) 0.19 (0.22) 0.19*** (0.06) 0.01 (0.06) 0.16 (0.05)

0.00 (0.00) 0.20 (0.10) 0.21* (0.10) 0.22 (0.22) 0.21*** (0.06) 0.00 (0.06) 0.15 (0.05)

Activity prior to current job Unemployed

0.15 (0.12) 0.12 (0.11) 0.03 (0.14) 0.21 (0.11) 0.03 (0.12)

Employee Self-employed Casual Student Employment history Unemployment Government Employee Domestic Family worker Self-employed Cooperative Casual N LR v2 Pseudo R2

661 101.6(17) 0.1109

658 110.44(23) 0.1282

0.14 (0.12) 0.14 (0.12) 0.13 (0.17) 0.27 (0.11) 0.07 (0.12) 0.21*** (0.04) 0.13 (0.06) 0.01 (0.04) 0.20 (0.07) 0.10 (0.13) 0.09 (0.09) 0.32*** (0.08) 0.13 (0.07) 661 153.69(25) 0.1686

0.20*** (0.05) 0.10 (0.07) 0.05 (0.06) 0.17* (0.08) 0.11 (0.13) 0.08 (0.12) 0.31*** (0.08) 0.17 (0.07) 658 161.12(25) 0.1774

Note: Marginal effects, evaluated at the mean. * Indicate significance at the 10% level. ** Indicate significance at the 5% level. *** Indicate significance at the 1% level.

assumption that the returns to schooling are homogeneous is relaxed and interactions between program participation and returns to schooling are included as additional regressors. These interaction terms are not significant suggesting the effect does not vary by education level. In column 3 we allow for the endogeneity of education by including the predicted residual of a model of educational attainment and interactions of this residual with program participation as additional regressors. The model of educational attainment is presented in Table A2 in Appendix and uses distance to school at the age of six and parental occupation as exclusion restrictions. The terms are not jointly significant

and parameter estimates do not change very much; the endogeneity of schooling is unlikely to be a serious problem as is also evidenced by the Hausman test, which does not reject the null hypothesis of no endogeneity of schooling. We would like to know why program constructors pay more than nonprogram constructors and consider firm-size, capital intensity, and input intensity as candidate explanations. 17 Adding a control for firm-size in column 4 returns the point estimate to that in column 1, suggesting that the program premium is not due to a positive association between firm-size and wages. 18 To test whether capital- and input intensity can explain the program premium, the capital–labor ratio

WHO BENEFITS FROM PROMOTING SMALL ENTERPRISES? SOME EMPIRICAL EVIDENCE FROM ETHIOPIA

533

Table 8. Earnings regressions – OLS. Dependent variable: log of monthly earnings.

Days per month (log) Casual dummy Age Age2 Sex Schooling Schooling2

(1) Coef./sd

(2) Coef./sd

(3) Coef./sd

(4) Coef./sd

(5) Coef./sd

(6) Coef./sd

(7) Coef./sd

0.525*** (0.104) 0.377*** (0.092) 0.005 (0.021) 0.007 (0.030) 0.383*** (0.084) 0.000 (0.032) 0.002 (0.002)

0.519*** (0.104) 0.385*** (0.092) 0.002 (0.021) 0.010 (0.030) 0.398*** (0.085) 0.020 (0.040) 0.004* (0.002) 0.022

0.522*** (0.128) 0.384*** (0.079) 0.004 (0.023) 0.006 (0.032) 0.399*** (0.077) 0.021 (0.061) 0.004 (0.003) 0.057

0.542*** (0.107) 0.425*** (0.095) 0.001 (0.021) 0.011 (0.030) 0.403*** (0.086) 0.008 (0.041) 0.004 (0.002) 0.003

0.215 (0.275) 0.321 (0.244) 0.022 (0.026) 0.041 (0.035) 0.350*** (0.107) 0.010 (0.042) 0.003 (0.003) 0.039

0.233 (0.275) 0.324 (0.244) 0.025 (0.026) 0.044 (0.035) 0.361*** (0.107) 0.009 (0.042) 0.003 (0.003) 0.036

0.157 (0.275) 0.340 (0.245) 0.015 (0.025) 0.032 (0.035) 0.337*** (0.107) 0.025 (0.035) 0.001 (0.002)

