WPS8096 Policy Research Working Paper 8096 Rural Non-Farm Employment and Household Welfare Evidence from Malawi Guigonan Serge Adjognon Saweda Lenis Liverpool-Tasie Alejandro de la Fuente Rui Benfica Poverty and Equity Global Practice Group June 2017 Policy Research Working Paper 8096 Abstract This paper uses nationally representative panel data and methods, dealing effectively with time invariant hetero- a combination of econometric approaches, to explore geneity, coupled with geographical covariate adjustments, linkages between rural non-farm activities (wage and controlling for time varying differences in local market con- self-employment) and household welfare in rural Malawi. ditions and employment opportunities. The results suggest The paper analyzes the average treatment effects and distri- that non-farm wage employment and non-farm self-em- butional effects on participants’ welfare indicators, such as ployment are welfare improving and poverty reducing. households’ per capita consumption expenditures. Then it However, households at the lower tail of the wealth distri- investigates the effects of non-farm activities on the use of bution benefit significantly less from participation than the agricultural inputs, one channel through which non-farm wealthiest. Although the results support the promotion of employment might improve the welfare of rural households. the rural non-farm economy for poverty reduction purposes, Although participation in non-farm activities is not ran- they indicate that targeted interventions that improve poor domly assigned in the data, the identification strategy relies households’ access to high-return non-farm opportunities on fixed effects and correlated random effects estimation are likely to lead to bigger successes in curbing rural poverty. This paper is a product of the Poverty and Equity Global Practice Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at gadjognon@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Rural Non-Farm Employment and Household Welfare: Evidence from Malawi Guigonan Serge Adjognon1, Saweda Lenis Liverpool-Tasie2, Alejandro de la Fuente1, Rui Benfica3 1 The World Bank Group (WBG) 2 Michigan State University (MSU) 3 International Fund for Agricultural Development (IFAD) Keywords: Poverty, Rural Non-Farm Employment, Household Enterprises, Sub Saharan Africa, Malawi. JEL Classification: J43, J71, O13, O55, Q12 Acknowledgements: This paper builds on and expands analysis done for the Malawi Poverty Assessment report (World Bank, 2016). The authors wish to thank comments received throughout the review process of the Poverty Assessment. The findings, interpretations, and conclusions of this paper are those of the authors and should not be attributed to the World Bank Group or its member countries. 1. Introduction The contribution of non-farm activities (such as non-farm wage employment and non-farm enterprises) to household income in Sub-Saharan Africa (SSA) is substantial and has increased over time (Haggblade et al., 2010, Start, 2001, Lanjouw and Lanjouw, 2001, Lanjouw and Shariff, 2004, Reardon et al., 1998). Recent estimates indicate that 44 percent of rural African households (on average) participate in non-farm wage employment or self-employment. The average income share from non-farm sources is 23%, with an overall positive correlation between diversification and GDP per capita (Davis et al., 2014). Consequently, rural non-farm employment (RNFE) has become an essential part of discussions on poverty reduction in rural Africa, being a potential pathway out of poverty for many. Despite the extent and growth in importance of the RNFE in SSA, there are still limited empirical analyses of the welfare effects of the subsector and how this varies across different kinds of rural households. Though evidence of positive correlations between RNFE participation and income exists, many studies are outdated while an ongoing debate still questions whether RNFE improves welfare or if indeed it is the wealthy who are able to engage in RNFE. Consequently, this paper uses nationally representative data and a combination of empirical approaches to examine links between RNFE participation and welfare. We explore a key mechanism through which welfare effects of RNFE are likely to operate in rural SSA and the heterogeneity of such effects across different types of rural households. The paper significantly enriches the discussion on RNFE and rural development in SSA in three main ways. First, we use panel data methods to address time invariant sources of heterogeneity in the estimation of the welfare effects of non-farm employment participation. Bezu et al. (2012) also use panel data from Ethiopia to show positive linkages between RNFE and consumption expenditure growth. But most of the remaining literature on the subject uses cross sectional data, leaving open the important policy question of whether non-farm employment is really a route out of poverty for the millions of poor in rural Africa (Owusu et al., 2011, Ackah, 2013).1 Using a rich nationally representative panel data set from Malawi,2 we are able to use panel data estimation techniques to at least address the endogeneity of RNFE participation due to time invariant unobserved characteristics, such as innate ability. We include a rich set of household and geographical control variables to capture potential time varying confounding factors such as differences in labor market conditions. Second, this paper goes beyond average effects and explores the heterogeneous effects of RNFE participation. Although the average treatment effects are useful measures of the link between participation in non-farm employment and welfare outcomes, they often provide an incomplete view of the relationship. Numerous studies on the determinants of household participation in non-farm activities (over the past two decades) have found evidence of entry barriers for marginalized socioeconomic groups such as women or the poor (Abdulai and CroleRees, 2001, Woldenhanna and Oskam, 2001, Barrett et al., 2001a, Smith et al., 2001, Lanjouw et al., 2001, Reardon et al., 2000). The marginalized groups are either totally excluded or restricted to the least lucrative non-farm activities, despite their relatively greater need for 1 Kijima et al. (2006) would be another exception as they used panel data collected from 894 rural Ugandan households in 2003 and 2005; but they focused more on off-farm labor supply and its determinants, rather than welfare effects. 2 https://www.malawi.gov.mw 2 diversification. These barriers that restrict the participation of marginalized groups in RNFE imply that the size of the welfare effects of participation in non-farm activities might be lower for marginalized groups compared to the more privileged groups with access to high return opportunities. Despite the widely recognized existence of entry barriers, studies that investigate the distributional effects of RNFE participation at different points of the wealth distribution are still very few. Such studies have very important implications in terms of informing the likely policies (and their targeting) that would be effective in maximizing potential benefits of RNFE for rural households. To explore such heterogeneity, Bezu et al. (2012) used subgroup analysis, comparing average consumption growth effects among the poor to the average effects among the non-poor. This paper goes a step further by using a quantile regression (QR) approach to test for heterogeneous welfare effects of RNFE participation on household consumption expenditure at different points of the conditional distribution of household consumption expenditure (Cameron and Trivedi, 2010). Applying the QR approach within a panel framework, we are able to still control for time invariant unobservable factors correlated with welfare and participation in RNFE. Third, this paper revisits the relationship between RNFE participation and agricultural investments, given that rural households are likely to at least partially be involved in agriculture. In addition to the direct effects of income for consumption that RNFE provides for households, RNFE can also serve as a source of cash for investment in agriculture (Adjognon et al., 2016, Oseni and Winters, 2009), or rather take resources (such as family labor) away from agricultural activities (Smale et al., 2016). In rural Malawi, more than 90 percent of households are involved in agricultural crop production activities, which provide about 55 percent of total household income on average. Thus, we go a step further and explore if participation in RNFE has positive (or negative) effects on investments in agricultural inputs. Apart from Holden et al. (2004) who investigated the effects of off-farm income on household welfare and agricultural production, most previous studies have focused either solely on the effect of RNFE on agricultural investments (Oseni and Winters, 2009, Smale et al., 2016) or welfare directly (Ackah, 2013, Owusu et al., 2011, Matsumoto et al., 2006). We explore both in the context of Malawi and go a step beyond Holden et al. (2004) to consider the heterogeneous welfare effect of RNFE for different kinds of rural households. We find that participation in RNFE is associated with higher household per capita consumption expenditure. However, the effects are larger for those at the top of the welfare distribution compared to those at the bottom. Moreover, we find that farm households involved in RNFE appear more likely to invest in inputs purchases. The rest of the paper is organized as follows. Section 2 describes the data while section 3 discusses the econometric and empirical framework used for our analyses. Section 4 presents and discusses the study results. Section 5 concludes with suggestions for policy consideration and future research. 2. Data sources and definition of key concepts Study context This study was conducted in Malawi, a small country located in eastern Africa. Malawi is one of the poorest countries in the world, with an estimated 71 percent of its population living below $1.90 per day of purchasing power parity in 2010 (World Bank, 2016). It was ranked 173 of 188 countries by the 2015 United Nations Human Development Index (HDI). According to the 3 2016 World Development Indicators (WDI), the country had a GNI per capita of US$350 in 2015. It has a relatively young and growing population (17.2 million people in 2015); and is one of the most densely populated countries in the world, with 177 people per sq. km of land area. Similar to most poor countries, the Malawian economy relies primarily on the agricultural sector which contributes over 30 percent of its GDP and employs 85 percent of its workforce (World Bank, 2016). Given the low return from agriculture and the persistent poverty facing the farm households, policy makers in Malawi would find it useful to know whether facilitating households’ movement into new, more productive, and more lucrative activities outside agriculture is a potential way for reducing poverty in Malawi. Data sources This study uses the panel component of the Malawi Third Integrated Households Survey (IHS3) and the Integrated Household Panel Survey (IHPS) implemented through a joint effort of the Government of Malawi through the National Statistical Office (NSO; www.nso.malawi.net), and the World Bank Living Standards Measurement Study – Integrated Surveys on Agriculture (LSMS-ISA) initiative. These are multi-topic surveys with detailed information about households’ characteristics, activities and livelihood, agricultural practices and community-level information. The full balanced panel sample includes 3,104 households successfully tracked and interviewed in both IHS3 in 2010 and IHPS in 2013. The sample frame includes all three geographical regions of Malawi: North, Centre and South. The survey stratified the country into rural and urban strata. The urban stratum includes the four major urban areas: Lilongwe, Blantyre, Mzuzu, and the Municipality of Zomba. All other areas including Bomas are considered as rural areas. In this study, we are interested in livelihood diversification strategies among rural population. Therefore, our analysis uses only the rural subsample, which represents about 70 percent of the sample (2,766 households). The IHPS data are representative at the national, urban/rural and regional levels. We also use auxiliary data such as the consumption aggregates developed by the World Bank poverty and LSMS teams for poverty analysis in Malawi.3 The use of these pre-generated variables allows us to compare and link our analyses to other reports and studies based on the publicly available LSMS data, thereby contributing coherently to the large discussion about livelihoods in rural Malawi, and Africa more generally. Definition of main concepts As suggested by Barrett et al. (2001b), it is important to distinguish between various terminologies such as “off-farm”, “non-agricultural” , and “non-farm” employment, often used synonymously in the literature to describe the RNFE. Using an adaptation of the sectoral classification in Barrett et al. (2011),4 we define non-farm employment in this paper to mean, all 3 The consumption aggregates data as well as a detailed discussion on how each component is calculated are in the dissemination documentation available for download along with the consumption aggregate data from the LSMS website. 4 Barrett et al. (2011) exclude completely the primary sector from non-farm employments. 4 activities outside of crop and livestock production. Agricultural wage employment such as ganyu wage5 is thus excluded from our non-farm employment definition. However, we include other activities of the primary sector such as forestry, hunting, fishing, mining and quarrying, etc. The main reason for excluding agricultural wage employment from our analysis is because it is a special category of employment that attracts generally the poorest, and has been shown to be of limited significance in household income (e.g. Matsumoto et al. (2006) in Eastern Africa). Following the functional classification, we distinguish between non-farm wage (involving a wage or salary contract) and non-farm self-employment (entrepreneurial activity). In line with our definition and classification of RNFE, we consider a household to be participating in non-farm wage employment if at least one member of the household holds or held a non-agricultural job involving a wage or salary contract, in the 12-month period prior to the survey. For non-farm self-employment, participation implies at least one member of the household has owned a business or worked on their own-account during the 12-month period prior to the survey.6 Note that by this definition we do not require the non-farm activity to be the primary activity of the household members. As mentioned by Winters et al. (2009), this avoids underreporting participation rates. However, our definition is not equivalent to the definition used in Winters et al, (2009) as we do not include transfers and remittances, and we keep non-farm wage employment and self-employment separate throughout our analyses. We lay out, in the later section of this paper, the main reasons for keeping both types of non-farm employment separate. 3. Conceptual framework and empirical strategy Participation in rural non-farm activities: Push and pull factors Observed patterns of participation in non-farm activities, such as wage employment and self-employment result from the combination of “pull” and “push” factors. Push factors relate to the need for ex-ante income smoothing strategies in the presence of binding financial constraints and limited risk mitigating solutions. This includes cases where households diversify in order to satisfy the need for cash to finance agricultural activities in the absence of rural financial services, or the need to feed a large household on a limited amount of land in case of crop failure. On the other hand, pull factors relate the desire by economically rational households to take advantage of opportunities generated by the transformation of the rural economy as a whole. Increased agricultural productivity from the use of modern production techniques, coupled with diminishing marginal return of labor in agricultural use, free labor for use in more productive non-farm alternatives. Also, increased urbanization, and income rise as part of the structural transformation underway in most developing countries, generates demand for non-agricultural goods and services, thereby offering remunerative opportunities in the non-agricultural sector for the surplus labor squeezed out from the agricultural sector (Barrett et al., 2001b, Haggblade et al., 2010). While in 5 Ganyu labor is short-term labor hired on a daily or other short-term basis. Most commonly, piecework weeding or ridging on the fields of other smallholders or on agricultural estates. 6 We have preferred using the binary participation variable instead of continuous participation variable (such as household income from non-farm activities, or number of hours spent in non- farm activities) mainly because the continuous forms are more prone to measurement errors. Also the household income from participation is available only for the participating households. Assuming zero for the non-participants is likely to lead to biases. 