Professional Documents
Culture Documents
net/publication/305349018
CITATIONS READS
48 616
1 author:
Hiroyuki Takeshima
International Food Policy Research Institute
84 PUBLICATIONS 912 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Hiroyuki Takeshima on 23 December 2018.
1
Custom-hired tractor services and returns to scale in smallholder agriculture: A production
function approach
Hiroyuki Takeshima
2033 K Street, NW
E-mail: H.takeshima@cgiar.org
Phone: +1-202-862-8195
2
Abstract: Historically, agricultural transformation has often accompanied the
increase in the returns to scale. Little direct evidence exists, however, on what
actually causes such increase, despite the knowledge of many factors that are
associated with this increase. We fill this knowledge gap by testing whether hiring
in tractor services has raised returns to scale in agriculture at the household level
in Nepal Terai that has undergone rapid growth in the tractor use through custom-
whether to hire in tractor services; (b) use of inputs in the production function.
agricultural production by about 0.2 ~ 0.3 among farm households not owning
tractors, for which suitable control groups are found. Findings are robust under
3
Acknowledgments
Kathmandu, Nepal, on September 2015, two anonymous referees and the editor of the journal for
providing constructive comments. The paper has also benefited from constructive dialogues with
Xinshen Diao, Nazaire Houssou, Scott Justice, Anjani Kumar and Xiaobo Zhang. This work was
undertaken as part of and funded by the CGIAR Research Program on Policies, Institutions, and
Markets (PIM) led by the International Food Policy Research Institute (IFPRI) and the United
States Agency for International Development (USAID) Food Security Policy project (FSP). The
4
1. Background
Estimating a cross-country Cobb-Douglas production function, Hayami and Ruttan (1985 pp.
148) shows that the production functions differ between developed and developing countries by
the larger scale factor for the former, while input shares are relatively similar. Constant or
increasing RTS is often observed for developed countries’ agriculture, while diminishing RTS is
prevalent in the developing countries’ agriculture (Adesina and Djato, 1997; Craig et al., 1997).
In Japan, agricultural RTS had increased despite the dominance of small farms (Hayami and
Kawagoe, 1989). The change in RTS has significant implications for agricultural policies. With
the same total factor productivity, greater RTS often induces greater intensification, as input use
often becomes more price responsive, so that price policies can have greater effects. The RTS
also affects the relationship between farm size and productivity (Foster and Rosenzweig, 2011;
Deininger and Byerlee, 2012), and thus the structure of the agricultural sector.
What raises RTS in agriculture has, however, not been well understood in the literature.
Studies associate the increase in RTS with various factors, including specialization, land
consolidation (Wan and Cheng, 2001), a change of crop mix (Cramb, 2011), or breeding of pest-
resistant varieties that reduce labor monitoring costs (Deininger and Byerlee, 2012), as well as
mechanization (Foster and Rosenzweig, 2011; Christiaensen, 2013). Few studies, however, have
established the direct causality. It is possible that RTS is increased by other factors, and such
increase in RTS may be inducing demand for specialization, change in crop mix, development of
We fill this knowledge gap by testing the hypothesis that the adoption of services
provided with tractors (tractor services, TS hereafter) through custom-hiring (hiring-in of TS,
5
HTS hereafter) “increases” the agricultural RTS among non-owners of tractors. We focus on
non-owners of tractors, because many smallholders using tractors in developing countries rely on
HTS instead of investing in own machines (Diao et al., 2014; Benin, 2015; Takeshima et al.,
2015a), and also because of methodological constraints as is described in later sections. Our
analysis therefore does not cover tractor owners whose hiring out services are often behind the
growths of smallholder mechanization through HTS in developing countries (Benin et al., 2013;
Diao et al., 2014; Takeshima et al., 2015a). Future studies investigate the contributions of tractor
Our analysis focuses on the lowland (Terai) zone of Nepal in 2010, using the Nepal
Living Standard Survey collected in 2010 (NLSS 2010 hereafter) by the Central Bureau of
Statistics of Nepal. In Nepal Terai, the share of tractor-using farm households increased
relatively fast, from 9% in 1995 to almost 49% in 2010, primarily through HTS (Takeshima et
al., 2015b). During this time, key production factors among nonhirers might not have adjusted
fast enough, and the production functions of nonhirers in 2010 may still exhibit significant
differences from that of TS-hirers. Focusing on Nepal Terai in 2010 therefore can detect the
potential difference in RTS due to HTS. For example, the trends in agricultural capital (see
online Appendix A for its definition) and area cultivated, which we show later as major inputs in
the production functions, provide partial evidence for this, based on author’s calculations using
the NLSS 2010 as well as two other rounds (1995, 2003) not used in this study. Between 1995
and 2010, the average value of agricultural capital holdings of farm households in Nepal Terai,
when measured in the equivalent of kg worth of cereals, had changed only slightly from 2900 in
1995 to 3000 in 2003 and 3300 in 2010. Average area cultivated in Nepal Terai during this
period had declined from 2.4 ha in 1995 to 2.0 in 2003 and 1.4 in 2010. These patterns are
6
consistent across both TS-hirers and nonhirers. Therefore, neither agricultural capital
accumulation nor farm size expansion in Nepal Terai is driving the growth in RTS that we show.
