Hope for Change: Is Diversifying Production Portfolios an Ideal Strategy to Boost Farming Efficiency in Afghanistan?

Abstract

In this article, I present empirical evidence on the extent of crop diversification and assess its merits as a strategy for improving production efficiency in Afghanistan. The transformed Herfindahl–Harshman index is used to measure the scale and magnitude of crop diversification. I find a compelling evidence that diversifying production portfolios significantly improves production efficiencies. This finding is critical, given that the data show that nearly a third of the farm households do not diversify, achieving, on average, about 52% of potential revenues. The estimated efficiency scores reveal that, on average, the farm households in our analytical sample of over 7,000 households achieve 74% of potential revenue, with nearly 15% of households realizing less than 50% and about 23% between 50% and 70% of potential revenue. These results infer that there exist substantial inefficacies in agricultural production that can be eliminated by employing improved management practices without having to use additional inputs and production resources and rising cost of production. Our results are robust to potential endogeneity bias in crop diversification; I account for the endogeneity problem in the stochastic frontier analysis, by employing a recent estimation approach, using instrumental variable techniques. Mapping the spatial distribution of crop diversification index and estimated efficiency scores across the country revealed that districts with higher diversification levels correspond to higher efficiency indices. Aside from crop diversification, other socio-economic factors also have critical implications for efficiency; households with access to farm assets (such as land, cattle, oxen and tractor) and extension services appear to realize substantially higher production efficiencies. A direct policy recommendation that can be generated from the findings of this study is that crop diversification should be given more recognition by policymakers to enhance productivity and resilience in agriculture.

Keywords

Crop diversification production efficiency endogenous stochastic frontier models Afghanistan

I. Introduction

Measuring the economic performance of a farm requires an understanding of the production decisions that influence the levels of production efficiency. Technical efficiency (TE) is a principal element in economic profitability as it measures the ability of a farm to produce maximal output from a given set of inputs. TE as a precondition for economic efficiency safeguards the economic viability and sustainability of a farm. Studies find that technical efficiencies complement household welfare. That is, improved efficiency at the farm level is crucial for farmers’ sustainable growth (Danso-Abbeam and Baiyegunhi, 2020). Farm productivity can be enhanced by adopting production technology. Alternatively, productivity can be improved by changing how production factors are combined to improve the efficiency by which inputs are being transformed into outputs such that higher outputs are produced from the same level of inputs and technology (Coelli, 1995). Hence, the level of efficiency and overall productivity is affected by farmer’s management decisions, including, for instance, decisions by farmers to shift away from mono-cropping towards adopting a diversified production system.

Empirical research suggests that farmers in developing countries are often unlikely to exploit fully the production technology and resources and often make inefficient decisions (Mekonnen et al., 2015; Tenaye, 2020). In a study of 85 low- and middle-income countries, Mekonnen et al. (2015) report an average TE of about 44–62% over the period from 2004 to 2011. To achieve the overarching goals of eliminating rural poverty and hunger, policymakers have paid greater attention to improve agricultural production through the introduction and adoption of modern farming technologies and production inputs. However, the use of modern technologies is costly and, in some instances, may not be affordable, especially by smallholders as they face severe financial and liquidity constraints with no access to low-cost credit. The use of productivity-enhancing technologies can be more efficient in improving food deficiencies if coupled with cost-minimizing management practices; hence, due attention should be given to improving the current level of inefficiencies in the farming sector.

Since emerging out of conflict and establishing a market-led economy in 2001, Afghanistan has undergone drastic economic policy change. The economy was on the verge of collapse due to conflict and political instability, lack of a sound economic policy and inefficiencies of public institutions. However, when international aid agencies began to pledge aid to support the economy, particularly the agricultural economy, the transition towards a fully market-led system began. This transition from subsistence food production to a market-oriented system typically involves diversification in farming systems (Minot et al., 2006; Ibrahim et al., 2009). Many challenges and uncertainties have resulted from policy changes made over the past 15 years, all of which have influenced farming practices and production decision-making in the country’s faltering progress towards a market economy. In line with Afghanistan’s National Comprehensive Agriculture Development Priority Program 2016–2021, achieving self-sufficiency in basic food crops and promoting economic growth and development via improvement in agricultural production and productivity remain on the top of the government’s strategic priorities.

Agriculture is dominated by small-scale farm households with an average farm size of 7 Jeribs (about 1.5 ha). Wheat occupies a significant portion of the agricultural land, and as the main staple food crop and a major source of calories, it plays a critical role in food security. Other important crops include maize, rice, barley, fodder crops, potato and other high-value crops (HVCs) like vegetables and fruits. Although wheat is critical for food security, market-oriented production requires diversifying production into HVCs. Meanwhile, the decline in farm productivity among farmers who grow staple grains, including wheat and rice, over the years, has triggered a change from monocropping towards crop diversification (CD)—a shift in production from grains to higher-value cash crops (Ahmadzai, 2020; Oushy, 2010). This is consistent with broader evidence that suggests that production systems, which are heavily reliant on grain production, may not continue to be as efficient and remunerating, especially in countries with a policy focus on raising incomes, generating employment opportunities and alleviating poverty (Joshi et al., 2007). Farmers in Afghanistan, especially those in remote areas which have limited access to extension services, generally lack information on market-led production, production technologies and inputs, and best management practices. As a result, they fail to exploit fully the production technology and resources and often end up making inefficient production decisions (e.g., adoption of low-input production systems, lack of crop rotation and over-reliance on the production of a single crop instead of diversifying, etc.). Accordingly, this study investigates CD strategies in Afghanistan and their implications for farm-level production efficiency.

II. The Concept of Crop Diversification

Economic theory offers different arguments about why firms diversify. These arguments broadly fall under three conceptual frameworks: the market power approach, the concept of synergy or jointness of production technology and the risk-averse behaviour of farmers’ perspective. The market power approach proposes that possible anti-competitive strategies (i.e., cross-subsidization or reciprocal buying) employed by diversified firms in pursuit of rising profits (Weiss and Briglauer, 2000). While this theory might be more pertinent in many industries, it is of little relevance when applied to the farming sector commonly characterized by a large number of small subsistence farms, especially in the context of developing and emerging economies (Duffy, 2009; Weiss and Briglauer, 2000). The concept of synergy or jointness of production technology is primarily based on the cost complementarities within a multiple output production system, suggesting a lower cost during the concurrent production of multiple outputs by a given farm. That is, jointness brings cost complementarity (often referred to as economies of scope) among outputs; hence, producing them jointly might be more inexpensive than producing them separately (Chaiechi and Stoeckl, 2013). The risk-averse perspective on diversification is grounded in the volatile nature of agricultural production. Smallholders are often risk-averse, favouring diversification as a potential strategy against production and market risks (e.g., weather shocks, pest and diseases outbreak, market failures, etc.). From this viewpoint, diversity in crop production generates more stable streams of income by enabling farm households to spread their risks across a broader portfolio of production (Makate et al., 2016).

From a development perspective, there are different trajectories that theorize multifunctionality and diversification at the farm level as a pathway to rural economic development. The emergence of these pathways is grounded in the direct economic benefits that diversification offers, and other spillover effects of diversification on regional economic development (Huttunen, 2019). From an economic development point of view, maintaining de facto diversification on farm has merits as an entry point for fostering agricultural innovations because it offers multiple direct benefits to smallholders’ well-being and livelihoods. Moreover, as a rural and regional development strategy, CD opens up new market opportunities, while contributing to self-consumption (Bellon et al., 2020). The empirical analysis, based on the data from a research-for-development project by Bellon et al. (2020), showcased that smallholders in Ghana, where production is for home consumption and income generation, are generally better off when they diversify rather than when they specialize.

Huttunen (2019) refers to diversification as a particularly desirable development pathway for resource-scarce smallholders as it offers innovative ways to combine and utilize farm-based resources more efficiently in production, while helping preserve the natural resource base. For instance, McNamara and Weiss (2005) and Weltin et al. (2017) mention the availability and efficient use of labour resources as the main driver in farmers’ decisions related to the type and level of production diversity. Diversification is supported within the European Common Agriculture Policy (CAP) to escape the crisis of a productivist model of agriculture, where the focus was mainly on raising farm output. The recent shift in the European Union (EU’s) agricultural policy model places multifunctionality and diversification at the centre of their policy paradigm to help local farmers effectively respond to diverse demands and to survive tough market conditions with emphasis on green box measures,¹for example, agri-environmental support scheme (Meraner et al., 2015).

The notion of diversification has been interpreted in different ways within the farming sector. Diversification might imply a shift away from monoculture to producing multiple crops on a single farm, or it could be viewed as having many enterprises at the farm (i.e., a larger mix of crops or a combination of livestock and crop units). Throughout this study, diversification is defined as adding multiple crops (especially HVCs such as vegetables, fruits, potatoes, etc.) to the production portfolio at the farm level.

Wheat, as the main staple food crop, occupies a major portion of the agricultural land in Afghanistan. However, the share of wheat in total revenue is proportionally lower than HVCs. This difference might indicate higher profits per unit of area for HVC, suggesting that adding high-value horticulture crops to the production portfolio may be positively associated with the farmer’s income as broadly observed in the literature. A study in Punjab, India, confirms that incorporating horticultural crops in the production mix increases net expected returns, while increasing the labour and working capital requirements. Van den Berg et al. (2007) concluded that diversification into high-value vegetable crops would enable Chinese farms to sustain a reasonable income level. Guvele (2001) argued that CD reduces variability in income.