5 poor agrarian economies, and for poor households, push factors might trigger the need for diversification (Bardhan and Udry, 1999), resource constraints and entry barriers faced by the poor often restrict their participation in high-return non-farm activities (Bezu and Barrett, 2012). Whether it is by necessity or by opportunity, it is a widely held view that participation in non-farm activities is likely to have a positive effect on the welfare of a participant through the above described channels (Owusu et al., 2011, Ackah, 2013, Oseni and Winters, 2009), though strong empirical evidence of this seems to still be lacking in the currently available literature. However, because poor and non-poor may have differing motives for participating in RNFE (opportunity versus necessity), their benefits from participation are also likely to differ. In this paper, we allow for non-farm wage employment and non-farm self-employment to affect household welfare differently. Assets that facilitate engagement in either type of non-farm employment, the timing of their availability, the returns from participation, and the costs imposed on participating households are likely to be different. For example, non-farm self-employment often comes from businesses owned and managed by the household and therefore may offer more flexibility for the household to manage them in parallel with other activities such as agricultural activities without creating too much competition for labor. Wage employment on the other hand may imply more competition with agricultural activities due to a less flexible schedule, but does not require households to make an upfront capital investment. While the data set does not allow us to fully investigate those differences between non-farm wage employment and non-farm self- employment, we still consider it important to maintain the separation between them in our analyses. Following Bardhan and Udry (1999), adapted by Abdulai and CroleRees (2001) we formalize households’ participation in non-farm self-employment and wage employment by assuming that households allocate resources such as time and land across various activities including farm and non-farm activities. Households choose consumption to maximize the following lifetime expected utility function Max = ∑ ( ) subject to the following constraints: Inter temporal budget constraint: = ( + )( − +∑ ( , ; )) (1) Time endowment constraint: ∑ ≤ (2) Non-negativity constraint: ≥ , k=1…K (3) Where T is life expectancy, and is the discount factor. is the amount of labor allocated to activity k at time t. ( , ; ) is the technology constraint that characterizes the returns from investing units of labor in alternative k. X captures household’s individual and location characteristics that influence the returns to labor use in each of the K uses. The first-order conditions for the above maximization problem imply that households allocate labor between K activities in order to equate the marginal utility of allocating one unit of labor to each of them. Mathematically this implies: [ ( ). ′ ( , ; )] = [ ( ). ′ ( , ; )] (4) where –k refers to activities other than k. The household labor allocation decision to activity k takes into account the expected return from that activity, and the maximum expected returns from all the other possible activities. If for any activity, the household’s endowments in human, financial, and physical capitals imply a low expected return from that activity compared to the 6 others, then no labor would be allocated to that activity. This implies that low expected returns from participation in non-farm employment might justify, at least partly, the limited participation observed for the poor relative to the non-poor. Even though, for the poorest, expected marginal utility from investing labor in agriculture is low given their land constraint and limited access to other productive agricultural assets, the low expected returns from the best non-farm employment opportunity they have access to, given their resources, may still be very low. This illustrates why econometric estimations of the impacts of non-farm employment participation needs to take seriously the endogeneity of the participation decision. A similar reasoning may apply when households are allocating labor between non-farm wage employment and non-farm self-employment. Different types of households’ resources might matter for the non-farm self-employment compared to wage employment and therefore will affect households revealed preference for each of them. We investigate empirically, the main factors that determine households’ participation in non-farm wage employment and non-farm self- employment in rural Malawi. Determinants of participation in non-farm employment Consider the following latent variables ∗ which characterize the differential benefit from participation in an activity k relative to the next best alternative. We drop the subscripts k for succinctness. As mentioned above, households’ individual and location characteristics play an important role in determining this differential benefit. We allow for both observable and unobservable household characteristics to affect this decision. Therefore, the latent variable model is specified as follows (Green William, 2000, Wooldridge, 2010, Owusu et al., 2011): ∗ = [ + + > ] (5) where is the observed participation decision of a household at time t. is the vector of explanatory variables included in the model. Assuming a standard normal distribution for the error term yields a Probit model for the participation decision. captures unobservable characteristics of the households such as ability, networks and preferences, that may affect their employment choice and may also be correlated with some explanatory variables such as education. is the vector of parameters of interest. The estimation of this model with pooled Ordinary Least Square (OLS) or Probit or Logit model without taking into account the unobserved parameter will likely lead to inconsistent estimates due to omitted variables. If this was a linear model, we could use a fixed effects (FE) approach to get rid of the unobserved time invariant effects. But in our case, we have non-linear (binary decision) models. The use of the fixed effects method is, thus, problematic because of the incidental parameters problem (Wooldridge, 2010). Instead we use the Mundlak (1978) special case of (Chamberlain, 1982) correlated random effect (CRE) to model the relation between the unobserved time invariant heterogeneity parameter and the explanatory variables , and we get the following unobserved effects Probit response function (Green William, 2000, Wooldridge, 2010): ( = | )= ( + ) (6) = + + , | ~ ( , ) (7) The full model becomes: 7 ( = | , )= ( + + ), (8) / with = /( + ) , =( , , ) Partial effects of variable xj, from the model above, are defined as: ( | , ) = = . ( + + ) (9) The parameters of equations (8) and (9) can be consistently and efficiently estimated using random effects conditional maximum likelihood (or full MLE). However, consistency of the full MLE relies on the conditional independence assumption, thereby ruling out serial correlation in the error terms. In our cases, there are reasons to believe that participation in non-farm employment in the current year is not totally independent from participation in previous years even after controlling for observable characteristics, creating a violation of the conditional independence assumption. An alternative approach uses the pooled binary model or partial (pooled) maximum likelihood estimation. Though such estimator is usually less efficient, it is consistent, even under the violation of the conditional independence assumption. A more efficient estimation is possible using the Generalized Estimating Equation (GEE) technique (Liang and Zeger, 1986). This estimator is also robust to violation of the conditional independence assumption, but panel-robust standard errors should be used for inference. Given its advantage over the full MLE and pooled probit approaches, we use mainly the GEE approach to estimate APEs. The GEE estimator is a more efficient version of the moment based GMM estimator, and is asymptotically equivalent to the Weighted Non Linear Square (WNLS) estimation approach (See Cameron and Trivedi (2005), p790 for a brief discussion). The main approach described above uses single equations and models separately the determinants of participation in non-farm self-employment and non-farm wage employment. This assumes that both decisions are made in complete independence from each other, implying that the residuals terms in both equations have zero correlation (See Green William (2000), chapter 17.5 for a discussion of multivariate Probit models): = , = (10) This does not have to be the case. It is very plausible that farmers make decisions about labor allocation to various activities simultaneously. In which case, there is some efficiency gain in estimating simultaneously both non-farm employment equations. The jointness of the decisions can be justified by the fact that households have a limited amount of labor available, and labor markets are not fully functional. Alternatively, there are probably common unobserved factors that affect participation in both non-farm activities. In the same spirit as seemingly unrelated regressions, a system of binary response equations can be estimated that take into account the potentially positive correlation between the residuals from both equations. Multivariate probit models rely on the joint normal distribution of the errors. Maximum likelihood is applied based in the joint density function. , ~ , (11) 8 While this method is expected to yield efficiency gains compared to the single equations, the joint normality of the residuals from both non-farm employment equations is a very restrictive assumption; because, joint normality is not always guaranteed even when each marginal distribution is normally distributed (Wooldridge, 2010). Under joint normality, bivariate Maximum Likelihood estimator algorithms are available in most statistical packages such as STATA. When we have several equations, simulation-based methods such as the Geweke– Hajivassiliou–Keane (GHK) smooth recursive conditioning simulator are recommended to go around the difficulty of managing a complex likelihood function (Cappellari and Jenkins, 2003). Because of the potential loss of consistency in case of violation of the joint distributional assumption, we maintain single equations as our main approach, but we estimate the bivariate Probit model using traditional bivariate probit Maximum Likelihood as well as the GHK Simulated Maximum Likelihood for robustness purposes. Average treatment effects of participation in RNFE on participants The core equation to be estimated, for each outcome of interest in this analysis, takes the form of the following unobserved effect model:7 = ( + + + + + ) (12) where is the vector of idiosyncratic errors, and is a vector of controls. We allow for an unobservable household heterogeneity effect (in equation 1) to be correlated with the explanatory variables in the model. G(.) is a positive function that links the explanatory variables to the dependent variable Yit. It could be a linear or non-linear function (for example: standard normal density, or Tobit function) depending on the dependent variable. , k=0, 1, 2, 3 are the parameters to be estimated. Our estimation approach uses the fixed effects (FE) approach or the correlated random effects (CRE) approach, as appropriate. Indeed, some of our outcome variables are better represented using non-linear models (e.g. poverty incidence (1/0), poverty gap (0-1), input purchase (1/0), value of input purchases (pile up at zero due to corner solution), and therefore we estimate both the CRE model and the linear FE model and compare results. A weakness of both the FE and the CRE approaches is that they only account for time invariant unobservable heterogeneity. Any remaining time-varying unobservable heterogeneity might still lead to inconsistent estimates unless they are effectively taken into account using appropriate instrumental variables. However, instrumental variable methods suffer the difficulty of finding a good instrument that satisfies both identification and rank conditions (See Wooldridge 2010): (i) they should have no partial effect on the outcome variables and should not be correlated with other factors that affect the outcomes variable; and (ii) they must be related, either positively or negatively, to the treatment indicator. Keeping in mind that “the cure may be worse than the disease” when a weak instrument is used (Baser, 2009), we use the FE and CRE approaches without instrumental variable, but we include a large set of time varying explanatory variables to hopefully capture and proxy for as many sources of heterogeneity as possible. In particular we control for geographical location using dummy variables thereby capturing local market conditions that may affect both the participation in RNFE and also the welfare outcome of interest. Distributional effects of non-farm employment participation: Quantile regression 7 See the list of outcome variables in table 1. 9 Though the average treatment effect is a useful way to summarize the link between participation in RNFE and welfare outcomes, it may hide a lot of heterogeneity in terms of how different groups benefit from non-farm employment participation. Quantile regression methods (Koenker, 2005) can address this limitation by informing about the relationship between non-farm employment participation and outcome variables at different points of the conditional distribution of the outcome (Cameron and Trivedi, 2010). As such, QR methods can inform the distributional effects of engaging in non-farm activities. This is usually overlooked in treatment effects estimation discussions and seems largely absent in the literature on RNFE and welfare effects. Distributional effects have very important implications for informing likely effective policies and the extent to which finer targeting might be necessary to improve effectiveness of policies. Another advantage of quantile regressions over conditional mean regressions is that the former is robust to outliers because quantiles are not sensitive to extreme values of the outcome while the mean is. For values of q between 0 and 1, let’s define the conditional quantile ( (. )| , ) = (.)| , ( ), as the value of outcome Y (HHPCE in our case) that splits the data into the proportions q below, and 1-q above, where F-1(.) represents the cumulative distribution function CDF of potential outcome Y(.). As described by Imbens and Wooldridge (2008) as well as Cameron and Trivedi (2010) we specify the q-th quantile treatment effect as the average difference between quantiles of the two marginal potential outcome distributions: ( )= ( )| ( )− ( )| ( ) (13) Assuming linearity, the following conditional quantile equation is estimated for various quantile of Household Consumption Expenditure Per Capita (Koenker, 2005): ( (. )| , , ) = + + + ( | , , ) (14) where = + is the composite error. We use to represent the vector of treatment variables ( , ). Again, we model the unobserved household time invariant characteristics using the CRE. We then apply pooled quantile regression to the resulting equation.8 Using a generalized version of the least absolute deviation (LAD) estimation approach (Wooldridge, 2010, Green William, 2000), we estimate and graphically present the quantile treatment effects for several quantiles (10th, 20th, 25th, median, 75th, 80th, and 90th) of household per capita consumption expenditures. Estimation strategy and description of key variables Table 1 presents the description and summary statistics of the main outcome variables analyzed in this paper, along with the methods used for estimating their equations. 8 While this way of introducing time averages as regressors is a practical way of introducing CRE into quantile regressions, it is not totally consistent with the unobserved effects quantile model in equation 12. Because the quantile of a sum is different from the sum of quantiles, the independence assumption after including time averages of explanatory variable is still restrictive. We recognize this as a limitation of our estimates. We explore a model without the time averages to check consistency of the results (Wooldridge, 2010). 10 Beyond our two treatment variables describing participation in non-farm wage employment and non-farm self-employment, the first main outcome variable considered is household per capita consumption expenditure (HHPCE) as a direct measure of household welfare. It includes annual aggregate expenditures on food, non-food, durable goods and housing, deflated by spatial and temporal prices indices to adjust for cost of living differences, and adjusted for household size. We prefer this consumption-based measure as opposed to its main competitor (income-based), because consumption is considered less prone to seasonal variations in living standards, especially in rural areas of developing countries. Deaton and Zaidi (2002) mention that consumption aggregates, over a relatively short period of time such as annual, offer a practical advantage over income aggregates by informing more about longer-term living standards. As indicated in Table 1, average HHPCE over both survey years is 57,750MKW.9 The conditional expectation model of HHPCE is a linear model that we estimate using the FE approach. We also estimate the treatment effects on Log HHPCE using the FE, which gives a sense of percentage effects of our treatment variable, and allows interpretation in relative terms. The conditional quantile model is also a linear model in which the quantiles of the conditional distribution of HHPCE are expressed as linear functions of the independent variables (see equation 14). This model includes the same explanatory variables as in the conditional expectation model, and is estimated using generalized least absolute deviation (LAD) approach. As suggested by equations 12 and 14, the independent variables in the estimating equations include primarily our two treatment variables, in addition to which we include the following set of control variables guided by on the basis of theory and the available literature: - household characteristics (such as education, age of head, household composition); - household ownership of productive capital (captured by normalized agricultural assets, normalized landholdings,10 as well as normalized total livestock unit11); - distance to infrastructures (road and markets to capture transaction costs faced by households); - community level characteristics that may affect productive activities (prices of fertilizer in the EA, price of maize in the EA,12 rainfall, temperature, etc.). We also include a time dummy variable in the equation to account for time specific shocks that could affect welfare directly, but also affect RNFE participation. In the CRE model, the time averages of all the explanatory variables are included to capture time invariant unobserved heterogeneity. In order to reduce remaining unobserved characteristics that might affect our estimates of the treatment effects, we include a set of 26 dummy variables to capture district 9 This is equivalent to about 346.5 USD given the prevailing exchange rate (0.006USD/MKW) during the survey period. See http://www.xe.com. 10 An alternative specification includes also a dummy variable for landless households. But adding this variable does not change the results at all because the proportion of landless represents only 7 percent of the sample. So our reported results exclude this variable purposely. 