Transient rural livelihoods and poverty in Ghana

Journal of Rural Studies 40 (2015) 102e110

Contents lists available at ScienceDirect

Journal of Rural Studies journal homepage: www.elsevier.com/locate/jrurstud

Transient rural livelihoods and poverty in Ghana Fred M. Dzanku Institute of Statistical, Social and Economic Research, University of Ghana, Ghana

a r t i c l e i n f o

a b s t r a c t

Article history: Received 30 April 2014 Received in revised form 2 June 2015 Accepted 22 June 2015 Available online xxx

Although agriculture remains the main economic livelihood activity for the majority of rural households in sub-Saharan Africa it has also been observed that livelihood diversity is the norm. Using household panel data from Ghana this article argues that aside from farming many rural economic livelihood options are transient because of a large gap between livelihood activity and professional vocation development. Further, the household welfare effect of being livelihood diversity transient is examined under the premise that of utmost relevance is the welfare implication of such behaviour. Evidence suggests that a substantial proportion (55%) of households exhibit considerable instability in livelihood diversity behaviour. More importantly, being livelihood diversity transient imposes statistically significant albeit economically marginal household welfare cost. The findings also reveal that spatial location of household, demographics factors, education, and consumereproducer price differentials were the most important determinants of rural household welfare. The key message from the findings is the need to develop entrepreneurial skills of rural households to acquire gainful employment opportunities, which leads to more stable livelihood diversity behaviour and poverty reduction. © 2015 Elsevier Ltd. All rights reserved.

Keywords: Ghana Transient Livelihood diversity Household welfare Panel data

1. Introduction This study examines the transience versus stability of rural livelihood diversity, and the extent to which such economic livelihood behaviour affects the welfare of rural households. Also, the study essentially offers a test of the position that ‘the process of trial-and-error [associated with rural economic livelihood diversity] can be costly’ in welfare terms (Bryceson, 2002b, pp. 736). Specifically, two main hypotheses are put forward and tested: (i) rural households exhibit transitory economic livelihood diversity behaviour, (ii) rural household welfare is decreasing with being livelihood diversity transient. Rural economies of most sub-Saharan Africa (SSA) countries are still largely agrarian, agriculture employs the largest proportion of the workforce and contributes the largest share of household income (Zezza et al., 2009; Davis et al., 2010). In Ghana, the most recent population census (Ghana Statistical Service, 2012) shows that about 42 percent of the economically active labour force are employed in agriculture. The literature has also documented evidence of a growing rural nonfarm economy. Indeed, for Ghana, the 42 percent of the labour force employed in agriculture represented a 9 percentage point decline over the figure recorded during the

E-mail address: [email protected]. http://dx.doi.org/10.1016/j.jrurstud.2015.06.009 0743-0167/© 2015 Elsevier Ltd. All rights reserved.

2000 population census. The extant literature suggest that rural households have an extremely diversified portfolio of income generating activities (Ellis, 2000b, pp. 3e27; Barrett et al., 2001b). The concept of livelihood diversity connotes, ex ante, the existence of two or more livelihood options per individual or household. Empirical evidence over the past decade from developing countries in general (e.g. Ellis, 2000b; Foster and Rosenzweig, 2004; Minot et al., 2006; Lanjouw and Murgai, 2009), SSA (e.g. Barrett et al., 2001a; Wouterse and Taylor, 2008; Lay et al., 2009; Stifel, 2010) and Ghana (e.g. Zezza et al., 2009; Anriquez and Daidone, 2010; Davis et al., 2010) point to household involvement in pluriactivity. Thus, ‘household diversification, not specialisation, is the norm’ (Davis et al., 2010, pp. 56). However, limited markets and entry barriers into rural nonfarm employment (Barrett et al., 2001b; Haggblade, 2007) means that even with the claim of ‘collapse of agriculture as the primary source of rural livelihoods in SSA’ (Ellis, 2010, pp. 54) seeking an alternative or supplementary income generating activity is not an easy task. Due to policy and institutional failures there is often inadequate attention paid to developing the capacity of rural households to engage in sustainable pro-poor activities outside agriculture (Bryceson, 2002a). As noted by Bryceson (1996, 2002b), occupational or labour specialisation is often nonexistent, making the search for survival in nonfarm rural employment one of trial-and-error.

F.M. Dzanku / Journal of Rural Studies 40 (2015) 102e110

Ellis (2000a; 2000b) has argued that rural livelihood diversity is pervasive and enduring rather than temporary engagements in nonfarm related activities in addition to agriculture for the purpose of overcoming shocks as claimed by Saith (1992). As Frank Ellis, Peter Timmer also had claimed that “the declining importance of agriculture is uniform and pervasive” (Timmer, 1988, pp. 276). This paper argues that there is a sense in which economic livelihood activities in rural Africa meet Saith's description and that the importance of agriculture to rural households may not be consistently declining in rural SSA. Aside from farming, many individual economic livelihood options in rural SSA are often transient because of the large gap between livelihood activity and professional vocation development (Bryceson, 2002a). This paper puts forward that rural economic livelihood options are often transient such that, over time, no sustained increase in diversification behaviour is observable. A substantial body of literature exists on the issue of whether or not diversification is driven by ‘push’ (survival) or ‘pull’ (accumulation) motives (e.g. Reardon et al., 1998; Ellis, 2000b; Barrett et al., 2001b; Little et al., 2001; Haggblade et al., 2002; Dimova and Sen, 2010). However, the welfare effect of transient versus stable rural livelihoods has neither been explored nor has the transience versus pervasiveness of rural livelihood diversity has not been previously studied empirically. This is where the current study contributes to the rural livelihoods and poverty literature. It is not obvious what the poverty implications will be for having a transient rather than a non-transient economic livelihood. This is because on the one hand economic livelihood mobility will be important in an uncertain economic environment and could be seen to have positive household welfare outcomes (Barrett et al., 2001a). On the other hand straddling economic livelihood options could be a sign that none provides security of livelihood (Grawert, 1998). Successful households are often those engaged in not just higher rewarding livelihood activities but those engaged that diversify over a longer period of time, not the transient (Tellegen, 1997). ‘The process of trial-and-error can be costly in time and money’ (Bryceson, 2002b, pp. 736). In order to better understand these research issues, Haggblade et al. (2007) highlights the need for panel data evidence. Among other advantages, such data allows controlling for unobserved preference for economic livelihood options as well as differences in attitudes towards risk. This paper contributes to the understanding of these issues using household panel data from Ghana. 2. Rural livelihood diversity: hypotheses and evidence from previous literature Rural household diversification behaviour is driven by either ‘distress’ (or ‘push’) or proactive (or ‘pull’) factors (Reardon et al., 1998; Haggblade et al., 2002). The so called push factors result from a search for survival due to precarious economic livelihood circumstances. Due to initial conditions (e.g. low private capital endowments), diversification under the ‘push’ motive involves engagement in low-return income generating activities which have low or no entry barriers. On the other hand, ‘pull’ motives are associated with attractive opportunities for wealth accumulation (Reardon et al., 1998; Barrett et al., 2001b), and often have barriers of entry. Survival or necessity driven diversification has poverty trap implications while that motivated by accumulation or choice may increase inequality if entry barriers persist or move households out of poverty if policy interventions reduce these barriers to entry (Dimova and Sen, 2010). In effect, ex ante asset rich households often have more income generation options than asset poor households (Barrett et al., 2001a). The determinants of the pull and push motives include factors