Nevertheless, there is limited empirical evidence that explicitly studies the impact of CD on TE, with mixed conclusions. Manjunatha et al. (2013), Ogundari (2013), Rahman (2009), and Coelli and Fleming (2004) concluded that diversification significantly improves the efficiency of farms in India, Bangladesh, Nigeria and Papua New Guinea, respectively. On the other hand, Haji (2007) found no significant relationship between diversification and TE but has found that it reduced allocative efficiencies in Ethiopia. Although the literature offers several measures for CD, in this study, the transformed Herfindahl–Hirschman index (THI) is used as a measure for diversification or specialization. The reverse of the THI gives due weightage to the proportional revenue share of individual crops.

\begin{matrix} T H I = 1 - H H I = 1 - \sum_{j = 1}^{J} {(\frac{Y_{j}}{\sum_{j = 1}^{J} Y_{j}})}^{2} \\ 0 \leq T H I_{i} \leq 1 \end{matrix}

(1)

where Y_j represents the revenue share occupied by the jth crop (for j = 1, 2, …, J) in total annual revenue. The computed THI index ranges from (close to) 0, reflecting complete specialization (i.e., just one crop) to one, reflecting complete diversification (i.e., the maximum number of crops).

III. Theoretical Framework

Since Farrell’s pioneering work (1957), a number of approaches to efficiency measurement have emerged. The two main approaches that have been extensively used in the efficiency literature are (a) parametric stochastic frontier analysis (SFA) initially proposed by Aigner et al. (1977) and Meeusen and Van den Broeck (1977) and (b) non-parametric data envelopment analysis (DEA) initially developed by Charnes et al. (1978).

Choosing between the SFA and DEA approaches has been controversial and depends upon the objective of the research, the type of industry and the availably of data (Wadud and White, 2000). The non-parametric DEA does not rely on the definition of a functional form characterizing the underlying technology and therefore avoids misspecification problems. However, a drawback of this technique is that it is deterministic and ignores the stochastic error term, implying that deviations from the frontier are entirely attributed to inefficiency effects (Kumbhakar and Lovell, 2000; Kumbhakar and Wang, 2015; Wadud and White, 2000).

In contrast, the main advantage of the parametric SFA approach is that it incorporates a composed error structure with a two-sided symmetric term and a one-sided component, which permits distinction between inefficiency and exogenous shocks (Aigner et al., 1977; Meeusen and Van den Broeck, 1977). In addition, SFA allows hypothesis testing and the construction of confidence intervals (Wadud and White, 2000). The disadvantages of this approach are the need to assume a functional form for the frontier technology and for the distribution of the technical inefficiency component of the composite error term. This study adopts the SFA since agricultural crop production exhibits random shocks, and there is a need to separate the influence of stochastic factors (random shocks and measurement errors) from the effects of other inefficiencies.

Kumbhakar and Lovell (2000), Battese and Coelli (1995), and, more recently, Kumbhakar and Wang (2015) have advocated a single-stage simultaneous estimation approach in which explanatory variables are incorporated directly into the inefficiency error component. In this approach, either the mean or the variance of the inefficiency error component is hypothesized to be a function of the explanatory variables. Following Aigner et al. (1977) and Meeusen and Van den Broeck (1977), the formulation of the stochastic frontier (SF) model in terms of general production function could be specified as:

Y_{i} = f (X_{i}, β) + ε_{i} = f (X_{i}, β) + v_{i} - u_{i}

(2)

where Y_i is a scalar output of the ith farmer, X_i is the vector that collects direct inputs and β is a vector of parameters to be estimated. ε_i is a composed error term where v_i is a two-sided ‘noise’ component assumed to be independently and identically distributed (iid), symmetric and distributed independently from u_i. It captures the effects of random shocks beyond the control of farmers (i.e., measurement errors as well as other noise). u_i is a non-negative (u_i ≥ 0) technical inefficiency component of εi that captures the factors that are under the control of the producer (i.e., determinants of inefficiency to be defined in the inefficiency model). u_i is assumed to be independently and identically distributed as normal–half-normal distribution (Aigner et al., 1977). There are other possible specifications of the distributional assumptions on u_i (i.e., truncated–normal distribution) suggested by Greene (1980) and Lee (1983), which are widely used in empirical work.

The basic concept of the SF models is illustrated in Figure 1. The SF estimates farm-level TE by measuring the distance between its observed output (represented by the dark circles in Figure 1) and the maximum feasible or the frontier output (i.e., large grey circles in Figure 1) that it could obtain if it were located on the so-called production frontier. This production frontier (represented by the concave curve in Figure 1) defines the highest possible amount of output that can be obtained from any given amount of input. Hence, any points located on the frontier curve represent the most efficient farms, whereas any points below the frontier curve represent farms that lag behind (Kumbhakar and Wang, 2015). Hence, TE is defined as the ratio of the empirically observed output to the corresponding maximum feasible output.

Deviations between individual farms’ observed output and the corresponding frontier output could occur for two reasons. First, technical inefficiency effects could occur due to the presence of systematic shortfalls in managerial capabilities—which mean that some farmers are less successful than others at putting the inputs at their disposal to the best possible use. Second, these deviations could be random or stochastic effects which represent production variations attributable to random shocks beyond the control of the producer like weather shocks and floods. Figure 1 shows this distinction in causes of inefficiency for farms A and B with input levels x_a and x_b, where we see how their observed output deviates from their respective ‘frontier output’ $y_{b}^{*}$ due to inefficiency and stochastic effects. It is important to allow both components in the estimation to avoid confounding the technical inefficiency factors with random errors (Lakner et al., 2012). The SF approach accommodates this by using the composed error structure (as in Equation (2)), where the one-sided error component u_i captures TE, and a second error, symmetric error component v_i, captures random deviations from the frontier. The SF techniques estimate the location of the stochastic production frontier from data on the input use and output production of a set of farms and decompose the deviation of each individual farm’s actual production from the frontier into its inefficiency and stochastic components. For a detailed exposition of the SF techniques, see Kumbhakar and Lovell (2000).

Figure 1.

Source: Adopted from Lakner et al. (2012).

IV. Empirical Specification and Estimation Strategy

In Equation (2), the inefficiency component (u_i) of the error term is the log difference between the maximum (frontier) and the actual (observed) output (i.e., u_i = In $y_{i}^{*}$ – In Y_i); hence, u_i × 100% is the percentage by which actual output can be increased using the same inputs if production is fully efficient (Kumbhakar and Wang, 2015). In other words, it is the percentage of output that is lost due to TE. The estimated value of u_i is referred to as the output-oriented (technical) inefficiency, with a value close to 0 implying fully efficient. Rearranging Equation (2), I derive the following equation for TE:

T E_{i} = \exp (- ​ u_{i}) = \frac{Y_{i}}{Y_{i}^{*}} = \frac{Y_{i}}{f (x_{i}; β ​) \exp {v_{i}}}

(3)

Equation (3) defines farm-specific TE as the ratio of observed output (Y_i) to the frontier output f(x_i; β) exp{v_i}, which is the maximum output feasible (under the current technology used) in an environment characterized by the stochastic elements specified by (v_i). Because u_i ≥ 0, the ratio is bounded between 0 and 1; hence, a farm achieves maximum efficiency if, and only if, TE_i = 1. Otherwise, TE_i ≤ 1 is a shortfall of observed output from the maximum feasible output that is stochastic and varies across farmers (Kumbhakar and Lovell, 2000). Based on Equation (2), the econometric specification of the Translog SF model can be written as:

\begin{matrix} \ln Y_{i} = \sum_{k = 1}^{7} β_{k} \ln X_{i k} \\ + \frac{1}{2} \sum_{k = 1}^{7} \sum_{j = 1}^{7} β_{j k} \ln X_{i k} \ln X_{i j} + v_{i} - u_{i} \end{matrix}

(4)

where Y_i represents aggregate revenue of the ith producer, k represents the number and X_ij represents a set of the production inputs applied by the ith farmer, β is a vector that collects unknown parameters to be estimated and ε_i is the composed error term where ε_i = v_i – u_i with u_i ≥ 0.

Two distributional assumptions on the inefficiency component of the error term are made and tested: (a) the half-normal distribution imposing restrictions on ε_i such that $v_{i} = i i d N (0, σ_{v}^{}_{}^{2}), u_{i} = i i d N^{+} (0, σ_{u}^{}_{}^{2}),$ and v_i and u_i are distributed independently of each other and of the regressors and (b) a truncated–normal distribution for the ε_i assuming $v_{i} = i i d N (0, σ_{v}^{}_{}^{2}), u_{i} = i i d N^{+} (μ, σ_{u}^{}_{}^{2}),$ and v_i and u_i are distributed independently of each other and of the regressors, and μ is the non-zero mean for u_i. Given the distributional assumption on the inefficiency component (u_i) of the composed error term, the Battese and Coelli (1995) inefficiency model could be specified as:

u_{i} = δ_{0} + \sum_{i = 1}^{14} δ_{i} Z_{i} + w_{i}

(5)

where u_i is the inefficiency score, Z_i represents a set of variables that are likely to affect efficiency, δ_s are the parameters to be estimated and w_i is the error term of the (in)efficiency model. As the dependent variable u_i is defined in terms of technical inefficiency, a farm-specific variable associated with the negative (positive) coefficient will have a positive (negative) impact on TE.