11 Households’ wealth and productive capital indices such as wealth index and agricultural assets index are generated using factor analysis as in FILMER, D. & PRITCHETT, L. H. 2001. Estimating wealth effects without expenditure data—or tears: An application to educational enrollments in states of india*. Demography, 38, 115-132. The index was then normalized such that norm_index = ((index - min(index)) / (max(index)-min(index))). 12 Fertilizer is a key modern input in agriculture production, and maize is the most popular food crop grown and consumed in Malawi. 11 specific factors that may affect welfare but also determine the economic opportunities available in the area and affect welfare. The description of all these variables, as well as the test of covariates balancing between treatment and control groups can be found in Appendices 1 through 3. Those balancing tests indicate significant differences in education, assets endowment, and resource constraints between participants and non-participants in non-farm wage employment and non-farm self-employment. We also notice significant differences between the poor and the non-poor in terms of household covariates (Appendix 4). Second, we investigate whether participation in RNFE influences household consumption expenditure enough to affect poverty status. For the poverty analysis, the class of poverty metrics proposed by Foster, Greer and Thorbecke (FGT, 1984) is used. Poverty incidence is based on the HHPCE, relative to the 85,852 MKW (real 2013 prices) local poverty line per person per year in Malawi. As opposed to the $1.90/day international poverty line, this local poverty line indicates the cost of maintaining a reference welfare level (here defined as satisfying necessary energy and nutritional requirements to have a healthy and active life) to a given person, at a given time in the specific context of Malawi (Ravallion, 1998).13 Household poverty incidence takes value 1 if household consumption falls below the poverty line, and 0 otherwise. The summary statistics presented in table 1 indicate that the proportion of rural households with consumption expenditure below the poverty line is 37 percent over the 2010-13 pooled sample, confirming that poverty is relatively widespread in rural Malawi.14 The main specification used to explore the effect of the RNFE on poverty is a CRE Probit model, estimated using the Generalized Estimating Equation (GEE) approach.15 Poverty gap (defined as the consumption shortfall relative to the poverty line, as a fraction of the poverty line) takes a value 0 for all non-poor households, creating a censoring at 0 and a continuous set of values between 0 and 1. The squared poverty gap (sensitive to extreme poverty) takes continuous values between 0 and 1 with a pile up at zero. On average, the poor in rural Malawi appear to be about 30 percent short of the poverty line, and the squared poverty gap is 0.13. We use the CRE fractional Probit approach to estimate responses of both poverty variables to RNFE participation. As demonstrated by Gallani et al. (2015), Fractional Response Model (FRM) is preferable since it overcomes the limitations of other approaches for the statistical analysis of dependent variables that are bounded in nature and present a significant number of observations at one of the boundary points. For all these three poverty measures, we also estimate the linear FE model to serve as benchmark for comparison as it requires fewer distributional assumptions and generally gives reasonable approximation of APEs (Wooldridge, 2010). The same set of covariates used in the estimation of the HHPCE response model are used for the poverty analyses. Third, we explore effects of RNFE on subjective measures of food security, captured by self-reported measures of food insecurity and food consumption adequacy (see Table 1). The food 13 More details about this can be found in the LSMS documentation for Malawi. 14 This refers to the average proportion of poor rural households in the pooled sample. Over the period 2010 to 2013, the proportion of poor rural households dropped from 40% to 33%. Overall, the proportion of poor rural people, the rural poverty headcount, fell from 44% in 2010 to 41% in 2013. 15 The GEE approach consistently identifies average partial effects (APE); it is robust to violation of the conditional independence assumption as opposed to the full maximum likelihood approach; it is more efficient than pooled Probit approach but clustered robust standard errors have to be used. (See Wooldridge (2002) for a full discussion.) 12 insecurity variable is a binary variable taking value one if the household head responded yes to the question “Did you worry that your household would not have enough food in the past 7 days?” This is modeled using the CRE Probit approach and estimated using GEE technique. The food consumption adequacy variable is an ordered response variable describing household food consumption over the past month on a scale of 1(less than adequate for household needs); to 3 (it was more than adequate for household needs). It is modeled using the CRE ordered Probit approach with maximum likelihood estimation approach. The ordered probit model enables us to account for the ordered nature of the food insecurity measure as well as time invariant unobserved household characteristics (See Wooldridge 2010 for details about ordered response models). For both these variables, we also estimate the linear FE for comparison purpose. Again, the same set of control variables are included in the model as in the HHPCE and poverty equations. Finally, we turn our attention to agricultural investments. Several studies in the RNFE literature have explored agricultural investment as a pathway through which non-farm employment participation can increase welfare, and concluded a positive effect. For example, Oseni and Winters (2009) concluded a positive effect of both non-farm wage employment and non-farm self- employment participation on Nigerian farmers’ agricultural investments (especially labor and fertilizer). Other studies such as Smale et al. (2016) found negative effects on fertilizer application (specifically nitrogen application on maize by Kenyan farmers), and concluded that engagement in the RNFE could be a distraction with trade-offs in labor allocation and farm investments. We investigate, in the case of Malawi, the effects of non-farm employment on agricultural investments. We focus on household purchases of fertilizer, and all inputs, as well as area of land cultivated. The effects of non-farm employment on the binary purchase decisions are first considered with a CRE Probit model and GEE estimation technique. Then, the effects on the quantity of inputs/fertilizer purchases per acre of land cultivated are modeled with a CRE Tobit approach and estimated using partial Maximum Likelihood. The linear FE estimation of partial effects is used for comparison purposes. 4. Results and discussion 4.1. The Rural Non-Farm Economy in Malawi Like many other parts of the developing world, the rural economy in Malawi is transforming as its economy grows. A sign of such transformation is the increasingly important place that non-farm activities are occupying for households in rural Malawi. Though, agriculture remains the main source of income for most households, participation rates in non-farm wage and self-employment, as well as income shares from those activities, have reached non-negligible levels. Table 2 shows that the proportion of households earning income from non-farm self- employment was about 18 percent in 2010, and 27 percent in 2013. Meanwhile, average households’ income share coming from non-farm self-employment was estimated at about 7 percent and 11 percent in 2010 and 2013 respectively. As for non-farm wage employment, participation rates sat around 17 percent and 15 percent, in 2010 and 2013 respectively, with associated income shares around 8 and 6 percent on average. These estimates are about average compared to other Sub-Saharan African (SSA) countries. The estimation by Davis et al. (2014) for a group of SSA countries, indicates on average 15 percent participation rate for non-farm wage employment and 34 percent for non-farm self-employment, with associated income shares of 8 percent and 15 percent. 13 Consistently with the existing literature, the data from Malawi suggest a divide in participation in the non-farm economy along poverty line. Our descriptive statistics by poverty status indicate that, while participation in self-employment increased for both the poor and the non-poor between the two years, participation rates for the poor are significantly lower than for the non-poor households in both years. In 2013, about 20% of the poor were engaged in self- employment activities, up from 15% in 2010. Meanwhile the same statistic for the non-poor increased from 20% in 2010 to 30% in 2013. The difference between poor and non-poor is even sharper for wage employment, where the participation rates for the non-poor are more than double the participation rates among the poor in both survey years. The lower gap observed for non-farm self-employment compared to wage employment likely reflects differences in entry barriers between the two types of employment. Non-farm wage employment jobs require more education and better social networks. While the lower participation rates among the poor call for further attention to participation barriers, the fact that participation rates among the poor are still almost 20% justifies the need to understand if and how such households benefit from non-farm activities. In a more detailed analysis, participation rates and returns by sectors of non-farm self- employment (Table 4) and non-farm wage employment (Table 5) for the poor and non-poor confirms further patterns of dualism. Our descriptive analyses indicate the existence of high and low return sectors of non-farm activities and the high return sectors are almost exclusively available to a handful of privileged, due to entry barriers that prevent the more marginalized groups (the poor and sometimes women) from accessing those opportunities. In the following paragraphs, we describe the main features and types of non-farm employment in which rural households in Malawi are involved, by poverty status. Non-farm self-employment in rural Malawi Table 3 summarizes the main traits of non-farm businesses observed in rural Malawi. As most rural non-farm enterprises in SSA, described by Nagler and Naude (2014), the characteristics of rural non-farm enterprises in Malawi are consistent with little potential for job creation as they are mostly informal, have low productivity, and short life spans. Almost half of all the non-farm businesses in rural Malawi are reportedly operated from home, with only 13 percent having access to electricity, and 6 to 8 percent with some formal registration. Business owners are relatively young (38 to 39 years old) and mostly uneducated (75% of business owners have no formal education). Businesses owned by poor household members are less likely to be registered formally compared to those owned by non-poor household members. Also poor business owners appear less educated on average compared to non-poor business owners. These differences might well affect productivity and returns to participation for the poor. As for the types of household enterprises operated, participants in non-farm self- employment are involved in a variety of sectors or industries.16 The distribution of non-farm self- 16 The classification of non-farm enterprise activities into industry categories used here closely follows the 1992 United Nations International Standard Industrial Classification (ISIC) standards into 5 main groups. The groups include: (a) Primary sector, which comprises agriculture, livestock, hunting, fishing, and mining; (b) Food, Beverage, and Tobacco Manufacturing; (c) Non-food Manufacturing, (d) Commerce and Tourism (wholesale and retail, and restaurants and 14 employment by sector, as well as average profit17 earned by participants, for the poor, and the non- poor are summarized in Table 4. The analysis of table 3 indicates that about half of household enterprises in rural Malawi are in the commerce and tourism sector (wholesale, retail trade, restaurants, and hotels). These are typically small businesses involving people selling or reselling a wide variety of products from groceries and food products to clothes, shoes, etc., and earning on average 6,000 to 7,000 MKW18 (or 8 to 10USD) monthly. The second most prominent sector is the manufacturing sector (food and non-food combined), which accounts for approximately 40 percent of all household enterprises. The manufacturing sector is dominated by food, beverage, and tobacco manufacturing, which accounts for more than half of the manufacturing sector, represents more than 25 percent of all household enterprises, and includes primarily street vendors of various food and drinks and making a profit in the range of 4,000 to 6,000MKW (or 5.5 to 6USD) monthly. In general, returns to non-farm self-employment are pretty low (around 6,000MKW ~ 8USD) in rural Malawi, and are lowest in the most popular sectors (Table 4). The commerce and tourism sector, which contains about half of the non-farm businesses, generates on average 6,000 to 7,000KW (8 to 10USD) monthly, as opposed to 9,000 to 12,000 KW (12 to 17USD) of profit reported in the construction and services sector, which represents only 9 to 10 percent of non-farm enterprises. The manufacturing sector, second most important sector generates even lower revenues. Returns from the Food, beverage, and tobacco manufacturing sector are about 3,000MKW monthly, and the non-food manufacturing sector generates about 5,000 to 6,000 KW monthly. With the sectors of non-farm enterprises defined here, no evidence of dualism becomes particularly apparent from simply observing the distribution of non-farm enterprises by sectors, for the poor versus the non-poor. However, a Kolmogorov-Smirnov test of equality of distribution does indicate a significant difference (at 1%) between poor and non-poor. We do observe a large heterogeneity in returns to participation within each category of non-farm enterprises. And profits earned by the poor are significantly lower compared to the non-poor across all categories of self- employment. The t-test of mean difference in profit (Table 4) show p-values less than 1% for most categories of self-employment. Non-farm wage employment in rural Malawi Non-farm wage employment activities are also distributed across several sectors19 presented in Table 5. The service sector appears to be the largest with about a third of all non- hotel businesses); and (e) Other sectors, which include construction, electricity and utilities, transportation, and other services. 17 This is a direct measure self-reported by the respondent. The reference period for profits is the latest month of operation prior to the interview. Might suffer measurement errors. Therefore, should serve only as rough estimation. 18 Kwachas (MWK), local currency used in Malawi. 1 USD = 723.4319 MWK according to currency.me.uk. 19 We follow the occupation codes used in the Malawi LSMS survey instruments, which includes: Relatively skilled labor jobs such as: (1) Professional, technical, & related workers; (2) Administration and managerial workers; (3) Clerical and related worker; (4) Sales workers. And 15 farm wage employment, closely followed by the category of transport equipment operators, and laborers not elsewhere classified which represents approximately a quarter of all wage employment. Then comes the sector of Professional, technical, & related worker groups, representing also about a quarter of the rural non-farm wage employment. Together, the remaining sectors represent each less than 7% of non-farm jobs in both 2010 and 2013. Notice that approximately 75% of all rural non-farm jobs are in sectors that do not require highly skilled labor. Wages in those sectors are significantly lower than wages in sectors that require more skilled labor. As a result, the average monthly wage across all non-farm jobs in rural Malawi is pretty low, around 8,000 Kwachas (~ 11USD). Dualism in the categories of non-farm wage employment is obvious compared to non-farm self-employment. The skilled labor jobs, which generate the highest returns, are almost exclusively accessible to the non-poor while the non-skilled labor jobs appear available for all. Table 5 indicates that about 80 percent of the jobs taken by the poor are in the non-skilled labor sectors, compared to about 60% for the non-poor. In addition, and non-surprisingly, the poor earn significantly less than the non-poor in most sectors of non-farm wage employment. Monthly wage earned by participants in non-farm wage employment, and t-test of difference in wages between the poor and non-poor by category of employment, reveals significantly lower wage for the poor. The only sector in which we fail to reject the absence of a significant difference between wages earned by poor and non-poor is the service sector, which is the most popular sector for both poor and non-poor. 4.2. Determinants of participation in non-farm activities Table 6 summarizes the estimation results of the determinants of participation in non-farm wage and self-employment in rural Malawi based on the unobserved effects Probit model presented in equation 8 and 9. Several important points emerge from those results. Education is an important driver of non-farm wage employment participation, but not a significant determinant of non-farm self-employment participation. Table 5 shows that the average partial effect of household head education increases with the level of education until MSCE level, beyond which educational attainment level does not appear significant anymore, probably due to the low proportion (1.3%) of people in that education level category in the rural Malawi. This is consistent with several previous findings such as in Oseni and Winters (2009), Winters et al. (2007) as well as Lanjouw and Shariff (2004), and De Janvry and Sadoulet (2001), though they have often lumped non-farm wage employment and self-employment together in their analyses, thereby losing out on some nuances between the two types of activities. Indeed, education may play a double role, influencing households’ participation in non-farm-self-employment and wage employment. On one hand, since returns to non-farm wage employment are higher in general than non-farm self-employment (See Table 4 and Table 5), the more educated may occupy the few non- farm wage employment, leaving the non-farm self-employment sector for the least educated. This would imply a negative effect of education on non-farm self-employment participation. On the other hand, those who are more educated also have higher expected returns from participation in non-farm self-employment due to skills acquired from their formal training. relatively unskilled labor jobs such as: (1) Service workers; (2) Agricultural, animal husbandry and forestry workers, fishermen and hunters; (3) Other sectors including production and related workers, transport equipment operators and laborers not elsewhere classified. 