Export horticulture and household welfare : Evidence from Zambia

Evidence is very scanty in Africa on the welfare effects of the recent shift of the horticulture industry from involving poor households through outgrower arrangements towards employing them in consolidated production entities. This study determines the impact of large-scale export vegetable production on the welfare of the employees in Zambia. It uses data from a survey of a random sample of farm worker households and comparison households in nine villages around one of the four largest estate vegetable farms in Zambia. Evidence from control function, propensity score matching, and odds-weighted regression models suggest huge and significant welfare effects as measured by per capita consumption expenditure. Estimated at 44 and 45% for non-food and food expenditure, respectively, the impact is not affected by the households' initial wealth in any statistically significant manner. This means that the recent industry changes might need to be supported and better understood, as opposed to being admonished.


INTRODUCTION
With a per capita income of USD 423 and 64% of the Zambia's population (which is estimated at 12.5 million), living below the poverty line (World Bank, 2011), Zambia ranks among the poorest countries in the world.About 56% live in rural areas (World Bank, 2002) of which 97.4% are engaged in agriculture (CSO, 2000).Within a labor force of 3.4 million, 85% are employed in agriculture, 6% in industry and 9% in services.With unemployment at 7.9% (CSO, 2008), agriculture is often the only potential source of livelihood or income within the informal sector.The Zambian agricultural sector contributes about 20% to real GDP and 39% of earnings from non-traditional exports (IMF, 2011).The sector mainly consists of smallholder farmers who make up about 52% of the country's farmers (Chipokolo, 2006) and contribute 80% of the nation's food.However, despite their substantial aggregate contribution to national food supply and GDP, smallholder farmers constitute a third of the nation's hungry and poor (CSO, 2004).Several factors have been cited for the low welfare levels among smallholder farmers, including low productivity, frequent droughts, and unsatisfactory access to markets, market information and credit facilities (Chiwele, 2004;USAID, 2005).Export horticulture (flowers and vegetables) in Zambia developed in the early 1980s and growth in the sector in the last decade is seen as one of the opportunities for raising welfare levels among the rural poor, while also generating foreign exchange *Corresponding author.E-mail: tembogel@gmail.com.(Hichaambwa, 2010).In the 1990s and early 2000s, the industry expanded rapidly as numbers of companies increased, raising export volumes of vegetables and cut flowers from US $6 million in 1994 to over $33 million in 2001 when the sector employed about 10,000 people (ZEGA, 2002).However, most of these gains in both export volume and earnings were lost in 2005 (Hichaambwa, 2010).
The decline was due mainly to the demise in 2004 of the largest horticultural export company, Agriflora.
1 For smallholder farmers who participate in the industry under contract with larger firms, the bankruptcy of Agriflora deprived them of reliable income, transport logistics and technical support.Although, donors tried to cushion part of the shock, their support could not be sustained indefinitely.The industry also faces a number of other challenges, including exchange rate fluctuations (Sergeant and Sewadeh, 2006), inelastic prices in traditional markets (Mataa and Hichaambwa, 2010) and high air freight costs (ZEGA, 2002) that exceed levels observed in most other countries in the region.
2 Tightening standards in the UK and other EU export destinations in recent years have also served a major blow to the Zambian horticultural sector, especially among smallholder producers.It is argued that the cost of compliance to the European retailers' private standards for Good Agricultural Practices (EurepGAP) cut farmers' incomes in half between 2002and 2006(AgriFood Standards Project, 2007).
As a result, less than 3% of the smallholder farmers involved in supplying European markets in 2000 were still doing so in 2006.Most large-scale exporters have employed three major coping strategies.The first is to significantly reduce output and concentrate on the highervalue and higher-margin lines.The second strategy is to increase output and try to reduce costs, that is, improve efficiency to increase margins.The third strategy has been to reduce or eliminate all outgrower arrangements with smallholder farmers, increasingly placing greater emphasis on consolidation and "own" production.
3 However, the shift from smallholder contract-based farming to large-scale estate production may imply more employment for rural households (Dolan and Humphrey, 2000).A number of studies have considered the impact of export horticulture on household income and poverty in Africa.In Kenya, McCulloch and Ota (2002) found that households involved in export horticulture were better off, particularly in rural areas.They further contend that enabling more households to participate in the sector could reduce poverty substantially in both rural and urban areas.Maertens and Swinnen (2009) used company and household survey data from the vegetable export chain in Senegal to quantify income and poverty effects of high standards trade through labor markets.They found that horticultural exports from Senegal to the EU had grown sharply despite strongly increasing food standards, and that these exports had strong positive effects on poor households' incomes, reducing regional poverty by about 12% and extreme poverty by half.Third, tightening food standards induced structural changes in the supply chain including a shift from smallholder contract-based farming to large-scale integrated estate production.These studies offer valuable lessons on the poverty-reducing effects of export horticulture.However, low-income countries are characterized by varying cost structures, levels of development and institutional sophistication, and experiences, all of which leave part of the debate for much of Africa still open.
The Zambian industry has faced relatively greater challenges adjusting to tightening standards due to a number of other unique structural constraints, including, as already outlined, the collapse of the largest market player, higher transport costs 4 , and macroeconomic factors (high agricultural taxes and unstable exchange rates).This study uses data from 41 farm worker households and 64 comparison households to determine the impact of large-scale export vegetable production on the welfare of employees.It also seeks to determine whether the household's initial wealth has significant effects on the level of impact.Most of the prior studies cited earlier use income as a proxy of welfare.We use consumption expenditure as the outcome variable.As a proxy of welfare, consumption expenditure is often argued to be more reliable and less prone to underreporting errors than income.We find huge and significant effects on consumption.At least, 49% of the farm workers' consumption can be attributed to participation in large scale estate horticultural farm activities.This is consistent with recent similar studies (Maertens and Swinnen, 2009;McCulloch and Ota, 2002) and challenges conventional arguments that consolidation of large-scale farms is bad to poor households.