103

relating to seasonality, attitudes to risk, coping behaviour, labour market behaviour, credit market behaviour, and asset holding strategies (Ellis, 2000b). The classical political economy and neoclassical notion of surplus labour also provides an explanation for rural household diversification behaviour (Bryceson, 1996). Other hypothesised reasons for diversification include diminishing returns to labour, market failure, ex ante risk management, ex post copping with adverse shocks, availability of social insurance, and economies of scope in production (Ellis, 2000b; Barrett et al., 2001b). These factors are influenced by contextual factorsdclimate/agroecology, macroeconomic policies, political systems, and institutional arrangements (Scoones, 1998; Barrett et al., 2001b; Bryceson, 2002a; Bryceson, 2002b; Havnevik et al., 2007; Zezza et al., 2009; Ellis, 2010). On empirical evidence of the determinants of diversification, Canagarajah et al. (2001) analysed data from the first and third rounds of the Ghana Living Standards Survey (GLSS) and found that households in remote areas were less diversified than those located near urban areas. Abdulai and CroleRees (2001) reached the same conclusion using data from Mali. Block and Webb (2001) on Ethiopia, Abdulai and CroleRees ^te d'Ivoire and Kenya (2001) on Mali, Barrett et al. (2001a) on Co have all identified wealth as an important determinant of diversification into high-return income generating activities. Poorer households have generally been found to have limited opportunities in non-crop oriented income generating activities and hence have less diversified incomes. More recent evidence is provided by Dimova and Sen (2010) using panel data from Tanzania. Smith et al. (2001), however, found an inverted U-shaped relationship between wealth and diversification in two rural districts of Uganda: households at the lowest and highest ends of the wealth distribution were relatively less diversified than those in the middle. The poorest third are constrained due to entry barriers while the richest third choose to specialise in high return livelihood activities. Barrett et al. (2001a) provide evidence of macroeconomic policy effects on rural income diversification using data from Cote d'Ivoire and Kenya. They found that devaluation of the exchange rate induced significant shift into agriculture and reduced income shares derived from nonfarm activities. The policy also induced mobility between economic livelihood strategies through reallocation of labour and household assets across activities. Poorer households remained stuck in unskilled labour and the production of nontradables and hence suffered real income losses. Asset rich households on the other hand gained through asset-enhanced diversification strategy mobility. Addressing issues of occupational skill acquisition and liquidity constraints are thus critical to avoiding being trapped in low-return, high-risk economic livelihood activities. The current research develops two key research questions derived from previous literature. The first is whether or not livelihood diversity is temporary or pervasive in rural areas of developing countries where farming tends to be the dominant economic livelihood option, but also where engagement in rural nonfarm income-earning activities has been observed to be growing. The second is whether or not being livelihood diversity transient imposes significant household welfare costs. These questions have not been previously addressed systematically although claims have been made (Timmer, 1988; Saith, 1992; Ellis, 2000a, 2000b). Answers to these questions should provide insight into whether or not the on-going agriculture-led rural poverty reduction renaissance is pragmatic in terms of both analytical and policy thinking of rural economic development. In addition, the analysis contributes to evidence on the issue of whether the notion of a shift away from farm oriented rural livelihoods as suggested by some (e.g. Ellis, 2010) is real or hyperbole.

104

F.M. Dzanku / Journal of Rural Studies 40 (2015) 102e110

3. Data and indicators 3.1. Source of data The study employs a balanced three-period panel dataset on 464 households (1392 observations) drawn from eight villages in the Eastern and Upper East administrative regions of Ghana. These regions are located in two distinct agroecological zones: the Semideciduous Forest and the Sudan Savanna. The surveys were conducted in 2002, 2006 and 2010. The sampling procedure was multistage and involved the selection of region, district and village based on agroecological potential and proximity or isolation from national, regional and district capitals. Based on the above criteria, two districts were selected from each region and then two villages from each of the four districts. At the village level, between 50 and 60 households were randomly selected. Two main instruments were used for data collection, a household questionnaire and village level interview guide which contained issues for focused group discussions and key informant interviews. The household questionnaire solicited detailed information on general household characteristics, human resource use, nonfarm activity and incomes, non-labour incomes and transfers, physical assets and amenities, social network capital, food and nonfood consumption expenditure, agricultural production (including livestock) and marketing activities, among others. 3.2. Indicators The purpose of this article is to study the transience and welfare implications of rural household livelihood diversity. In the rural livelihoods and related literature, the term diversification has been ambiguously used. Ellis (2000b, pp. 14) defined diversification in relation to the holding of multiple income sources. However, the phenomenon of movement away from (or relying less on) agriculture has also been labelled diversification (e.g. Abdulai and CroleRees, 2001; Minot et al., 2006). Most of the empirical literature cited earlier measured diversification simply by variations in the amount of nonfarm income generated across households. Except in a few cases (Ersado, 2006; Dimova and Sen, 2010; Baird and Gray, 2014) where the inverse of the Herfindahl-Index has been used the majority of studies have not examined both the number of income sources and magnitude of income generated by each income source within a household's livelihood portfolio. In this paper, we apply the Simpson's index for constructing an income diversity index for household i in time t as

div ¼ 1 

8 < X N :

i¼1

y2i

!2 !2 31 9 N = X 4 yi 5 ; i¼1

where N is the number of income sources per household and y is the amount of income generated from each source at each time period. This index considers both the number and magnitude of each income source, and lies between zero (if a household obtain income from only one source) and one (if a household has infinite income sources). A binary livelihood diversity status variable is useful for the current analysis; a household is defined as diversified if the value of its labour income diversity index is positive. This indicator is important for examining the transience versus pervasiveness of rural livelihood diversity. From this indicator a household can be classified as having a transient economic livelihood status if over the panel the household switches between being diversified and specialised, otherwise the household has a pervasive livelihood

diversity status. Two welfare indicators are employed for examining the welfare effects of transient versus stable economic livelihoods. The first is price adjusted adult equivalent consumption expenditure per year.1 The second is a composite welfare index which takes into account the multidimensional nature of welfare by combining money and non-money metric indicators such as income, household durable assets, household dwelling characteristics (number of bedrooms per adult equivalent, type of roof and walls of dwelling, electricity, and toilet facilities) using principal component analysis (PCA) techniques (see Finan et al., 2005). We use both money-metric and multidimensional measures because an improvement in well-being in the money-metric dimension does not necessarily translate to an improvement in other dimensions of welfare (Grosse et al., 2008). Using both helps establish the robustness of our findings.