16 We find that access to credit is an important determinant of households’ participation in non-farm self-employment. More precisely, access to credit for relevant households increases participation in non-farm self-employment by about 8 to 10 percentage points on average, ceteris paribus (see table 6). These results challenge the common view that rural household enterprises require little to no capital investment and thus can provide a source of cash for the poorest who are lacking access to financial services in presence of rural credit market failures (Poulton et al., 2006, Bardhan and Udry, 1999). From our analysis, it appears that viable household enterprises in the context of rural Malawi require some level of investments, significant enough to preclude the poorest from engaging in those activities. Wealth of the household (captured by a normalized asset index) significantly increases participation in both non-farm wage employment and self-employment. This is consistent with the findings in Oseni and Winters (2009), and add to reasons why the poorest have a more restricted access than the non-poor to non-farm employment opportunities. Assets such as tables, sewing machine, TV, bicycle, etc., can be used to operate household enterprises, commute to a non-farm wage or business place, etc., thereby increasing households’ likelihood (and may be expected returns) of participating in non-farm wage and self-employment. In addition, asset ownership is correlated with social status in most rural areas in Africa. So households with more assets may have stronger networks through which they can receive information about non-farm wage employment opportunities. Depending on the type of business, households with large networks may also have higher expected returns from non-farm enterprises, which increases participation. Besides, our results reveal generally stronger evidence of a pull scenario explaining the participation in non-farm activities in rural Malawi. The participation rate in non-farm wage employment, among neighboring households of the same geographical area, affects, positively and significantly, each household’s participation in both non-farm wage employment and non-farm self-employment. Similarly, the proportion of other households owning a business in the same geographical area significantly and positively affects each household’s likelihood of participation in non-farm self-employment and non-farm wage employment. This is consistent with the pull scenario described in Haggblade et al. (2010), which stresses that the economic environment greatly determines access and participation in non-farm activities. Having a lot of households participating in non-farm wage employment or non-farm self-employment in the same locality may reflect a more vibrant economic environment with more jobs or business opportunities, which therefore increase participation in those employment categories. From a social network effects point of view, this also corresponds with what was dubbed correlated neighborhood effects by (Manski, 1993). People behave like one another when they face similar shocks or environment. However, the observed correlation might also suggest the existence of social effects (endogenous and contextual peer effects), which imply that households’ decision to participate in non-farm wage activities depend on their peers’ behaviors and characteristics. Disentangling these effects is beyond the scope of this paper. However, the findings that social network effects might influence participation in non-farm self-employment (similarly to peer effects in agricultural technology adoption) are intrinsically interesting and deserve further investigation. In the social network literature and technology adoption literature, Bramoulle et al. (2009) followed by Krishnan and Patnam (2014) addressed the Manski reflection problem inherent to the identification of peer effects by using average characteristics of neighbors as instruments for participation rates among neighbors. On the estimation methods, we find large consistency between our linear model results and the CRE Probit estimation results. The APE estimates have usually similar magnitudes as well as 17 consistent signs and significance levels in both the non-farm wage employment and non-farm self- employment cases. More importantly, the results from the more efficient multivariate system estimates do not add any particular insight (see appendix 5). While the Simulated Maximum Likelihood (SML) estimates and the Bivariate Probit Maximum Likelihood (ML) estimates both do indicate a negative correlation (ρ=-0.15) and significant (at 1%) between the residuals of the non-farm wage employment and non-farm self-employment models, the coefficient estimates are remarkably consistent in signs and significance with the main Probit results described above. This robust consistency of our estimates across methods increases our confidence in these results, which indicates that policies geared to increase rural household participation in non-farm employment might want to focus on factors such as education and access to credit for the poor, in addition to improving infrastructure and growth motors. 4.3. Impacts of participation in the RNFE Impact on HHPCE Does participation in non-farm activities increase households’ welfare, ceteris paribus? Are the poverty reduction motives of policies promoting the RNFE in developing countries justified? Our APE estimation results, summarized in Table 7, indicate a consistently positive response to both questions. Appendices 6 through 11 report the full set of results. We discuss below the main points that come out of those results. First, non-farm wage employment participation increases households’ per capita consumption expenditure by a margin of 4,500MKW, though only significant at 16%. The log linear model indicates a 10 percent increase effects, statistically significant at 5%. As for non-farm self-employment, the APE estimates on the levels of HHPCE is 7,000MKW corresponding approximately to a 13 percent increase in HHPCE, and statistically significant at 1%. These numbers are not directly comparable with previous studies such as Owusu et al. (2011) and Ackah (2013) since they lumped both types of non-farm activities together and also focused on different outcome variables (notably household income instead of consumption expenditure). However, they confirm the general conclusion of a positive effect of non-farm employment participation on direct measures of household welfare. Besides, the t-test of comparison of coefficients of non-farm self-employment and non-farm wage employment indicates, at 1% significance level, that the effect of non-farm self-employment on participants is higher on average than the effects of non- farm wage employment participation. While we find positive effects of participation in both non-farm wage employment and non-farm self-employment, we also find significant heterogeneity of HHPCE response to participation in the RNFE. Everything else held constant, households at the bottom of the wealth distribution seem to benefit significantly less from participating in non-farm wage employment or in non-farm self-employment than households at the top of the distribution (Table 8 and Figure 1). Though the effects of engaging in non-farm wage employment or non-farm self-employment remain positive for all classes of the welfare distribution, there is a generally increasing trend in the size of the effect as we go from lower percentiles to the top of the distribution of HHPCE. The quantile effects of non-farm self-employment go from a low of about 2,300MKW for the 10th percentile of HHPCE, and increases to a high of over 16,000MKW for the 90th percentile (almost 10 times the effect on the lowest quantile). As for non-farm wage employment, the effect is not significant for the 10th and 20th percentiles. It starts being significant for the 25th percentile, from 18 a low of 4,600MKW to a high of about 6,500 MKW for the 75th and 80th percentiles, before dropping again to a non-significant size of 2,000MKW at the 90th percentile. This indicates that, for non-farm wage employment, the middle segments of the population benefit the most; while, for non-farm self-employment, the upper segments benefit the most from participating. However, in both cases the poorest appear to benefit the least from participating, confirming the findings in (Bezu et al., 2012). This heterogeneity of effects also illustrates the fact that resource constraints faced by the poorest likely consign them to low return activities with less potential to increase their income considerably, potentially putting them into a poverty trap. Meanwhile the middle and upper segments earn significantly higher returns from participating in non-farm activities because they have access to the most lucrative employment opportunities. These results are largely consistent with our descriptive analysis, and also support the evidence in the existing literature on the existence of entry barriers that confine marginalized groups to low returns and non-lucrative employment categories (Abdulai and CroleRees, 2001, Woldenhanna and Oskam, 2001, Barrett et al., 2001a, Smith et al., 2001, Lanjouw et al., 2001, Reardon et al., 2000). Impacts on poverty and food security We extend beyond HHPCE to analyze several other welfare outcomes. Our results indicate that both non-farm wage employment and non-farm self-employment have a negative and significant effect on all three FGT poverty metrics. Both the fractional Probit model and the linear FE model show that, on average, a household’s engagement in non-farm wage employment reduces the likelihood of its consumption expenditure falling below the poverty line by 7 percentage points. Similarly, self-employment reduces poverty incidence by a statistically significant margin of about 8.5 percentage points. This is an important margin relative to the pooled sample rural household poverty incidence of about 37 percent (see summary statistics in Table 1). For the depth and severity of poverty measured by poverty gap and squared poverty gap respectively, our CRE fractional Probit, and linear FE models produce consistent results indicating significant and negative effects of both non-farm wage employment and nonfarm self- employment. This implies that though the poorest seem to benefit less from engaging in non-farm activities, the size of the effect is still enough to reduce significantly the depth and severity of poverty confirming that the rural non-farm sector can well serve as a pathway out of poverty for poor rural households in Malawi. The positive effect of non-farm activities persists in the analysis of household subjective perceptions of food security. In particular, estimates from the CRE Probit model and linear FE model consistently indicate that participation in non-farm wage employment reduces the likelihood that a household feels food insecure by about 5 to 6 percentage points. Similarly, the ordered Probit results show that households engaged in non-farm wage employment perceive their food consumption level to be more adequate for the needs of the household, compared to non- participating households. For non-farm self-employment, the estimated effect does not appear statistically significant. RNFE and agricultural investment Several studies in the RNFE literature have explored agricultural investment as a pathway through which non-farm employment increases welfare, and concluded a positive effect. For example, Oseni and Winters (2009) concluded a positive effect of both non-farm wage 19 employment and non-farm self-employment participation on Nigerian farmers’ agricultural investments (especially labor and fertilizer). Other studies such as Smale et al. (2016) found negative effects on nitrogen application on maize by Kenyan farmers, and concluded that engagement in the RNFE could be a distraction with trade-offs in labor allocation and farm investments. Our results (at bottom panel of Table 7) indicate that participation in non-farm self- employment increases likelihood (but not intensity) of inputs use, especially inorganic fertilizer purchases among rural households in Malawi by about 5 percentage points. Though restricted to the purchase decision only, these findings tend to align more with the strand of existing literature finding evidence that RNFE reduces the cash constraint for rural households at least partially engaged in agriculture such as Owusu et al. (2011) and Ackah (2013) as well as Barrett et al. (2001b) on the subject. We do not find a statistically significant effect of wage employment on input purchases. The lack of significance of the effects of non-farm wage employment could imply that the negative effect on agriculture due to labor displacement counters the positive effect on financial constraints for agricultural investments. We also do not find any effects of RNFE on the area of land cultivated in the season. It is possible that decisions to participate in non-farm employment and decision to use inputs are made simultaneously, and therefore a system estimation of non-farm employment participation and inputs use equations would improve the efficiency of the estimates; though identification relies on joint normality of the residuals from the equations in the system. As an additional robustness check, we estimate a series of recursive trivariate Probit models as an alternative way to explore the effects of non-farm employment participation on inputs purchase decisions (Appendix 11). In each system, we have one equation for each treatment variable (wage- employment and self-employment), and one equation for the outcome variable (fertilizer purchase, or inputs purchase), making a total of three equations. The equation for the outcome variable includes both treatment variables as explanatory variables, while the equations for the treatment variables do not include the outcome variable as explanatory variable. This recursive formulation (see equation 17-49 in Green William (2000)) implies a one-way causality between the treatments and outcome variables. A similar approach was used by (Smale et al., 2016) to investigate the same question in Kenya. We estimate this system, for input purchase and for fertilizer purchase decisions using GHK Simulated Maximum Likelihood (Cappellari and Jenkins, 2003). The estimated coefficients are consistent with the single equation results presented above suggesting a positive effect of non-farm employment participation on inputs purchase decision in general, and fertilizer purchase in particular. More interestingly while the effect of non-farm wage employment was not significant in the single equation models, the more efficient system estimations indicates a rather positive and significant effect of non-farm wage employment on inputs purchase decisions as well. The likelihood ratio tests of joint significance of the residuals from the equations are always significant at 1%, justifying the relevance of a system estimation approach to improve efficiency in our case. Overall, the study results indicate positive welfare effects of RNFE on objective and subjective measures of household welfare and these findings are consistent across different estimation techniques. The results are also largely consistent with the existing literature. The magnitude of the welfare effects of self-employment is on average larger than that from wage- employment. One mechanism through which RNFE likely improves household welfare is through the provision of funds for investment in agricultural production; the mainstay of most rural households. Finally, the quantile regression analysis reveals considerable heterogeneity in the 20 welfare effects of RNFE, with the poor benefitting significantly less than the non-poor from both non-farm wage-employment and non-farm self-employment. 