(0.067) 0.003

(0.101) 0.005

(0.068) 0.000

(0.081) 0.001

(0.081) 0.001

0.214** (0.090) 0.219*** (0.074) 0.085

(0.004) 0.213** (0.090) 0.216*** (0.074) 0.085

(0.005) 0.213** (0.096) 0.218*** (0.078) 0.086

(0.000) 0.243*** (0.091) 0.221*** (0.074) 0.084

(0.004) 0.296*** (0.112) 0.155* (0.093) 0.129

(0.004) 0.293*** (0.112) 0.153* (0.092) 0.146*

0.306*** (0.112) 0.158* (0.093) 0.136

(0.068)

(0.069)

(0.069)

(0.085)

(0.086)

(0.086)

0.255*** (0.074) 0.711*** (0.126) 0.204 (0.152)

0.253*** (0.074) 0.685*** (0.127) 0.171 (0.153)

(0.061) 0.005 (0.024) 0.000 (0.004) 0.003 (0.042) 0.003 (0.004) 0.228*** (0.086) 0.685*** (0.133) 0.210 (0.167)

0.447 (0.411) 0.668*** (0.137) 0.028 (0.438) 0.024 (0.036)

0.152 (0.101) 0.719*** (0.203) 0.129 (0.242) 0.085 (0.052)

Program * schooling (demeaned) Program * schooling2 (demeaned) Prior: casual work Prior: employee firm Apprenticeship completed v v2 v * program v2 * program Program constructor Contractor Program contractor Workers Capital intensity Input intensity Constant N R2 Adjusted R2 Joint sign. schooling and schooling2 Joint sign. program * schooling and program * schooling2

0.160 0.143 (0.101) (0.100) 0.656*** 0.652*** (0.212) (0.214) 0.146 0.170 (0.242) (0.243) 0.095* 0.103* (0.053) (0.053) 0.009 0.005 (0.021) (0.021) 0.049* 0.049* (0.027) (0.027) 3.028*** 2.997*** 2.857*** 5.551*** 5.214*** 4.856*** 2.955*** (0.516) (0.533) (0.666) (0.543) (0.998) (1.020) (1.009) 550 550 550 536 356 356 356 0.353 0.357 0.358 0.367 0.258 0.266 0.252 0.325 0.327 0.322 0.334 0.199 0.203 0.193 F(2,526) = 11.84 F(2,524) = 11.43 v2(2) = 1.01 F(2,509) = 11..81 F(2,329) = 9.56 F(2,327) = 9.90 F(2,329) = 7.19 P > F = 0.00 P > F = 0.00 P > F = 0.00 P > F = 0.00 P > F = 0.00 P > F = 0.00 P > v2 = 0.03 F(2,524) = 1.63 v2(2) = 2.35 F(2,509) = 1.86 F(2,329) = 2.92 F(2,327) = 2.98 P > F = 0.16 P > F = 0.06 P > F = 0.05 P > F = 0.20 P > v2 = 0.31

Note: Sub city dummies included but not reported. Hausman test of endogeneity for column 3: v2(24) = 0.72. v is the residual from the education model. Standard errors in column 3 bootstrapped. * Indicate significance at the 10% level. ** Indicate significance at the 5% level. *** Indicate significance at the 1% level.

534

WORLD DEVELOPMENT Table 9A. Quantile Earnings Regressions – No Firm Controls. Dependent variable: Log of monthly earnings.

Days per month (log) Casual dummy Age Age2 Sex Schooling Schooling2 Prior: casual work Prior: employee firm Apprenticeship completed Contractor Program contractor Program constructor Constant Pseudo R2 N

p10 (1) Coef./sd

p25 (2) Coef./sd

p50 (3) Coef./sd

p75 (4) Coef./sd

0.278 (0.179) 0.074 (0.102) 0.017 (0.018) 0.004 (0.021) 0.303*** (0.084) 0.065* (0.036) 0.002 (0.002) 0.002 (0.085) 0.129* (0.077) 0.023 (0.068) 0.859*** (0.169) 0.282 (0.188) 0.263*** (0.066) 4.240*** (0.740) 0.21 345

0.146 (0.244) 0.313 (0.259) 0.002 (0.022) 0.006 (0.028) 0.273** (0.111) 0.013 (0.042) 0.001 (0.002) 0.060 (0.123) 0.084 (0.099) 0.085 (0.091) 0.642*** (0.194) 0.247 (0.237) 0.065 (0.095) 5.460*** (0.859) 0.10 345

0.064 (0.150) 0.283 (0.173) 0.002 (0.016) 0.003 (0.022) 0.324*** (0.073) 0.033 (0.028) 0.004** (0.001) 0.176** (0.079) 0.095 (0.064) 0.071 (0.059) 0.607*** (0.136) 0.086 (0.168) 0.093 (0.064) 5.552*** (0.556) 0.13 345