In the robustness check section, we also show that we are capturing the effects of HTS per se,
and not the effects of other endogenous inputs uses that may be correlated with HTS.
We specifically test whether the HTS raises the RTS measured as the sum of factor
coefficients in production functions, address endogeneity from two sources. First, the decision of
HTS may be endogenous to the changes in RTS. Second, inputs in the production functions may
be endogenous to the output. We address them jointly through switching regression Generalized
We contribute to the literature in various ways. The mechanization literature suggests that
agricultural mechanization often leads to farm size expansion (ex., Houssou and Chapoto, 2014),
or is associated with larger farm size (ex., Binswanger, 1986; Pingali, 2007; Otsuka et al., 2013;
Yamauchi, 2016). It has, however, not investigated whether mechanization simply complements
land and substitutes land-saving inputs within the fixed production function (and thus not
affecting the RTS), or it affects the production function itself (thus affecting the RTS). We partly
fill this knowledge gap by showing evidence for HTS. Lastly, we contribute to the literature on
impact evaluation methods, by showing how the causal effect of technology adoption on the
HTS may increase the agricultural RTS through various mechanisms. Mechanical power
may mitigate the rapid decrease in marginal productivity inherent in human or animal powers
7
due to drudgery and supervising costs (Foster and Rosenzweig, 2011). Timely land preparation
with mechanical power may mitigate the decrease in the marginal land productivity due to the
(Haggblade, 2005). Using tractors for transporting harvests may allow timely market
transactions, mitigating the decrease in returns to agricultural capital like processing machines
and private storage facilities. This might also increase feed crop harvests. Combined with the
timeliness in transporting fresh animal products, RTS in livestock holdings may increase.
Alternatively, HTS may increase the RTS through greater backward and forward linkages
and outsourcing of the supply of inputs and services. Outsourcing of TS may keep the marginal
costs of land preparation relatively constant. Timely land preparation may allow planting of
homogeneous varieties that grow, flower and mature at the same time. This may induce sourcing
of larger quantities of homogeneous seeds from the market, rather than relying on diverse
varieties that are more suitable for varying land preparation dates. With an increase in RTS of
physical production, farmers may demand more inputs and services from the market, which
exhibit relatively constant marginal costs, relative to internally supplied inputs and services
which tend to exhibit increasing marginal costs. The use of tractors for transportation may allow
greater use of purchased inputs (rented draft animal for secondary tillage, small threshing
machines, rented irrigation pump, etc.), while transport of harvests may increase the demand for
However, these mechanisms may not always raise RTS if the primary function of
mechanical power by tractor is simply to substitute a relatively small fraction of overall labor
used. The rise in RTS is also limited if the benefit from timeliness is minimal due to either the
low influence on planting date (because of sufficient irrigation access, or climatic conditions),
8
the existence of alternative transportation methods, among others.
3. Empirical framework
RTS measures the elasticity of output increases with respect to increases in all inputs at a
fixed proportion. It is in one way measured as the sum of the output elasticities of all inputs in a
production function (Kislev and Peterson, 1996). The effect of HTS on RTS can be investigated
in the following framework. A farm household belongs to one of two production systems, (a)
with TS (denoted as regime 𝑅 = 1), and (b) without TS (𝑅 = 0). Output 𝑌 is realized in one of
𝑓1 (𝐾1 ; 𝐴) if 𝑅 = 1
𝑌={ } (1)
𝑓0 (𝐾0 ; 𝐴) if 𝑅 = 0
in which 𝑓𝑅 ’s are regime specific production functions that relate inputs / services vector 𝐾𝑅 ,
The farm household chooses 𝑅 and 𝐾𝑅 that maximize utility U that depends on the
agricultural profit 𝜋;
where 𝑐𝑅 (𝑤, 𝐾𝑅 ) is the cost of using 𝐾𝑅 given 𝑤, a vector of factors affecting inputs costs
(including opportunity costs of family labor) standardized by setting output price at 1, and a set
where 𝜂 is a set of factors affecting the liquidity of the households. Though our interest is in 𝑓𝑅 ,
9
above conditions also allow 𝑐𝑅 and 𝑔𝑅 to vary across R.