5. Conclusion This paper makes several contributions to the debate about the poverty reduction potential of the rural non-farm employment sector. The paper first revisits the determinants of participation in non-farm wage employment and non-farm self-employment, using recent and nationally representative panel data from Malawi. The results offer little support for the early view that the poor and least endowed households are more likely to participate in non-farm activities, in order to insure against production risks and meet capital constraints (Bardhan and Udry, 1999). Our results rather indicate the presence of significant barriers to participation, faced by the poorest. These barriers include human, financial, and physical constraints such as education, assets and credit. Different sets of resources seem to matter for participation in either type of non-farm activity. Educated and wealthy households are pulled into non-farm wage employment, while participation in non-farm self-employment is mostly driven by access to credit and wealth. Market access and proximity to better infrastructure and opportunities are important for both non-farm wage employment participation and non-farm- self-employment. Then, using panel data estimation techniques, we find consistent evidence that the RNFE can serve as a potential mechanism to increase welfare and reduce poverty in rural Malawi. Our results are consistent across a wide range of econometric approaches and broad suite of objective and subjective welfare and poverty measures. We conclude that while any causal interpretation of the usually positive correlation between non-farm engagement and wealth requires some care, the context of rural Malawi offers consistent evidence of a positive welfare impact of participation in non-farm activities, and thus the RNFE can be considered as a pathway out of poverty for rural households. However, our results suggest that more attention should be paid to improving not just access, but also the quality of the non-farm opportunities available for the poorest. Though on average, participation in non-farm wage-employment and non-farm self-employment has a consistently positive effect on welfare, we find strong evidence that even when the poor participate in the RNFE, they benefit significantly less than the non-poor. The existing literature and observations from our data indicate that human, physical, and financial capital are key factors that limit participation in non-farm employment for the poor in rural areas. Addressing dualism along poverty and gender lines as well as other dimensions, is of utmost importance if the RNFE is to be an effective tool for broad-based rural development. Failing to take this into consideration will likely lead to widening inequality instead of shared prosperity. Moreover, given the notable heterogeneity of activities within non-farm employment, future research should delve even deeper into these different types of activities and explore participation in and roles of various sub- categories of non-farm wage employment and non-farm self-employment for household welfare. 21 REFERENCES ABDULAI, A. & CROLEREES, A. 2001. Determinants of income diversification amongst rural households in Southern Mali. Food policy, 26, 437-452. ACKAH, C. 2013. Nonfarm employment and incomes in rural Ghana. Journal of International Development, 25, 325-339. ADJOGNON, S. G., LIVERPOOL-TASIE, L. S. O. & REARDON, T. A. 2016. Agricultural input credit in Sub-Saharan Africa: Telling myth from facts. Food Policy. BARDHAN, P. & UDRY, C. 1999. Development microeconomics, Oxford University Press. BARRETT, C. B., BEZUNEH, M. & ABOUD, A. 2001a. Income diversification, poverty traps and policy shocks in Côte d’Ivoire and Kenya. Food Policy, 26, 367-384. BARRETT, C. B., REARDON, T. & WEBB, P. 2001b. Nonfarm income diversification and household livelihood strategies in rural Africa: concepts, dynamics, and policy implications. Food policy, 26, 315-331. BASER, O. 2009. Too much ado about instrumental variable approach: is the cure worse than the disease? Value in Health, 12, 1201-1209. BEZU, S. & BARRETT, C. 2012. Employment dynamics in the rural nonfarm sector in Ethiopia: Do the poor have time on their side? Journal of Development Studies, 48, 1223- 1240. BEZU, S., BARRETT, C. B. & HOLDEN, S. T. 2012. Does the nonfarm economy offer pathways for upward mobility? Evidence from a panel data study in Ethiopia. World Development, 40, 1634-1646. CAMERON, A. C. & TRIVEDI, P. K. 2005. Microeconometrics: methods and applications, Cambridge university press. CAMERON, A. C. & TRIVEDI, P. K. 2010. Microeconometrics using stata, Stata Press College Station, TX. CAPPELLARI, L. & JENKINS, S. P. 2003. Multivariate probit regression using simulated maximum likelihood. The Stata Journal, 3, 278-294. CHAMBERLAIN, G. 1982. Panel data. National Bureau of Economic Research Cambridge, Mass., USA. DAVIS, B., DI GIUSEPPE, S. & ZEZZA, A. 2014. Income diversification patterns in rural sub- Saharan Africa: reassessing the evidence. World Bank Policy Research Working Paper. 22 DEATON, A. & ZAIDI, S. 2002. Guidelines for constructing consumption aggregates for welfare analysis, World Bank Publications. FILMER, D. & PRITCHETT, L. H. 2001. Estimating wealth effects without expenditure data— or tears: An application to educational enrollments in states of india*. Demography, 38, 115-132. GALLANI, S., KRISHNAN, R. & WOOLDRIDGE, J. 2015. Applications of Fractional Response Model to the Study of Bounded Dependent Variables in Accounting Research. Harvard Business School Accounting & Management Unit Working Paper. GREEN WILLIAM, H. 2000. Econometric analysis. Forth Edition, Prentice Hall International, Inc, New York University. HAGGBLADE, S., HAZELL, P. & REARDON, T. 2010. The rural non-farm economy: Prospects for growth and poverty reduction. World Development, 38, 1429-1441. HOLDEN, S., SHIFERAW, B. & PENDER, J. 2004. Non-farm income, household welfare, and sustainable land management in a less-favoured area in the Ethiopian highlands. Food Policy, 29, 369-392. KOENKER, R. 2005. Quantile regression, Cambridge university press. LANJOUW, J. O. & LANJOUW, P. 2001. The rural non‐farm sector: issues and evidence from developing countries. Agricultural economics, 26, 1-23. LANJOUW, P., QUIZON, J. & SPARROW, R. 2001. Non-agricultural earnings in peri-urban areas of Tanzania: evidence from household survey data. Food policy, 26, 385-403. LANJOUW, P. & SHARIFF, A. 2004. Rural non-farm employment in India: Access, incomes and poverty impact. Economic and Political Weekly, 4429-4446. LIANG, K.-Y. & ZEGER, S. L. 1986. Longitudinal data analysis using generalized linear models. Biometrika, 13-22. MATSUMOTO, T., KIJIMA, Y. & YAMANO, T. 2006. The role of local nonfarm activities and migration in reducing poverty: evidence from Ethiopia, Kenya, and Uganda. Agricultural Economics, 35, 449-458. NAGLER, P. & NAUDE, W. 2014. Non-farm enterprises in rural Africa: new empirical evidence. World Bank Policy Research Working Paper. OSENI, G. & WINTERS, P. 2009. Rural nonfarm activities and agricultural crop production in Nigeria. Agricultural Economics, 40, 189-201. 23 OWUSU, V., ABDULAI, A. & ABDUL-RAHMAN, S. 2011. Non-farm work and food security among farm households in Northern Ghana. Food policy, 36, 108-118. REARDON, T., STAMOULIS, K., BALISACAN, A., CRUZ, M., BERDEGUÉ, J. & BANKS, B. 1998. Rural non-farm income in developing countries. The state of food and agriculture, 1998, 283-356. REARDON, T., TAYLOR, J. E., STAMOULIS, K., LANJOUW, P. & BALISACAN, A. 2000. Effects of non‐farm employment on rural income inequality in developing countries: an investment perspective. Journal of agricultural economics, 51, 266-288. SMALE, M., KUSUNOSE, Y., MATHENGE, M. K. & ALIA, D. 2016. Destination or Distraction? Querying the Linkage Between Off-Farm Work and Food Crop Investments in Kenya. Journal of African Economies. SMITH, D. R., GORDON, A., MEADOWS, K. & ZWICK, K. 2001. Livelihood diversification in Uganda: patterns and determinants of change across two rural districts. Food Policy, 26, 421-435. START, D. 2001. The rise and fall of the rural non‐farm Economy: Poverty Impacts and Policy options. Development policy review, 19, 491-505. WOLDENHANNA, T. & OSKAM, A. 2001. Income diversification and entry barriers: evidence from the Tigray region of northern Ethiopia. Food Policy, 26, 351-365. WOOLDRIDGE, J. M. 2010. Econometric analysis of cross section and panel data, MIT press. WORLD BANK. 2016. Malawi Poverty Assessment. World Bank, Washington, DC. 24 TABLES & FIGURES The Heterogeneous Welfare Effects of Rural Non-Farm Employment: Recent Evidence from Malawi 25 Table 1: Description and summary statistics of the main variables used in this paper, household level, 2010-2013, rural Malawi Variables Definition Types of variables Overall SD N Estimation approach Mean Treatment variables Non-farm wage employment A member of household is engaged in non-farm wage employment 15.11 35.82 5532 GEE CRE Probit (0/1) Non-farm self-employment A member of household is engaged in non-farm self-employment 21.58 41.14 5532 GEE CRE Probit (0/1) Outcome variables HHPCE Household total real consumption expenditure per capita Continuous 57,759.59 46,702.55 5532 FE, CRE, Quantile regression Log of HHPCE Log of Household total real consumption expenditure per capita Continuous 10.75 0.64 5532 FE, CRE Poverty incidence 1[HHPCE<=Poverty Line] Binary (0/1) 37.20 48.34 5532 GEE CRE Probit, Linear FE Poverty gap Consumption shortfall relative to poverty line as a fraction of poverty line Fractional [0, 1] 11.2 18.8 5532 GLM CRE FRM, Linear FE Poverty severity Squared poverty gap Fractional [0, 1] 0.05 .11 5532 GLM CRE FRM, Linear FE Food insecurity (0/1) Did you worry that your household would not have enough food in the past Binary (0/1) 32.57 46.87 5532 GEE CRE Probit, Linear FE 7 days Food consumption adequacy Which of the following describes best your food consumption over the past Categorical ordered 5532 ML CRE Ordered Probit, month {1,2,3} Linear FE 1 1== less than adequate for household needs 41.45 49.27 2 2== just adequate for household needs 51.93 49.97 3 3== It was more than adequate for household needs 6.62 24.86 Fertilizer purchase decision The household purchased fertilizer for agricultural production in the rainy Binary (0/1) 38.85 48.75 5305 GEE CRE Probit, Linear FE, (0/1) season system ML Inputs purchase decision (0/1) The household purchased seeds, fertilizer or other chemicals for Binary (0/1) 60.13 48.97 5305 GEE CRE Probit, Linear FE, agricultural production in the rainy season system ML Fertilizer purchase per acre Value of fertilizer purchased per acre of land cultivated Corner solution [0, +) 3.59 9.79 5305 CRE Tobit, Linear FE (1000MKW) Inputs purchase per acre Total Value of seed, fertilizer, and chemicals purchased per acre of land Corner solution [0, +) 4.10 10.01 5305 CRE Tobit, Linear FE (1000MKW) cultivated Land cultivated (acres) Number of acres of cultivated land by household Corner solution [0, +) 4.43 52.31 5305 CRE Tobit, Linear FE Source: Generated by authors using LSMS data. 26 Table 2: Household participation in non-farm employment in rural Malawi, 2010/2013 Malawi Non Poor Poor P-value t-test poor vs. non poor Share of Share of Share of Share of Participation Households Participation Households Participation Households Participation Households rates (%) Income from rates (%) Income from rates (%) Income from rates (%) Income from Source (%) Source (%) Source (%) Source (%) 2010 2013 2010 2013 2010 2013 2010 2013 2010 2013 2010 2013 2010 2013 2010 2013 Non- Agricultural 17.1 15.1 8 6.5 21.6 19 10.6 8.5 9.7 7.4 3.9 2.3 0.000 0.000 0.000 0.000 wage Self 18.2 26.8 7.2 11.1 20.3 30.5 9 12.9 14.9 19.4 4.4 7.4 0.002 0.000 0.076 0.000 Employment Source: Malawi IHS3 panel sample (2010) and IHPS (2013) 27 Table 3: Selected characteristics of household enterprises in rural Malawi Rural Malawi Non Poor Poor p-value difference Characteristics of households’ enterprises 2010 2013 2010 2013 2010 2013 2010 2013 Age of enterprises (years) 11.1 9.0 11.7 9.0 12.9 9.1 0.250 0.821 Outside partner (%) 3.4 3.2 4.5 3.0 0.7 3.7 0.008 0.727 Business operating premises (%) Home 47.0 43.3 48.1 44.4 44.6 39.8 0.592 0.356 Market place and commercial area shop 31.5 38.0 29.8 36.6 35.5 42.8 0.285 0.198 Roadside and other areas 21.4 18.6 22.1 19.0 19.9 17.4 0.671 0.712 Formal registration (%) 7.1 7.9 9.1 9.3 2.4 3.3 0.003 0.008 FBPEa (%) 14.3 14.3 10.1 13.4 23.7 17.5 0.001 0.282 Access to electricity (%) 8.0 13.6 13.8 15.8 0.0 0.0 0.141 0.069 Number of enterprises per household (1 to 4) 1.1 1.1 1.0 1.1 1.1 1.0 0.907 0.027 Owner if household head (%) 73.6 69.5 75.6 72.3 69.1 60.4 0.233 0.008 Age of owner (%) 38.3 38.5 37.4 38.4 40.4 39.0 0.056 0.651 Education of enterprise owner (%) None 77.2 74.1 74.7 70.3 82.8 86.6 0.116 0.000 PSLCb 10.2 12.7 9.6 14.2 11.5 8.0 0.673 0.009 JCEc 8.1 7.9 9.6 8.7 4.7 5.1 0.078 0.091 MSCEd 3.7 4.5 4.9 5.7 1.0 0.4 0.010 0.000 Non-university diploma 0.5 0.5 0.7 0.6 0.0 0.0 0.117 0.044 University diploma 0.4 0.4 0.5 0.5 0.0 0.0 0.311 0.135 Post-graduate degree 0.0 0.0 0.0 0.0 0.0 0.0 0.322 Source: Malawi IHS3 panel sample (2010) and IHPS (2013). Note: a Forest Based Products Enterprise. b Primary School Leaving Certificate. c Junior Certificate Examination. d Malawi School Certificate of Education Examination 28 Table 4: Distribution of non-farm household enterprises and returns by Sector, by poverty status, in Rural Malawi, 2010/2013 Sector Rural Malawi Non Poor Poor p-value difference 2010 2013 2010 2013 2010 2013 2010 2013 Participation Proportion of enterprises by sector of industry (%) Primary sector 1.6 2.7 0.8 1.9 3.4 5.4 0.161 0.183 (mostly charcoal related activities and stone quarrying) Food, beverage, tobacco manufacturing 26.1 25.6 22.3 25.3 34.8 26.6 0.036 0.81 (mostly food and drinks street vendors) Non-food manufacturing 16.2 13.6 15.9 12.7 16.9 16.7 0.786 0.248 (mostly mats, baskets, and pottery, tailor, metal tinsmith) Wholesale and retail trade + restaurant + hotels 46.4 49.2 50.4 50.9 37.3 43.4 0.016 0.121 (include a wide range of commerce & tourism related activities) Construction, services, and other sectors 9.7 8.9 10.7 9.1 7.6 8 0.329 0.674 (mostly electrical and other repair shops, barber shop, hair salon) Returns Profit of enterprises by industry (1000 KW) Primary sector 4.6 3.2 6.1 3.4 3.8 3.1 - 0.651 Food, beverage, tobacco manufacturing 3.8 3.7 5.1 4.3 1.7 2.1 0.038 0.006 Non-food manufacturing 4.8 5.8 5.9 6.7 2.4 3.1 0.052 0.086 Wholesale and retail trade and restaurant & hotels 6.6 7 7.6 8 3.6 3.3 0 0 Construction, services, and other sectors 8.6 12.2 10 13.8 2.8 6.5 0.017 0.109 Overall 5.7 6.4 7 7.3 2.6 3.2 0.000 0.000 Source: Malawi IHS3 panel sample (2010) and IHPS (2013). Note: The information in this table is at the enterprise level. Profit is the net profit generated by the enterprise over the month of operation prior to the interview, as reported by the enterprise owner. Kwachas (MWK), local currency used in Malawi. 29 Table 5: Distribution of non-farm wage employment and returns by Sector, by poverty status, in Rural Malawi, 2010/2013 Sector Rural Malawi Non Poor Poor p-value difference 2010 2013 2010 2013 2010 2013 2010 2013 Participation Proportion of participants by non-farm wage employment sector (%) Professional, technical, & related workers 24.3 24.8 26.7 28 15.3 7.1 0.037 0.000 Administration and managerial workers 1.3 1.7 1.5 2 0.8 0 0.441 0.022 Clerical and related worker 5.7 4.6 7.2 5.4 0 0 0.000 0.000 Sales workers 4.8 5 4.9 4.6 4.3 7.4 0.848 0.496 Service workers 34.4 30.2 30.3 27.6 50.1 44.4 0.002 0.046 Forestry workers, fishermen and hunters 1.8 3.8 2.2 2.7 0.2 9.6 0.074 0.165 Transport equipment operators and laborers not elsewhere classified 27.6 30 27.2 29.6 29.3 31.5 0.745 0.785 Return Average monthly wages of participants by Sector (1000 KW) Professional, technical, & related workers 16.3 16.8 16.8 17.3 12.8 6.2 0.281 0.000 Administration and managerial workers 10.6 2.9 11.5 2.9 4 - 0.012 - Clerical and related worker 10.4 10.9 10.4 10.9 - - - - Sales workers 5.6 6.8 6.5 7.4 2 4.8 0.000 0.065 Service workers 4.7 4.9 4.6 5.2 4.9 4.1 0.659 0.131 Forestry workers, fishermen and hunters 14.6 3.3 14.8 5 4.4 0.7 0.072 - Transport equipment operators and laborers not elsewhere classified 7.3 5.7 7.6 6.2 5.9 3 0.239 0.001 Overall 8.1 7.6 8.8 8.4 5.5 3.7 0.000 0.000 Source: Malawi IHS3 panel sample (2010) and IHPS (2013). Note: Wage is the estimated monthly salary using the last payment reported by the participant member and the period of time covered by that payment. Kwachas (MWK), local currency used in Malawi. 30 Table 6: Average partial effects estimates of determinants of non-farm employment participation Non-farm wage employment Non-farm self-employment VARIABLES CRE Probit Linear FE CRE Probit Linear FE model model model model Age of the household head -0.000 -0.000 0.000 0.001 [0.863] [0.751] [0.988] [0.631] Male-headed household (0/1) 0.072*** 0.072*** 0.017 0.008 [0.001] [0.002] [0.552] [0.783] Highest level of formal education acquired by household head PSLC 0.035 0.039 0.003 0.010 [0.194] [0.202] [0.937] [0.803] JCE 0.130*** 0.120*** 0.046 0.055 [0.007] [0.008] [0.365] [0.290] MSCE 0.236*** 0.257*** 0.033 0.029 [0.004] [0.000] [0.651] [0.697] Non-University Diploma and above 0.198 0.222 0.286 0.249+ [0.294] [0.228] [0.184] [0.102] Maximum level of formal education acquired in the household PSLC 0.017 0.016 0.017 0.014 [0.478] [0.495] [0.578] [0.680] JCE 0.016 0.015 -0.041 -0.058 [0.579] [0.628] [0.224] [0.165] MSCE 0.079+ 0.102** -0.052 -0.056 [0.105] [0.030] [0.271] [0.294] Non-University Diploma and above 0.262 0.263+ -0.184*** -0.294* [0.167] [0.127] [0.000] [0.095] Number of infant (<5yo) in HH 0.001 0.006 0.022** 0.019+ [0.903] [0.535] [0.036] [0.128] Number of children (5-14yo) in the household 0.015*** 0.017** 0.001 -0.000 [0.006] [0.011] [0.930] [0.974] Number of prime adults (15-60yo) in HH -0.003 -0.002 0.022** 0.021* [0.669] [0.793] [0.022] [0.065] Number of elderly (60yo+) in HH -0.019 -0.023+ -0.006 -0.004 [0.383] [0.149] [0.844] [0.872] Household access to loan (0/1) -0.006 -0.001 0.083*** 0.105*** [0.695] [0.931] [0.000] [0.000] Normalized wealth index 0.301*** 0.519*** 0.450*** 0.567*** [0.003] [0.001] [0.000] [0.000] Normalized TLU index -0.432* -0.268 0.503+ 0.176 [0.086] [0.506] [0.117] [0.717] Total land area owned by HH in Acres 0.006* 0.000 -0.000 -0.000 [0.100] [0.735] [0.825] [0.860] Household was affected negatively by some income 0.017 0.021 -0.018 -0.037+ Shock During the last 12 months (0/1) 31 [0.297] [0.195] [0.420] [0.118] Household has a migrant network (0/1) 0.010 0.008 0.005 0.007 [0.532] [0.669] [0.799] [0.774] Rain - EA level CoV of Dec-Jan rainfall from -0.886*** 0.229 1983/84 - 2012/13 [0.006] [0.526] Average 12-month total rainfall (mm) -0.000 0.000 0.000 0.000 [0.500] [0.721] [0.227] [0.741] Annual Mean Temperature (∞C * 10) 0.000 0.000 -0.000 -0.002 [0.838] [0.909] [0.839] [0.419] HH Distance in (KMs) to Nearest Road -0.002 -0.004 -0.002 0.000 [0.326] [0.263] [0.589] [0.948] HH Distance in (KMs) to Nearest Population Center -0.000 -0.001 -0.000 0.001 with +20,000 [0.953] [0.542] [0.851] [0.574] Value of daily ganyu wage salary in the EA -0.000 -0.000 0.000 0.000** [0.465] [0.282] [0.301] [0.023] Price of fertilizer in the EA -0.000+ -0.000 -0.000 -0.000* [0.106] [0.233] [0.638] [0.090] Price of maize grains in the EA -0.000 -0.000 0.000 0.000* [0.449] [0.457] [0.299] [0.070] EA neighbor’s wage employment participation 0.002*** 0.003*** 0.001* 0.002** [0.001] [0.003] [0.088] [0.026] EA neighbor’s self-employment participation 0.001*** 0.001** 0.004*** 0.004*** [0.010] [0.016] [0.000] [0.000] Time dummy (year 2010=1) -0.002 -0.004 -0.008 -0.022 [0.929] [0.888] [0.787] [0.363] District dummies (27 -1 dummies) Included Included Included Included Time average of explanatory variables Included Included Constant -0.856 0.805 [0.276] [0.369] Number of observations 5,286 5,286 5,286 5,286 Number of households 2,751 2,751 Source: Generated by authors using LSMS data Note: ***, **, *, and + indicate that the corresponding regression APE is statistically significant at the 1%, 5%, 10%, and 15% levels, respectively. The CRE model is estimated using GEE approach with cluster robust standard errors. The linear FE results are presented for comparison. Pvalues between brackets. 