0.008 (0.119) 0.121 (0.133) 0.049*** (0.014) 0.083*** (0.018) 0.260*** (0.067) 0.010 (0.025) 0.002 (0.001) 0.206*** (0.070) 0.154** (0.063) 0.078 (0.056) 0.742*** (0.122) 0.031 (0.148) 0.077 (0.058) 6.206*** (0.461) 0.20 345

Note: Sub city dummies included but not reported. * Indicate significance at the 10% level. ** Indicate significance at the 5% level. *** Indicate significance at the 1% level.

and input intensity are included as explanatory variables in columns 6 and 7. Unfortunately, information on these variables is not available for all workers, forcing us to drop about 40% of our sample. The result is to make our point estimates less precise but we cannot reject the hypothesis that the average program effect on earnings is the 28% implied by the first column. In addition, the interactions between program participation and schooling are significant at the 10% level when firm characteristics are controlled for, suggesting program participation may be especially beneficial for those with low levels of educational attainment, who are typically concentrated at the bottom of the earnings distribution. Appendix B demonstrates that these results are robust to correcting for selection bias using Lee selection corrections based on a multinomial logit approach (see Lee, 1983). As a robustness check to test whether the AAIHDP indeed benefits those at the bottom of the earnings distribution most, conditional quantile regressions are estimated using the number of days, human capital characteristics, dummies controlling for prior experience as a casual worker and employee and, crucially, firm-type dummies as explanatory variables. The results are presented in Table 9A. The premium on working for a program constructor is positive and statistically significant at the 5% level only at the bottom of the earnings distribution. In Table 9B, firm-size, capital-intensity, and input-intensity are controlled for. The estimated program

premium declines across the earnings distribution, yet remains significant and highest for those at the lowest decile of the earnings distribution, whereas it is negligible and statistically insignificant at higher quantiles. It would thus appear that the premium is highest for the poorest workers. 6. DISCUSSION Our analysis is based on a comparison of program and nonprogram firms. If the program was of sufficient size to impact on the overall market then such a comparison might well be misleading as the program would have affected the nonprogram firms. In our analysis we have used nonprogram firms as the basis for our analysis. If the AAIHDP had spillover effects, that would undermine our ability to assess its impact. One possibility is that the AAIHP may have contributed to increased competition, thus lowering prices for outputs and by exacerbating inputs shortages driving up input prices. This may have undermined the profitability of the existing firms. Moreover, the program may have biased the composition of nonprogram firms operating in the construction sector toward small firms and affected their factor choices. Not all small firms created by the program were allowed to participate in the program and the AAIHDP may have induced incumbent firms to exit, to grow slower or to invest less. The program

WHO BENEFITS FROM PROMOTING SMALL ENTERPRISES? SOME EMPIRICAL EVIDENCE FROM ETHIOPIA

535

Table 9B. Quantile earnings regressions with firm controls. Dependent variable: log of monthly earnings.

Days per month (log) Casual dummy Age Age2 Sex Schooling Schooling2 Prior: casual work Prior: employee firm Apprenticeship completed Contractor Program contractor Program constructor Workers Capital intensity Input intensity Constant Pseudo R2 N

p10 (1) Coef./sd

p25 (2) Coef./sd

p50 (3) Coef./sd

p75 (4) Coef./sd

0.107 (0.16) 0.031 (0.09) 0.019 (0.02) 0.006 (0.02) 0.325*** (0.08) 0.124*** (0.02) 0.004*** (0.00) 0.038 (0.08) 0.081 (0.07) 0.001 (0.07) 0.686*** (0.10) 0.231* (0.14) 0.172** (0.08) 0.067* (0.04) 0.029** (0.01) 0.027 (0.02) 2.087*** (0.69) 0.22 345

0.078 (0.29) 0.214 (0.32) 0.015 (0.03) 0.011 (0.04) 0.334** (0.14) 0.038 (0.05) 0.0001 (0.00) 0.057 (0.16) 0.077 (0.12) 0.022 (0.12) 0.310 (0.24) 0.201 (0.28) 0.030 (0.13) 0.166*** (0.06) 0.036 (0.03) 0.077** (0.04) 3.455*** (1.12) 0.14 345

0.107 (0.23) 0.236 (0.26) 0.007 (0.03) 0.001 (0.03) 0.346*** (0.11) 0.019 (0.04) 0.003 (0.00) 0.112 (0.12) 0.012 (0.10) 0.063 (0.09) 0.302 (0.22) 0.186 (0.24) 0.029 (0.11) 0.141** (0.06) 0.019 (0.02) 0.035 (0.03) 4.838*** (0.87) 0.15 345