The optimization problems (2) through (4) lead to the Lagrangian function ℒ𝑅 ,
where 𝜆𝑅 is the Lagrange multiplier. Given (4) and the nonnegativity constraints of 𝐾𝑅 , the
optimal solutions 𝐾𝑅∗ (asterisks indicate the solution values hereafter) satisfy the following Kuhn-
Tucker conditions for each R; (i) 𝜕ℒ𝑅∗ /𝜕𝐾𝑅 ≤ 0, (ii) 𝐾𝑅∗ (𝜕ℒ𝑅∗ /𝜕𝐾𝑅 ) = 0, (iii) 𝐾𝑅∗ ≥ 0, (iv)
𝜕ℒ𝑅∗ /𝜕𝜆𝑅 ≥ 0, (v) 𝜆∗𝑅 (𝜕ℒ𝑅∗ /𝜕𝜆𝑅 ) = 0, and (vi) 𝜆∗𝑅 ≥ 0, where the negative sign in (i) holds if
𝐾𝑅∗ = 0. 𝐾𝑅∗ satisfies the above Kuhn-Tucker conditions for each R. The farmer chooses R* = 1 if
𝑐𝑅 and 𝑔𝑅 are also functions of exogenous variables 𝐴, 𝑤, 𝜂. Our empirical models therefore are
where r and k are functions. 𝑌𝑅∗ is related to 𝐾𝑅∗ through structural equation 𝑓𝑅 , so that,
Our empirical approach is to estimate 𝑓𝑅 and corresponding RTS, 𝜌𝑅 , for each R given
the observed 𝑌𝑅∗ , 𝑅 ∗ , 𝐾𝑅∗ , 𝐴, w and 𝜂. If 𝜌1 > 𝜌0 conditional on exogenous variables 𝐴, 𝑤, and 𝜂,
then we interpret that the HTS increases the RTS. However, if the RTS increases due to other
factors and HTS is simply a response to such an increase, then we expect to find 𝜌1 = 𝜌0
conditional on 𝐴, 𝑤, and 𝜂. While some studies assess the impact of R* on the production
10
function parameters by adding the interaction terms between R* and input variables (Wan and
Cheng, 2001), we depart from such approach to address potential endogeneity problems
associated with (6) through (8). First, a non-zero correlation between idiosyncratic error terms in
(6) and (8) causes the sample selection problem. Second, 𝐾𝑅∗ can be correlated with idiosyncratic
Several sample selection models address the first type of endogeneity, including
switching regression model (SR) (Goldfeld and Quandt, 1976; Maddala, 1983), and inverse
probability weighting model (IPW) (Horvitz and Thompson, 1952; Wooldridge, 2007). For
impact evaluation in the agricultural sector, IPW is applied to the estimation of the average
treatment effects (ATE) (for example Cunguara and Moder, 2011), or the effect on production
function (Cavatassi et al., 2011). However, production function in Cavatassi et al. (2011) treats
Generalized method of moments (GMM) (Hansen, 1982) has long been used to address
the second type of endogeneity issues. Compared to other instrumental variable (IV) approaches
like the two-stage least squares, GMM is more efficient in the presence of heteroskedasticity,
including serial correlation of error terms, which is common in agriculture due to locally
The literature addressing both types of endogeneity is relatively thin. Recent studies
modify the SR by incorporating the potential endogeneity in the second stage equations
(Bellemare and Barrett, 2006; Takeshima and Winter-Nelson, 2012). Their specifications
provide a tool to estimate the effect of R* on 𝜌𝑅 even when the estimation of 𝑓𝑅 involves
11
endogeneity. However, SR usually does not fully address the issue of the comparability between
treatment and control groups, in the same way as IPW does. Also, SR is susceptible to the
multicollinearity problem between Inverse Mills Ratio (IMR) and the other exogenous variables
(Puhani, 2000). Therefore, an additional empirical method needs to complement the SR.
(ignorability), GMM weighted by the inverse of the probability is consistent on the subsample of
responding units (Nicoletti, 2006). Recent studies extend this approach to obtain consistent
GMM estimates when samples are non-randomly selected (Abowd et al., 2001; Chen et al.,
2008). We contribute to the literature by exploiting the fact that IPW-GMM is a useful tool to
estimate the case like ours in which the effect of an endogenous treatment on endogenous
We estimate (6), (7) and (8) with IPW- and SR-GMM. For both methods, we first
estimated. Φ is the standard normal distribution function, while 𝜙 and 𝑣 are the standard normal
in which 𝛽’s are parameters to be estimated (with corresponding subscripts), and 𝜀 is the
idiosyncratic error terms, and i, j = chemical fertilizer in nutrient, family labor (adult male, adult
female, child), hired labor, land area cultivated, irrigation, agricultural capital and other cash
̂𝑅 = ∑𝑖(𝜕 ln 𝑌 /𝜕 ln 𝐾𝑖 ) = ∑𝑖 𝛽̂𝑖 +
expenses (which includes expenses on HTS). Then 𝜌
12
∑𝑖[(𝛽̂𝑖𝑖 + ∑𝑗≠𝑖 𝛽̂𝑖𝑗 ) ln 𝐾𝑖 ] (Kim, 1992).
There are trade-offs between these specifications. The Cobb-Douglas specification provides
reasonable estimates of production function parameters when samples are small and endogeneity
of multiple 𝐾𝑗 must be addressed, but it forces the elasticity of substitution between all 𝐾𝑗 to be
one, and forces the RTS to be the same across all agents. The Translog specification drops these
restrictions, but often complicates the endogeneity of multiple 𝐾𝑗 , and leads to inefficient
estimates if multicollinearity is severe (In our case, doing so in Translog specification involves
52 endogenous variables, and may lead to unreliable estimates.). This is especially true when, as
is described in the later section, the sample size is kept small to ensure comparability of TS-
hirers and nonhirers. We therefore estimate both specifications (Translog forms are estimated
observations in the Translog specification. In the Translog specification, we focus on the IPW
sample medians of 𝜌
̂,
𝑅 as they are likely to be both representative and robust against potential
outliers in 𝜌
̂.