32 Table 7: Effect of participation in the non-farm activities on various outcomes in rural Malawi Effects of non-farm wage employment participation Effects of non-farm self-employment participation VARIABLES CRE Probit/fractional Probit / CRE Tobit Linear FE CRE Probit/fractional Probit / CRE Tobit Linear FE ordered Probit ordered Probit Objective welfare outcomes (N=5321) HHPCE (1000 MKW) 4.476 7.003*** [0.158] [0.000] Log HHPCE 0.102** 0.129*** [0.022] [0.000] Poverty incidence -0.074** . -0.072** -0.085*** -0.083*** [0.012] [0.022] [0.000] [0.000] Poverty gap -0.034*** -0.033** -0.034** -0.035*** -0.035*** -0.030*** [0.007] [0.011] [0.018] [0.000] [0.000] [0.002] Squared poverty gap -0.018** -0.017** -0.019** -0.018*** -0.017*** -0.014** [0.021] [0.017] [0.045] [0.000] [0.000] [0.011] Subjective welfare outcomes (N=5321) Food insecurity (0/1) -0.047+ -0.065** -0.034+ -0.027 [0.125] [0.041] [0.140] [0.281] Food consumption adequacy (1-3) 0.177** 0.104** 0.057 0.025 [0.037] [0.021] [0.379] [0.482] Agricultural outcomes (N=5038) Fertilizer purchase decision (0/1) 0.039 0.037 0.045** 0.055** [0.190] [0.303] [0.043] [0.036] Inputs purchase decision (0/1) 0.046 0.037 0.053** 0.060** [0.160] [0.260] [0.031] [0.039] Value of fertilizer purchase (1000 0.156 -0.638 0.388 0.136 MKW) [0.745] [0.419] [0.225] [0.786] Value of inputs purchase (1000 MKW) -0.124 -0.874 0.234 0.085 [0.802] [0.305] [0.495] [0.870] Land cultivated (1000 acres) -0.292 0.274 0.710 -0.576 [0.699] [0.800] [0.561] [0.893] Source: Generated by authors using LSMS data Note: ***, **, *, and + indicate that the corresponding regression APE is statistically significant at the 1%, 5%, 10%, and 15% levels, respectively. The table presents only APEs on our main treatment variables. The APEs from the CRE Probit models are estimated using GEE approach. The CRE fractional Probit APEs are from GLM estimation. The unconditional APEs from CRE Tobit are estimated using pooled Tobit regression methods. The linear FE results are presented for comparison purpose. The full results tables are in appendices 2, 4, 5, and 6. P-values between brackets are based on cluster robust standard errors. 33 34 Table 8: Distributional effects of participation in the non-farm activities on HHPCE in Malawi (Quantile regression results) VARIABLES p10 p20 p25 p50 p75 p80 p90 Effects of non-farm wage 2.611+ 2.634+ 4.639*** 6.198*** 6.386** 6.214* 2.709 employment participation [0.114] [0.138] [0.006] [0.005] [0.016] [0.086] [0.714] Effects of non-farm self- 2.347** 4.209*** 4.303*** 6.503*** 10.804*** 13.041*** 16.227** employment participation [0.018] [0.001] [0.000] [0.000] [0.000] [0.000] [0.016] Other controls included Yes Yes Yes Yes Yes Yes Yes Time averages included Yes Yes Yes Yes Yes Yes Yes Observations 5321 5321 5321 5321 5321 5321 5321 Source: Generated by authors using LSMS data Note: ***, **, *, and + indicate that the corresponding regression coefficient is statistically significant at the 1%, 5%, 10%, and 15% levels, respectively. The table presents quantile partial effects of our main treatment variables from 10th to the 90th percentile. The full results tables are in appendix 3. P-values between. Also the results without inclusion of time averages of explanatory variables yields the same conclusions. For the sake of conciseness, they are not presented here, but are available from authors upon request. Figure 1: Quantile effects of non-farm wage and self-employment on HHPCE in rural Malawi Source: Generated by authors using LSMS data Note: The graph presents quantile partial effects of our main treatment variables. Other controls were included in the estimation 35 Appendix 1: Description and summary statistics of the covariates used in this paper, household level, 2010-2013, rural Malawi Variables Definition Overall SD Mean Age of the household head Age of the household head 44.04 16.52 Male headed-household Male headed-household 76.16 42.62 Highest level of formal education acquired by household head None The household head never attended formal school 77.17 41.98 PSLC PSLC is the highest formal level of education of household head 10.03 30.05 JCE JCE is the highest formal level of education of household head 6.87 25.30 MSCE MSCE is the highest formal level of education of household head 4.68 21.13 Non-Univ Diploma and above Non-Univ Diploma and above 1.25 11.10 Highest level of formal education acquired in the household None No member of the household ever attended formal school 63.41 48.17 PSLC PSLC is the highest formal level of education of the most educated 15.20 35.91 household member JCE JCE is the highest formal level of education of the most educated household 12.22 32.75 member MSCE MSCE is the highest formal level of education of the most educated 7.66 26.61 household member Non-Univ Diploma and above Non-Univ Diploma and above 1.50 12.16 Size of the household Number of people living in the household at the time of the interview 5.22 2.42 Number of infant (<5yo) in HH Number of infant living in the household at the time of the interview 0.84 0.85 Number of children (5-14yo) in the household Number of children living in the household at the time of the interview 1.62 1.41 Number of prime adults (15-60yo) in HH Number of prime adults living in the household at the time of the interview 2.46 1.39 Number of elderly (60yo+) in HH Number of elderly living in the household at the time of the interview 0.29 0.57 Household access to loan (0/1) A member of the household received a loan in the year prior to the interview 16.63 37.24 Normalized wealth index Principal component analysis estimate of asset index 0.05 0.08 Normalized TLU index Total livestock unit index tlu=cattle*0.5+pigs*0.2+sheep*0.1+goats*0.1 0.01 0.03 Total land area owned by HH in Acres Total land area owned by HH in Acres 3.35 49.13 Household was affected negatively by some income Shock During Household was affected negatively by some income Shock During the last 12 86.37 34.31 the last 12 months (0/1) months (0/1) Household has a migrant network (0/1) A member of the household is living outside of the EA 35.65 47.90 Rain variability Rain - EA level Coefficient of Variation of Dec-Jan rainfall from 1983/84 - 0.25 0.04 2012/13 Rainfall (mm) Average 12-month total rainfall (mm) 850.32 89.24 Temperature (∞C * 10) Annual Mean Temperature (∞C * 10) 216.31 19.12 36 HH Distance to Nearest Road Household Distance in (KMs) to Nearest Road 9.84 9.85 HH Distance to Population Center Household Distance in (KMs) to Nearest Population Center with +20,000 37.67 17.95 ganyu wage salary in the EA (MKW) Value of daily ganyu wage salary in the EA (MKW) 566.22 599.87 Price of maize grains in the EA (MKW) Price of maize grains in the EA (MKW) 198.59 112.66 Price of fertilizer in the EA (MKW) Price of fertilizer in the EA (MKW) 62.27 39.34 EA neighbor’s wage employment participation Participation rate in non-farm wage employment by other household in the 15.11 14.90 same EA EA neighbor’s self-employment participation Participation rate in non-farm self-employment by other household in the 21.58 15.27 same EA Share of total land cultivated in crops (crop mix) Grains Share of total land cultivated in grains 64.53 26.22 Legumes Share of total land cultivated in legumes 23.00 22.86 Tubers Share of total land cultivated in tubers 1.55 7.75 Oils crops Share of total land cultivated in oil crops 0.43 3.86 Horticulture crops Share of total land cultivated in horticulture crops 4.03 11.10 Cotton Share of total land cultivated in cotton 1.29 7.29 Tobacco Share of total land cultivated in tobacco 4.19 11.79 Source: Generated by authors using LSMS data Note: PSLC=Primary School Leaving Certificate. JCE=Junior Certificate Examination. MSCE=Malawi School Certificate of Education Examination. 37 Appendix 2: Test of balancing of covariates between non-farm wage employment participants and non-participants households, 2010- 2013, rural Malawi Wage employment Non participants Wage employment participants t-test difference Variables Mean SD Mean SD pvalue Treatment variables Non-farm wage employment (0/1) 0.00 0.00 100.00 0.00 Non-farm self-employment (0/1) 21.53 41.11 21.89 41.37 0.668 Outcome variables HHPCE 53,903.41 40,631.58 79,420.64 67,908.37 0.000 Log of HHPCE 10.70 0.62 11.04 0.67 0.000 Poverty incidence 40.03 49.00 21.29 40.96 0.000 Poverty gap Poverty severity Food insecurity (0/1) 34.01 47.38 24.52 43.05 0.001 Food consumption adequacy 1 43.51 49.58 29.90 45.81 0.000 2 50.45 50.00 60.29 48.96 0.002 3 6.05 23.84 9.81 29.76 0.003 Seeds purchase decision (0/1) 39.52 48.89 50.07 50.03 0.000 Fertilizer purchase decision (0/1) 36.71 48.21 51.79 50.00 0.000 Inputs purchase decision (0/1) 58.35 49.30 70.92 45.45 0.000 Seed purchase per acre (1000MKW) 0.46 1.56 0.91 2.26 0.000 Fertilizer purchase per acre (1000MKW) 3.10 8.51 6.57 15.06 0.001 Inputs purchase per acre (1000MKW) 3.57 8.92 7.31 14.60 0.000 Land cultivated (acres) 4.63 55.98 3.20 18.04 0.194 Covariates Age of the household head 44.53 17.00 41.30 13.22 0.000 Male headed-household 74.25 43.73 86.84 33.82 0.000 Highest level of formal education acquired by household head None 82.54 37.97 47.01 49.94 0.000 PSLC 9.71 29.61 11.84 32.33 0.012 38 JCE 5.64 23.08 13.76 34.46 0.000 MSCE 1.96 13.86 19.98 40.01 0.000 Non-Univ Diploma and above 0.15 3.86 7.42 26.22 0.000 Highest level of formal education acquired in the household None 68.70 46.38 33.73 47.31 0.000 PSLC 15.63 36.32 12.80 33.43 0.926 JCE 11.07 31.38 18.66 38.98 0.000 MSCE 4.34 20.39 26.32 44.06 0.000 Non-Univ Diploma and above 0.26 5.05 8.49 27.89 0.000 Size of the household 5.17 2.40 5.47 2.55 0.701 Number of infant (<5yo) in HH 0.85 0.85 0.82 0.82 0.027 Number of children (5-14yo) in the household 1.62 1.41 1.66 1.41 0.826 Number of prime adults (15-60yo) in HH 2.39 1.36 2.84 1.50 0.000 Number of elderly (60yo+) in HH 0.32 0.59 0.16 0.45 0.000 Household access to loan (0/1) 15.48 36.18 23.09 42.16 0.010 Normalized wealth index 0.04 0.06 0.11 0.13 0.000 Normalized TLU index 0.01 0.03 0.01 0.02 0.901 Total land area owned by HH in Acres 3.55 52.84 2.26 16.94 0.202 Household was affected negatively by some 86.71 33.95 84.45 36.26 0.104 income Shock During the last 12 months (0/1) Household has a migrant network (0/1) 36.52 48.15 30.74 46.17 0.028 Rain - EA level CoV of Dec-Jan rainfall from 0.25 0.04 0.26 0.04 0.017 1983/84 - 2012/13 Average 12-month total rainfall (mm) 850.57 87.09 848.91 100.55 0.017 Annual Mean Temperature (∞C * 10) 215.97 18.97 218.22 19.84 0.587 HH Distance in (KMs) to Nearest Road 10.43 10.05 6.53 7.89 0.000 HH Distance in (KMs) to Nearest Population 38.11 17.87 35.17 18.26 0.012 Center with +20,000 Value of daily ganyu wage salary in the EA 561.67 599.41 592.09 602.18 0.820 (MKW) Price of maize grains in the EA (MKW) 201.75 112.90 180.81 109.70 0.001 Price of fertilizer in the EA (MKW) 62.12 38.98 63.14 41.31 0.731 EA neighbor’s wage employment 13.26 13.52 25.50 17.79 0.000 participation EA neighbor’s self-employment participation 20.74 14.92 26.32 16.28 0.000 39 Share of total land cultivated in crops (crop mix) Grains 64.31 26.17 65.90 26.52 0.370 Legumes 22.81 22.73 24.16 23.63 0.173 Tubers 1.55 7.65 1.55 8.33 0.035 Oils crops 0.48 4.09 0.13 1.90 0.107 Horticulture crops 4.14 11.28 3.38 9.93 0.271 Cotton 1.29 7.29 1.34 7.31 0.900 Tobacco 4.48 12.20 2.41 8.64 0.000 Source: Generated by authors using LSMS data Notes: PSLC=Primary School Leaving Certificate. JCE=Junior Certificate Examination. MSCE=Malawi School Certificate of Education Examination. 40 Appendix 3: Test of balancing of covariates between non-farm self-employment participants and non-participants, 2010-2013, rural Malawi Self-employment Non participants Self-employment participants t-test difference Variables Mean SD Mean SD p-value Treatment variables Non-farm wage employment (0/1) 15.05 35.76 15.33 36.04 0.668 Non-farm self-employment (0/1) 0.00 0.00 100.00 0.00 Outcome variables HHPCE 54,741.65 43,462.77 68,724.30 55,600.92 0.000 Log of HHPCE 10.70 0.63 10.91 0.66 0.000 Poverty incidence 39.81 48.96 27.72 44.78 0.000 Poverty gap Poverty severity Food insecurity (0/1) 33.24 47.11 30.15 45.91 0.213 Food consumption adequacy 1 42.81 49.49 36.52 48.17 0.063 2 50.90 50.00 55.70 49.70 0.369 3 6.29 24.29 7.79 26.81 0.094 Seeds purchase decision (0/1) 39.27 48.84 47.51 49.96 0.000 Fertilizer purchase decision (0/1) 36.73 48.21 46.71 49.91 0.000 Inputs purchase decision (0/1) 57.48 49.44 69.98 45.85 0.000 Seed purchase per acre (1000MKW) 0.49 1.66 0.65 1.78 0.011 Fertilizer purchase per acre (1000MKW) 3.48 9.62 4.00 10.38 0.215 Inputs purchase per acre (1000MKW) 3.96 9.96 4.60 10.16 0.078 Land cultivated (acres) 4.09 51.38 5.67 55.60 0.903 Covariates Age of the household head 44.72 17.01 41.58 14.37 0.000 Male headed-household 74.92 43.35 80.65 39.52 0.014 Highest level of formal education acquired by household head None 78.33 41.20 72.95 44.44 0.010 PSLC 9.70 29.61 11.22 31.58 0.194 41 JCE 6.27 24.25 9.05 28.69 0.106 MSCE 4.52 20.77 5.28 22.37 0.354 Non-Univ Diploma and above 1.18 10.78 1.51 12.19 0.126 Highest level of formal education acquired in the household None 65.42 47.57 56.11 49.65 0.000 PSLC 14.29 35.00 18.51 38.85 0.000 JCE 11.60 32.02 14.49 35.21 0.246 MSCE 7.26 25.95 9.13 28.81 0.248 Non-Univ Diploma and above 1.43 11.87 1.76 13.15 0.211 Size of the household 5.12 2.39 5.56 2.50 0.000 Number of infant (<5yo) in HH 0.82 0.83 0.94 0.89 0.004 Number of children (5-14yo) in the household 1.59 1.41 1.75 1.41 0.001 Number of prime adults (15-60yo) in HH 2.40 1.39 2.67 1.38 0.000 Number of elderly (60yo+) in HH 0.32 0.59 0.19 0.48 0.000 Household access to loan (0/1) 13.53 34.21 27.89 44.86 0.000 Normalized wealth index 0.05 0.07 0.08 0.11 0.000 Normalized TLU index 0.01 0.02 0.01 0.06 0.097 Total land area owned by HH in Acres 3.19 49.77 3.97 46.75 0.598 Household was affected negatively by some 85.75 34.96 88.61 31.78 0.027 income Shock During the last 12 months (0/1) Household has a migrant network (0/1) 36.75 48.22 31.66 46.53 0.070 Rain - EA level CoV of Dec-Jan rainfall from 0.25 0.04 0.26 0.04 0.002 1983/84 - 2012/13 Average 12-month total rainfall (mm) 853.06 89.51 840.36 87.59 0.012 Annual Mean Temperature (∞C * 10) 215.84 19.15 218.01 18.92 0.026 HH Distance in (KMs) to Nearest Road 10.11 9.92 8.87 9.54 0.013 HH Distance in (KMs) to Nearest Population 38.05 18.05 36.28 17.55 0.343 Center with +20,000 Value of daily ganyu wage salary in the EA 548.59 564.61 630.35 710.21 0.005 (MKW) Price of maize grains in the EA (MKW) 196.70 111.45 205.43 116.73 0.014 Price of fertilizer in the EA (MKW) 60.87 39.22 67.36 39.37 0.000 EA neighbor’s wage employment participation 14.20 14.52 18.43 15.76 0.000 EA neighbor’s self employment participation 19.74 14.32 28.28 16.68 0.000 Share of total land cultivated in crops (crop 42 mix) Grains 65.45 26.17 61.15 26.15 0.001 Legumes 22.18 22.64 26.00 23.42 0.007 Tubers 1.51 7.71 1.68 7.88 0.758 Oils crops 0.47 4.01 0.28 3.26 0.589 Horticulture crops 3.92 10.98 4.42 11.53 0.183 Cotton 1.22 7.23 1.57 7.52 0.149 Tobacco 4.36 12.04 3.57 10.79 0.078 Note: PSLC=Primary School Leaving Certificate. JCE=Junior Certificate Examination. MSCE=Malawi School Certificate of Education Examination. 