0.023 (0.17) 0.124 (0.19) 0.045** (0.02) 0.077*** (0.03) 0.278*** (0.10) 0.014 (0.03) 0.003 (0.00) 0.201* (0.11) 0.233** (0.09) 0.165** (0.08) 0.554*** (0.21) 0.113 (0.23) 0.009 (0.10) 0.097* (0.06) 0.006 (0.02) 0.038 (0.03) 5.579*** (0.68) 0.21 345

Note: Sub city dummies included but not reported. * Indicate significance at the 10% level. ** Indicate significance at the 5% level. *** Indicate significance at the 1% level.

may also have discouraged the entry of new nonprogram firms. Our ability to quantitatively assess such possible spillover effects of the AAIHDP is hampered by the lack of data. Almost a quarter of all housing construction firms in Addis participate in the program. This suggests that spillover effects are likely to be moderate, though nonnegligible. The limited available evidence is consistent with this inference since only 12% of the managers of nonparticipating firms in the AACES believe that termination of the AAIHDP would improve their business. Somewhat surprisingly, program MSEs were on average larger than nonprogram MSEs. On the basis of the AACES dataset, the AAIHDP has not led to the formation of smaller firms. Yet, it could be that the program has led to a reduction in the proportion of firms with more than 50 employees. Unfortunately, it is difficult to draw definitive conclusions regarding the impact of the program on the evolution of the size distribution of housing construction firms without more comprehensive data.

That the program has benefitted workers is consistent with the limited literature on the impact of small enterprise support programs. Aivazian and Santor (2008) assess a microcredit program in Sri Lanka and find that firms receiving credit expanded their labor force and paid higher wages than firms which did not. The authors speculate that this may be because workers were able to appropriate rents. The existence of a program premium and pay differentials across contractors and constructors are consistent with the existence of such rents. 7. SUMMARY AND CONCLUSIONS Active Labor Market Programs, such as the AAIHDP, are widely viewed as a means to provide employment opportunities for the poor by providing more, relatively unskilled, jobs. To do this, they must either use a more unskilled labor intensive technology or more unskilled labor per unit of capital for a given technology. We have found that the technol-

536

WORLD DEVELOPMENT

ogy of program firms is not significantly different from the technology used by firms outside the program, a conclusion which is robust to controlling for selection bias and endogeneity of factor inputs. Program firms are not more labor intensive and thus this is not a route by which more jobs

per unit of investment will have been created. We do find there is an earnings premium associated with program participation which again is unlikely to be driven by selection bias and it does appear to be larger for the poorer paid workers.

NOTES 1. This paper focuses on these aspects of MSEs and does not consider other possible benefits, such as their contributions to competition, entrepreneurship and innovation, creation of products which are more suitable for the poor as well as providing social and political dividends. See, e.g. Biggs (2002) and Hallberg (2001) for reviews of the empirical evidence regarding such benefits. 2. The supply of housing has also been constrained by restrictive land policies and a legacy of marginalizing the private sector in housing development. Housing demand has been increasing as a result of population growth, immigration, progressively increasing diaspora demand for housing, a lack of alternative investment opportunities and speculation, see World Bank (2007). 3. Conventionally, all firms with a contracting license grade – an indicator of technological capability – are referred to as contractors. Firms with a contracting license grade between 1 and 6 can execute foundational and structural works, while those with a license grade between 7 and 10 cannot. For expositional purposes, in this paper we only refer to a subset of this group, namely those with a license grade between 1 and 6, as “contractors”. 4. Alternatively, program participants might have been employed in a different sector altogether had the program not been implemented, though this is arguably less likely since the program explicitly targeted construction workers. 5. The AACES survey also included questions asking workers whether or not they had taken an AAIHDP test and if so, whether or not they had passed and with what score. Unfortunately, there was a lot of nonresponse as participants were hardly ever capable of recollecting the results of their tests; consequently, it was not feasible to exploit variation in test scores to assess the impact of the program. 6. One potential caveat with our sample arises from having had to request the approval of the manager in order to interview workers. Given the manager’s control over whom we could interview, the sampling of workers within occupational categories may not be entirely random. In addition, response bias cannot be ruled out a priori. Yet, most of our questions are factual in nature and most managers were cooperative. It is our impression that the assignment of workers for interviews was more due to chance, i.e. being around at the time of our interview, than to strategic selection on the manager’s part. 7. In addition, there was one firm engaging in pre-cast beam production which did not meet any of these criteria but was classified as a program firm since the pre-cast beam technology is not widely applied by nonprogram firms as yet. 8. Our starting-point for assessing technology differences was the more general translog production function. Since the translog specifications did not reject Cobb–Douglas restrictions, we proceeded with Cobb–Douglas production functions. Results are omitted to conserve space, but are available upon request.