𝑅 In the results section, we show that the evidence is robust to the choice of
specification.
The estimation approaches for (10) differ between IPW-GMM and SR-GMM. IPW-
13
GMM. Weights are the inverse of “square root” of the propensity scores, because the weights
apply to the objective function that is a square of the moment conditions. 𝑚(⋅) is the moment
condition,
IPW-GMM is “doubly-robust” (Robins and Rotnitzky, 1995); the overall model is consistent as
long as either the model of the propensity score 𝑝̂ , or the model of the production function is
In SR-GMM, we first construct IMR for TS-hirers (𝜔1 = 𝜙(𝑍𝛾̂)/Φ(𝑍𝛾̂)) and non-hirers
(𝜔0 = −𝜙(𝑍𝛾̂)/[1 − Φ(𝑍𝛾̂)]) from (9). We then estimate 𝑓𝑅 for TS-hirers and non-hirers using
GMM, removing √𝑝̂ and √1 − 𝑝̂ from (11) and (12), while adding 𝜔
̂𝑅 to A in (13).
While some studies using SR jointly estimate the selection equation and the main
equation, we estimate them in two steps because the joint estimation is more complicated due to
the endogeneity in the main equation. In addition, unlike the joint estimation method, our two-
step estimation does not require the assumption that idiosyncratic errors in the main equation are
normally identically distributed. This is appropriate when idiosyncratic errors are serially
correlated within the locality or heteroskedastic, as is often the case in the agricultural sector.
Our analyses are based on the cross-section data as described in the next section.
Although panel samples can be constructed with earlier rounds of the survey, the sample size is
small, particularly after extracting comparable samples among TS-hirers and non-hirers, and
preventing us from getting reliable results. In addition, various scholars note that the RTS is a
long-run rather than a short-run concept, and cross-section data (with methods addressing
14
endogeneity) may be more appropriate than fixed effects and first difference models which are
more suitable for capturing short-run effects (Basu, 2008; Sheng et al., 2015; Gollin et al., 2016).
Arguably the primary difference between cross-sectional specification and typical panel
specification with fixed effects is that the former focuses on the environment where various
factors that do not change in the short-run (for example, land endowments, demographics as well
as unobserved individual specific effects including ability or management skills) change over
time, and considers each cross-sectional household as a hypothetical state which a certain
household can evolve to in the long-run. While the virtue of a panel data specification is to
isolate unobserved fixed effects that are correlated with exogenous variables, and the absence of
such mechanism in cross sectional specification leads to potential endogeneity, our specifications
Standard errors for our estimators are adjusted accounting for the fact that 𝑝̂ is an
estimated variable. Unlike standard SR or IPW, our specification involves endogenous variables
in the second stage regressions. To our knowledge, the formulas of standard errors for this type
of estimators have not been developed in the literature. We therefore use paired bootstrap (Efron,
1979); we repeat 100 times the entire estimation procedures using randomly resampled data, and
use the standard deviations of bootstrapped coefficients as finite estimates of standard errors for
each parameter. Resampling is conducted treating the enumeration areas (EAs), described in the
next section, as clusters so that the estimated standard errors reflect the potential serial
The NLSS 2010 were collected through two-stage stratified random sampling methods
15
(Nepal CBS, 2011). First, the country was divided into 14 strata across three agro-ecological
belts, Terai, Hills and Mountains. They consist of 3 urban strata (urban areas in the Terai, Hills
outside Kathmandu valley, and Kathmandu Valley), 10 rural strata from Terai and Hills (5
developing regions each; Eastern, Central, Western, Mid-western and Far-western), and
Mountains. Second, a total of 500 EAs were randomly selected, with a pre-determined number of
EAs within each stratum (Nepal CBS, 2011, Box 1.2). Then, 12 households were randomly
selected from each EA. In addition, NLSS 2010 randomly surveyed 12 households each from
100 EAs randomly selected from the second round of NLSS (NLSS, 2003). In total, the NLSS
2010 interviewed 7200 households (Nepal CBS, 2011, Box 1.3). 2856 are the primary sample for
the Terai, among which 1965 are farm households, reporting the cultivation of some plots. As is
described in 4.2, we focus on the subset 600 of this sample that satisfies good overlap properties,
and later check the robustness of results against this subsampling. Other sources of data used to
4.1. Variables
Variables used as determinants of HTS, output and inputs in the agricultural production
function, and IVs for endogenous inputs variables, are selected and constructed based on the
relevant literature. These variables capture technological, economic, and biophysical resource
endowment of the households. Due to the space limitation, definitions and how these variables
are constructed are provided in the Appendix A. All monetary values are deflated by average
local prices of rice and wheat, which are the most commonly consumed staples in Nepal, to
control for the potential variations in price index across space. Throughout the paper, when
variables are transformed into the natural log form, the values of 0.0001 were added so that zero
16
values can also be transformed into the natural log. We also estimate the models using 0.01
instead of 0.0001, and find that the results are robust. Similar approaches are used in various
studies (see Jacoby, 1993, and Michalopoulos and Papaioannou, 2014, for example).
important for IPW (Busso et al., 2014). Overlap properties are poor when the covariate
distributions are substantially different between the two groups (Imbens and Wooldridge, 2009).