The decision to adopt a diversified production system is likely to depend on unobservable or omitted variables, leading to the endogeneity bias, thereby resulting in the estimation of inconsistent effects of CD on efficiency. Due to its voluntary nature, farmers self-select or choose whether to produce a single crop or diversify into multiple crops. That is, farmers who are relatively wealthier and have more technical knowledge of diversification as a viable strategy might be more likely to adopt a diversified production portfolio as compared to their counterparts (i.e., those without access to extension); thus, this unobserved selection bias may overstate the impact of diversification. On the other hand, to the extent that diversification is maybe measured with error, the basic SF models may underestimate the impact of diversification due to attenuation bias. In either case, there are unobserved factors in the error term (u_i) that could be correlated with the endogenous variable (CD). I use the mean value of the CD index for other farm households in the district as an instrumental variable (IV) to correct for the endogeneity bias in crop diversity (CD), constructed as:

I n s t r u m e n t a l var i a b l e (I V) = (\frac{\sum^{​} D_{d} - D_{i}}{N_{d} - 1})

where $\sum^{} D_{d}$ is the sum of the diversification index in the district, D_i is the CD value of the ith farm in the respective district and N_d is the number of observations in the respective district (so n – 1 is the number of observations in the district excluding the ith farm itself). This implies that the value of IV will differ slightly for each farm, depending on the variance relative to that farm. Note that, on average, there are 40 farm households in each district.

The extent or degree of CD may be magnified through social interactions between farmers in the local neighbourhood. Farms that face similar demographic characteristics and preferences are likely to adopt similar production systems. For instance, a farm household located in a district, where farmers have greater access to information and markets and are therefore more likely to diversify, is more likely to adopt a diversified production system than a farm in a less diversified district. Observing that neighbours’ diversification would encourage a farmer to follow their example, so even relatively ‘low-ability’ farmers are more likely to diversify. However, the fact that neighbours’ diversification should not in itself affect the efficiency of the farmer as factors associated with efficiency, since factors like farm size or ownership of farm assets are not affected by neighbours’ diversification decisions. Such unobserved heterogeneity is likely to be correlated with the random error, leading to an estimation bias. The instrumental variable is used to break that correlation between the error term and independent variables to uncover the causal effect of the endogenous explanatory variable on the dependent variable. This is because the instrumental variable induces changes in the endogenous variable but has no independent effect on the dependent variable (Wooldridge, 2019).

Addressing the endogeneity issue is relatively more complicated in the SF models due to the unique nature of the error term. Guan et al. (2009) employed a two-step estimation methodology to handle endogenous regressors in the frontier framework. However, the efficiency estimates are inconsistent if the two-sided and one-sided error terms are correlated. Gronberg et al. (2015) attempt to solve the endogeneity problem in frontier models through pseudo-IV methodologies. Amsler et al. (2016) present a copula method in which the more general correlation structures are allowable when modelling endogeneity. However, the copula approach is computationally intensive and complex, and it requires choosing a copula correctly. Besides, this approach does not allow variables that affect inefficiency, which makes it less applicable. The IV estimator used in this study follows the recent work of Karakaplan and Kutlu (2017), who developed a general maximum likelihood (ML)–based framework to handle the endogeneity problem in the SF models. For further discussion and mathematical derivation, see Karakaplan and Kutlu (2017) and Karakaplan (2017).

V. Data and Description of Variables

This study uses data from the Afghanistan Living Condition Survey (ALCS) conducted by the National Statistics and Information Authority (NISA) in 2013–2014. The data are nationally representative as it covers all 34 provinces across the country. The survey work follows a two-stage sampling strategy in which a total of 35 strata were identified, 34 for the provinces of Afghanistan and 1 for the nomadic or Kuchi inhabitants. Subsequently, farm households, as the ultimate sampling units, are selected on the basis of the two-stage cluster design within each stratum. For details on the sampling strategy and geographical coverage, see the ALCS final report (CSO, 2014).

The data are representative at the national and provincial levels and covered 157,262 individuals within 20,786 households across the country. The data are unique in the sense that they also include the nomadic (Kuchi) population and because they comprise continuous data collection over a cycle of 12 months, so capture important seasonal variations. Using structured questionnaires, data were collected on a number of indicators, including agricultural production, labour market, household assets, education and other household characteristics. Descriptive analysis of the data showed that about 50% (as many as 9,642 households) reported some involvement in agriculture. However, after accounting for missing values on key variables, the total number of usable observations reduced to 7,052 households.

Referring to Equation (4), the dependent variable is the household’s annual aggregate revenue (measured in Afghani symbolized as AFN throughout this study) from farming. The land measured in Jeribs (1 Jerib is equivalent to 0.2 ha) is the total land cultivated throughout the year, including both irrigated and rain-fed land owned or leased by the household. Total labour applied in farming include both own family and hired labour. Persons aged 14 and over are adult labourers and those below this threshold are considered child labourers and are excluded from the regression to avoid mixing child and adult labourers to reduce concerns due to productivity differences. Hired labour includes only those who were hired in by the farm. Other variables included in the production function are expenditures on seed, chemical fertilizers, chemicals (i.e., pesticides and herbicides), tractor rental and other expenditures (i.e., irrigation water and additional miscellaneous costs) measured in Afghani (Table 1).

Table 1.
Summary Statistics for Variables Used in the Analysis

Variable Description Mean SD Min Max

Dependent variable (Y)