9. Given the existence of a fixed price system, prices for products of program firms and nonprogram firms might differ, causing revenue to be a misleading indicator of productivity. We investigated this possibility by regressing price deflators on activity dummies and program participation but could not find evidence for systematic price differences between program and nonprogram firms. In addition, deflating revenue and inputs by output and input price deflators did not affect the overall pattern of results. Results are omitted to conserve space, but available upon request. 10. These observations are not missing entirely at random; contractors are much less likely to report full information on material inputs usage then constructors, presumably because it was more difficult for them to provide aggregated subtotals for each input. 11. The F-test for the joint significance of being a contractor, contractor * workers, contractor * capital and contractor * workers is F(4,147) = 2.22, P > F = 0.070). 12. We also experimented with control function approaches. The results we obtained were similar to the results obtained using IV methods; they are omitted to conserve space, but are available upon request. 13. We are grateful to an anonymous referee for pointing this out. 14. Of course it could be that that different aspects of the program offset each other, yet this possibility cannot be rigorously assessed with the data at hand. 15. This Lee selection correction (see Lee, 1983) is based on the multinomial logit model, which is appropriate since workers can be employed in four different types of firms; program contractors, nonprogram contractors, program constructors, and nonprogram constructors. We also used the Dubin–McFadden selection corrections but this led to very imprecise estimates (given our small sample size, Lee selection corrections are expected to result in better estimates – see Bourguignon, Fournier, and Gurgand (2007)). In addition, we explored the possibility of doing IV estimation but could not find sufficiently convincing instruments. We also experimented with matching estimators, but do not present these as they do not allow for selection on unobservables and because the results were very similar to those obtained by means of OLS. 16. Note that, if returns to program participation are heterogeneous and both program participation and schooling are endogenous, we should also control for the endogeneity of the interaction of schooling and participation. In the context of our model this is done by interacting the terms correcting for selection bias with the terms which correct for the endogeneity of schooling. 17. We also experimented with other firm characteristics, such as the age of the firm, the organizational structure of the firm (for example whether the firm is a cooperative or not), and the activities the firms engage in, but these variables could not explain why program firms pay more. 18. Note that the sample has been reduced by 14 workers. However, differences in sample composition are not driving the results; estimates available upon request.

WHO BENEFITS FROM PROMOTING SMALL ENTERPRISES? SOME EMPIRICAL EVIDENCE FROM ETHIOPIA

537

REFERENCES Ackerberg, D., Caves, K., & Frazer, G. (2005). Structural identification of production functions. University of Colorado, Working Paper. Ackerberg, D., Lanier Benkard, C., Berry, S., & Pakes, A. (2007). Econometric tools for analyzing market outcomes. In J. Heckman, & E. E. Leamer (Eds.). Handbook of econometrics (Vol. 6, A, pp. 4171–4276). Springer. Aivazian, V. A., & Santor, E. (2008). Financial constraints and investment: Assessing the impact of a World Bank credit program on small and medium enterprises in Sri Lanka. Canadian Journal of Economics, 41(2), 475–500. Betcherman, G., Olivas, K., & Dar, A. (2004). Impacts of active labor market programs: New evidence from evaluations with particular attention to developing and transition countries. World Bank Social Protection Discussion Paper Series No. 0402. Biggs, T. (2002). Is small and beautiful worthy of subsidy? IFC Discussion Paper. Available at: . Blundell, R., & Costa Dias, M. (2002). Alternative approaches to evaluation in empirical microeconomics. CEMMAP Working Paper CWP 10/02, Institute for Fiscal Studies. Bourguignon, F., Fournier, M., & Gurgand, M. (2007). Selection bias corrections based on the multinomial logit model: Monte Carlo comparisons. Journal of Economic Surveys, 21(1), 174–205. Card, D. (2001). Estimating the returns to schooling: Evidence on some persistent econometric problems. Econometrica, 69(5), 1127–1160. Construction Ahead (2005). Low cost housing: Would this be the solution? Construction Ahead: Bi-monthly interface with the construction industry September–October. Butterfly Publishing Addis Ababa. Dar, A., & Tzannatos, Z. (1999). Active labor market programs: A review of the evidence from evaluations. World Bank Social Protection Discussion Paper No. 9901. Fretwell, D., Benus, J., & O’Leary, C. (1999). Evaluating the impact of active labor market programs: Results of cross country studies in Europe and Asia. World Bank Social Protection Discussion Paper No. 9915.