We therefore select the subset of samples based on several exogenous criteria. First, we limit the
samples to Terai, and exclude those in the Hills or Mountains where agro-ecological
characteristics are very different. Samples are also limited to the households that own draft
animals (bullock, cow, buffalo, horse, donkey, mule), because owning draft animals may
significantly affect the use of animal traction that can substitute the HTS. As was mentioned
We also limit the samples to the village district committees (VDCs), Nepal’s
administrative units below districts, where at least 20% of the samples are TS-hirers, and at least
10% are nonhirers, so that 𝑝̂ may be distributed away from 0 and 1, a key overlap condition
(Busso et al., 2014). We also exclude households whose heads have at least 12 years of formal
education, are located in areas with terrain ruggedness index greater than 300, without assets
(other than farmland, agricultural capital, livestock), or with such assets worth more than 30,000
kg of cereals evaluated at local cereal prices. After these procedures as well as dropping
observations with missing values, the total samples were reduced to 600.
17
We find that main results are robust when the analyses include all samples. However, those
samples lead to poorer overlap properties, indicating that our primary sample leads to more
As was mentioned earlier, we also exclude tractor owners from the analysis. Tractor-
owner farm households in our sample differ considerably from non-owner farm households,
particularly in financial endowments (Takeshima et al., 2015b). Combining them with non-
owner farm households in estimating production function imposes strong assumptions that
tractor-owner farm households have the same production function structure as non-owner farm
households. Tractor owners are also few in our sample relative to nonowner farm households,
preventing us from testing a separate hypothesis that tractor ownership affects the RTS.
Table A3 in online Appendix A summarizes the mean values of each variable used in the
analysis. It shows the descriptive statistics with and without weights which are estimated in the
next section, because it demonstrates how the IPW process improves the matching properties of
the samples. While sample means of exogenous variables are relatively similar across TS-hirers
and non-hirers without weights, they are still statistically significantly different across groups for
several variables. However, with weights, the differences become statistically insignificant for
most exogenous variables. Meanwhile, mean values of endogenous variables are still statistically
The IPW procedure therefore effectively reduces the differences between two groups in
our sample, so that any differences in production behaviors are more reliably attributed to HTS
rather than the differences in other observed characteristics. Descriptive statistics in Table A3
18
therefore further motivate our empirical analyses.
5. Empirical results
We first briefly discuss the determinants of the propensity of HTS, 𝑝̂ . We then discuss
more in detail the estimated production functions, RTS, and robustness of results under various
alternative specifications.
Appendix B. It is important to note that Probit results in Table B1 are applicable only within a
narrowly defined population as described in section 4.2, and should not be generalized to the
broader farm population of Nepal. For example, 𝑝̂ is generally unaffected by factors like more
widespread presence of tractor owners within VDCs in the sample, or the wages. This is possibly
because our sample is limited to the Terai with relatively well integrated labor market (Maystadt
et al., 2014) that reduces wage variations across locations, and to VDCs with at least 20% of the
𝑝̂ is higher if the household is located in areas with the smoother terrain, has a more
educated household head. 𝑝̂ is higher for households with fewer male members, possibly
reflecting the demand for TS to substitute family labor shortage. 𝑝̂ is higher the closer the nearest
ARS is, possibly reflecting the positive roles of agricultural extension or improved varieties. 𝑝̂ is
higher if DAP fertilizer price is higher, possibly indicating the substitution effect between
mechanization and more land saving factors like fertilizer. 𝑝̂ is also higher in areas more distant
from the nearest market center. This might be because these areas may be more sparsely
19
populated with greater land-to-labor ratios. For example, in Ghana, Benin (2015) shows that the
potential for mechanization expansion can be greater in those areas if tractor service providers
are present.
The distributions of 𝑝̂ are illustrated in the left panel of Figure 1. For comparison, the
right panel of Figure 1 illustrates 𝑝̂ obtained from the full sample. Compared to the right panel in
which a substantial mass of propensity scores are distributed around 0 or 1, the two samples of
TS-hirers and non-hirers in the left panel have a fairly good overlap without concentrations of 𝑝̂
around 0 or 1. Such good overlap can ensure IPW-estimators to be reliable (Busso et al., 2014).