Impact identification strategy
Program impact can be defined as the expected value of the difference between the level of the outcome variable attained by participating households and that which they would have attained had they not participated in the program (Wooldridge, 2002;Ravallion, 2001).That is: Where ATT is the average treatment effect on the treated, Y1i is per capita consumption expenditure (the outcome of interest) for the treatment group (that is, households supplying labor to the large horticultural farm), Y0i is the outcome of interest for the comparison group, wi is a dichotomous variable equal to one if the household has at least one of its members supplying labour to the large-scale horticultural farm and zero otherwise, and E(.) is the expectations operator.
Consumption expenditure was computed by adding together the values of all goods and services consumed by the household (purchased or own-produced) during the 12-month period prior to the survey.This was divided by household size to obtain per capita consumption expenditure.When the i th individual participates in wage estate employment, their level of consumption expenditure is i Y 1 and if they do not their income is Y0i.This is the conditional mean impact, conditional on participation, also known as the treatment effect or the average effect on the treated (Wooldridge, 2002).However, if there is a difference in the mean of the outcome variable between participants and non-participants in the absence of the program, a bias will arise and this bias is given by: (2) This bias could be corrected if were known.
Unfortunately, the level of participants' consumption expenditure had they not participated cannot be observed.However, had the program been assigned randomly, the participants and nonparticipants could have the same expected income in the absence of the program.In this case, the expected income of nonparticipants will correctly reveal the counterfactual.For most programs, randomization is not possible due to ethical, cost and other pragmatic reasons.In the case of vegetable estate employment, treatment households either self-select themselves and/or are deliberately chosen on the basis of their individual characteristics.Under such a quasi-experimental design, statistical controls must be used to address the differences between the treatment and control groups (Barker, 2000).Under some form of exogeneity (Imbens, 2004), most quasi-experimental impact studies estimate the conditional average treatment effect on the treated as: Where x is a vector of covariates.
The assumption implied by (Equation 3) is that conditioning on carefully selected covariates renders the household's treatment status independent of potential outcomes, such that the unobserved This makes it possible to attribute any systematic differences in the outcome variables between treated and control units with the same values of the covariates to the program in question.A more dimensionally appealing but equivalent version of 'selection on observables' involves replacing x in (Equation 3) with the estimated conditional probability of participation, or propensity score, defined (Rosenbaum and Rubin, 1983).