4. Analytical methods This section explains how the hypotheses of the current research derived from the research questions are tested. For the first, panel tabulation of the binary diversity indicator is used to identify (in)stability in household diversification behaviour over time. In addition, transition probabilities are computed to answer the question of transience or pervasiveness of livelihood diversity. Defining household i as diversified or otherwise, let d and s be indicative of the household's position in year t. The estimated probability that a diversified household in time t has become specialised (or remained diversified) over a given period of the panel can be computed as

   Pr divi;tþ1 ¼ sdivi;t ¼ d ; c i ¼ 1; …; N; t ¼ 1; …; T where div is a dichotomous livelihood diversity indicator defined earlier. We also use cumulative density plots for verifying the presence of first-order stochastic dominance in the distribution of livelihood diversity over time. The second research question involves the determination of whether being livelihood diversity transient imposes significant welfare costs on a rural household. To do this, we model household welfare as a function of a transient economic livelihood measure:

welfit ¼ Xit0 b þ dtrani þ ci þ uit

i ¼ 1; …; N; t ¼ 1; …; T

(1)

where i and t indexes household and time respectively; welf is a continuous household welfare indicator (i.e. consumption expenditure and a composite welfare index); X is a vector of exogenous explanatory variables including a constant term; b is a vector of unknown coefficients to be estimated; tran is a binary variable that takes on the value 1 if the household has a transient livelihood diversity status, 0 if otherwise; d is the main coefficient of interest; c is the household specific unobserved effect; and u is the idiosyncratic error term. By construction tran is time invariant which means that d is identified only under the mean independence assumption (i.e. ci and the explanatory variables are uncorrelated). This leads to estimating equation (1) using the random effects (RE) estimator. Since the mean independence assumption may fail in many empirical applications we allow correlation between the unobserved effect and the explanatory variables within the RE framework by specifying

1 Consumption expenditures include both actual and imputed. The imputed expenditures involve consumption of own produced goods. These were aggregated using market prices. Total consumption expenditure was adjusted using the Paasche price indices.

F.M. Dzanku / Journal of Rural Studies 40 (2015) 102e110

0

welfit ¼ Xit0 b þ X i l þ dtrani þ ci þ uit

i ¼ 1; …; N; t ¼ 1; …; T (2)

where X is the time average of time-varying explanatory variables which allows correlation with the household specific effect (Mundlak, 1978; Wooldridge, 2010). Another issue to contend with is whether the estimate of d in equations (1) or (2) is a consistent estimate of the marginal effect of having a transient livelihood diversity status on household welfare. If households elect to be transient or not then the error term in a transient livelihood diversity status model is clearly correlated with the error term in the household welfare equation, meaning the estimate of d will be biased and inconsistent. Within the panel data framework, Vella and Verbeek (1998) derived an endogeneity correction term from a first-stage random effects probit model to be included in the primary regression of interest which is then estimated by OLS. In the current application, tran is time-invariant and so we specify the equations within the framework of the treatment-effects model (Maddala, 1983, pp. 117e122). Thus, together with the household welfare equation we specify trani as an outcome of the latent variable tran*i as

105

Previous literature (Karbasioun et al., 2005) observed higher job stability and satisfaction among households with such informal vocations. Bryceson (2002a) suggests that it is the lack of such occupational training that makes the difference between occupational stability and trail-and-error economic livelihood behaviour. It is also observed that a higher percentage of transient than nontransient households view farming as their dominants economic livelihood activity. The discussion in this section has focused on mean differences between transient and non-transient households with respect to the main outcome variables of interest as well as other household and household head related characteristics. In the next section, we control for factors that may confound the observed effect of transient livelihood behaviour on household welfare to determine whether such behaviour has implications, positive or negative, for rural household welfare. 6. Empirical results 6.1. Transience of economic livelihood diversity

Descriptive statistics of all variables used in the analysis are presented in Table 1. They are summarised by the transient livelihood diversity status variable. For 55 percent of households in our sample, livelihood diversity is transient, meaning that they are diversified at one point and specialised at another; the rest have non-transient livelihoods. We observe that transient households are on average less diversified than non-transient households but the average difference in the diversity index is small (0.04). Observed mean welfare is significantly lower for transient than non-transient households both in terms of consumption expenditure and the composite welfare index, but at a higher significance level for the later (0.01 versus 0.1). There are also differences between transient and non-transient households in terms of formal education of spouse, urban social network capital, informal vocational training, previous outmigration, and self-reported dominant sector of economic activity.2 A particularly interesting observation is the statistically significant difference in informal vocational training between the two groups.

This subsection presents results on the first research question: is livelihood diversity a transient phenomenon? First, we note that the overall average value of the livelihood diversity index is 0.32, which is lower than the 0.35 and 0.36 reported by Dimova and Sen (2010) and Baird and Gray (2014), respectively for different Tanzanian samples.3 The average values of the indices are 0.313, 0.315 and 0.325 in 2002, 2006 and 2010 respectively, showing very little change over time. Indeed, two sample t-tests show no statistically significant change in the index over time for each pair of years. Similarly, cumulative density plot of the index (Fig. 1) exhibit no first order stochastic dominance ordering over time. Given suggestions from the rural livelihoods literature, one would have expected that for every possible value of the diversity index, the diversity rates in years t ¼ 2 and t ¼ 3 will be above that in year t ¼ 1. If the argument of transient livelihood diversity holds, one will expect considerable instability in the diversity status variable. First, note that our sample is made up of 464 households but the between total frequency count is 718 (Table 2). This implies some instability of livelihood diversity status over time. Indeed, conditional on a household having a diversified status in an initial period, about 23 percent changed status in a next period. There is greater instability in specialisation behaviour of these rural households. Transition probabilities provide further evidence of transient livelihood diversity status. We observe that 77 percent of households ever having a diversified livelihood status remained so during a next period, the rest had become specialised (Table 3). Diversified households had a 23 percent chance of becoming specialised in a next period whereas specialised households had a 77 percent chance of becoming diversified. Further, the sample is split into pair of time periods and transition probabilities computed (Table 4). The results show more stability in diversification behaviour than specialisation behaviour. Nonetheless, as noted from the descriptive statistics in Section 5, taking all the three panel years together, more households (55%) change economic livelihood status, meaning that a substantial number of households are livelihood diversity transient. This result lends some support to Saith's (1992) argument and casts some doubt on the pervasive and enduring rural livelihood diversity

2 Household heads were asked to indicate whether they view farming, nonfarm work, or both equally as their dominant economic activity.