differentiated by the TS-hiring status, and specifications used (due to space limitations,
coefficients on other variables are not shown). Hausman test (1978) rejects the null hypothesis
that input variables are jointly exogenous, justifying our choice of GMM over OLS. All models
̂),
The estimated RTSs (𝜌𝑅 our main interests, are statistically significantly higher among
TS-hirers (around 0.9) than among nonhirers (around 0.6 ~ 0.7), consistent with the hypothesis
that HTS is causing an increase in RTS. The difference in RTS may be driven by the output
elasticity (the estimated coefficient) of area cultivated, and to a lesser extent that of agricultural
20
capital, while that of labor is generally higher among non-hirers (albeit marginally). While
endogeneity is suggested, estimated patterns are qualitatively similar between OLS and GMM. In
the SR-GMM, the coefficients on IMR are statistically insignificant. Also, the results between
GMM and IPW-GMM are generally fairly similar. Endogeneity due to self-selection may be
therefore relatively limited in our sample. In any case, the evidence that HTS is causing an
The relative magnitudes of estimated output elasticities are generally consistent with the
previous studies. For both TS-hirers and nonhirers, the area cultivated exhibits the highest output
elasticities among all factors, consistent with previous studies on Nepal (Abdulai and Regmi,
2000; Devkota and Upadhyay, 2013). Relatively high output elasticities for agricultural capital
are consistent as well with Devkota and Upadhyay (2013). Low or slightly negative output
elasticity of labor tends to be observed in environments with surplus labor (Wan and Cheng,
2001), which might apply to Terai of Nepal, experiencing excessive labor in-migrations in recent
years (Maystadt et al., 2014). This is contrasting to estimates in Africa (for example, Adesina
The implications on RTS are similar in Translog specifications. Table 2 presents the
estimated RTS evaluated at the IPW-sample medians, while estimated individual coefficients are
shown in the Appendix B (Tables B2 through B7). First, full Translog specifications lead to
magnitudes of RTS generally similar to the Cobb-Douglas specification (0.9 ~ 1.0 for TS hirers
and 0.7 for nonhirers) although, due to higher standard errors, the difference in RTS is somewhat
imprecise with p-values of 0.13 and 0.154 (insignificant at 10% level). To obtain more precise
results, we use the findings that all interactive terms containing family labor are jointly
insignificant for TS-hirers (p-value = .614 and .722). We re-estimate for TS-hirers a restricted
21
Translog function in which all interactive terms containing family labor are excluded. This
reduces the standard errors of estimated RTS (from .139 to .085 in SR-GMM and .130 to .083 in
IPW-GMM). The difference in RTS is statistically significant at 10% (p-value = .075 and .097)
The F-test rejects the Cobb-Douglas specification in favor of the Translog specification
(p-value = .000). However, this is only suggestive because the exact test must take into account
the endogeneity of input variables, which cannot be feasibly incorporated in the Translog
specification as discussed above. The implication is that the effects on RTS may vary across
households, and future studies need to examine such heterogeneity more closely.
We further conducted robustness checks (Appendix C provides the detail). Our results are
found unlikely to be driven by (a) price index used to convert all monetary terms into real terms;
(b) samples constructed; (c) the distinction of permanently hired labor from temporarily hired
labor; nor (d) other potential sources like specialization or differential access to improved
varieties. Our results are also unlikely to be driven by the effects of other endogenous inputs uses
that our results are likely to be robust against the small changes in 𝑝̂ due to the violation of
22
ignorability assumption. However, to our knowledge, no formal procedure has been developed to
check the robustness of IPW-GMM estimates against the violation of ignorability assumption.
6. Conclusions
transformation around the world has often accompanied the increase in the RTS, and may also be
importance, literature is thin on the direct causes of changes in agricultural RTS. We attempted
to fill this knowledge gap. We found that HTS significantly increased the RTS among some
Nepal Terai farm households. We did so by jointly addressing two forms of endogeneity; the
endogeneity in decisions to hire in tractor services, and in decisions on inputs use in production
functions.
particularly mechanization supplied through custom-hiring. Also, it raises some important new
questions for the future studies regarding the technology adoption, such as the choice of
technologies in the face of not only the relative marginal profitability, but also the differing
understand better the implications of the findings. Promoting HTS may shift the comparative
advantage in agricultural production to more land and capital-endowed households, rather than
less-endowed households. Further research should be conducted to assess how these less land
and capital-endowed households are affected by the mechanization growths through own HTS
23
and HTS by more land and capital-endowed households.
Lastly, there are also limitations to our analysis. The chief limitation of our analysis is
that, due to the data limitation, it does not investigate the effect of tractor ownership and their
service provisions on the RTS, which is potentially one of the substantial sources of the growth
of RTS in agriculture. Such data limitation arises because tractor-owning farm households,
despite such potential importance, tend to account for a small share of the population and their
sample sizes tend to remain too small in nationally representative household surveys like NLSS.