Data and data sources
This study uses data from a cross-sectional survey conducted in 2009 in nine villages around Borassus Estate, one of the three largest export horticulture producers, located about 25 km west of Lusaka, the capital of Zambia.A total sample of 41 treatment (that is, farm worker households) and 64 comparison households (that is, poor households located in the same neighbourhood as farm worker households but with no members working on the large vegetable farm) was drawn using stratified random sampling.Selection of farm worker households was based on a sampling frame developed out of a farm register, whereas the sampling frame for the comparison households was developed through comprehensive listing of non-worker households within the same neighbourhood.The simple random sampling applied to each stratum/frame ensured that, within the stratum, every listed household had an equal chance of being selected into the sample.
Although, the households in the two strata looked similar on the basis of visible characteristics (save for participation status), we also used matching techniques to ensure comparability.The 41:64 (or roughly 2:3) sample allocation ratio between the treatment and comparison strata was deliberately done to provide more matching options for each treatment households.Among other things, the household questionnaire elicited information about participation in the horticultural industry, other livelihood activities, as well as standard demographic and human capital status.It also collected detailed information regarding food, nonfood and durable goods consumption expenditures, which was used in the computation of consumption-based measures of welfare.
The study benefited from secondary data and publications obtained from various organizations, including the Ministry of Agriculture and Cooperatives (MACO), the Central Statistical Office (CSO), the World Bank (WB), NZTT, and other relevant publications.Discussions with personnel from the Zambia Export Growers Association (ZEGA), the Natural Resources Development College (NRDC)/ZEGA Training Trust (NZTT), and management of the three major horticultural farms provided valuable information on the sub-sector. 5The resultant expanded understanding of the subsector also helped in the interpretation of the quantitative results.

Estimation of the propensity scores
Program impacts are measured by assessing whether a program changes the mean value of an outcome variable among participants compared with what the outcome would have been had they not participated.The central evaluation problem then is that participants cannot be simultaneously observed in the alternative state of no participation (referred to as the counterfactual) (Shahidur et al., 2010).Evaluators typically simulate the counterfactual by comparing program participants with a control with similar characteristics.Construction of the counterfactual determines the evaluation design, which is broadly classified as experimental or quasi-experimental.A key feature of the experimental design is complete randomization, which ensures that households in treatment and control groups are, on average, similar and that any observed systematic differences in the outcome variables after the intervention are attributable to the intervention (Table 2).However, randomization is not always possible in observational studies such as ours.Ravallion (2001Ravallion ( , 2003) ) characterizes the various methods used to estimate impact under quasi-experimental conditions.As a second-best alternative for these conditions, for example, comparison can be facilitated by statistically constructing comparable treatment and comparison strata.Propensity score matching (PSM) presents a unique set of techniques for 5 In general, the discussions provided a picture of a once-prosperous sub-sector that was unfortunately on a decline at the time of the study.
reconstructing an experimental environment out of non-random, quasi-experimental conditions.We use variants of propensity-scorebased methods to estimate the impact of employment in estate horticultural firms on household consumption, where the propensity scores (PS), or conditional probabilities of participation (given the observed characteristics), were estimated using a probit specification: Where  is a standard normal cumulative distribution function (CDF),  is an error term,  is the intercept to be estimated, δ is a vector of slope parameters also to be estimated, and x is a vector of covariates.Equation 4 was estimated using maximum likelihood (ML) procedures in Stata (StataCorp, 2008).
In general, participation can be explained by the household's observable characteristics associated with access to resources (land, capital, and labor) and information, skills and ability (age, education), preferences (age, ethnicity, demographic structure), and geographic location (Maertens and Swinnen, 2009).To avoid endogeneity, we use initial ( 2005) values of variables such as asset endowment and livestock ownership.To ensure consistency of the PSM, only covariates that exhibited significant correlation with the participation variable and/or the outcome variable were included in x.Propensity-score-based models are only as good as the quality of the matching and are valid only under certain identifying assumptions.The balancing effects of the propensity scores were tested using a number of procedures including stratification, t tests for the differences in covariate means between the two groups (participants and non-participants) before and after the matching (Rosenbaum and Rubin 1985), effectiveness in reducing standardized bias, and ability to drive the overall probit relationship to insignificance as measured by a joint likelihood ratio (LR) test and pseudo R 2 (Caliendo and Kopeinig, 2008).6