Y Aggregate Annual Revenue (10K AFN) 5.83 9.04 0.011 12.80

Production inputs

X1 Land (Jeribs) 7.038 9.14 0.10 90.00

X2 Labour (hours) 63.85 62.09 1.00 413.85

X3 Seed expenditure (AFN) 2,343 3,602 0.00 45,000

X4 Fertilizer expenditure (AFN) 4,743 8,307 0.00 99,000

X5 Chemicals expenditure (AFN) 363.0 1,194 0.00 10,000

X6 Tractor rental (AFN) 2,494 4,285 0.00 60,000

X7 Other expenditure (AFN) 2,163 5,751 0.00 90,000

Factors or sources of efficiency/inefficiency

Z1 Diversification index (1-hhindex) 0.296 0.23 0.00 0.82

Z2 Opium share by province (%) 0.033 0.10 0.00 0.48

Z3 Extension services (1 = access, 0 = otherwise) 0.209 0.41 0.00 1.00

Z4 Land type (0 = irrigated, 1 = rainfed) 0.231 0.42 0.00 1.00

Z5 Household size (Persons) 8.335 3.46 1.00 36.0

Z6 Household head Age (Years) 44.38 13.8 14.00 95.0

Z7 HH head sex (0 = female, 1 = male) 0.996 0.06 0.00 1.00

Z8 HH head literacy (0 = no, 1 = literacy) 0.321 0.47 0.00 1.00

Z9 HH education (years) 1.829 4.24 0.00 19.0

Z10 Off-farm employment (1 = yes, 0 = no) 0.125 0.33 0.00 1.00

Z11 Own cattle (herds) 1.601 2.04 0.00 31.0

Z12 Own tractor (number) 0.053 0.23 0.00 3.00

Z13 Own oxen (number) 0.232 0.62 0.00 9.00

Z14 Farm size 1 (0.1–2 Jeribs) 0.318 0.47 0.00 1.00

Farm size 2 (>2 to 5 Jeribs) 0.293 0.46 0.00 1.00

Farm size 3 (>5 to 10 Jeribs) 0.226 0.42 0.00 1.00

Farm size 4 (>10 to 20 Jeribs) 0.103 0.30 0.00 1.00

Farm size 5 (>20 Jeribs and above) 0.060 0.24 0.00 1.00

Z15 AEZ 1: North eastern mountains (NEM) 0.023 0.15 0.00 1.00

AEZ 2: Central mountains (CM) 0.137 0.34 0.00 1.00

AEZ 3: Heart–Farah lowlands (HFL) 0.043 0.20 0.00 1.00

AEZ 4: Southern mountains and foothills (SMF) 0.202 0.40 0.00 1.00

AEZ 5: Helmand Valley and Sistan Basins (HVSB) 0.121 0.33 0.00 1.00

AEZ 6: Turkistan plains (TP) 0.064 0.24 0.00 1.00

AEZ 7: Northern mountains and foothills (NMF) 0.169 0.37 0.00 1.00

AEZ 8: Eastern mountains and foothills (EMF) 0.241 0.43 0.00 1.00

N 7,066

Variable	Description	Mean	SD	Min	Max
Dependent variable (Y)
Y	Aggregate Annual Revenue (10K AFN)	5.83	9.04	0.011	12.80
Production inputs
X1	Land (Jeribs)	7.038	9.14	0.10	90.00
X2	Labour (hours)	63.85	62.09	1.00	413.85
X3	Seed expenditure (AFN)	2,343	3,602	0.00	45,000
X4	Fertilizer expenditure (AFN)	4,743	8,307	0.00	99,000
X5	Chemicals expenditure (AFN)	363.0	1,194	0.00	10,000
X6	Tractor rental (AFN)	2,494	4,285	0.00	60,000
X7	Other expenditure (AFN)	2,163	5,751	0.00	90,000
Factors or sources of efficiency/inefficiency
Z1	Diversification index (1-hhindex)	0.296	0.23	0.00	0.82
Z2	Opium share by province (%)	0.033	0.10	0.00	0.48
Z3	Extension services (1 = access, 0 = otherwise)	0.209	0.41	0.00	1.00
Z4	Land type (0 = irrigated, 1 = rainfed)	0.231	0.42	0.00	1.00
Z5	Household size (Persons)	8.335	3.46	1.00	36.0
Z6	Household head Age (Years)	44.38	13.8	14.00	95.0
Z7	HH head sex (0 = female, 1 = male)	0.996	0.06	0.00	1.00
Z8	HH head literacy (0 = no, 1 = literacy)	0.321	0.47	0.00	1.00
Z9	HH education (years)	1.829	4.24	0.00	19.0
Z10	Off-farm employment (1 = yes, 0 = no)	0.125	0.33	0.00	1.00
Z11	Own cattle (herds)	1.601	2.04	0.00	31.0
Z12	Own tractor (number)	0.053	0.23	0.00	3.00
Z13	Own oxen (number)	0.232	0.62	0.00	9.00
Z14	Farm size 1 (0.1–2 Jeribs)	0.318	0.47	0.00	1.00
	Farm size 2 (>2 to 5 Jeribs)	0.293	0.46	0.00	1.00
	Farm size 3 (>5 to 10 Jeribs)	0.226	0.42	0.00	1.00
	Farm size 4 (>10 to 20 Jeribs)	0.103	0.30	0.00	1.00
	Farm size 5 (>20 Jeribs and above)	0.060	0.24	0.00	1.00
Z15	AEZ 1: North eastern mountains (NEM)	0.023	0.15	0.00	1.00
	AEZ 2: Central mountains (CM)	0.137	0.34	0.00	1.00
	AEZ 3: Heart–Farah lowlands (HFL)	0.043	0.20	0.00	1.00
	AEZ 4: Southern mountains and foothills (SMF)	0.202	0.40	0.00	1.00
	AEZ 5: Helmand Valley and Sistan Basins (HVSB)	0.121	0.33	0.00	1.00
	AEZ 6: Turkistan plains (TP)	0.064	0.24	0.00	1.00
	AEZ 7: Northern mountains and foothills (NMF)	0.169	0.37	0.00	1.00
	AEZ 8: Eastern mountains and foothills (EMF)	0.241	0.43	0.00	1.00
	N	7,066

Source: The authors (calculations of the ALCS 2013–2014 data).

The computed average value of the THI) for the sample is 0.30 with a standard deviation of 0.23, implying the presence of a relatively low level of CD in Afghanistan. Rahman (2009) reported the average Herfindahl index of 0.60 for Bangladesh, Ogundari (2013) reported a Herfindahl index of 0.46 in Nigeria and Manjunatha et al. (2013) reported 0.55 in India.

An essential aspect of farming in the context of Afghanistan is opium cultivation that should be considered in the farm efficiency analysis as a potential source of income that could affect household behaviour and production portfolio. The ALCS surveys attempt to provide information on opium production from the households; however, only a small number reported growing opium (98% did not), and the reliability of the data might be a concern as production and trade of narcotics are illegal according to the constitution; therefore, households might have refrained from providing data, or those that did provide data may have misled interviewers over the extent of opium production. I use Afghanistan’s Ministry of Counter Narcotics (MCN) annual data to construct an intensity variable capturing opium cultivation by province. Mostly, opium production may have a connection with the security situation in the country; since the areas where opium is most commonly grown are likely to be more insecure, opium production may proxy for provincial-level insecurity. It may also capture unreported revenue from opium as a cash crop.

Geographical or spatial variation is another crucial aspect in our analysis. Based on Humlum’s (1959) early work, which was, later on, revived by Dupree (1973), Afghanistan was divided into 11 geographical zones. However, recently, a study by Maletta and Favre (2003) concluded that not all 11 have agricultural significance since some zones were classified as deserts. Based on the ecological properties of land and climate and some supplementary criteria about accessibility and prevailing agricultural activities, Maletta and Favre (2003) adopted the eight agro-ecological zone scheme, which were constructed by aggregating districts.

Cattle ownership is used as a proxy for the availability of animal manure for the farm. Animal waste is an essential source of organic fertilizer, especially in the context of Afghanistan, and is generally believed to improve soil fertility. Similarly, oxen and tractors are the two primary sources of traction power used on the farm for ploughing and other farming activities. It is generally believed that a household’s ownership of a tractor or oxen or both might be a cost-effective investment in production and therefore might have an influence on the TE. On the other hand, oxen or tractor ownership may substitute for farm labour, especially since some of the activities that are traditionally carried out by hand may be completed by oxen or tractor, implying that oxen or tractor ownership may have a positive impact on TE.

Most empirical evidence suggests an inverse relationship between the farm size and TE (Oladeebo and Oyetunde, 2013). In Afghanistan, where agriculture holding is relatively small, farm size is included in the inefficiency model to account for potential variability due to the farm size. Besides farm size, land quality is also expected to affect efficiency. Farm households cultivate either irrigated, rain-fed or a combination of both irrigated and rain-fed land to produce crops. Initial descriptive statistics showed higher annual aggregate revenue for households that cultivate irrigated land alone. To capture this variation attributed to the quality of land, a binary variable was included in the analysis.

Access to extension services is vital in assisting farmers in the production decision-making process since it can be a reliable source of information, technical advice, training and improved farm management practices. Access to extension services is broadly believed in the literature to have a positive impact on the farm output and on the level of CD (Ogundari, 2013). However, only 21% of the sample farmers reported being able to access extension services. Although relatively few farmers can avail of them, extension visits and training are essential sources of information, farm management techniques, use and dissipation of innovation and technology. A descriptive summary of the farm characteristics by access to extension services revealed that households with access earn slightly higher revenues and are more diversified than those without access (Table A1).

Several recent studies have identified off-farm employment in the inefficiency effect model (Yang et al., 2016; Zhang et al., 2016), showing mixed evidence of the impact of off-farm employment on TE. On the one hand, off-farm employment shrinks the availability of labour for on-farm activities, especially if hiring agricultural labour incurs transaction costs and therefore may negatively affect TE. On the other hand, off-farm employment enables households to increase their incomes, to overcome credit and insurance constraints, and to increase their use of industrial inputs.

Socio-economic characteristics such as household size, household head literacy and education (formal schooling), and household head sex are generally included in the inefficiency effect model. These household-level socio-economic characteristics are widely believed in the literature to affect efficiency. For instance, household size may affect labour supply. In addition, household head literacy and education are used as proxies for farming experience and necessary skills of management.

VI. Empirical Results and Discussion

Before presenting my econometric results, I check the robustness and validity of the SF specification. Schmidt and Lin (1984) and Coelli (1995) proposed that in specifying the SF model, a pretest of the skewness of the ordinary least squares (OLS) residual based on the third moment (M3T) should be carried out to test the null hypothesis of no skewness. The theory behind the test is that, for a production-type SF model with the composed error (ε_i = v_i – u_i) with u_i ≥ 0 and vidistributed symmetrically around 0, the residuals from the corresponding OLS should skew to the left (i.e., negative skewness). A negative skew of the third moment is an indication that efficiency effects exist. The Coelli (1995) test is given by:

M 3 T = m_{3} / \sqrt{\frac{6 m_{2}^{3}}{N}}

where m₂ and m₂ are the second and the third sample moments of the OLS residuals, respectively. If the value of M3T is statistically significant at the 1% level, the frontier framework is supported. In my case, the computed value of the test statistic is −6.51. Because it has a normal distribution, the critical value is 1.96, rejecting the null hypothesis of no skewness in the OLS residuals. This result is also confirmed by the significance of variance parameters (γ and σ²) presented later. The parametric SFA approach imposes several assumptions, including the assumptions on the functional form and distributional assumption on the inefficiency term of the composed error term. It is therefore of interest to test the following hypothesis:

Hypothesis 1: H0: β_jk = 0. This null hypothesis identifies an appropriate functional form between the restrictive Cobb–Douglas and the Translog production function. It specifies that the coefficients on square and interaction terms of input variables in Equation (4) are not statistically different from 0.

Hypothesis 2: H0: γ = 0. In Equation (5), the null hypothesis suggests that the inefficiencies are not stochastic and that the technical inefficiency effects are not present in the model at every level. Hence, the combined effect of these variables on TE is statistically insignificant. If this null hypothesis is not rejected, the SFA model could be reduced to the OLS specification. In this case, the output difference among farmers, given equal inputs, is purely due to the differences in random shocks that are outside of the control of the farmer.

Hypothesis 3: H0: δ0 = δ1 = δ2 … δn = 0. In Equation (5), the null hypothesis specifies that the influence of identified inefficiency factors (i.e., household socio-economic, farm-specific and geographical factors) is 0.

Hypothesis 4: H0: $u_{i} = i i d N^{+} σ_{u}^{}_{}^{2}$ . In Equation (4), the null specifies that the half-normal distribution better fits the model as opposed to the alternative truncated–normal distribution for the u_i.

Hypothesis 5: H0: $\sum_{1}^{7} β_{i} = 1$ . In Equation (4), the null hypothesis specifies a constant return to scale in the production function.