Future studies should investigate such effect of tractor ownership on the RTS, using specially
designed surveys which properly stratify the sample framework so that information is collected
for a sufficient number of tractor owning farm households with characteristics that are
comparable to those of nonowners, as well as panel data collected over sufficiently long time-
span of new tractor owners, which can offer critical insights into the process of the change in
RTS. With such data, our methodology can be extended to investigate the effect of tractor
Similarly, our results apply only to a segment of tractor service-hirers whose suitable
counterfactuals can be found among the nonhirers. Our findings may not be readily extended to
households whose characteristics differ considerably from those studied here. In addition, we can
conduct only informal analysis on the robustness of the results against the possible violation of
ignorability assumption inherent in IPW-GMM. Future studies need to address these issues to
24
References
Abdulai, A., Regmi, P., 2000. Estimating labor supply of farm households under nonseparability:
Abowd, J., Crépon, B., Kramarz, F., 2001. Moment estimation with attrition: an application to
Adesina, A., Djato, K., 1997. Relative efficiency of women as farm managers: Profit function
Basu, S., 2008. Returns to scale measurement, in S. Durlauf and L. Blume, eds., The New
Palgrave Dictionary of Economics, 2nd edition. The Palgrave Macmillan. New York.
Bellemare, M., Barrett, C., 2006. An Ordered Tobit Model of Market Participation: Evidence
Benin, S., 2015. Impact of Ghana’s agricultural mechanization services center program. Agric.
Benin, S., Johnson, M., Abokyi, E., Ahorbo, G., Jimah, K., Nasser, G., Owusu, V., Taabazuing,
J., Tenga, A., 2013. Revisiting agricultural input and farm support subsidies in Africa:
The case of Ghana's mechanization, fertilizer, block farms, and marketing programs.
Busso, M., DiNardo, J., McCrary, J., 2014. New evidence on the finite sample properties of
propensity score reweighting and matching estimators. Rev. Econ. Stat. 96(5), 885–897.
Cavatassi, R., Salazar, L., González‐Flores, M., Winters, P., 2011. How do agricultural
programmes alter crop production? Evidence from Ecuador. J. Agr. Econ. 62(2), 403–
25
428.
Chen, X., Hong, H., Tarozz, A., 2008. Semiparametric efficiency in GMM models with auxiliary
Christiaensen, L., 2013. When China runs out of farmers, in C. Barrett, eds., Food Security and
Craig, B., Pardey, P., Roseboom, J., 1997. International productivity patterns: accounting for
input quality, infrastructure, and research. Amer. J. Agr. Econ. 79(4), 1064–1076.
Cramb, R., 2011. Re-inventing dualism: policy narratives and modes of oil palm expansion in
Cunguara, B., Moder, K., 2011. Is Agricultural Extension Helping the Poor? Evidence from
Deininger, K., Byerlee, D., 2012. The Rise of Large Farms in Land Abundant Countries: Do
Devkota, S., Upadhyay, M., 2013. Agricultural Productivity and Poverty Reduction in Nepal.
Diao, X., Cossar, F., Houssou, N., Kolavalli, S., 2014. Mechanization in Ghana: Emerging
demand, and the search for alternative supply models. Food Policy 48, 168–181.
Efron, B., 1979. Bootstrap methods: another look at the jackknife. Ann. Stat. 7, 101–118.
Foster, A., Rosenzweig, M., 2011. Are Indian farms too small? Mechanization, agency costs, and
farm efficiency. Economic Growth Center, Yale University, New Haven CT. Accessed
Seminars/Development/rosenzweig-111114.pdf.
Gollin, D., Jedwab, R., Vollrath, D., 2016. Urbanization with and without Industrialization. J.
26
Econ. Growth 21(1), 35–70.
Goldfeld, S., Quandt, R., 1976. Studies in nonlinear estimation. Ballinger. Cambridge, MA.
Haggblade, S., 2005. From roller coasters to rocket ships: The role of technology in African
Hansen, L., 1982. Large sample properties of generalized method of moments estimators.
Hayami, Y., Kawagoe, T., 1989. Farm mechanization, scale economies and polarization: The
Hayami, Y., Ruttan, V., 1985. Agricultural development: An international perspective. John
Horvitz, D., Thompson, D., 1952. A generalization of sampling without replacement from a
Houssou, N., Chapoto, A., 2014. The Changing Landscape of Agriculture in Ghana: Drivers of
Farm Mechanization and Its Impacts on Cropland Expansion and Intensification. IFPRI
Imbens, G., Woolridge, J., 2009. Recent developments in the econometrics of program
Jacoby, H., 1993. Shadow wages and peasant Family Labour Supply: An Econometric
Kim, H. Y., 1992. The translog production function and variable returns to scale. Rev. Econ.
27
Kislev, Y., Peterson, W., 1996. Economies of scale in agriculture: A reexamination of the
Kleibergen, F., Paap, R., 2006. Generalized reduced rank tests using the singular value
Maystadt, J., Mueller, V., Sebastian, A., 2014. Environmental Migration and Labor Markets in
Michalopoulos, S., Papaioannou, E., 2014. National Institutions and Subnational Development in
Nepal CBS (Central Bureau of Statistics), 2011. Nepal Living Standards Survey 2011/11:
Nicoletti, C., 2006. Nonresponse in dynamic panel data models. J. Econometrics 132(2), 461–
489.
Otsuka, K., Liu, Y., Yamauchi, F., 2013. Factor Endowments, Wage Growth, and Changing
Pingali, P., 2007. Agricultural mechanization: adoption patterns and economic impact, in R.