Global Development Solutions (2006). Integrated value chain analysis for the housing construction sector in Ethiopia. Mimeo. Hallberg, K. (2001). A market-oriented strategy for small and medium-scale enterprises. International Finance Corporation Discussion Paper No. 40. HDPO (Housing Development Program Office) (2004). Brief Project Profile, Mimeo. Lee, J. (1983). Generalized econometric models with selectivity. Econometrica, 51(3), 507–512. Little, I. (1987). Small manufacturing enterprises in developing countries. World Bank Economic Review, 1(2), 203–235. Little, I., Mazumdar, D., & Page, J. M. Jr., (1987). Small manufacturing enterprises: A comparative analysis of India and other economies. New York, NY: Oxford University Press. Rijkers, B. (2008). Small enterprise performance and labour market outcomes in Ethiopia. Unpublished doctoral dissertation, Oxford University, Oxford, UK. So¨derbom, M., Teal, F., Wambugu, A., & Kahyarara, G. (2006). The dynamics of returns to education in Kenyan and Tanzanian manufacturing. Oxford Bulletin of Economics and Statistics, 68(3), 261–288. Teal, F. (2009). Using firm surveys for policy analysis in low income countries. Journal of Development Perspectives, 4(1), 1–31. World Bank (2007). The Addis Ababa Integrated Housing Program – Background Paper to a survey of firms and workers in the construction sector in Addis. Mimeo.

APPENDIX A. ADDITIONAL TABLES See Tables A1.1, A1.2 and A2.

Table A1.1. First stage-IV – columns 1–4. Corresponding column in Table 5 Dependent variable

Capital per worker (log) Firm age Input price

(1)

(2)

(3)

(4)

(4)

(4)

Inputs per worker (log) Coef./sd

Program participation Coef./sd

Program participation Coef./sd

Inputs per worker (log) Coef./sd

Capital per worker * program participation Coef./sd

Program participation Coef./sd

0.163*** (0.061) 0.007 (0.035) 0.120*** (0.033)

0.008 (0.019) 0.003 (0.011)

0.009 (0.020) 0.016 (0.011)

0.044 (0.065) 0.012 (0.014)

0.156 (0.431) 0.025 (0.093)

0.060 (0.051) 0.001 (0.011)

0.191*** (0.056)

0.284*** (0.047)

3.492*** (0.279) 0.037

0.509 (1.851) 0.009

0.429* (0.220) 0.039

(0.034) 0.426***

(0.224) 0.066

(0.027) 0.007

(0.016) 8.458*** (0.529) Yes 147 0.905 0.892

(0.103) 0.085 (3.504) Yes 147 0.420 0.338

(0.012) 0.428 (0.417) Yes 147 0.420 0.338

Workers at start-up (log) Workers at start-up (log) * capital per worker (log) Inputs per worker (log) * workers at start-up (log) Constant Activity dummies Number of observations R2 Adjusted R2 * ***

7.003*** (0.555) Yes 147 0.394 0.319

Indicate significance at the 10% level. Indicate significance at the 1% level.

0.087 (0.205) Yes 147 0.410 0.337

0.009 (0.182) No 147 0.245 0.229

538

WORLD DEVELOPMENT Table A1.2. First stage-IV – columns 5 and 6.

Corresponding column in Table 5 Dependent variable

Capital per worker (log) Firm age Input price Workers at start-up (log)

(5) Inputs per worker (log) Coef./sd

(5) Program participation Coef./sd

(6) Inputs per worker (log) Coef./sd

(6) Program * capital per worker (log) Coef./sd

(6) Program * inputs per worker (log) Coef./sd

(6) Program participation Coef./sd

0.163*** (0.061) 0.008 (0.035) 0.118*** (0.034) 0.098 (0.180)

0.008 (0.019) 0.003 (0.011) 0.002 (0.011) 0.190*** (0.057)