A generalized log-likelihood ratio (LR) test can be used to test these hypotheses:

\begin{matrix} L R = - 2 [\ln {L (H_{0})} / \ln {L (H_{1})}] = \\ - 2 [\ln {L (H_{0})} - \ln {L (H_{1})}] \end{matrix}

where L(H0) and L(H1) are the values of the likelihood functions under the null (H0) and alternative (H1) hypothesis, respectively. The computed test statistics should be compared with the mixed chi-square distribution’s critical values proposed by Kodde and Palm (1986). The LR and Wald tests are applied using the lrtest and test commands in STATA, and results are presented in Table 2.

Table 2.
Testing Hypothesis and Model Specification

Null hypotheses Test statistic p-Value Decision

Functional form (translog vs. Cobb–Douglas)

H0: β_jk = 0 LR = 583.43 0.000 Reject H0

Specification tests

H0: γ = 0 LR = 167.76 0.000 Reject H0

H0: δ0 = δ1 = δ2 =… δn = 0 LR = 1,163.40 0.000 Reject H0

H0: $u_{i} = i i d N^{+} (0, σ_{u}^{2})$ LR = 8.42 0.004 Reject H0

Testing for CRS

H0: $\sum_{1}^{7} β_{i} = 1$ Wald (χ²) = 0.07 0.784 Fail to reject H0

Null hypotheses	Test statistic	p-Value	Decision
Functional form (translog vs. Cobb–Douglas)
H0: β_jk = 0	LR = 583.43	0.000	Reject H0
Specification tests
H0: γ = 0	LR = 167.76	0.000	Reject H0
H0: δ0 = δ1 = δ2 =… δn = 0	LR = 1,163.40	0.000	Reject H0
H0: $u_{i} = i i d N^{+} (0, σ_{u}^{2})$	LR = 8.42	0.004	Reject H0
Testing for CRS
H0: $\sum_{1}^{7} β_{i} = 1$	Wald (χ²) = 0.07	0.784	Fail to reject H0

Source: The authors (calculations of the ALCS 2013–2014 data).

The calculated LR statistic rejects the Cobb–Douglas (restricted) functional form, indicating that square and interaction terms in the Translog model are significantly different from 0, and that Translog cannot be reduced to the Cobb–Douglas specification. The second null hypothesis is tested using the LR test based on the log-likelihood function’s value under OLS and ML estimation of the SF model. The LR test rejects the null; hence, the traditional average (OLS) production function is not an appropriate representation of the sample data (confirming the results of M3T test). To test the third hypothesis, the LR test statistic is calculated based on the log-likelihood value of SF model without explanatory variables of the inefficiency effect model (H0) and the full frontier model with all explanatory variables included in the inefficiency effect model (H1). The test rejects the H0, indicating that the explanatory variables associated with the inefficiency effect model are jointly different from zero. To test the fourth hypothesis, two models are constructed, corresponding to the distributional assumptions of half and truncated–normal for the one-sided error term (under H0 and H1, respectively). The LR test rejects the null, implying that the truncated–normal model is preferred to half-normal. The computed Wald test statistic fails to reject the null hypothesis of constant return to scale (CRS), implying that the production function exhibits CRS.

The ML estimates of parameters of the stochastic production frontier and inefficiency model given by two-equation systems (4) and (5) are simultaneously obtained using STATA and are reported in Tables 3 and 4. The estimated value of σ² is statistically significant at the 1% level, indicating that there exists sufficient evidence to suggest that technical inefficiencies are present in the data, and that the differences between the observed (actual) and frontier (potential) output are due to inefficiency and not chance alone. Gamma (γ) is the variance ratio explaining the total variation in output from the frontier level of output attributed to TE. The estimated value of γ is 0.902 for the preferred truncated–normal model, indicating that about 90% of the difference between the observed and frontier output are primarily due to the inefficiency factors that could be controlled by farms (Table 3). The estimated square and several of the interaction terms are significantly different from zero, indicating that there exists important interactions among the variables, thereby rejecting the appropriateness of the Cobb–Douglas model as an adequate representation of the data; hence, the non-linear functional form is favoured.

Table 3.

ML Estimates of the Stochastic Frontier Model

Variable	Exogenous		Endogenous
Variable	Coefficient	SE	Coefficient	SE
Dependent variable (total aggregate revenue in AFN)
Constant	0.145***	0.041	0.120***	0.045
Ln Land (X₁)	0.440***	0.025	0.420***	0.025
Ln Labour (X₂)	0.050**	0.020	0.052**	0.021
Ln Seed expenditures (X₃)	0.136***	0.015	0.130***	0.015
Ln Fertilizer expenditures (X₄)	0.193***	0.015	0.195***	0.015
Ln Chemical expenditures (X₅)	0.038**	0.015	0.044***	0.015
Ln Tractor rental (X₆)	0.116***	0.017	0.112***	0.017
Ln Other expenditures (X₇)	0.021*	0.012	0.035***	0.012
0.5 × Ln Land (X₁)²	0.014	0.016	−0.008	0.016
0.5 × Ln Labour (X₂)²	0.057***	0.014	0.057***	0.014
0.5 × Ln Seed expenditures (X₃)²	0.035***	0.004	0.033***	0.004
0.5 × Ln Fertilizer expenditures (X₄)²	0.037***	0.004	0.038***	0.004
0.5 × Ln Chemical expenditures (X₅)²	0.018***	0.006	0.019***	0.006
0.5 Ln Tractor rental (X₆)²	0.032***	0.005	0.031***	0.005
0.5 Ln Other expenditures (X₇)²	0.006*	0.003	0.010***	0.003
Ln Land × Ln Labour	−0.021**	0.01	−0.019*	0.01
Ln Land × Ln Seed	−0.007**	0.003	−0.006**	0.003
Ln Land × Ln Fertilizer	0.006**	0.003	0.009***	0.003
Ln Land × Ln Chemicals	0.001	0.004	0.001	0.004
Ln Land × Ln Tractor rental	−0.005	0.003	−0.004	0.003
Ln Land × Ln Other expenses	0.009***	0.003	0.008***	0.003
Ln Labour × Ln Seed	0.008***	0.003	0.009***	0.003
Ln Labour × Ln Fertilizer	−0.010***	0.003	−0.010***	0.003
Ln Labour × Ln Chemicals	−0.001	0.004	−0.003	0.004
Ln Labour × Ln Tractor rental	0.003	0.003	0.004	0.003
Ln Labour × Ln Other expenses	−0.004*	0.003	−0.005*	0.003
Ln Seed × Ln Fertilizer	0.001	0.001	0.001	0.001
Ln Seed × Ln Chemicals	−0.003**	0.001	−0.002**	0.001
Ln Seed × Ln Tractor rental	−0.001**	0.001	−0.001**	0.001
Ln Seed × Ln Other expenses	−0.002***	0.001	−0.002***	0.001
Ln Fertilizer × Ln Chemicals	0	0.001	0	0.001
Ln Fertilizer × Ln Tractor rental	−0.003***	0.001	−0.003***	0.001
Ln Fertilizer × Ln Other expenses	0.002***	0.001	0.002***	0.001
Ln Chemicals × Ln Tractor rental	−0.002*	0.001	−0.002*	0.001
Ln Chemicals × Ln Other expenses	−0.001	0.001	−0.001	0.001
Ln Tractor rental × Ln Other expenses	−0.002***	0.001	−0.003***	0.001
(σ)2	0.371***	0.011
c	0.904***	0.038
Log-likelihood		−7,519.80		−5820.30
Wald χ² (prob. χ²)	4,688.85 (0.000)
eta (η) Endogeneity test (p-value)			26.56*** (0.000)
N		7,066		7,066

Source: The authors (calculations of the ALCS 2013–2014 data).

Note: Significances is indicated by *p < 0.10, **p < 0.05, and ***p < 0.010.

Since input and output variables were transformed to their corresponding log values and were normalized by their respective sample means, the estimated parameters were directly interpreted as partial elasticities at the sample mean. All slope coefficients or output elasticities of inputs showed the expected signs and were found to be highly significant. Production inputs such as land, labour, fertilizers, chemicals and seed were significant as expected a priori with large partial elasticities, indicating they are important determinants of revenue. The sum of the partial elasticities with respect to the production inputs estimated by the ML estimator of the translog stochastic production function is 0.99, roughly consistent with constant returns to scale (CRS), which implies that an increase in all inputs leads to an equal proportional increase in farm revenues.

The ML estimator is employed to estimate the δ coefficients in the inefficiency model (Table 4). A negative sign of the estimated parameters indicates a reduction in technical inefficiency or an increase in TE. The estimated coefficients of all variables are significant, except for gender, age and education of household head, household size and off-farm employment.

Table 4.