Elsevier.
Puhani, P., 2000. The Heckman correction for sample selection and its critique. J. Econ. Surveys
14, 53–68.
28
Robins, J., Rotnitzky, A., 1995. Semiparametric Efficiency in Multivariate Regression Models
Sheng, Y., Zhao, S., Nossal, K., Zhang, D., 2015. Productivity and farm size in Australian
agriculture: reinvestigating the returns to scale. Aust. J. Agr. Resour. Ec. 59(1), 16–38.
Takeshima, H., Winter-Nelson, A., 2012. Sales location among semi-subsistence cassava farmers
Takeshima, H., Edeh, E., Lawal, A., Isiaka, M., 2015a. Characteristics of private-sector tractor
Takeshima, H., Adhikari, R., Poudel, M., Kumar, A., 2015b. Farm Household Typologies and
Washington DC.
Wan, G., Cheng, E., 2001. Effects of land fragmentation and returns to scale in the Chinese
Wooldridge, J., 2007. Inverse probability weighted estimation for general missing data problems.
Yamauchi, F., 2016. Rising real wages, mechanization and growing advantage of large farms:
29
Table 1. Production functions by tractor hiring status among non-tractor-owners under various econometric methods
OLS
TS- Non- TS- Non- TS- Non- TS- Non- TS- Non-
hirers hirers hirers hirers hirers hirers hirers hirers hirers hirers
ln (Fertilizer nutrient)b -.009 -.016 -.010 -.018 .002 -.016 .000 -.014 .013 -.016
(.017) (.014) (.024) (.014) (.027) (.016) (.046) (.026) (.043) (.024)
ln (adult male family labor)b -.006 .014* -.006 .013 -.003 .003 -.002 .012 -.004 .004
(.007) (.008) (.009) (.009) (.005) (.009) (.011) (.014) (.011) (.013)
ln (adult female family labor)b -.026** .024*** -.028** .026*** -.015 .060*** -.014 .062*** -.023 .063***
(.011) (.008) (.014) (.010) (.013) (.007) (.025) (.018) (.024) (.017)
ln (child family labor)b .007 .001 .007 .000 .019*** .019*** .019* .018* .016* .019**
(.006) (.006) (.007) (.007) (.006) (.006) (.010) (.010) (.010) (.009)
ln (hired labor)b -.007 .001 -.008 .000 -.010 -.015* -.010 -.020 -.012 -.014
(.006) (.008) (.008) (.009) (.009) (.009) (.014) (.019) (.012) (.019)
30
ln (area cultivated)b .450*** .238*** .463*** .231*** .529*** .397*** .519*** .378*** .564*** .382***
(.064) (.065) (.077) (.077) (.058) (.061) (.132) (.117) (.125) (.116)
ln (irrigation (= 1 if used .006 .029*** .005 .028*** .009 .016 .010 .015 .002 .015
irrigation))b (.007) (.008) (.008) (.009) (.008) (.012) (.015) (.020) (.015) (.019)
ln (agricultural capital) .309*** .262*** .299*** .298*** .291*** .161** .281*** .123 .305*** .168*
(.059) (.075) (.060) (.074) (.055) (.064) (.078) (.094) (.082) (.092)
ln (other cash expenditure) .164*** .045** .157** .042* .092 .084*** .095 .097** .071 .091**
(.053) (.020) (.067) (.024) (.064) (.024) (.125) (.040) (.119) (.039)
(.223) (.341)
̂)
Estimated RTS (𝜌𝑅 .888*** .597*** .878*** .620*** .914*** .709*** .898*** .671*** .933*** .712***
(.065) (.069) (.073) (.071) (.056) (.061) (.097) (.098) (.094) (.093)
p-value
H0 = covariates jointly .000 .000 .000 .000 .000 .000 .000 .000 .000 .000
insignificant
31
H0 = weakly identifiedd .000 .000 .000 .000 .000 .000
Number of observations 291 309 291 309 291 309 291 309 291 309
Source: Author. Asterisks indicate the statistical significance; *** 1%, ** 5%, * 10%.
described in Appendix A, and intercept. bPotentially endogenous variables. cBased on Hausman test (Hausman, 1978). dBased on
Kleibergen and Paap (2006) rk Wald statistics. eBased on Hansen J-statistics (Hansen, 1982).
32
Table 2. Robustness check on the differences in RTS based on Translog production
SR-GMM IPW-GMM
translogb translogb
a
̂)
Estimated RTS (𝜌𝑅 .991 .977* .702 .998 .979* .723
interaction variables
jointly insignificant)
p-value: (H0: 𝜌
̂𝑅 not .106 .077 .130 .075
nonhireres)
Source: Author. Asterisks indicate the statistically significant difference from the estimate
33
2.5
2
2
1.5
1.5
Density
Density
1
1
.5
.5
0
0 .2 .4 .6 .8 1 0 0 .2 .4 .6 .8 1
propensity score propensity score
Figure 1. Kernel density estimations of propensity scores 𝑝̂ (left = matched samples; right = raw
samples)
Source: Author.
34