Estimation of impact
We use three broad categories of models to estimate the impact of participation on the outcome variablethe control function approach, propensity score matching, and propensity score weighting.Heckman and Robb (1985) showed that selection bias can be controlled by including a vector of covariates as control functions: Where i y is the outcome variable (in our case per capita consumption expenditure) for household i,  and  are parameters to be estimated,  is a vector of parameters to be estimated, x is as defined above, and  is a random error term.Wooldridge (2002) contends that (equation 5) could be consistently estimated by OLS as long as the outcome variable is not correlated with the unobservable characteristics, also known as selection on observables.However, robust standard errors were used due to failure to reject heteroskedasticity.In the second specification of the control function approach, we replace x with the propensity score, a method pioneered by Rosenbaum and Rubin (1983): Where the propensity score (PS) is as defined in (equation 1), that is, , and  is a parameter to be estimated.
In a more general version of correction on propensity score, we also include an interaction term between participation and the demeaned propensity score (Rosenbaum and Rubin, 1983;Wooldridge, 2002): Where PS  is the mean of the propensity score, and  is a parameter to be estimated.
The results from the control function models (equation 5) through (equation 7) were corroborated with ones obtained through propensity score matching (PSM), which involves for each treatment unit finding matches in the control group based on observable characteristics (Abadie and Imbens, 2002;Dehejia and Wahba, 2002).Thus, the ATT was computed as the weighted average of the difference in the outcome variable between treatment households and matched control ones, where matching was done by kernel functions and ATT computation was restricted to the region of common support.The kernel matching estimator is given as (Heckman et al., 1997;Smith and Todd, 2005;Gilligan and Hoddinott, 2007): Where T is the treatment group participants; C refers to the comparison group, K is the kernel function, and n a is the kernel bandwidth.Inferences were made possible by bootstrapping standard errors. 7hile matching produces consistent estimates, Hirano et al. (2003) show that the odds-weighted regression approach to PSM, or propensity score weighting (PSW), results in fully efficient estimates.Under this framework, impact is the estimated slope coefficient 1  in the simple regression model: But with the observations weighted by 1 for treatment households and by the estimated odds ratio,       is the estimated conditional probability of participation.

Heterogeneous impact
The Hirano et al. (2003) framework can be extended to the case where the impact of the treatment is differentiated by some defined household categorization.We use this framework to estimate disaggregated impact: Where D is a dummy variable based on the household's initial wealth status.
A household was categorized as poor (D = 1) if the initial wealth index was negative, where the wealth index was computed from assets data using principal components analysis (Filmer and Pritchett, 2001).Thus, the impact of participation is equal to  ˆ for the relatively less poor households (D = 0) and    for the poor ones.Thus, ˆ is the additional impact that a poor household would experience relative to its relatively richer counterparts.

RESULTS AND DISCUSSION
Table 1 presents selected sample characteristics, comparing control and treatment households.The results indicate that the two sub-samples were generally wellbalanced with respect to most characteristics.Significant differences between control and treatment households were evident only with respect to the age of the household head, location and initial wealth of the households.Although, the age of the household head was generally low (averaging 40 years), households with at least one estate farm worker had generally younger heads compared to their non-worker counterparts.Treatment households were also more likely to be maleheaded, to have more educated members, and to be further away from the main road and schools; although, these differences were not statistically significant.Not only did treatment households have greater initial wealth but they also were well-off as indicated by a positive mean wealth index.On the other hand, households in the control group were generally poor (windex < 0).At the time of the survey, treatment households also had almost twice as much consumptions as their counterparts in the control group, and this was true even if consumption was disaggregated into its components (Figure A2).These general descriptive results were further confirmed by probit analysis of participation (Table 2).The marginal effects (column 2) show that an additional year to the age of the household head was associated with a 1.3% drop in the household's probability to participate in estate wage employment.Surprisingly, the probability to participate was inversely and significantly correlated with the number of members in the active age group (15 to 55 years).Location and initial wealth were the largest determinants of participation.
The PS balancing test results confirm the existence of strong bias for most covariates and that balancing successfully eliminated this bias (Table A1). 8In general, matching produces consistent estimates as long as the unobserved factors are equally distributed between the two groups. 9The estimated PS was also inspected for the common support requirement.This was found to be satisfied, as indicated by the fact that 0 < PS < 1 and by a large PS overlap (0.07, 0.86) between the control and treatment groups (Figure A1).