43 Appendix 4: Test of balancing of covariates between poor and non-poor households, 2010-2013, rural Malawi NON POOR POOR t-test difference Variables Mean SD Mean SD p-value Treatment variables Non-farm wage employment (0/1) 18.94 39.19 8.65 28.12 0.000 Non-farm self-employment (0/1) 24.84 43.22 16.08 36.75 0.000 Outcome variables HHPCE 76,672.85 49,807.94 25,833.13 7,228.67 0.000 Log of HHPCE 11.13 0.45 10.11 0.33 0.000 Poverty incidence 0.00 0.00 100.00 0.00 Poverty gap Poverty severity Food insecurity (0/1) 25.07 43.35 45.24 49.78 0.000 Food consumption adequacy 1 33.02 47.03 55.69 49.69 0.000 2 58.29 49.32 41.21 49.23 0.000 3 8.69 28.18 3.11 17.36 0.000 Seeds purchase decision (0/1) 42.10 49.38 39.26 48.84 0.177 Fertilizer purchase decision (0/1) 45.93 49.84 27.30 44.56 0.000 Inputs purchase decision (0/1) 63.68 48.10 54.34 49.82 0.000 Seed purchase per acre (1000MKW) 0.61 1.94 0.39 1.16 0.000 Fertilizer purchase per acre (1000MKW) 4.57 11.40 1.99 6.00 0.000 Inputs purchase per acre (1000MKW) 5.15 11.59 2.38 6.28 0.000 Land cultivated (acres) 4.93 62.06 3.60 30.21 0.599 Covariates Age of the household head 43.51 16.88 44.93 15.87 0.025 Male headed-household 77.49 41.77 73.91 43.93 0.026 Highest level of formal education acquired by household head None 71.53 45.13 86.69 33.98 0.000 PSLC 11.34 31.71 7.82 26.86 0.000 JCE 8.72 28.22 3.74 18.98 0.000 44 MSCE 6.51 24.67 1.60 12.56 0.000 Non-Univ Diploma and above 1.90 13.65 0.15 3.82 0.000 Highest level of formal education acquired in the household None 57.11 49.50 74.05 43.85 0.000 PSLC 16.41 37.04 13.17 33.82 0.001 JCE 14.16 34.87 8.94 28.54 0.000 MSCE 10.02 30.03 3.69 18.86 0.000 Non-Univ Diploma and above 2.30 15.00 0.15 3.82 0.000 Size of the household 4.68 2.35 6.12 2.27 0.000 Number of infant (<5yo) in HH 0.73 0.80 1.03 0.89 0.000 Number of children (5-14yo) in the 1.32 1.33 2.12 1.40 0.000 household Number of prime adults (15-60yo) in HH 2.34 1.37 2.65 1.41 0.000 Number of elderly (60yo+) in HH 0.28 0.56 0.31 0.59 0.141 Household access to loan (0/1) 18.42 38.77 13.61 34.29 0.003 Normalized wealth index 0.07 0.10 0.02 0.03 0.000 Normalized TLU index 0.01 0.04 0.00 0.01 0.000 Total land area owned by HH in Acres 3.88 59.96 2.46 20.45 0.448 Household was affected negatively by 86.41 34.27 86.30 34.40 0.640 some income Shock During the last 12 months (0/1) Household has a migrant network (0/1) 36.30 48.09 34.55 47.56 0.737 Rain - EA level CoV of Dec-Jan rainfall 0.25 0.04 0.26 0.04 0.706 from 1983/84 - 2012/13 Average 12-month total rainfall (mm) 846.63 83.60 856.55 97.75 0.158 Annual Mean Temperature (∞C * 10) 215.16 18.38 218.25 20.17 0.030 HH Distance in (KMs) to Nearest Road 9.05 9.26 11.19 10.65 0.000 HH Distance in (KMs) to Nearest 37.05 17.71 38.70 18.31 0.121 Population Center with +20,000 Value of daily ganyu wage salary in the EA 596.44 628.35 515.51 545.18 0.042 (MKW) Price of maize grains in the EA (MKW) 204.00 113.01 189.44 111.50 0.006 Price of fertilizer in the EA (MKW) 64.59 39.75 58.36 38.34 0.000 EA neighbor’s wage employment 16.49 15.74 12.79 13.04 0.000 participation EA neighbor’s self-employment 22.60 15.62 19.87 14.49 0.000 45 participation Share of total land cultivated in crops (crop mix) Grains 63.34 26.30 66.47 25.99 0.000 Legumes 23.73 22.88 21.82 22.79 0.004 Tubers 1.69 8.23 1.31 6.89 0.577 Oils crops 0.41 3.85 0.47 3.88 0.403 Horticulture crops 3.95 10.95 4.17 11.35 0.862 Cotton 1.21 7.18 1.44 7.47 0.493 Tobacco 4.76 12.69 3.26 10.08 0.004 Source: Generated by authors using LSMS data Note: PSLC=Primary School Leaving Certificate. JCE=Junior Certificate Examination. MSCE=Malawi School Certificate of Education Examination. 46 Appendix 5: Seemingly unrelated system equation estimates of non-farm wage employment and non-farm self-employment participation model GHK Simulated bivariate ML Seemingly unrelated bivariate ML Probit estimates estimates VARIABLES Wage Self Wage employment Self employment employment employment Age of the household head -0.000 -0.000 -0.000 -0.000 [0.982] [0.977] [0.984] [0.983] Male-headed household (0/1) 0.423*** 0.061 0.423*** 0.061 [0.003] [0.599] [0.003] [0.596] Highest level of formal education acquired by household head PSLC 0.192 0.003 0.192 0.001 [0.270] [0.984] [0.270] [0.993] JCE 0.621*** 0.158 0.622*** 0.156 [0.004] [0.425] [0.004] [0.430] MSCE 0.991*** 0.117 0.991*** 0.118 [0.001] [0.685] [0.001] [0.683] Non-University Diploma and above 0.894 0.942+ 0.896 0.934+ [0.169] [0.127] [0.169] [0.130] Maximum level of formal education acquired in the household PSLC 0.097 0.071 0.098 0.072 [0.525] [0.568] [0.524] [0.565] JCE 0.102 -0.157 0.102 -0.154 [0.579] [0.332] [0.580] [0.341] MSCE 0.408* -0.211 0.408* -0.211 [0.094] [0.360] [0.094] [0.360] Non-University Diploma and above 1.099* -1.195** 1.098* -1.192** [0.066] [0.037] [0.067] [0.038] Number of infant (<5yo) in HH 0.007 0.086* 0.007 0.087* [0.906] [0.064] [0.904] [0.062] Number of children (5-14yo) in the household 0.087** 0.003 0.087** 0.003 47 [0.024] [0.925] [0.024] [0.923] Number of prime adults (15-60yo) in HH -0.022 0.087** -0.022 0.087** [0.604] [0.015] [0.602] [0.015] Number of elderly (60yo+) in HH -0.103 -0.016 -0.102 -0.015 [0.499] [0.897] [0.502] [0.905] Household access to loan (0/1) -0.042 0.329*** -0.042 0.329*** [0.674] [0.000] [0.671] [0.000] Normalized wealth index 1.813*** 1.776*** 1.810*** 1.771*** [0.005] [0.002] [0.005] [0.002] Normalized TLU index -2.788+ 1.882 -2.791+ 1.873 [0.117] [0.192] [0.117] [0.194] Total land area owned by HH in Acres 0.042** -0.000 0.042** -0.000 [0.011] [0.794] [0.011] [0.788] Household was affected negatively by some income Shock During the 0.104 -0.064 0.104 -0.063 last 12 months (0/1) [0.364] [0.501] [0.364] [0.510] Household has a migrant network (0/1) 0.042 0.021 0.042 0.021 [0.711] [0.819] [0.713] [0.821] Rain - EA level CoV of Dec-Jan rainfall from 1983/84 - 2012/13 -5.370*** 1.022 [0.002] [0.454] Average 12-month total rainfall (mm) -0.001 0.002 -0.001 0.002 [0.571] [0.265] [0.570] [0.266] Annual Mean Temperature (∞C * 10) 0.002 -0.002 0.002 -0.002 [0.828] [0.814] [0.825] [0.816] HH Distance in (KMs) to Nearest Road -0.014 -0.008 -0.014 -0.008 [0.318] [0.544] [0.316] [0.548] HH Distance in (KMs) to Nearest Population Center with +20,000 0.001 -0.001 0.001 -0.001 [0.936] [0.860] [0.938] [0.871] Value of daily ganyu wage salary in the EA -0.000 0.000 -0.000 0.000 [0.544] [0.329] [0.544] [0.322] Price of fertilizer in the EA -0.001+ -0.000 -0.001+ -0.000 [0.120] [0.764] [0.120] [0.754] Price of maize grains in the EA -0.001 0.001 -0.001 0.001 [0.445] [0.404] [0.446] [0.398] EA neighbor’s wage employment participation 0.012*** 0.006* 0.012*** 0.006* [0.003] [0.091] [0.003] [0.087] EA neighbor’s self employment participation 0.007** 0.016*** 0.007** 0.016*** [0.028] [0.000] [0.028] [0.000] Time dummy (year 2010=1) -0.029 -0.020 -0.029 -0.019 [0.852] [0.878] [0.852] [0.883] 48 District dummies (27 -1 dummies) Included Included Included Included Time average of explanatory variables Included Included Included Included rho -0.145*** -0.150*** [0.000] [0.000] Constant -0.885 -2.354** -0.890 -2.358** [0.437] [0.013] [0.434] [0.012] Observations 5,286 5,286 5,286 5,286 Source: Generated by authors using LSMS data Note: The symbols ***, **, *, and + indicate that the corresponding regression coefficient is statistically significant at the 1%, 5%, 10%, and 15% levels, respectively. P-values between brackets. The GHK estimates are generated using the mvprobit command with 100 draws in STATA. The bivariate ML estimates are generated using the biprobit command in STATA. The parameters in the table are coefficients and not APEs. 49 Appendix 6: Effect of participation in the non-farm activities on HHPCE in Malawi, FE estimates (2) (4) VARIABLES HHPCE Log HHPCE FE FE Wage employment participation 4.476 0.102** [0.158] [0.022] Self-employment participation 7.003*** 0.129*** [0.000] [0.000] Age of the household head -0.053 -0.002 [0.633] [0.232] Male-headed household (0/1) 2.895 0.063+ [0.269] [0.101] Highest level of formal education acquired by household head PSLC 6.198** 0.036 [0.048] [0.401] JCE 15.719*** 0.187*** [0.005] [0.005] MSCE 20.869* 0.197** [0.054] [0.049] Maximum level of formal education acquired in the household PSLC -2.228 -0.011 [0.350] [0.765] JCE -0.755 0.020 [0.814] [0.667] MSCE -1.458 0.004 [0.774] [0.950] Number of infant (<5yo) in HH -9.266*** -0.139*** [0.000] [0.000] Number of children (5-14yo) in the household -8.872*** -0.136*** [0.000] [0.000] Number of prime adults (15-60yo) in HH -6.768*** -0.094*** [0.000] [0.000] Number of elderly (60yo+) in HH -7.836*** -0.118*** [0.002] [0.001] Household access to loan (0/1) 1.635 0.034 [0.436] [0.221] Normalized TLU index 140.889** 1.326** [0.045] [0.032] Normalized agricultural asset index + Normalized land holdings 41.423*** 0.689*** [0.000] [0.000] Household was affected negatively by some income Shock During the last 12 1.644 0.027 months (0/1) [0.473] [0.446] Household has a migrant network (0/1) -1.124 -0.004 [0.585] [0.906] Rain - EA level CoV of Dec-Jan rainfall from 1983/84 - 2012/13 Average 12-month total rainfall (mm) 0.064 0.001+ 50 [0.228] [0.134] Annual Mean Temperature (∞C * 10) 0.413 0.000 [0.261] [0.931] HH Distance in (KMs) to Nearest Road 0.006 -0.001 [0.987] [0.784] HH Distance in (KMs) to Nearest Population Center with +20,000 -0.085 -0.003 [0.689] [0.160] Price of fertilizer in the EA -0.013 -0.000 [0.403] [0.604] Price of maize grains in the EA 0.083** 0.001** [0.024] [0.029] Time dummy (year 2010=1) -2.343 -0.037 [0.567] [0.623] District dummies Included Included Time average of explanatory variables Constant -100.016 9.223*** [0.354] [0.000] Observations 5,321 5,321 Number of households 2,764 2,764 Source: Generated by authors using LSMS data Note: PSLC=Primary School Leaving Certificate. JCE=Junior Certificate Examination. MSCE=Malawi School Certificate of Education Examination. The dependent variable in column 1 and 2 are levels of HHPCE. In column 3 and 4, the dependent variable is Log (HHPCE). The symbols ***, **, *, and + indicate that the corresponding regression coefficients are statistically significant at the 1%, 5%, 10%, and 15% levels, respectively. P-values based on clustered robust standard errors between brackets. 51 Appendix 7: Effects of participation in the RNFE on quintiles of HHPCE p10 p20 p25 p50 p75 p80 p90 VARIABLES Wage employment participation 2.611+ 2.634+ 4.639*** 6.198*** 6.386** 6.214* 2.709 [0.114] [0.138] [0.006] [0.005] [0.016] [0.086] [0.714] Self-employment participation 2.347** 4.209*** 4.303*** 6.503*** 10.804*** 13.041*** 16.227** [0.018] [0.001] [0.000] [0.000] [0.000] [0.000] [0.016] Age of the household head -0.042 0.005 -0.013 -0.079 -0.161 -0.192+ -0.392* [0.524] [0.939] [0.829] [0.341] [0.198] [0.144] [0.057] Male-headed household (0/1) 1.870 2.858* 1.598 2.211 6.659* 6.856* 2.388 [0.268] [0.063] [0.335] [0.399] [0.060] [0.091] [0.740] Highest level of formal education acquired by household head PSLC -0.349 -1.262 -0.916 0.929 9.709* 10.191* 6.786 [0.828] [0.423] [0.658] [0.717] [0.055] [0.070] [0.466] JCE 5.945** 8.658*** 7.995** 9.316** 21.622*** 18.744** 20.544 [0.011] [0.004] [0.011] [0.033] [0.006] [0.034] [0.210] MSCE 2.274 4.883 5.615 16.777** 36.827*** 43.716** 28.356 [0.479] [0.444] [0.348] [0.014] [0.009] [0.016] [0.424] Maximum level of formal education acquired in the household PSLC 0.626 0.779 0.259 -0.171 -4.342 -6.652 -1.599 [0.734] [0.693] [0.915] [0.943] [0.316] [0.223] [0.829] JCE 0.325 1.269 0.372 1.340 -5.894 -4.363 -1.536 [0.887] [0.539] [0.885] [0.703] [0.265] [0.521] [0.881] MSCE 2.153 1.823 -0.258 -0.767 -15.890+ -10.826 -0.373 [0.397] [0.678] [0.957] [0.896] [0.135] [0.405] [0.983] Number of infant (<5yo) in HH -2.766*** -4.092*** -4.592*** -6.785*** -9.154*** -10.031*** -9.751*** [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] Number of children (5-14yo) in the household -3.433*** -4.063*** -4.556*** -5.737*** -6.848*** -7.026*** -8.640*** [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] Number of prime adults (15-60yo) in HH -2.939*** -3.856*** -4.221*** -4.003*** -4.337*** -5.659*** -8.076*** [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] Number of elderly (60yo+) in HH -3.744** -7.800*** -6.678*** -4.867** -7.138* -6.759 -1.831 [0.022] [0.000] [0.001] [0.024] [0.086] [0.199] [0.780] Household access to loan (0/1) 2.727** 1.593 1.089 0.723 0.871 0.757 2.754 [0.048] [0.318] [0.452] [0.728] [0.632] [0.784] [0.549] Normalized TLU index 32.917 12.544 48.236 89.698* 99.853+ 19.958 13.738 52 [0.242] [0.790] [0.281] [0.097] [0.138] [0.759] [0.912] Normalized agricultural asset index + Normalized land holdings 19.665*** 22.493*** 23.349*** 27.503*** 32.530*** 32.966*** 46.026*** [0.000] [0.000] [0.000] [0.000] [0.001] [0.002] [0.006] Household was affected negatively by some income Shock During the last 12 months 0.662 0.474 1.783 0.960 4.360+ 6.643+ 5.915+ (0/1) [0.593] [0.644] [0.167] [0.647] [0.131] [0.115] [0.110] Household has a migrant network (0/1) -0.970 -1.881+ -1.215 0.127 0.220 2.956 2.504 [0.340] [0.121] [0.396] [0.948] [0.938] [0.361] [0.583] Rain - EA level CoV of Dec-Jan rainfall from 1983/84 - 2012/13 -13.093 -22.098 -28.994 -39.669* -60.347+ -71.742 34.300 [0.515] [0.348] [0.247] [0.087] [0.141] [0.187] [0.677] Average 12-month total rainfall (mm) 0.019 0.013 0.011 0.019 0.026 0.025 0.014 [0.443] [0.634] [0.733] [0.561] [0.508] [0.566] [0.857] Annual Mean Temperature (∞C * 10) -0.096 -0.076 -0.022 -0.003 0.201 0.528 1.146* [0.406] [0.537] [0.886] [0.988] [0.683] [0.244] [0.077] HH Distance in (KMs) to Nearest Road -0.100 -0.151 -0.148 -0.034 -0.189 -0.365 -0.212 [0.640] [0.382] [0.437] [0.859] [0.694] [0.483] [0.772] HH Distance in (KMs) to Nearest Population Center with +20,000 0.052 0.062 0.026 -0.109 -0.045 -0.043 -0.555+ [0.635] [0.642] [0.817] [0.431] [0.758] [0.870] [0.116] Price of fertilizer in the EA -0.006 -0.012** -0.011* -0.014* -0.018 -0.017 -0.010 [0.390] [0.038] [0.064] [0.086] [0.173] [0.295] [0.731] Price of maize grains in the EA 0.056** 0.033 0.026 0.073*** 0.082*** 0.097** 0.195** [0.024] [0.154] [0.323] [0.009] [0.003] [0.011] [0.010] Time dummy (year 2010=1) 0.005 -3.124+ -3.459* -2.524 -3.625 -2.797 4.653 [0.998] [0.138] [0.075] [0.341] [0.297] [0.388] [0.571] [.] Time average of explanatory variables District dummies Constant 36.094** 33.481** 30.597** 44.849*** 62.689* 65.833+ 34.729 [0.011] [0.016] [0.028] [0.002] [0.079] [0.100] [0.489] 5,321 5,321 5,321 5,321 5,321 5,321 5,321 Source: Generated by authors using LSMS data Note: PSLC=Primary School Leaving Certificate. JCE=Junior Certificate Examination. MSCE=Malawi School Certificate of Education Examination. ***, **, *, and + indicate that the corresponding regression coefficient is statistically significant at the 1%, 5%, 10%, and 15% levels, respectively. The table presents quantile partial effects of explanatory variables from 10th to the 90th percentile. P-values between. 53 Appendix 8: Effects of participation in the non-farm activities on poverty incidence, gap and severity in rural Malawi Poverty incidence Poverty gap Poverty severity VARIABLES CRE Probit Linear FE CRE Fractional CRE CRE Linear FE CRE Fractional CRE CRE Linear Probit Probit Tobit Probit Probit Tobit FE Wage employment participation -0.074** -0.072** -0.034*** -0.034** -0.033** -0.034** -0.018** -0.018* -0.017** -0.019** [0.012] [0.022] [0.007] [0.019] [0.011] [0.018] [0.021] [0.056] [0.017] [0.045] Self-employment participation -0.085*** -0.083*** -0.035*** -0.035*** -0.035*** -0.030*** -0.018*** -0.018*** -0.017*** -0.014** [0.000] [0.000] [0.000] [0.000] [0.000] [0.002] [0.000] [0.001] [0.000] [0.011] Age of the household head 0.003** 0.002* 0.001+ 0.001 0.001* 0.001 0.000+ 0.000 0.000* 0.000 [0.027] [0.069] [0.110] [0.154] [0.064] [0.279] [0.142] [0.195] [0.058] [0.301] Male-headed household (0/1) -0.044+ -0.045 -0.025** -0.026** -0.023** -0.022+ -0.010* -0.011* -0.010* -0.008 [0.120] [0.206] [0.020] [0.020] [0.035] [0.103] [0.091] [0.077] [0.076] [0.315] Highest level of formal education acquired by household head PSLC 0.043 0.060+ 0.005 0.003 0.008 0.016 -0.000 -0.001 0.003 0.007 [0.288] [0.142] [0.775] [0.849] [0.611] [0.272] [0.990] [0.939] [0.676] [0.392] JCE -0.096* -0.061 -0.044*** -0.046** -0.042** -0.015 -0.019** -0.020** -0.018** -0.001 [0.051] [0.282] [0.004] [0.012] [0.017] [0.431] [0.013] [0.030] [0.031] [0.954] MSCE -0.075 -0.010 -0.037 -0.040 -0.035 0.004 -0.014 -0.015 -0.013 0.008 [0.347] [0.888] [0.185] [0.176] [0.217] [0.869] [0.376] [0.342] [0.318] [0.576] Maximum level of formal education acquired in the household PSLC -0.002 -0.016 0.008 0.008 0.004 -0.004 0.004 0.004 0.002 -0.003 [0.949] [0.644] [0.554] [0.565] [0.729] [0.750] [0.577] [0.598] [0.783] [0.694] JCE -0.070* -0.069+ -0.014 -0.015 -0.020 -0.015 -0.009 -0.009 -0.011+ -0.009 [0.077] [0.119] [0.328] [0.345] [0.162] [0.347] [0.280] [0.312] [0.113] [0.331] MSCE -0.090+ -0.077 -0.026 -0.026 -0.030 -0.026 -0.012 -0.012 -0.015 -0.012 [0.122] [0.201] [0.227] [0.299] [0.166] [0.304] [0.377] [0.422] [0.174] [0.394] Number of infant (<5yo) in HH 0.074*** 0.078*** 0.028*** 0.028*** 0.030*** 0.030*** 0.013*** 0.013*** 0.015*** 0.014*** [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] Number of children (5-14yo) in the household 0.074*** 0.081*** 0.029*** 0.029*** 0.030*** 0.030*** 0.014*** 0.014*** 0.015*** 0.014*** [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] Number of prime adults (15-60yo) in HH 0.052*** 0.047*** 0.024*** 0.024*** 0.024*** 0.021*** 0.012*** 0.012*** 0.012*** 0.010*** [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] Number of elderly (60yo+) in HH 0.104*** 0.103*** 0.040*** 0.040*** 0.042*** 0.034*** 0.019*** 0.019*** 0.020*** 0.015** [0.001] [0.003] [0.000] [0.000] [0.000] [0.010] [0.001] [0.002] [0.000] [0.048] Household access to loan (0/1) -0.000 -0.003 -0.015* -0.015* -0.010 -0.019** -0.013*** -0.013*** -0.008* - 0.016*** [0.984] [0.921] [0.091] [0.091] [0.277] [0.030] [0.008] [0.006] [0.066] [0.001] Normalized TLU index -1.535** -0.921* -0.426 -0.434 -0.537* -0.102 -0.224 -0.222 -0.262* -0.018 [0.045] [0.086] [0.181] [0.201] [0.056] [0.554] [0.217] [0.256] [0.057] [0.835] Normalized agricultural asset index + Normalized land -0.478*** -0.465*** -0.195*** -0.197*** -0.198*** -0.194*** -0.102*** -0.103*** -0.101*** - holdings 0.100*** [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] Household was affected negatively by some income Shock -0.015 -0.023 -0.006 -0.005 -0.005 -0.003 0.001 0.001 -0.000 0.004 54 During the last 12 months (0/1) [0.541] [0.477] [0.523] [0.608] [0.613] [0.818] [0.844] [0.856] [1.000] [0.593] Household has a migrant network (0/1) 0.007 0.016 -0.006 -0.006 -0.001 -0.007 -0.004 -0.005 -0.001 -0.006 [0.769] [0.551] [0.517] [0.528] [0.875] [0.519] [0.367] [0.375] [0.787] [0.368] Rain - EA level CoV of Dec-Jan rainfall from 1983/84 - 0.900** 0.171 0.164 0.239 0.057 0.056 0.098 2012/13 [0.018] [0.223] [0.431] [0.249] [0.448] [0.603] [0.317] Average 12-month total rainfall (mm) -0.000 -0.001 -0.000** -0.000* -0.000 -0.000* -0.000** -0.000* -0.000 -0.000+ [0.745] [0.209] [0.020] [0.061] [0.201] [0.100] [0.013] [0.055] [0.190] [0.130] Annual Mean Temperature (∞C * 10) 0.001 -0.000 0.001 0.001 0.001 0.001 0.000 -0.000 0.000 -0.000 [0.659] [0.919] [0.515] [0.554] [0.494] [0.632] [0.975] [0.988] [0.667] [0.968] HH Distance in (KMs) to Nearest Road -0.001 -0.000 -0.001 -0.001 -0.000 -0.002 -0.001 -0.001 -0.000 -0.001 [0.835] [0.949] [0.466] [0.502] [0.787] [0.249] [0.355] [0.360] [0.660] [0.246] HH Distance in (KMs) to Nearest Population Center with -0.001 0.000 0.000 0.000 -0.000 0.001+ 0.000 0.000 0.000 0.000 +20,000 [0.510] [0.897] [0.890] [0.885] [0.863] [0.128] [0.539] [0.527] [0.959] [0.193] Price of fertilizer in the EA 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 [0.632] [0.892] [0.341] [0.413] [0.486] [0.792] [0.450] [0.520] [0.484] [0.767] Price of maize grains in the EA -0.001** -0.001+ -0.000*** -0.000* -0.000* -0.000* -0.000*** -0.000* -0.000* -0.000* [0.022] [0.112] [0.003] [0.055] [0.054] [0.059] [0.008] [0.083] [0.063] [0.084] Time dummy (year 2010=1) 0.010 0.011 0.015 0.015 0.013 0.012 0.010 0.011 0.008 0.010 [0.786] [0.857] [0.273] [0.436] [0.513] [0.586] [0.177] [0.332] [0.409] [0.462] Time average included Included Included Included Included Included Included Included District dummies included Included Included Included Included Included Included Included Constant 1.578+ 0.629+ 0.388* [0.137] [0.131] [0.099] Observations 5,317 5,321 5,317 5,321 5,321 5,321 5,317 5,321 5,321 5,321 R-squared 0.130 0.141 0.107 Number of households 2,764 2,764 2,764 Source: Generated by authors using LSMS data Note: PSLC=Primary School Leaving Certificate. JCE=Junior Certificate Examination. MSCE=Malawi School Certificate of Education Examination. ***, **, *, and + indicate that the corresponding regression APE is statistically significant at the 1%, 5%, 10%, and 15% levels, respectively. The table presents APEs on all explanatory variables. The APEs from the CRE Probit models are estimated using GEE approach. The CRE fractional Probit APEs are from GLM estimation. The unconditional APEs from CRE Tobit are estimated using pooled Tobit regression methods. The linear FE results are presented for comparison purpose. P-values between brackets are based on cluster robust standard errors 55 Appendix 9: Effects of participation in the non-farm activities on food security and subjective well being Food insecure Adequacy of food consumption VARIABLES CRE Probit Linear FE CRE ordered Probit Linear FE Wage employment participation -0.047+ -0.065** 0.177** 0.104** [0.125] [0.041] [0.037] [0.021] Self-employment participation -0.034+ -0.027 0.057 0.025 [0.140] [0.281] [0.379] [0.482] Age of the household head 0.000 -0.000 -0.002 -0.000 [0.743] [0.952] [0.591] [0.793] Male-headed household (0/1) 0.016 0.009 0.103 0.045 [0.596] [0.792] [0.268] [0.372] Highest level of formal education acquired by household head PSLC 0.071+ 0.074+ -0.068 -0.027 [0.121] [0.146] [0.576] [0.691] JCE -0.053 -0.067 0.055 0.038 [0.355] [0.252] [0.730] [0.607] MSCE -0.141** -0.119 0.344+ 0.180+ [0.036] [0.151] [0.116] [0.108] Maximum level of formal education acquired in the household PSLC -0.029 -0.029 0.107 0.042 [0.385] [0.430] [0.290] [0.442] JCE 0.087* 0.095* 0.177 0.083 [0.072] [0.055] [0.169] [0.236] MSCE 0.092 0.113+ -0.069 -0.066 [0.188] [0.137] [0.698] [0.504] Number of infant (<5yo) in HH 0.023* 0.026* -0.063* -0.032+ [0.073] [0.056] [0.093] [0.102] Number of children (5-14yo) in the household 0.019** 0.024*** -0.057** -0.031** [0.029] [0.007] [0.025] [0.032] Number of prime adults (15-60yo) in HH 0.008 0.010 -0.073** -0.034* [0.461] [0.506] [0.011] [0.089] Number of elderly (60yo+) in HH 0.035 0.021 -0.157* -0.073+ [0.301] [0.613] [0.095] [0.117] Household access to loan (0/1) 0.022 0.024 -0.135* -0.067* [0.377] [0.403] [0.052] [0.087] Normalized TLU index -1.001+ -0.904* 1.198 1.257 [0.103] [0.068] [0.295] [0.163] Normalized agricultural asset index + Normalized land holdings -0.179*** -0.173** 0.876*** 0.414*** 56 [0.006] [0.022] [0.000] [0.000] Household was affected negatively by some income Shock During the last 12 months (0/1) 0.112*** 0.107*** -0.274*** -0.147*** [0.000] [0.000] [0.000] [0.002] Household has a migrant network (0/1) 0.019 0.030 0.023 -0.001 [0.465] [0.296] [0.754] [0.977] Rain - EA level CoV of Dec-Jan rainfall from 1983/84 - 2012/13 0.425 -0.531 [0.270] [0.616] Average 12-month total rainfall (mm) 0.001 0.001 -0.001 -0.001 [0.291] [0.184] [0.613] [0.399] Annual Mean Temperature (∞C * 10) -0.000 0.001 0.006 0.003 [0.832] [0.784] [0.330] [0.450] HH Distance in (KMs) to Nearest Road 0.006+ 0.004 0.001 0.005 [0.123] [0.293] [0.919] [0.331] HH Distance in (KMs) to Nearest Population Center with +20,000 0.002 0.000 -0.004 -0.003 [0.337] [0.909] [0.404] [0.444] Price of fertilizer in the EA 0.000 0.000 0.000 0.000 [0.222] [0.203] [0.740] [0.895] Price of maize grains in the EA 0.002*** 0.002*** -0.002** -0.001* [0.000] [0.002] [0.026] [0.078] year_2011 0.101*** 0.122** -0.114 -0.066 [0.009] [0.036] [0.263] [0.335] Constant -2.209* 2.298+ [0.061] [0.115] Observations 5,317 5,321 5,321 5,321 R-squared 0.058 0.044 Number of Households 2,764 2,764 2,764 Source: Generated by authors using LSMS data Note: PSLC=Primary School Leaving Certificate. JCE=Junior Certificate Examination. MSCE=Malawi School Certificate of Education Examination. ***, **, *, and + indicate that the corresponding regression APE is statistically significant at the 1%, 5%, 10%, and 15% levels, respectively. The table presents APEs except in the ordered Probit cases. The APEs from the CRE Probit models are estimated using GEE approach. The coefficients form the CRE ordered Probit are from pooled ordered Probit. The linear FE results are presented for comparison purpose. P-values between brackets are based on cluster robust standard errors. 57 Appendix 10: Effects of participation in non-farm activities on inputs purchases for farm households only Fertilizer purchase Inputs purchase decision Value of fertilizer purchase per Value of inputs purchase per acre Land cultivated in decision (0/1 (0/1) acre of land cultivated of land cultivated acre VARIABLES CRE Probit Linear FE CRE Probit Linear FE CRE Tobit Linear FE CRE Tobit Linear FE CRE Linear Tobit FE Wage employment participation 0.039 0.037 0.046 0.037 0.156 -0.638 -0.124 -0.874 -0.292 0.274 [0.190] [0.303] [0.160] [0.260] [0.745] [0.419] [0.802] [0.305] [0.699] [0.800] Self-employment participation 0.045** 0.055** 0.053** 0.060** 0.388 0.136 0.234 0.085 0.710 -0.576 [0.043] [0.036] [0.031] [0.039] [0.225] [0.786] [0.495] [0.870] [0.561] [0.893] Age of the household head -0.004*** -0.003** -0.001 -0.000 -0.034+ -0.001 -0.007 -0.011 0.031 0.303 [0.003] [0.015] [0.654] [0.777] [0.111] [0.961] [0.714] [0.639] [0.241] [0.312] Male-headed household (0/1) -0.013 -0.003 -0.035 -0.027 -0.707 -1.346** -1.105** -1.417** -1.272 3.526 [0.689] [0.942] [0.311] [0.535] [0.211] [0.026] [0.030] [0.023] [0.179] [0.623] Highest level of formal education acquired by household head PSLC -0.006 -0.017 0.022 0.017 -0.395 -0.887 -0.226 -0.764 0.895+ 13.556 [0.882] [0.693] [0.620] [0.708] [0.458] [0.291] [0.702] [0.392] [0.138] [0.270] JCE 0.075 0.077 -0.026 -0.061 0.976 1.177 0.496 1.481 1.209 6.053 [0.182] [0.165] [0.684] [0.287] [0.260] [0.423] [0.602] [0.339] [0.270] [0.264] MSCE 0.062 0.065 -0.033 -0.028 0.755 2.555 0.489 3.071 1.690 8.861 [0.442] [0.385] [0.712] [0.719] [0.590] [0.346] [0.760] [0.279] [0.273] [0.158] Maximum level of formal education acquired in the household PSLC 0.015 0.017 -0.003 -0.010 0.243 0.151 0.058 0.199 -1.708* -17.457 [0.644] [0.644] [0.931] [0.806] [0.647] [0.803] [0.909] [0.756] [0.065] [0.267] JCE -0.032 -0.002 0.059 0.067 -1.003* -1.702** -0.428 -1.653* -1.362 -9.212 [0.457] [0.973] [0.221] [0.171] [0.079] [0.046] [0.506] [0.074] [0.159] [0.299] MSCE -0.013 0.038 0.092+ 0.122* 0.288 1.060 1.623 1.620 -2.040* -11.930 [0.822] [0.550] [0.148] [0.078] [0.819] [0.617] [0.310] [0.472] [0.096] [0.217] Number of infant (<5yo) in HH -0.024* -0.022* -0.001 0.003 -0.214 -0.023 -0.047 -0.039 2.165 9.402 [0.054] [0.090] [0.953] [0.849] [0.319] [0.938] [0.818] [0.904] [0.296] [0.304] Number of children (5-14yo) in the household 0.006 0.003 0.004 0.006 -0.095 -0.314+ -0.149 -0.301 0.525 2.975 [0.529] [0.765] [0.664] [0.549] [0.505] [0.125] [0.345] [0.194] [0.460] [0.477] Number of prime adults (15-60yo) in HH 0.027*** 0.021** 0.014 0.011 0.189 -0.109 0.040 -0.106 0.243 -0.202 [0.008] [0.049] [0.221] [0.326] [0.232] [0.642] [0.809] [0.665] [0.248] [0.776] Number of elderly (60yo+) in HH -0.003 -0.011 -0.007 -0.014 -0.161 -0.376 -0.312 -0.365 -0.151 -6.750 [0.927] [0.734] [0.837] [0.687] [0.743] [0.549] [0.555] [0.604] [0.777] [0.340] Household access to loan (0/1) 0.013 0.017 0.030 0.046+ 0.833** 1.291* 0.918** 1.170* -0.166 1.353 [0.566] [0.504] [0.263] [0.111] [0.035] [0.065] [0.025] [0.080] [0.896] [0.702] Normalized TLU index 0.519 0.941* 0.552 0.436 16.426** 47.688*** 14.816* 45.936*** 1.380 13.952 [0.278] [0.067] [0.473] [0.433] [0.013] [0.005] [0.075] [0.008] [0.678] [0.549] Normalized agricultural asset index + 0.260*** 0.252*** 0.210*** 0.219*** 2.757*** 1.438 2.326** 1.643 2.436 -7.698 Normalized land holdings 58 [0.000] [0.000] [0.001] [0.000] [0.001] [0.286] [0.013] [0.263] [0.198] [0.321] Household was affected negatively by some -0.006 0.000 -0.008 -0.005 0.030 0.712 0.194 0.588 1.490+ 4.532+ income Shock During the last 12 months (0/1) [0.815] [0.989] [0.765] [0.873] [0.940] [0.211] [0.653] [0.294] [0.131] [0.123] Household has a migrant network (0/1) 0.027 0.013 -0.035 -0.046+ -0.180 -0.794+ -0.622+ -0.692 0.555 4.839 [0.281] [0.639] [0.204] [0.129] [0.662] [0.104] [0.135] [0.175] [0.363] [0.174] Rain - EA level CoV of Dec-Jan rainfall from -1.141*** 0.253 -13.870+ 2.814 -25.396 1983/84 - 2012/13 [0.005] [0.559] [0.147] [0.749] [0.483] Average 12-month total rainfall (mm) -0.001 0.000 -0.001 -0.000 -0.020** -0.035 -0.026** -0.042+ -0.009 -0.081 [0.150] [0.899] [0.203] [0.938] [0.036] [0.154] [0.045] [0.145] [0.474] [0.175] Annual Mean Temperature (∞C * 10) -0.007*** -0.002 -0.006** -0.004+ -0.095*** -0.110+ -0.113*** -0.138** 0.026 0.278 [0.004] [0.479] [0.025] [0.102] [0.001] [0.107] [0.000] [0.035] [0.609] [0.212] HH Distance in (KMs) to Nearest Road -0.001 -0.003 -0.002 -0.005 -0.115* -0.266** -0.156** -0.302** 0.165 0.034 [0.764] [0.644] [0.691] [0.308] [0.057] [0.024] [0.011] [0.014] [0.279] [0.860] HH Distance in (KMs) to Nearest Population -0.003 -0.003 -0.003 -0.001 -0.000 0.067 0.023 0.095 -0.037 0.114 Center with +20,000 [0.171] [0.244] [0.217] [0.768] [0.994] [0.609] [0.669] [0.550] [0.267] [0.532] Price of fertilizer in the EA -0.000+ -0.000 -0.000 -0.000 -0.004* -0.002 -0.001 -0.002 0.003 0.072 [0.109] [0.162] [0.437] [0.442] [0.058] [0.416] [0.634] [0.428] [0.327] [0.231] Price of maize grains in the EA -0.000 0.000 0.000 0.000 0.006 0.011 0.010 0.013 -0.007 0.022 [0.982] [0.950] [0.541] [0.969] [0.371] [0.181] [0.211] [0.151] [0.669] [0.768] Share of total land cultivated in crops grains crops 0.001 -0.006** -0.004** -0.004 0.027 0.081* -0.001 0.088* 0.012 -0.124 [0.348] [0.045] [0.016] [0.189] [0.281] [0.083] [0.923] [0.065] [0.600] [0.371] legumes crops 0.001 -0.006** -0.003* -0.004 -0.003 0.032 -0.026* 0.037 0.018 -0.121 [0.650] [0.029] [0.053] [0.294] [0.916] [0.514] [0.096] [0.452] [0.460] [0.431] tubers crops -0.001 -0.008*** -0.005** -0.005+ -0.014 0.040 -0.030 0.054 -0.027+ -0.243 [0.647] [0.009] [0.013] [0.148] [0.654] [0.415] [0.207] [0.297] [0.108] [0.162] oil crops 0.007*** 0.001 0.044 -0.010 0.026 [0.007] [0.811] [0.372] [0.828] [0.497] Horticulture crops 0.001 -0.006** -0.002 -0.002 -0.004 0.033 -0.019 0.037 0.010 -0.210 [0.472] [0.038] [0.336] [0.503] [0.875] [0.493] [0.320] [0.450] [0.434] [0.371] cotton crops -0.002 -0.009*** -0.003 -0.004 -0.060* -0.020 -0.049*** -0.013 0.003 -0.149 [0.346] [0.006] [0.227] [0.313] [0.092] [0.690] [0.010] [0.789] [0.927] [0.290] tobacco crops 0.007*** 0.000 0.002 0.001 0.095*** 0.158*** 0.065*** 0.159*** 0.110 0.089 [0.000] [0.999] [0.366] [0.866] [0.001] [0.002] [0.000] [0.002] [0.294] [0.295] Other crops -0.007** -0.001 0.044 0.046 -0.475 [0.032] [0.820] [0.384] [0.372] [0.376] Total value of crop sales (MKW) 0.000 0.000 -0.000 -0.000 0.000*** 0.000*** 0.000*** 0.000*** -0.000 -0.000 [0.793] [0.891] [0.787] [0.409] [0.001] [0.000] [0.000] [0.000] [0.532] [0.621] Time dummy (year 2010=1) -0.045 -0.028 0.001 -0.015 -1.800*** -2.110*** -1.294** -2.397*** 0.088 15.212 [0.206] [0.490] [0.989] [0.741] [0.002] [0.002] [0.046] [0.002] [0.964] [0.339] District dummies included Constant 0.873 1.835+ 43.646+ 55.424* -46.400 [0.455] [0.105] [0.117] [0.076] [0.430] 59 Observations 5,038 5,038 5,038 5,038 5,038 5,038 5,038 5,038 5,038 5,038 R-squared 0.073 0.049 0.119 0.131 0.033 Number of households 2,675 2,675 2,675 2,675 2,675 Source: Generated by authors using LSMS data Note: PSLC=Primary School Leaving Certificate. JCE=Junior Certificate Examination. MSCE=Malawi School Certificate of Education Examination. ***, **, *, and + indicate that the corresponding regression APE is statistically significant at the 1%, 5%, 10%, and 15% levels, respectively. The table presents APEs on all explanatory variables. The APEs from the CRE Probit models are estimated using GEE approach. The unconditional APEs from CRE Tobit are estimated using pooled Tobit regression methods. The linear FE results are presented for comparison purpose. P-values between brackets are based on cluster robust standard errors. 60 Appendix 11: Multivariate recursive Probit estimation of effects of non-farm employment participation on activities on inputs purchases for farm households only. Fertilizer purchase Inputs purchase (1) (2) (3) VARIABLES Wage Self- Fertilizer Wage Self- Inputs employment employment purchased employment employment purchased Wage employment participation . 0.437*** . 0.271* [0.002] [0.053] Self-employment participation . 0.339*** . 0.237* [0.007] [0.051] Age of the household head 0.002 0.002 -0.013*** 0.002 0.001 -0.002 [0.722] [0.743] [0.008] [0.761] [0.775] [0.679] Male-headed household (0/1) 0.445*** 0.031 -0.062 0.447*** 0.026 -0.107 [0.003] [0.795] [0.585] [0.003] [0.826] [0.315] Highest level of formal education acquired by household head PSLC 0.199 0.075 -0.039 0.208 0.077 0.053 [0.274] [0.629] [0.793] [0.255] [0.623] [0.717] JCE 0.700*** 0.208 0.189 0.698*** 0.211 -0.085 [0.002] [0.305] [0.333] [0.002] [0.297] [0.666] MSCE 1.172*** 0.272 0.077 1.169*** 0.277 -0.140 [0.000] [0.334] [0.782] [0.000] [0.324] [0.624] Maximum level of formal education acquired in the household PSLC 0.127 0.044 0.038 0.122 0.041 -0.016 [0.427] [0.733] [0.760] [0.446] [0.749] [0.897] JCE 0.064 -0.219 -0.133 0.068 -0.219 0.168 [0.736] [0.188] [0.393] [0.724] [0.187] [0.275] MSCE 0.335 -0.365+ -0.074 0.343 -0.361+ 0.275 [0.175] [0.115] [0.737] [0.166] [0.119] [0.213] Number of infant (<5yo) in HH 0.030 0.087* -0.083* 0.031 0.089* -0.004 [0.607] [0.067] [0.070] [0.590] [0.063] [0.932] Number of children (5-14yo) in the household 0.094** 0.001 0.012 0.094** 0.002 0.011 [0.018] [0.975] [0.699] [0.018] [0.953] [0.719] 61 Number of prime adults (15-60yo) in HH -0.023 0.085** 0.084** -0.019 0.085** 0.041 [0.605] [0.021] [0.019] [0.671] [0.021] [0.239] Number of elderly (60yo+) in HH -0.079 -0.023 -0.004 -0.082 -0.019 -0.027 [0.612] [0.855] [0.969] [0.600] [0.880] [0.804] Household access to loan (0/1) -0.030 0.346*** 0.016 -0.025 0.346*** 0.076 [0.769] [0.000] [0.852] [0.812] [0.000] [0.369] Normalized TLU index -2.637+ 1.632 2.046 -2.536+ 1.661 1.912 [0.130] [0.252] [0.201] [0.149] [0.241] [0.354] Normalized agricultural asset index + Normalized land 0.846*** 0.614*** holdings [0.000] [0.003] Household was affected negatively by some income 0.158 -0.052 -0.036 0.163 -0.052 -0.027 Shock During the last 12 months (0/1) [0.191] [0.594] [0.683] [0.178] [0.596] [0.760] Household has a migrant network (0/1) 0.066 -0.017 0.083 0.076 -0.016 -0.103 [0.573] [0.855] [0.358] [0.517] [0.868] [0.230] Average 12-month total rainfall (mm) -0.002 0.001 -0.002 -0.001 0.001 -0.002 [0.416] [0.444] [0.199] [0.469] [0.426] [0.207] Annual Mean Temperature (∞C * 10) 0.007 -0.002 -0.022*** 0.006 -0.002 -0.017** [0.478] [0.843] [0.007] [0.498] [0.815] [0.036] HH Distance in (KMs) to Nearest Road -0.014 -0.006 -0.003 -0.015 -0.006 -0.006 [0.334] [0.643] [0.831] [0.303] [0.635] [0.626] HH Distance in (KMs) to Nearest Population Center 0.004 -0.000 -0.011* 0.004 -0.000 -0.008 with +20,000 [0.561] [0.965] [0.070] [0.547] [0.957] [0.208] Price of fertilizer in the EA -0.001* -0.000 -0.001 -0.001* -0.000 -0.000 [0.085] [0.473] [0.251] [0.084] [0.460] [0.513] Price of maize grains in the EA -0.002 0.002 0.000 -0.002 0.002 0.001 [0.229] [0.317] [0.893] [0.246] [0.290] [0.614] Total land area owned by HH in Acres -0.000 0.000 -0.000 0.000 [0.746] [0.755] [0.733] [0.804] value of ganyu salary in the area 0.000 0.000 0.000 0.000 [0.538] [0.373] [0.501] [0.370] EA neighbor’s wage employment participation 0.010*** 0.005*** 0.010*** 0.005*** [0.000] [0.005] [0.000] [0.005] EA neighbor’s self-employment participation 0.006*** 0.013*** 0.006*** 0.013*** [0.004] [0.000] [0.006] [0.000] Share of total land cultivated in crops 62 grains crops 0.005 -0.011** [0.364] [0.033] legumes crops 0.003 -0.009* [0.638] [0.088] tubers crops -0.003 -0.014** [0.714] [0.039] oil crops 0.024** 0.002 [0.023] [0.829] Horticulture crops 0.004 -0.005 [0.498] [0.386] cotton crops -0.007 -0.010 [0.343] [0.163] tobacco crops 0.022*** 0.005 [0.000] [0.377] Total value of crop sales (MKW) 0.000 -0.000 [0.439] [0.665] Time dummy (year 2010=1) -0.048 -0.050 -0.128 -0.048 -0.048 0.000 [0.763] [0.700] [0.339] [0.765] [0.714] [1.000] Normalized wealth index 2.175*** 2.298*** 2.082*** 2.207*** [0.000] [0.000] [0.000] [0.000] District dummies included included included included included included Time average of explanatory variables included included included included included included atrho21 -0.107*** -0.111*** [0.000] [0.000] atrho31 -0.157*** -0.068 [0.004] [0.191] atrho32 -0.100* -0.037 [0.089] [0.513] Likelihood ratio test of joint significance of rhos: 28.09*** 16.13*** Chi2(3) [0.000] [0.001] Constant -0.796 -2.522** 2.460** -0.816 -2.553** 0.558 [0.564] [0.025] [0.022] [0.555] [0.023] [0.583] Observations 5,018 5,018 5,018 5,018 5,018 5,018 Source: Generated by authors using LSMS data 63 Note: PSLC=Primary School Leaving Certificate. JCE=Junior Certificate Examination. MSCE=Malawi School Certificate of Education Examination. The symbols ***, **, *, and + indicate that the corresponding regression coefficient is statistically significant at the 1%, 5%, 10%, and 15% levels, respectively. P-values between brackets. The GHK estimates are generated using the mvprobit command with 100 draws in STATA. The parameters in the table are coefficients and not APEs. 64