ML Estimates of the Inefficiency Effect Model

Variable	Exogenous		Endogenous
Variable	Coefficient	SE	Coefficient	SE
Diversification index (THI)	−3.765***	0.346	−7.015***	1.076
Head sex (male)	−0.451	0.377	−0.436	0.437
Head age (years)	0.001	0.002	0.001	0.003
Head education (years)	0.006	0.102	0.011	0.012
Head literacy (can read and write)	−0.026	0.093	−0.058	0.114
Household size (persons)	0.001	0.011	0.004	0.013
Extension services (1 = yes)	−0.323***	0.098	−0.467***	0.123
Oxen and yaks (number)	−0.176***	0.062	−0.215***	0.078
Tractor/threshers (number)	−0.807***	0.255	−0.755**	0.298
Cattles (number)	−0.107***	0.022	−0.099***	0.027
Off-farm employment (1 = yes)	−0.069	0.1	−0.072	0.126
Opium share by province (%)	−1.025	0.724	−0.040	0.884
Farm size (>2 to 5 Jeribs)	−0.392***	0.094	−0.315***	0.111
Farm size (>5 to 10 Jeribs)	−0.351***	0.117	−0.293**	0.133
Farm size (>10 to 20 Jeribs)	−0.048	0.157	−0.015	0.183
Farm size (>20 and above Jeribs)	0.358*	0.203	0.228*	0.263
Land type (rainfed)	0.236***	0.09	0.053**	0.122
AEZ 1 (CM)	−0.079	0.191	0.029	0.22
AEZ 2 (HFL)	−0.062	0.225	−0.12	0.26
AEZ 3 (SMF)	−0.919***	0.209	−0.851***	0.239
AEZ 4 (HVSB)	−0.821***	0.234	−0.956***	0.269
AEZ 5 (TP)	−0.002	0.207	0.05	0.242
AEZ 6 (NMF)	−0.241	0.183	−0.106	0.213
AEZ 7 (EMF)	−0.503**	0.197	−0.315**	0.228
Constant	1.254***	0.43	1.073*	0.494
N	7,066			7,066

Source: The authors (calculations of the ALCS 2013–2014 data).

Note: Table reports estimates of Equation (5). The omitted categories are: none for education level, <2 Jeribs for farm size and AEZ 8, none for extension services, none for literacy, none for off-farm employment, and irrigated land alone for land quality. significance levels indicated by * p < 0.10, ** p < 0.05 and *** p < 0.010.

Before presenting the results from the multivariate inefficiency model to formally test the impact of CD, I carried out a t-test to show patterns and variations in the farm characteristics among farms that rely on mono-cropping and farms that adopt diversification. The results of the two-tailed t-test are provided in Table A2, revealing that farms with a higher diversification index earn significantly higher revenues. The multivariate regression results confirm this relationship. The estimated coefficient for the index of CD is negative and statistically significant at the 1% level under both half-normal and truncated–normal estimations of u_i. This indicates that a higher CD index² is associated with a higher level of TE at the farm level. Figure 2 illustrates this effect; efficiency increases with diversification. Although diversifying crops may require additional management skills, it has the advantages of greater utilization of inputs, producing marketable crops, and reducing reliance on production of a single staple crop mainly for subsistence.

Figure 2.

Source: The authors (calculations of the ALCS 2013–2014 data).

The causal link among CD, production efficiency and household welfare is well established. CD is viewed as a hedge against shocks due to extreme weather conditions, crop diseases and pests, and unexpected fall of market prices. Krupinsky et al. (2002) argued that CD reduces the potential risk against uncertainty by reducing high dependency on mono-cropping; reduces economic losses due to diseases, weed and infestation; and increases soil fertility through crop rotation. Lin (2011) describes CD as an environmentally sound and viable climate-smart agriculture practice widely perceived to significantly enhance farm productivity and increase resilience in rural farming systems. Makate et al. (2016) argued that diversified cropping systems primarily reduce crop production risks by providing more income options to the farmer, therefore, making production on the farm more stable. Our finding that more diversified farms are more efficient is consistent with Nguyen (2014), Manjunatha et al. (2013), Ogundari (2013), Rahman (2009) and Coelli and Fleming (2004) for Vietnam, India, Bangladesh, Nigeria and Papua New Guinea, respectively.

Karakaplan and Kutlu (2017) offer an endogeneity test similar to the Durbin–Wu–Hausman test as part of the sfkk estimation³ to test for endogeneity in SF models. The eta (η) endogeneity test examines the joint significance of the components of the η term. If the components are jointly significant, endogeneity is detected and must be corrected for; otherwise, the model can be fitted to traditional frontier models. Under the null hypothesis, the η test assumes that correction for the endogeneity is not necessary, and that the exogenous estimation of CD is valid. The estimated test statistic is χ² = 25.13 with a p-value of 0.000, which rejects the null hypothesis of no endogeneity in CD at 1% statistical significance. The estimated coefficient of the instrument (0.71) is significant at the 1% level and strongly correlated with the endogenous variable conditional on the other covariates, indicating that the instrument is valid (note that the results from the first stage are not reported here to save space).

Correcting for the endogeneity bias in CD increases its coefficient (in absolute terms) from 4.95 to 6.81 in the inefficiency model, implying that failing to account for the endogeneity underestimated the effect of CD in the standard exogenous SF model. This is consistent with attenuation bias due to measurement error in CD, causing a downward bias. As a result, the average estimated level of TE using the endogenous model is 4% (i.e., mean efficiency for the exogenous model is 69.9% and 73.9% from the endogenous model) higher than the estimated efficiency using the standard SF model. To illustrate this effect, I plot the distribution of the estimated TE by the exogenous and endogenous models (Figure 3). The efficiency estimates largely overlap for less efficient farms; however, the two estimates are different for farms with higher efficiency levels. This is perhaps because more efficient farms are highly diversified (as evidenced before). When endogeneity is not accounted for, the efficiency estimates are biased downwards towards zero, consistent with the measurement error.

Figure 3.

Source: The authors (calculations of the ALCS 2013–2014 data).

The estimated efficiency scores from the preferred endogenous model range from 1.5% to 98%, with a sample mean of 74.31%. The distribution of the estimated efficiency indices shows that about 15% of the sample farmers realized less than 50% of the potential revenues, whereas about 40% of farms achieved more than 80% efficiency. The remaining farmers were operating between the levels of 50% and 80%. About a third of the farms did not diversify (i.e., they grew only one crop), achieving, on average, 52% of the potential revenues (Table 5).

Table 5.

Distribution of Efficiency Over Index of Diversification

Diversification index (0 £ THI £ 1	Efficiency
Diversification index (0 £ THI £ 1	Mean	Std.	Percentage of Farms
0–<0.1	52.60	2,338	33
0.1–<0.2	68.11	321	5
0.2–<0.3	75.40	445	6
0.3–<0.4	81.86	638	9
0.4–<0.5	87.44	2,239	32
0.5–<0.6	90.01	566	8
0.6–<0.7	92.32	410	6
0.7–1	94.32	88
Overall	74.31	7,045	100%

Source: The authors (calculations of the ALCS 2013–2014 data).

The negative and significant effect of access to extension services on technical inefficiency implies that farmers who have had contact with extension services achieved higher TE, perhaps because they were helped to diversify. Although a relatively small proportion of farmers (about 21% of the sample) can access them, extension services have critical implications on diversification and production efficiency. Our findings of the positive impact of extension services on efficiency agree with Ogundari (2013) and Bozoğlu and Ceyhan (2007).

There seems to be an inverted U-shaped relationship between farm size and TE (efficiency level rises initially with farm size but appears to fall when farm size exceeds 20 Jeribs). That is, medium-sized farms are relatively more efficient, perhaps because these farms are more diversified. However, farms with the largest endowments of land are more market-oriented and commercialized, and therefore tend to specialize, aiming to achieve optimum economies of scale. Findings on the relationship between farm size and efficiency vary in the literature. Oladeebo and Oyetunde (2013) find an inverse relationship between farm size and TE. Manjunatha et al. (2013) conclude that increased farm size improves TE. Helfand and Levine (2004) concluded that the relationship between farm size and efficiency is non-linear, with efficiency first falling and then rising with size and falling again when farm size is too large. Narala and Zala (2010) found that medium-sized farms are the most efficient, presumably due to having agriculture as their primary occupation and allocating their resources more effectively.

The wide range of altitude in Afghanistan leads to a significant variation in climate, spatial and temporal biodiversity within relatively small distances that, in turn, affect households’ CD strategies. The map in Figure 4 illustrates the spatial variation in CD and efficiency, confirming that higher averages of efficiency scores correspond with the higher values of the diversification index. Both CD and efficiency significantly vary across the country. The distribution of efficiency brings together the relationship between diversification and the efficiency levels realized by the farmers. That is, the most efficient ecological zones are relatively more diversified, on average, as compared to the other zones that are somewhat less efficient.

Figure 4.

Source: The authors (calculations of the ALCS 2013–2014 data).

Ownership of farm assets, including cattle, oxen and tractors by the farm households, is positively correlated to the level of TE. Cattle and oxen ownership might imply the availability of animal manure, which is an important and cheap source of organic fertilizer (particularly in the small-scale farming system) and is widely believed to improve soil structure and fertility. Oxen and tractor ownership provide sources of more affordable traction power to farmers as compared to hiring tractors. Besides, tractor ownership may indicate that farm mechanization ensures timely land preparation, planting and weeding.

Land quality also appeared to have a significant impact on both CD and TE. Farmers operating a combination of rain-fed land and irrigated land were found to be more inefficient than those who cultivated irrigated land alone. This may suggest that rain-fed land may not allow a greater level of diversity.

Opium intensity in the province is positively but insignificantly linked with TE. The insignificance of this relationship may be due to a trade-off between the effects of access to cash and insecurity. Production in provinces where farmers grow opium may be relatively more efficient as compared to other regions because farmers can purchase inputs (and sales of opium may inflate reported revenue), but opium-affected provinces are likely to be more insecure.