0.004 (0.165) 0.012 (0.035) 0.013 (0.077) 0.858 (0.681) 0.091* (0.051) 0.084

0.094 (0.438) 0.033 (0.093) 0.247 (0.205) 0.595 (1.808) 0.182 (0.135) 0.042

0.304 (0.473) 0.031 (0.100) 0.381* (0.221) 2.613 (1.951) 0.308** (0.145) 0.187

0.053 (0.052) 0.002 (0.011) 0.025 (0.024) 0.442** (0.216) 0.019 (0.016) 0.035

7.201*** (0.666) Yes 147 0.395 0.315

0.095 (0.210) Yes 147 0.410 0.332

(0.084) 8.669*** (1.369) Yes 147 0.414 0.326

(0.223) 0.730 (3.634) Yes 147 0.426 0.340

(0.241) 1.448 (3.923) Yes 147 0.421 0.334

(0.027) 0.344 (0.434) Yes 147 0.425 0.339

Input price * workers at start-up (log) Workers at start-up (log) * capital per worker (log) Constant Activity dummies Number of observations R2 Adjusted R2 *

Indicate significance at the 10% level. Indicate significance at the 5% level. *** Indicate significance at the 1% level. **

Table A2. The determinants of educational attainment-OLS.

APPENDIX B. TACKLING SELECTION BIAS IN THE EARNINGS EQUATIONS Table A3 presents earnings functions estimates using Lee’s selection corrections based on the multinomial logit model. The underlying multinomial logit model for the selection of workers into different types of firms is presented in Appendix in Table A4 and uses age, educational attainment, prior activity, and employment history dummies as explanatory variables. Prior activities may help explain selection into different jobs as they may affect individual’s job search strategies and access to information, or because they reflect individual’s preferences. Since labor market experience is a key determinant of an individuals’ human capital stock, however, the exclusion restriction that these variables have no direct impact on earnings might well be violated. Columns 1 and 2 assume education is exogenous, while columns 3 and 4 allow for the endogeneity of schooling. All specifications control for individual characteristics. Columns 2 and 4 also control for firm characteristics. None of the selection terms are individually or jointly significant in any of the specifications. In addition, the resulting parameter estimates are very similar to OLS-estimates. This indicates that endogeneity is not a major issue. Once again, the estimated return to program participation is higher for the subsample of observations for whom information on capital stock, input usage and size of the labor force is available (see columns 2 and 4). When firm characteristics are not controlled for, the null hypothesis of homogeneous treatment effects is not rejected (columns 1 and 2). However, the null hypothesis is rejected at the 5% level when education is assumed exogenous and firm characteristics are controlled for (column 3). The parameter estimates suggest the benefits of program participation are highest for those with relatively low levels of education. Once the endogeneity of education is allowed for, however, the null hypothesis that the treatment effect is homogeneous cannot be rejected,

Coef./sd Dependent variable: years of schooling Age Age2 Sex Distance to school when 6 Father was a farmer Father was a civil servant Mother was a farmer Mother was a housewife Constant N R2 Adjusted R2 * ***

0.571*** (0.09) 0.745*** (0.12) 0.173 (0.35) 0.640*** (0.20) 2.513*** (0.38) 1.193*** (0.32) 1.250*** (0.56) 0.605* (0.34) 3.134* (1.61) 592 0.308 0.298

Indicate significance at the 10% level. Indicate significance at the 1% level.

though it is borderline significant at the 10% level (column 4). Given the weakness of our exclusion restrictions and the strong assumptions on functional form made in these specifications, our interpretation is that these results should be interpreted with caution (especially since the exogeneity of educational attainment could not be rejected).

WHO BENEFITS FROM PROMOTING SMALL ENTERPRISES? SOME EMPIRICAL EVIDENCE FROM ETHIOPIA

539

Table A3. Earnings regressions – Lee selection corrections.

Dependent variable: log of monthly earnings Days per month (log) Casual dummy Age Age2 Sex Schooling Schooling2 Prog * schooling (demeaned) Prog * schooling2 (demeaned) Prior: casual work Prior: employee firm Apprenticeship completed Contractor Program contractor Program constructor Workers Capital intensity Input intensity Lee v v2 vLee v2Lee Constant

(1) Coef./sd

(2) Coef./sd

(3) Coef./sd

0.519*** (0.132) 0.385*** (0.078) 0.002 (0.023) 0.010 (0.035) 0.398*** (0.084) 0.020 (0.038) 0.004* (0.002) 0.022 (0.068) 0.003 (0.003) 0.213** (0.093) 0.216*** (0.075) 0.084 (0.052) 0.681*** (0.144) 0.171 (0.145) 0.252*** (0.061)

0.518*** (0.109) 0.386*** (0.075) 0.005 (0.025) 0.005 (0.036) 0.401*** (0.074) 0.013 (0.047) 0.004 (0.003) 0.034 (0.088) 0.004 (0.005) 0.210** (0.092) 0.214*** (0.070) 0.086 (0.068) 0.676*** (0.157) 0.191 (0.168) 0.254*** (0.068)