3 The diversity index values from the cited literature are subtracted from unity to make them comparable to the value of the index calculated in the present study.

tran*i ¼ Wi g þ ei

(3)

with the observed decision to have a transient livelihood diversity status being

 trani ¼

1; 0;

if tran*i > 0 otherwise

where W is a vector of exogenous explanatory variables and g is the vector of corresponding unknown parameters; u and e are bivariate normal with zero mean and covariance matrix



s2 rs

 rs : 1

If the correlation between the error terms, r, equals zero, then the estimate of d in equations (1) or (2) is consistent.

5. Descriptive statistics

106

F.M. Dzanku / Journal of Rural Studies 40 (2015) 102e110

Table 1 Descriptive statistics by household livelihood diversity status. Variable definition

Livelihood diversity index Adult equivalent consumption expenditure (GH¢) Household welfare index Female headed household, dummy (%) Age of household head (years) Household dependent proportion (%) Head has no formal education, dummy (%) Head completed basic education, dummy (%) Head has higher than basic education, dummy (%) Spouse has some formal education, dummy (%) Livestock holding (tropical livestock units) No urban social network capital, dummy (%) Urban social network without remittance, dummy (%) Urban social network with remittance, dummy (%) Producer/consumer staple price differential (GH¢ kg1) Informal vocational training, dummy (%) Number of household members previously out-migrated Farm dominant livelihood, dummy (%) Nonfarm dominant livelihood, dummy (%) No dominant economic livelihood, dummy (%)

Transient household N ¼ 254 (55%)

Non-transient households N ¼ 210 (45%)

Mean

Std. Dev.

Mean

Std. Dev.

0.30 319.14 7.18 17.85 46.27 0.43 70.60 16.01 13.39 16.40 3.77 37.66 14.30 48.03 0.16 23.62 1.19 68.11 5.77 26.12

0.19 242.15 1.71 n.a 15.26 0.24 n.a n.a n.a n.a 13.47 n.a n.a n.a 0.11

0.34 343.25 7.60 18.41 46.64 0.45 67.94 17.46 14.60 22.06 3.54 39.52 18.73 41.75 0.15 35.71 1.12 58.89 9.37 31.75

0.19 254.56 2.20

0.67 n.a n.a n.a

14.50 0.23 n.a n.a n.a n.a 9.78 n.a n.a n.a 0.10 n.a 0.76 n.a n.a n.a

Difference

0.04*** 24.11* 0.43*** 0.56 0.37 0.01 2.67 1.45 1.22 5.66*** 0.23 1.86 4.43** 6.28** 0.01 12.09*** 0.07** 9.22*** 3.59** 5.63**

Note: GH¢ is the local currency unit which was equivalent to $0.86, $0.92 and $1.47 in 2002, 2006 and 2010 respectively; n.a denotes not applicable. The asterisks *, ** and *** denotes significance at 10%, 5% and 1%, respectively.

Fig. 1. Cumulative distribution of livelihood diversity in the study sample.

claim by Timmer (1988) and Ellis (2000a; 2000b). Next, we explore the important issue of household welfare implications of transient versus non-transient livelihood diversity behaviour. 6.2. Welfare effects of transient livelihood diversity The correlation coefficient between the two household welfare indicators used is 0.46 when measured in levels and 0.52 when computed in logarithms, and are both statistically significant at the 1 percent level. The value of the correlation coefficient also means that although positive and significant only about 27 percent of the variability between the two household welfare measures are shared together.4 So, the two indicators may not necessarily tell a consistent story and using both helps verify the robustness of our results.

4

Squaring the correlation coefficient and multiplying by 100 gives this result.

We first estimated equation (2) and tested the null hypothesis that the coefficients on the group means jointly equal zero.5 This is essentially testing for fixed or random effects. The Wald statistic on the tests were 34.74 and 51.29 for Models 1a and 2a, respectively (Table 5), which are far larger than the critical chi-squared values with four and six degrees of freedom, p-values equal zero to four decimal places. Since this provides evidence against the random effects estimator we retain the estimates from equation (2) which allows unobserved household heterogeneity to be correlated with the observed explanatory variables (Table 5). The main hypothesis of interest is on the coefficient of the transient diversity indicator, d. A negative sign is observed on the coefficients in both models 1a and 2a and are both statistically significant at the 5 percent level. Since we can interpret these coefficients like those in a linear regression model (Kohler and Kreuter, 2005, pp. 242), using Kennedy's approximation (Kennedy, 1981; Jan van Garderen and Shah, 2002), we estimate that average consumption expenditure is about 7 percent lower among transient than non-transient households, and composite welfare is approximately 4 percent lower among transient than non-transient households, all other variables held constant. Although the difference in welfare between the two groups is statistically significant the economic significance appears marginal. Models 1b and 2b takes into account the possibility that a household may elect to have a transient or non-transient economic livelihood, in which case the estimates of d reported under models 1a and 2a are biased and inconsistent due to the problem of endogeneity. Testing at even the 10 percent level, one will find insufficient evidence in the current data to reject a null hypothesis that the error terms in the welfare and transient diversity status equations are uncorrelated (Table 5). Even so, we still find a statistically significant negative effect of transient rural livelihood diversity on household welfare, all else held constant. Although previous research have not directed analysed the welfare implications of transient economic livelihood diversity

5 Group means of four and six time-varying variables were included in the consumption expenditure and composite welfare index models, respectively.

F.M. Dzanku / Journal of Rural Studies 40 (2015) 102e110

107

Table 2 Panel tabulation of household livelihood diversity status (N ¼ 464). Overall

Diversified Specialised Total

Between Percent

Freq.

Percent

Percent

1068 324 1392

76.72 23.28 100.0

460 258 718

99.14 55.60 154.7

77.39 41.86 64.62

Table 3 Transition probabilities for livelihood diversity status (N ¼ 464).