Off-farm employment has no significant impact on TE. Previous studies that assessed the effect of off-farm employment on TE produced mixed conclusions. Haji (2007) and Zhang et al. (2016) found that off-farm employment positively contributed to TE due to income effects. Other studies concluded that off-farm employment is negatively associated with TE due to a reduction in labour supply to farm activities, as hiring agricultural labour incurs transaction costs (O’Neill and Matthews, 2001).

VII. Robustness Checks

An important feature of farming in the context of Afghanistan is opium production, which could alter farm households’ production decisions and management strategies like CD, particularly since opium cultivation is relatively more widespread in some agro-ecological zones and provinces than others. To test whether such heterogeneities related to the geographical aspects of opium production would lead to the systematic differences in the household’s decision-making and CD levels that in turn may affect production efficiencies, I run a robustness check by dividing the total analytical sample into two sub-samples based on the intensity of opium production at the provincial level. Farm households located in the main opium-producing provinces are assigned to one sub-sample (periphery), whereas households in the opium-free provinces are assigned to another sub-sample (core). To do this, I use the MCN provincial-level data on the intensity of opium production. Subsequently, I run the analysis for each subgroup of households separately to test the extent to which CD and other households’ socio-economic characteristics differ among the core and periphery samples and whether these differences generate systematic difference in the estimated production efficiency levels.

My results from the two sub-samples suggest no dramatic qualitative differences in the estimates among the two subgroups, although there are some quantitative differences in the estimated parameters. CD remains positively and significantly associated with the farm’s TE. The results of these tests do not add additional insights and are therefore not reported here. I included a control variable that captures the intensity of poppy cultivation at the provincial level in my primary analysis, which was sufficient to handle variations due to poppy cultivation.

VIII. Discussion and Conclusion

The focus of the analysis in this article is to estimate farm-level production efficiency and to investigate whether CD strategies by farm households improve the efficiency of the farming sector in Afghanistan. I employ a new estimation strategy to allow for endogeneity bias in the SF models, a major econometric issue in traditional SF models. The results reveal that the farming sector in Afghanistan experiences significant technical inefficiencies. Nearly 15% of farm households achieve less than 50%, whereas about 23% realize 50%–70% of potential revenue. The overall TE is estimated at 72%, on average, indicating substantial room for improving farm revenues via employing improved farm management practices and without using additional production resources and rising production costs. The derived CRS implies that an increase in all inputs leads to an equal proportional increase in revenues, signifying the existence of inefficiencies and indicating that focusing only on applying additional inputs may not be economical.

CD was found to be a key factor associated with higher levels of efficiency. CD increased productivity by shifting the production portfolio from mono-cropping to a diversified high-value production system that also helped mitigate production and marketing risks. This outcome was particularly critical for the policy as the data suggested that nearly 33% of the households did not diversify (i.e., they were mono-cropping) and achieved less than 50% of production efficiencies.

A direct policy recommendation that can be generated from the findings of this study is that CD should be given more recognition by both farm households and policymakers, particularly in shifting production systems away from mono- and staple-crop production to a multiple- and HVC production system. CD is essential to boost efficiency, but it is also considered a climate-smart agricultural technique that can improve resilience and adaptive capacity in volatile agro-ecological conditions due to climate change. Diversification improves yield stability and brings more spatial and temporal biodiversity on the farm. As indicated by the significance of the extension services variable, one way to contribute to greater farm diversity is to educate farmers through expanding access to extension services in remote rural areas. Future policies should also facilitate access to financial credit through strengthening financial institutions to remove liquidity barriers and ensure the more efficient use of productive resources, especially for smallholders who excessively rely on conventional inputs like animal manure and oxen use.

Although the contribution of new technologies and inputs is indispensable, agricultural policies should incentivize efforts towards improving the current level of efficiency of farmers so that their use of existing technologies and inputs are cost-effective instead of entirely relying on introducing new technologies. Future agricultural policies should advocate for a more integrated and holistic market-led approach to facilitate access to the market and augment market-oriented systems, especially the production of high-value cash crops. More generally, government investment in the development of rural infrastructure will not only increase production efficiencies at the farm level but will also complement CD strategies by improving opportunities for technology diffusion, marketing, storage and resource supplies.

The ongoing conflict has had a devastating effect on agricultural production and food security. It is likely that the protracted crisis and the related violence will limit the feasibility of implementing agricultural policies in some areas of the country, especially in heavily affected areas. Agricultural policies, therefore, must go hand in hand with conflict resolution, coping mechanisms and efforts towards building resilience of the vulnerable communities to ensure the target impacts are achieved. More importantly, rural development interventions must be designed alongside conflict prevention mechanisms to avoid triggering and escalating conflicts between communities and farmers. Risk-sensitive management of agriculture and natural resources holds the key to social regeneration for recovery: building resilience for those more heavily affected by conflict will lay the groundwork for economic development programmes as relative stability becomes more established. Constructive engagement of government institutions to prevent and mitigate the impact of conflict will require profound understanding of the dynamics and vulnerabilities of local agricultural systems; good understanding of the drivers of violence both during ongoing conflict and when relative stability is established; as well as a willingness to work in conflict-affected areas with institutions providing humanitarian assistance, relief and support for those who are internally displaced.

Footnotes

Acknowledgements

The author would like to thank Professor Oliver Morrissey (University of Nottingham) for his guidance, support and advice during the development of an earlier version of this article. The author also wishes to extend his gratitude to Catherine Locke, editor-in-chief (Progress in Development Studies) for her thorough editorial review, which made this article more readable, engaging and clearer. The author acknowledges and thanks the anonymous reviewers for their constructive comments and feedback, which helped improve the contents of this study.

An earlier version of this article is available in the authors’ doctoral dissertation and as a working paper available online through the Centre for Research in Economic Development and International Trade (CREDIT) at the University of Nottingham (Ahmadzai, 2017).

Declaration of Conflicting Interests

The author declared no potential conflicts of interest with respect to the research, authorship and/or publication of this article.

Funding

The author received no financial support for the research, authorship and/or publication of this article.

Appendix

Table A2.

Characteristics of Households with CDI Below and Above the Median (0.37)

	Specialized	Diversified	Two-tailed T-test
	Mean	Mean	Difference	SE
Aggregate annual revenue (AFN)	46,884	69,621	−22,737***	−10.6
Land (Jeribs)	6.77	7.30	−0.530**	−2.44
Labour (hours)	60.16	67.41	−7.253***	−4.89
Seed expenditure (AFN)	2,389	2,320	69.33	−0.8
Fertilizer expenditure (AFN)	3,582	5,945	−2,363***	−12.1
Chemicals expenditure (AFN)	228.0	502.1	−274***	−9.68
Tractor rental (AFN)	2,460	2,549	−88.7	−0.87
Other expenditure (AFN)	797	749	47.98	−0.91
Herfindahl index (HHI)	0.91	0.50	0.414***	−163
Household size (persons)	7.91	8.76	−0.842***	−10.3
Head age (years)	44.22	44.55	−0.325	−0.99
Head sex (1 = male, 0 = female)	1.00	1.00	−0.0014	−1.00
Extension services (1 = access, 0 = No)	0.19	0.23	−0.043***	−4.49
Head literacy (1 = yes, 0 = otherwise)	0.31	0.34	−0.0315***	−2.83
Off-farm employment (1 = yes, 0 = No)	0.12	0.13	−0.0199**	−2.53
Cattle (number)	1.33	1.87	−0.533***	−11.1
Tractors (number)	0.05	0.06	−0.0102	−1.85
Oxen (number)	0.23	0.24	−0.00851	−0.58
N	7,052

Source: The authors (calculations of the ALCS 2013–2014 data).

Note: Significance levels indicated by *p < 0.10, **p < 0.05 and ***p < 0.010.

Notes

References

Ahmadzai

2017: Crop diversification and technical efficiency in Afghanistan. CREDIT Research Paper 17/04. www.nottingham.ac.uk/credit/news/papers/1704.aspx.

Ahmadzai

2020: How is off-farm income linked to on-farm diversification? Evidence from Afghanistan. Studies in Agricultural Economics, 122(1), 1–12. doi.org/10.7896/j.2010.

Aigner

, Lovell

C.K.

and Schmidt

1977: Formulation and estimation of stochastic frontier production function models. Journal of Econometrics, 6, 21–37.

Amsler

, Prokhorov

and Schmidt

2016: Endogeneity in stochastic frontier models. Journal of Econometrics, 190, 280–88.

Battese

G.E.

and Coelli

T.J.

1995: A model for technical inefficiency effects in a stochastic frontier production function for panel data. Empirical Economics, 20, 325–32.

Bellon

M.R.

, Kotu

B.H.

, Azzarri

and Caracciolo

2020: To diversify or not to diversify, that is the question. Pursuing agricultural development for smallholder farmers in marginal areas of Ghana. World Development, 125, 104682

Bozog˘lu

and Ceyhan

2007: Measuring the technical efficiency and exploring the inefficiency determinants of vegetable farms in Samsun province, Turkey. Agricultural Systems, 94, 649–56.

Central Statistical Organization (CSO). 2014: Afghanistan living conditions survey 2013–2014: National risk and vulnerability assessment. CSO.