Impact estimates
The descriptive statistics discussed earlier indicate that those who participate in estate horticulture firms as workers are better off as indicated by wealth and consumption.However, descriptive statistics are limited and may not imply causality as they fail to account for other sources of the observed differences.Table 3 presents impact estimates as determined by the various models discussed earlier.All the five models indicate huge positive and significant effects of participation.More specifically, employment in estate horticulture farms raises per capita household consumption by 49 to 53%.Although, the specific impact estimates vary from model to model, they are generally very close to each other.The control function models (columns 1 through 3) further confirm the importance of conditioning on the observables, either directly (column 1), or through the propensity score (columns 2 and 3).The interaction between the treatment indicator and the demeaned propensity score had a dampening but insignificant effect.Model 1 also shows that per capita consumption expenditure is directly correlated with education level attained by the members, and inversely related to household size.Village 2 households had 21% less consumption, just as they were less likely to participate compared to households in all other villages.Table 4 presents impact estimates disaggregated by initial wealth and category of consumed items based on the oddsweighted regression analysis (Hirano et al., 2003).
The results re-confirm the significance of consumption effects, ranging from 46% for relatively non-poor households, to 56% for poorer ones.The greatest difference between poor and non-poor households was with respect to food items; although, impact heterogeneity across wealth strata was generally not statistically significant.
blocks (Becker and Ichino, 2002).Estimation of the propensity score and generation of balancing tests were achieved through a combination of psmatch2 (Leuven and Sianesi, 2003), pscore and pstest (Leuven and Sianesi, 2003) procedures in Stata.

9
A key identifying assumption for the PSM is that there should be no unobserved factors that influence both participation and the outcome variable.This is variantly called in the literature as the conditional independence assumption (CIA), matching on observables, unconfoundedness, etc. 'Hidden bias' would be of concern if this assumption is violated (Rosenbaum and Rubin, 1983;Caliendo and Kopeinig, 2008;Jalan and Ravallion, 2003;Cameron and Trivedi, 2005).Test of statistical significance of mean differences between treatment and control/comparison households: *** p<0.01, ** p<0.05, * p<0.1 .Dependency ration was computed as the ratio of inactive members to household size.Asset wealth index was computed with principal components analysis as in Filmer et al. (2001).Villages 1, 3 to 6, and 9 had very low frequencies.Thus, they were grouped together into dvilr.Source: Data from estate horticulture worker survey (2009).

Conclusions
Poverty is widespread in low-income countries like Zambia.Encouragement of land and labor intensive industries such as export horticulture is seen by many as one way to reduce poverty.This study determined the impact of large-scale export horticulture on the welfare of the employees.Data were from a survey of rural households around one of the four major large-scale export horticultural farms about 25 km west of Lusaka.The results, based on eight alternative econometric specifications, consistently point to the existence of huge and positive consumption effects.On average, as much as 44 to 56% of the workers' per capita consumption expenditure could be attributed to their participation in the export horticultural industry.The impact was found to be greater for households that were poor to start with and especially with respect to food consumption; although, statistically, such differences were not significant.As the industry is undergoing structural transformation from contract farming towards consolidation, these results suggest that export horticulture could still play an

Figure A1 .
Figure A1.Distribution of propensity scores over comparison and treatment households.Notes: Common support requirement was satisfied within (0.070, 0.855).

Figure A2 .
Figure A2.Expenditure patterns for comparison and treatment households.

Table 4 .
Heterogeneous impact estimates based on odds-weighted regression analysis.poverty reduction.This is somewhat contrary to conventional, and largely anecdotal, arguments, that large-scale commercial farms are exploitative.It also calls for a re-orientation of public sector support and emphasis from enhancement of contract farming alone to a mix of strategies that also include ways to enhance large-scale export production.For example, domestic and export tax regimes that promote large-scale export horticulture could eventually translate into welfare gains for the poor households that live around those farms.