Diversified Specialised

Diversified

Specialised

540 (77.36) 178 (77.39)

158 (22.64) 52 (22.61)

Table 4 Rural livelihood diversity transitions by pair of years.

2002e2006 Diversified Specialised 2006e2010 Diversified Specialised 2002e2010 Diversified Specialised

Within

Freq.

Diversified

Specialised

75.14 74.56

24.86 25.44

79.60 80.17

20.40 19.83

78.29 84.21

21.71 15.79

combining our earlier observation (Section 5) that transient households are less diversified with the result in this section that they have lower average welfare agrees with the existing literature (e.g. Ersado, 2006; Dimova and Sen, 2010) that wealthier rural households tend to be more diversified. Parameter estimates from the endogenous transient livelihood equation can be found in Table 6.6 The probability of having a transient livelihood status is significantly decreasing with formal education of spouse, informal vocational training, and households' perception of the importance of nonfarm activities compared with farming in their economic livelihood portfolio. The observed effect of informal vocational training on the likelihood of having a transient livelihood is particularly an important one and is related to the argument of Bryceson (2002b) that the ‘uncertainty and wasted energy embedded in trial-and-error’ livelihoods is partly due to the lack of policy attention paid to developing long-term occupational futures in rural areas. Karbasioun et al. (2005) found evidence that informal technical vocational training led to job stability. We also observe that spatial location of a household significantly affect the likelihood of being livelihood diversity transient. For households located in the relatively dry semi-arid districts, the average probability of observing pervasive livelihood diversity behaviour is higher compared with households in the higher agroproductive potential district. Returning to Table 5 we explore other determinants of rural household welfare in our sample. We observe that welfare does not

6 Although not theoretically required, we included four variables in the endogenous transient livelihood diversity equation for the purpose of identification (Table 6). Test of the null that the overidentifying restrictions are valid (that these variables are rightly excluded from the welfare equations and are exogenous) yielded a chi-squared statistic value of 3.15 with three degrees of freedom, quite easily failing to reject the null (p-value ¼ 0.369). These variables strongly predict transient livelihood diversity status (F-statistic ¼ 21.24, p-value ¼ 0.000).

discriminate against living in a female headed household, which is consistent with findings from some developing country studies (e.g. Appleton, 1996; Justino et al., 2008) but contrary to results from others (e.g. Finan et al., 2005; Kijima and Gonzalez, 2013; Mensah et al., 2014). This means that the nature of the association between household headship and welfare is context specific. Other household demographic factors such as age and the proportion of dependants are important determinants of welfare in our study sample. We find that consumption expenditure is first decreasing with age of household head until average age 51 when a positive effect begins to manifest; approximately 63 percent of household heads in the sample are below age 51. As previous studies (e.g. Appleton, 1996; Minot and Baulch, 2005) have shown, having a high proportion of dependants in a household is associated with high welfare cost, particularly with respect to consumption expenditure. It is estimated that increasing the average household dependant proportion by an extra unit, all else held constant, leads to an average decrease in household consumption expenditure by about 33 percent; increasing the dependant proportion by one unit decreases average composite welfare by approximately 3 percent. Thus, having a high proportion of dependants is more costly in terms of consumption, as one might expect. As observed in other studies (e.g. Finan et al., 2005; Justino et al., 2008; Ogundari and Aromolaran, 2013; Dzanku et al., 2015), formal education has high welfare value. We observe that household head's education is associated with significant positive welfare benefits. For example, a household whose head completed secondary or higher education has, on average, 18 percent higher consumption than a household whose head has no formal education. Our estimate is comparably larger than the 8 percent reported by Justino et al. (2008) for Vietnamese household panel data but considerably lower than that reported by Ackah et al. (2014) who used nationally representative data from Ghana but employed income as the welfare measure. In addition to the positive welfare effect of head's education, having an educated spouse in the household significantly contributes to consumption welfare gains. No such advantage is observed if welfare was measured on the multidimensional scale. Given that previous research (e.g. Appleton, 1996; Justino et al., 2008; Ackah and Medvedev, 2012) have found remittances to be associated with household welfare gains our finding that average consumption welfare is significantly lower for households with remittance associated urban social networks compared with households with no such social capital appears counterintuitive. However, our finding suggests that households receiving such remittances are so deprived that they still have lower average consumption even after controlling for other covariates of household welfare. Food prices have important implications for household welfare (Minten and Barrett, 2008). Based on information from focus group discussions (FGD) in the study villages we test the hypothesis that household specific consumereproducer staple crop price differentials impose rural household welfare cost. The reason is that being cash constrained, asset poor rural households are often

108

F.M. Dzanku / Journal of Rural Studies 40 (2015) 102e110

Table 5 Household welfare regression estimates with transient livelihood diversity status effect. Variables

Log (Consumption expenditure equation)

Log (Composite welfare index)

Model 1a

Model 2a

Coef. Transient livelihood diversity status 0.077** Female headed household 0.059 Age of household head 0.028*** Age of household head squared 2.5e-04*** Household dependent proportion 0.408*** Head's education (reference is not completed primary or no education) Head completed basic education 0.038 Head has higher than basic education 0.164*** Spouse has some formal education 0.113** Livestock holding (tropical livestock units) 0.002 Urban social network capital (reference is no capital) Urban social network capital without remittance 0.050 Urban social network capital with remittance 0.109*** Consumer/producer staple price differential 0.534*** District dummies (reference district is Fanteakwa in the Eastern Region) Manya Krobo, Eastern Region 0.091** KassenaeNankana, Upper East Region 0.514*** TalensieNabdam, Upper East Region 0.546*** Time dummies (reference is 2002) 2006 0.029 2010 0.144*** Intercept 6.820*** Observations 1392 Number of households 464 sc 0.206 su 0.543 r 0.125 Wald test that r ¼ 0 p-value of Wald test

Model 1b

Model 2b

Std. Err.

Coef.

Std. Err.

Coef.

Std. Err.

Coef.

Std. Err.