Chaiechi

and Stoeckl

2013: Estimating the costs of conservation in multiple output agricultural setting. International Journal of Economics and Management Engineering, 7, 1840–47.

10.

Charnes

, Cooper

W.W.

and Rhodes

1978: Measuring the efficiency of decision-making units. European Journal of Operational Research, 2, 429–44.

11.

Coelli

1995: Estimators and hypothesis tests for a stochastic frontier function: A Monte Carlo analysis. Journal of Productivity Analysis, 6, 247–68.

12.

Coelli

T.J.

1995: Recent developments in frontier modeling and efficiency measurement. Australian Journal of Agricultural Economics, 39, 219–45.

13.

Coelli

and Fleming

2004: Diversification economies and specialisation efficiencies in a mixed food and coffee smallholder farming system in Papua New Guinea. Agricultural Economics, 31, 229–39.

14.

Danso-Abbeam

and Baiyegunhi

L.J.S.

2020: Do farm-level technical efficiency and welfare complement each other? Insight from Ghana’s cocoa industry. Journal of Economic Structures, 9, 23.

15.

Duffy

2009: Economies of size in production agriculture. Journal Hunger and Environmental Nutrition, 4, 375–92.

16.

Dupree

1973: Afghanistan, Oxford Pakistan paperbacks. Oxford University Press.

17.

Farrell

M.J.

1957: The measurement of productive efficiency. Journal of the Royal Statistical Society Series A (General), 120, 253.

18.

Greene

W.H.

1980: Maximum likelihood estimation of econometric frontier functions. Journal of Econometrics, 13, 27–56.

19.

Gronberg

T.J.

, Jansen

D.W.

, Karakaplan

M.U.

and Taylor

L.L.

2015: School district consolidation: Market concentration and the scale-efficiency tradeoff: School consolidation: Scale & efficiency. Southern Economic Journal, 82, 580–97.

20.

Guan

, Kumbhakar

S.C.

, Myers

R.J.

and Lansink

A.O.

2009: Measuring excess capital capacity in agricultural production. American Journal of Agricultural Economics, 91, 765–76.

21.

Guvele

C.A.

2001: Gains from crop diversification in the Sudan Gezira scheme. Agricultural Systems, 70, 319–33.

22.

Haji

2007: Production efficiency of smallholders’ vegetable-dominated mixed farming system in Eastern Ethiopia: A non-parametric approach. Journal of African Economies, 16, 1–27.

23.

Helfand

and Levine

2004: Farm size and the determinants of productive efficiency in the Brazilian Center-West. Agricultural Economics, 31, 241–49.

24.

Humlum

1959: La géographie de l’Afghanistan. Etude d’un pays aride. University Books.

25.

Huttunen

2019: Revisiting agricultural modernisation: Interconnected farming practices driving rural development at the farm level. Journal of Rural Studies, 71, 36–45.

26.

Ibrahim

, Rahman

S.A.

, Envulus

and Oyewole

S.O.

2009: Income and crop diversification among farming households in rural area of North Central Nigeria. Journal of Tropical Agriculture, Food, Environment and Extension, 8, 84–89.

27.

Joshi

P.K.

, Gulati

and Cummings

R.W.

2007: Agricultural diversification in South Asia: Beyond food security. Academic Foundation, 47–82.

28.

Karakaplan

U.M.

2017: Fitting endogenous stochastic frontier models in Stata. The Stata Journal, 17, 39–55.

29.

Karakaplan

U.M.

and Kutlu

2017: Handling endogeneity in stochastic frontier analysis. Economics Bulletin, 37, 889–901.

30.

Kodde

D.A.

and Palm

F.C.

1986: Wald criteria for jointly testing equality and inequality restrictions. Econometrica, 54, 1243–48.

31.

Krupinsky

J.M.

, Bailey

K.L.

, McMullen

M.P.

, Gossen

B.D.

and Turkington

T.K.

2002: Managing plant disease risk in diversified cropping systems. Agronomy Journal, 94, 198–209.

32.

Kumbhakar

S.C.

and Lovell

C.A.K.

2000: Stochastic frontier analysis. Cambridge University Press.

33.

Kumbhakar

and Wang

H.-J.

2015: A practitioner’s guide to stochastic frontier analysis using Stata. Cambridge University Press.

34.

Lakner

, Cramon-Taubadel

and von Brümmer

2012: Technical efficiency of organic pasture farming in Germany: The role of location economics and of specific knowledge. Renewable Agriculture and Food Systems, 27, 228–41.

35.

Lee

L.-F.

1983: The determination of moments of the doubly truncated multivariate normal tobit model. Economics Letters, 11, 245–50.

36.

Lin

B.B.

2011: Resilience in agriculture through crop diversification: Adaptive management for environmental change. BioScience, 61, 183–93.

37.

Makate

, Wang

, Makate

and Mango

2016: Crop diversification and livelihoods of smallholder farmers in Zimbabwe: Adaptive management for environmental change. SpringerPlus 5.

38.

Maletta

and Favre

2003: Agroecological zones. Food and Agriculture Organization of the United Nations.

39.

Manjunatha

A.V.

, Anik

A.R.

, Speelman

and Nuppenau

E.A.

2013: Impact of land fragmentation, farm size, land ownership and crop diversity on profit and efficiency of irrigated farms in India. Land Use Policy, 31, 397–405.

40.

McNamara

K.T.

and Weiss

2005: Farm household income and on- and off-farm diversification. Journal of Agricultural and Applied Economics, 37, 37–48.

41.

Meeusen

and Van den

Broeck

. 1977: Efficiency estimation from Cobb-Douglas production functions with composed error. International Economic Review, 18(2), 435–44.

42.

Mekonnen

D.K.

, Spielman

D.J.

, Fonsah

E.G.

and Dorfman

J.H.

2015: Innovation systems and technical efficiency in developing-country agriculture. Agricultural Economics, 46, 689–702.

43.

Meraner

, Heijman

, Kuhlman

and Finger

2015: Determinants of farm diversification in the Netherlands. Land Use Policy, 42, 767–80.

44.

Minot

, Epprecht

, Tran

T.T.A.

and Trung

L.Q.

2006: Income diversification and poverty in the Northern Uplands of Vietnam. International Food Policy Research Institute.

45.

Narala

and Zala

Y.C.

2010: Technical efficiency of rice farms under irrigated conditions in Central Gujarat. Agricultural Economics Research Review, 23, 375–81.

46.

Nguyen

2014: Crop diversification, economic performance and household behaviour Evidence from Vietnam. Paper presented at the 89th Annual Conference, Agricultural Economics Society, Warwick University, Coventry, 13–15 April.

47.

Ogundari

2013: Crop diversification and technical efficiency in food crop production: A study of peasant farmers in Nigeria. International Journal of Social Economics, 40, 267–87.

48.

Oladeebo

J.O.

and Oyetunde

O.T.

2013: Relationship between plot size and technical efficiency of Cassava Farms in Oyo State of Nigeria: A stochastic frontier analysis. International Journal Current Research, 5, 2782–86.

49.

O’Neill

and Matthews

2001: Technical change and efficiency in Irish agriculture. Economic and Social Review, 32, 263–84.

50.

Oushy

2010: Farm resource management in Afghanistan–Training course. United States Agency for International Development (USAID)/New Mexico State University (NMSU), Afghanistan Water, Agriculture, and Technology, Transfer (AWATT) Project.

51.

Rahman

2009: Whether crop diversification is a desired strategy for agricultural growth in Bangladesh? Food Policy, 34, 340–49.

52.

Schmidt

and Lin

T.-F.

1984: Simple tests of alternative specifications in stochastic frontier models. Journal of Econometrics, 24, 349–61.

53.

Tenaye

2020: Technical efficiency of smallholder agriculture in developing countries: The case of Ethiopia. Economies, 8, 34.

54.

Van den Berg

M.M.

, Hengsdijk

, Wolf

, Van Ittersum

M.K.

, Guanghuo

and Roetter

R.P.

2007: The impact of increasing farm size and mechanization on rural income and rice production in Zhejiang province, China. Agricultural Systems, 94, 841–50.

55.

Wadud

and White

2000: Farm household efficiency in Bangladesh: A comparison of stochastic frontier and DEA methods. Applied Economics, 32, 1665–73.

56.

Weiss

and Briglauer

2000: Determinants and dynamics of farm diversification. Paper presented at the 10th European Association of Agricultural Economists (EAAE) Congress on ‘Exploring Diversity in the European Agri-Food System’, Zaragoza, Spain, 28–31 August.

57.

Weltin

, Zasada

, Franke

, Piorr

, Raggi

and Viaggi

2017: Analysing behavioural differences of farm households: An example of income diversification strategies based on European farm survey data. Land Use Policy, 62, 172–84.

58.

Wooldridge

J.M.

2019: Introductory econometrics: A modern approach, 7 edition. MIT Press.

59.

Yang

, Wang

, Jin

, Chen

, Riedinger

and Peng

2016: Migration, local off-farm employment, and agricultural production efficiency: Evidence from China. Journal of Productivity Analysis, 45, 247–59.

60.

Zhang

, Su

, Eriksson

and Liu

2016: How off-farm employment effects technical efficiency of China’s Farms: The case of Jiangsu. China & World Economy, 24, 37–51.