0.263 (0.291) 0.409* (0.241) 0.019 (0.035) 0.037 (0.049) 0.366*** (0.077) 0.015 (0.042) 0.003 (0.003) 0.077 (0.094) 0.001 (0.005) 0.305** (0.134) 0.153 (0.108) 0.142* (0.080) 0.452*** (0.166) 0.167 (0.316) 0.114 (0.094) 0.115 (0.070) 0.002 (0.024) 0.045* (0.026) 0.265 (0.194)

(4) Coef./sd

0.265 (0.328) 0.405* (0.244) 0.007 (0.040) 0.021 (0.055) 0.371*** (0.098) 0.013 (0.071) 0.002 (0.004) 0.057 (0.095) 0.000 (0.005) 0.299** (0.130) 0.146 (0.096) 0.150* (0.086) 0.469** (0.212) 0.201 (0.369) 0.122 (0.115) 0.110 (0.074) 0.002 (0.023) 0.044* (0.026) 0.005 0.032 0.216 (0.130) (0.151) (0.247) 0.003 0.026 (0.038) (0.059) 0.003 0.003 (0.004) (0.009) 0.009 0.006 (0.029) (0.058) 0.005 0.005 (0.005) (0.011) 3.001*** 4.994*** 4.937*** 3.026*** (0.627) (0.655) (1.146) (1.285) 550 550 345 345 0.357 0.359 0.272 0.275 0.325 0.322 0.205 0.198 Chi(2) = 20.35 Chi(4) = 13.09 Chi(2) = 18.49 Chi(4) = 7.88 P > Chi = 0.00 P > Chi = 0.01 P > Chi = 0.00 P > Chi = 0.10

N R2 Adjusted R2 Joint significance schooling variables Columns 1 and 2: schooling, schooling2 Columns 3 and 4: schooling, schooling2, v, v2 Joint significance interaction schooling and program participation variables Chi(2) = 3.96 Chi(4) = 3.42 Chi(2) = 6.21 Chi(4) = 7.73 P > Chi = 0.14 P > Chi = 0.49 P > Chi = 0.04 P > Chi = 0.10 Columns 1 and 2: program * schooling and program * schooling2 Columns 3 and 4: program * schooling, program * schooling2, vLee, v2Lee Note: Sub city dummies included but not reported. v is the residual from the education model. Bootstrapped standard errors. * Indicate significance at the 10% level. ** Indicate significance at the 5% level. *** Indicate significance at the 1% level.

540

WORLD DEVELOPMENT Table A4. Multinomial logit selection model. Program constructor

Program contractor

Nonprogram contractor

Coef.

S.E.

Coef.

S.E.

Coef.

S.E.

Age Primary complete Secondary complete College TVET complete Sex Apprentice in the past

0.01 0.78 1.10** 0.57 0.70** 0.23 0.72

0.02 0.49 0.50 1.33 0.28 0.27 0.21

0.04 0.12 0.66 0.80 0.21 0.01 0.62

0.03 0.81 0.80 1.55 0.48 0.51 0.39

0.02 0.43 0.29 1.24 1.34*** 0.25 0.46

0.02 0.74 0.74 1.30 0.41 0.38 0.32

Employment history Ever unemployed Government employee Employee Domestic Self employed Working in a cooperative Casual worker

0.91 0.58 0.23 0.67 0.84 1.67 0.66*

0.23 0.36 0.28 0.41 0.57 0.49 0.35

0.35 1.04** 0.63 0.91 1.82 0.25 0.35

0.39 0.56 0.49 0.62 0.98 0.92 0.65

0.55 0.79 0.07 1.55** 0.18 0.92 1.71**

0.34 0.52 0.41 0.55 0.86 0.68 0.79

Activity prior to current job Unemployed Employee Self employed Casual worker Student Government employee N LR v2 (90) Pseudo R2

0.48 0.54 1.24 1.00* 0.34 0.32 661 912.41 0.207

0.52 0.54 0.76 0.54 0.53 0.74

0.90 0.39 1.63* 0.53 1.03 1.46**

1.18 1.18 1.45 1.29 1.17 1.61

2.03 2.05 0.46 2.20 1.46 1.92

1.21 1.21 1.70 1.22 1.25 1.38

* ** ***

Indicate significance at the 10% level. Indicate significance at the 5% level. Indicate significance at the 1% level.

Available online at www.sciencedirect.com