0.035 0.051 0.006 6.3e-05 0.082

0.095* 0.055 0.030*** 2.7e-04*** 0.407***

0.048 0.050 0.006 6.3e-04 0.081

0.036** 0.007 2.6e-04

0.014 0.012 2.3e-04

0.062** 0.007 4.4e-05

0.028 0.012 3.4e-05

0.031**

0.016

0.031*

0.016

0.047 0.047 0.047 0.002

0.035 0.174*** 0.101** 0.001

0.048 0.047 0.047 0.002

0.037*** 0.090*** 0.021 0.001

0.013 0.015 0.015 0.000

0.035*** 0.088*** 0.020 0.001

0.013 0.015 0.015 0.001

0.055 0.040 0.169

0.049 0.110*** 0.544***

0.054 0.040 0.170

0.008 0.004 0.117**

0.011 0.007 0.055

0.007 0.005 0.119**

0.017 0.011 0.054

0.045 0.056 0.056

0.095** 0.519*** 0.550***

0.045 0.056 0.057

0.001 0.091*** 0.097***

0.026 0.028 0.028

0.003 0.102*** 0.108***

0.026 0.029 0.029

0.045 0.041 0.174

0.030 0.144*** 6.938*** 1392

0.044 0.041 0.182

0.010 0.068*** 2.053*** 1392 464 0.141 0.107 0.635

0.010 0.007 0.042

0.008 0.068*** 2.128*** 1392

0.010 0.007 0.061

0.137 1.415 0.234

0.337 2.627 0.105

Notes: Results are based on coefficients whose standard errors are robust to data clustering and general forms of heteroscedasticity and serial correlation; Models 1a and 1b are estimates from the correlated random effects model; Models 2a and 2b are estimates from the treatment-effects model; *, ** and *** denotes significance at 10%, 5% and 1%, respectively; sc is the estimate of the between-household standard deviation; su is the estimate of the within-household standard deviation; in models 1 r is the estimate of the intraclass correlation; in models 2 r is the estimate of the correlation between the error terms in the welfare equation and the endogenous transient diversity status equation.

compelled to sell their crop output soon after harvest when prices are low. Under rainfed conditions, consumer prices rise substantially a few months after harvest, with many households having to purchase food. The majority of households being net buyers, the household specific price differential should have household welfare implications.7 We find evidence from our data in support of the above hypothesis with nontrivial magnitudes of effect in both the consumption expenditure and composite welfare equations. Increasing the average consumereproducer price differential by one local currency unit/kg of staple crops is estimated to reduce average household consumption expenditure by 53 percent, all other factors held constant. The same local currency unit increase reduces the average value of the composite welfare index by about 12 percent. Note, however, that increasing the price gap by one local currency unit is not realistic given that even at the 99th percentile the average price gap is 0.38 of a local currency unit. The historical North e South divide with respect to standards of living in Ghana is well known and documented (e.g. VanderpuyeOrgle, 2008; World Bank, 2011). As one will expect, we find strong evidence of substantial household welfare cost associated with locating in the Upper East Region districts compared with the reference district in the Eastern Region. Average consumption expenditures were approximately 40 and 42 percent lower for

7 The most widely produced and consumed cereals in the villages are used: maize, rice, sorghum and millet. Cassava is an important staple in some of the study areas but is not included due to measurement issues.

households residing in the two Upper East districts compared with the reference district. The estimated average composite welfare difference between the Upper East districts and the reference district was approximately 9 and 10 percent. No statistically significant differences in average welfare between districts in the same region were observed as Wald tests showed. We find differing evidence of time effects in the estimated household welfare equations. In the consumption expenditure model we find that average welfare was higher in 2010 than in 2002 by approximately 15 percent (the difference is statistically significant at the 1% level). That poverty has reduced over time in our study sample is generally consistent with the national picture which is also based on the consumption expenditure measure of welfare. Per contra, if the household welfare measure accounted for other dimensions of welfare (including non-money metric dimensions), then we observe that average household welfare in 2010 was about 7 percent below the average value in 2002, all other factors held constant. Clearly, choice of appropriate welfare indicator matters for assessing poverty differences across regions. Finally, we sort to assess the relative importance of the rural household welfare covariates by computing the standardised coefficients.8 The unreported results suggest that, irrespective of household welfare measure, spatial location of household and time effects are the most important determinants of household welfare in the sample. Aside from these, household demographic factors and

8 The standardised estimates are obtained by first transforming all variables before running the regressions.

F.M. Dzanku / Journal of Rural Studies 40 (2015) 102e110

109

Table 6 Estimates from the endogenous transient diversity status equation. Variables

Female headed household Age of household head Household dependent proportion Head's education (reference is not completed primary or no education) Head completed basic education Head has higher than basic education Spouse has some formal education Livestock holding (tropical livestock units) Informal vocational training Number of household members previously out-migrated Self-reported dominant economic livelihood (reference is farming) Nonfarm dominant No dominant economic livelihood (i.e. diversified) District dummies (reference district is Fanteakwa in the Eastern Region) Manya Krobo, Eastern Region KassenaeNankana, Upper East Region TalensieNabdam, Upper East Region Intercept

r l Standard error of l Log-likelihood value

Endogenous treatment from consumption equation

Endogenous treatment from composite welfare equation

Coef.

Std. Err.

Coef.

Std. Err.

0.118 4.8e-03 0.383

0.284 7.4e-03 0.443

0.032 3.4e-03 0.399

0.291 7.4e-03 0.436

0.451 0.115 0.444* 0.001 1.508*** 0.929***

0.335 0.360 0.239 0.007 0.252 0.205

0.388 0.045 0.411* 0.003 1.365*** 0.874***

0.337 0.364 0.238 0.006 0.327 0.224

1.343*** 1.539***

0.429 0.319

1.665*** 1.628***

0.495 0.303

0.081 0.606*** 0.592** 0.979** 0.137 0.079 0.064 2051

0.185 0.232 0.240 0.472

0.083 0.579** 0.548** 0.938** 0.337 0.061 0.037 390

0.183 0.234 0.247 0.468

Note: r is the estimate of the correlation coefficient between the error terms from the welfare and transient diversity status equations. The asterisks *, **, *** represent statistical significance at 10%, 5% and 1%, respectively.

human capital quality as measured by education of household head and spouse exerted the most influential effect on household welfare. For example, a one standard deviation increase in the proportion of dependants in a household decreases consumption expenditure by 0.14 standard deviation; a one standard deviation increase in the consumereproducer price differential decreases consumption expenditure by 0.08 standard deviation. This means that the same relative movement of the proportion of household dependants in our sample population has a larger effect on consumption poverty than the staple price gap does. In the composite welfare equation, transient livelihood diversity status had the second largest standardised effect aside location and time effects; the largest standardised effect aside from location and time effects is household head's completion of higher education. 7. Summary and conclusions This paper contributes to the literature on quantitative rural livelihoods and poverty analysis by testing two important hypotheses which have not been previously addressed systematically: that a rural household exhibits transitory economic livelihood diversity behaviour, and that being livelihood diversity transient imposes significant welfare costs on the rural household. Thus, this paper challenges the widely held view of increasing and pervasive livelihood diversity among rural households. It argues that aside from farming, many other livelihood options in rural areas are transient. In terms of policy relevance one could argue that the welfare implication of such diversification behaviour is what matters. Thus, we examine the household welfare effects of transient versus non-transient livelihood diversity behaviour using two welfare indicators. The panel data evidence provided by this paper suggests that there is no sustained increase in livelihood diversification behaviour over time. Although a substantial proportion of households maintain a diverse economic livelihood, a considerable proportion also exhibit instability overtime. Over the entire period of the panel, 55 percent of households were livelihood diversity transient. Most

of such households maintain a base in agriculture and only take advantage of transient opportunities outside the farm. On the household welfare implications of transient versus nontransient rural livelihoods, the evidence we found suggests that there are modest welfare costs associated with having a transient livelihood diversity status rather than being stable, regardless of welfare measure. Bryceson (2002b, pp. 726) attributes a transient rural livelihood diversity status to economic livelihood trial-anderror which results from the lack of policy focus on long-term occupational development in rural Africa. We find evidence in support of her thesis that such livelihood behaviour may indeed be costly. The implication is that if diversifying out of agriculture is viewed as a rural welfare improvement strategy then there is the need for policy to address the gap in apprenticeship and entrepreneurial skills of rural households so that they can develop gainful employment opportunities outside the farm as has been pointed out by Bryceson (2002a). The study found other important determinants of household welfare in the study sample. While some of the determinants depend on welfare measure, suggesting the importance of measuring welfare in multiple dimensions whenever possible, some are robust to the welfare measures. Among the later are the proportion of dependants in a household, and consumereproducer staple crop price differential (both associated with significant welfare cost to the household); formal education of household head is associated with welfare benefits. Having a spouse with some formal education in the household is associated with higher average consumption but not composite welfare. Similarly, on average, consumption expenditure is influenced by age of household head but composite welfare does not discriminate on the basis of household head's age. By far the most important determinant of poverty in our sample is spatial location of household. Poverty highly discriminates against living in villages in the Upper East Region of Ghana. This highlights the importance of taking location specific poverty reduction policy implementation efforts by both government and development agencies more seriously.

110

F.M. Dzanku / Journal of Rural Studies 40 (2015) 102e110

Acknowledgements Funding for the third round of data collection was supported by the Institute of Statistical, Social and Economic Research and the University of Reading through Prof. Chris Garforth. References Abdulai, A., CroleRees, A., 2001. Determinants of income diversification amongst rural households in Southern Mali. Food Policy 26 (4), 437e452. Ackah, C., Adjasi, C., Turkson, F., Acquah, A., 2014. Education, Skill, and Earnings: Further Evidence from Ghana. WIDER Working Paper 2014/073. WIDER. United Nations University. Available: http://www.econstor.eu/handle/10419/97123. Last accessed: 15 April 2015. Ackah, C., Medvedev, D., 2012. Internal migration in Ghana: determinants and welfare impacts. Int. J. Soc. Econ. 39 (10), 764e784. Anriquez, G., Daidone, S., 2010. Linkages between the farm and nonfarm sectors at the household level in rural Ghana: a consistent stochastic distance function approach. Agric. Econ. 41 (1), 51e66. Appleton, S., 1996. Women-headed households and household welfare: an empirical deconstruction for Uganda. World Dev. 24 (12), 1811e1827. Baird, T.D., Gray, C.L., 2014. Livelihood diversification and shifting social networks of exchange: a social network transition? World Dev. 60 (0), 14e30. Barrett, C.B., Bezuneh, M., Aboud, A., 2001a. Income diversification, poverty traps ^ te d'Ivoire and Kenya. Food Policy 26 (4), 367e384. and policy shocks in Co 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 (4), 315e331. Block, S., Webb, P., 2001. The dynamics of livelihood diversification in post-famine Ethiopia. Food Policy 26 (4), 333e350. Bryceson, D.F., 1996. Deagrarianization and rural employment in sub-Saharan Africa: a sectoral perspective. World Dev. 24 (1), 97e111. Bryceson, D.F., 2002a. Multiplex livelihoods in rural Africa: recasting the terms and conditions of gainful employment. J. Mod. Afr. Stud. 40 (1), 1e28. Bryceson, D.F., 2002b. The scramble in Africa: reorienting rural livelihoods. World Dev. 30 (5), 725e739. Canagarajah, S., Newman, C., Bhattamishra, R., 2001. Non-farm income, gender, and inequality: evidence from rural Ghana and Uganda. Food Policy 26 (4), 405e420. Davis, B., Winters, P., Carletto, G., Covarrubias, K., Quinones, E.J., Zezza, A., Stamoulis, K., Azzarri, C., Digiuseppe, S., 2010. A cross-country comparison of rural income generating activities. World Dev. 38 (1), 48e63. Dimova, R., Sen, K., 2010. Is Household Income Diversification a Means of Survival or a Means of Accumulation? Panel Data Evidence from Tanzania. BWPI Working Paper 122. Brooks World Poverty Institute, Manchester, UK. Available: http://www.bwpi.manchester.ac.uk/. Last accessed: 17 May 2012. € m, M., Marstorp, H., 2015. Yield gap-based poverty gaps in rural Dzanku, F.M., Jirstro Sub-Saharan Africa. World Dev. 67 (0), 336e362. Ellis, F., 2000a. The determinants of rural livelihood diversification in developing countries. J. Agric. Econ. 51 (2), 289e302. Ellis, F., 2000b. Rural Livelihoods and Diversity in Developing Countries. Oxford University Press, Oxford. Ellis, F., 2010. Strategic dimensions of rural poverty reduction in sub-Saharan Africa. In: Harriss-White, B., Heyer, J. (Eds.), The Comparative Political Economy of Development: Africa and South Asia. Routledge, New York, pp. 47e63. Ersado, L., 2006. Income Diversification in Zimbabwe: Welfare Implications from Urban and Rural Areas. World Bank Policy Research Working Paper 3964. World Bank, Washington, DC. Available: http://ssrn.com/abstract¼923271. Last accessed: 25 July 2011. Finan, F., Sadoulet, E., de Janvry, A., 2005. Measuring the poverty reduction potential of land in rural Mexico. J. Dev. Econ. 77 (1), 27e51. Foster, A.D., Rosenzweig, M.R., 2004. Agricultural productivity growth, rural economic diversity, and economic reforms: India, 1970-2000. Econ. Dev. Cult. Change 52 (3), 509e542. Ghana Statistical Service, 2012. Population and Housing Census: Summary Report of Final Results. Ghana Statistical Service, Accra, Ghana, p. 2010. Grawert, E., 1998. Making a Living in Rural Sudan: Production of Women, Labour Migration of Men, and Policies for Peasants' Needs. Macmillan, London. Grosse, M., Harttgen, K., Klasen, S., 2008. Measuring pro-poor growth in nonincome dimensions. World Dev. 36 (6), 1021e1047. Haggblade, S., 2007. Alternative perspectives of the rural nonfarm economy. In: Haggblade, S., Hazell, P.B.R., Reardon, T.A. (Eds.), Transforming the Rural Nonfarm Economy: Opportunities and Threats in the Developing World. Johns Hopkins University Press, Washington D.C, pp. 25e54. Haggblade, S., Hazell, P., Reardon, T., 2002. Strategies for Stimulating Povertyalleviating Growth in the Rural Non-farm Economy in Developing Countries. EPTD Discussion paper 92. International Food Policy Research Institute, Washington DC. Available: http://www.ifpri.org/. Last accessed: 20 February 2010.

Haggblade, S., Hazell, P.B.R., Reardon, T.A., 2007. Research perspectives and prospectives on the rural nonfarm economy. In: Haggblade, S., Hazell, P.B.R., Reardon, T.A. (Eds.), Transforming the Rural Nonfarm Economy: Opportunities and Threats in the Developing World. Johns Hopkins University Press, Washington D.C, pp. 381e395. Havnevik, K., Bryceson, D., Birgegård, Lars-Erik, Matondi, P., Beyene, A. (Eds.), 2007. African Agriculture and the World Bank: Development or Impoverishment? Policy Dialogue No.1. The Nordic Africa Institute, Uppsala, Sweden. Jan van Garderen, K., Shah, C., 2002. Exact interpretation of dummy variables in semilogarithmic equations. Econ. J. 5 (1), 149e159. Justino, P., Litchfield, J., Pham, H.T., 2008. Poverty dynamics during trade reform: evidence from rural Vietnam. Rev. Income Wealth 54 (2), 166e192. Karbasioun, M., Mirzaei, S., Mulder, M., 2005. Informal technical and vocational training programs and farming in the Province of Isfahan, Iran. J. Int. Agric. Ext. Educ. 12 (2), 43e53. Kennedy, P.E., 1981. Estimation with correctly interpreted dummy variables in semilogarithmic equations. Am. Econ. Rev. 71 (4), 801e801. Kijima, Y., Gonzalez, H., 2013. Does observance of religious holidays affect agricultural productivity and household welfare? Evidence from rural Ethiopia. J. Dev. Stud. 49 (9), 1188e1201. Kohler, U., Kreuter, F., 2005. Data Analysis Using Stata. Stata Press, Texas, USA. Lanjouw, P., Murgai, R., 2009. Poverty decline, agricultural wages, and nonfarm employment in rural India: 1983e2004. Agric. Econ. 40 (2), 243e263. Lay, J., Narloch, U., Mahmoud, T.O., 2009. Shocks, structural change, and the patterns of income diversification in Burkina Faso. Afr. Dev. Rev.eRevue Africaine De Developpement 21 (1), 36e58. Little, P.D., Smith, K., Cellarius, B.A., Coppock, D.L., Barrett, C.B., 2001. Avoiding disaster: diversification and risk management among east African herders. Dev. Change 32 (3), 401e433. Maddala, G.S., 1983. Limited-dependent and Qualitative Variables in Econometrics. Cambridge University Press. Mensah, E.J., Huchet-Bourdon, M., Latruffe, L., 2014. Infrastructure access and household welfare in rural Ghana. Afr. Dev. Rev. 26 (3), 508e519. Minot, N., Baulch, B., 2005. Spatial patterns of poverty in Vietnam and their implications for policy. Food Policy 30 (5e6), 461e475. Minot, N., Epprecht, M., Anh, T.T.T., Trung, L.Q., 2006. Income Diversification and Poverty in the Northern Uplands of Vietnam. Research Report 145. International Food Policy Institute, Washington, D.C. Available: www.ifpri.org/. Last accessed: 12 February 2009. Minten, B., Barrett, C.B., 2008. Agricultural technology, productivity, and poverty in Madagascar. World Dev. 36 (5), 797e822. Mundlak, Y., 1978. On the pooling of time series and cross section data. Econometrica 46 (1), 69e85. Ogundari, K., Aromolaran, A.B., 2013. Impact of education on household welfare in Nigeria. Int. Econ. J. 28 (2), 345e364. Reardon, T., Stamoulis, K., Cruz, M.-E., Balisacan, A., Berdegue, J., Banks, B., 1998. Rural non-farm income in developing countries. In: The State of Food and Agriculture 1998: Part III. Food and Agricultural Organisation, Rome. Saith, A., 1992. The Rural Non-farm Economy: Processes and Policies. International Labour Organization, Geneva. Scoones, I., 1998. Sustainable Rural Livelihoods: a Framework for Analysis. Institute of Development Studies, University of Sussex, Sussex, UK. 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 (4), 421e435. Stifel, D., 2010. The rural non-farm economy, livelihood strategies and household welfare. Afr. J. Agric. Resour. Econ. 4 (1), 82e109. Tellegen, N., 1997. Rural Enterprises in Malawi: Necessity or Opportunity? Ashgate, Aldershot. Timmer, P., 1988. The agricultural transformation. In: Chenery, H., Srinivasan, T.N. (Eds.), The Handbook of Development Economics, vol. 1. North Holland Press, Amsterdam, pp. 275e332. Vanderpuye-Orgle, J., 2008. The north-south devide and the disappearing of the middle class: an analysis of spatial inequality and polarization in Ghana. In: Amoako-Tuffour, J., Armah, B. (Eds.), Poverty Reduction Strategies in Action: Perspectives and Lessons from Ghana. Lexington Books, Plymouth, UK, pp. 49e71. Vella, F., Verbeek, M., 1998. Whose wages do unions raise? a dynamic model of unionism and wage rate determination for young men. J. Appl. Econ. 13 (2), 163e183. Wooldridge, J.M., 2010. Correlated Random Effects Models with Unbalanced Panels. Manuscript. Available: http://econ.msu.edu/faculty/wooldridge/docs/. Last accessed: 5 March 2012. World Bank, 2011. Tackling Poverty in Northern Ghana. World Bank, Washington DC. Report No. 53991-GH. Wouterse, F., Taylor, J.E., 2008. Migration and income diversification: evidence from Burkina Faso. World Dev. 36 (4), 625e640. Zezza, A., Carletto, G., Davis, B., Stamoulis, K., Winters, P., 2009. Rural income generating activities: whatever happened to the institutional vacuum? Evidence from Ghana, Guatemala, Nicaragua and Vietnam. World Dev. 37 (7), 1297e1306.