Abstract
A unique GIS data set from Indonesia that distinguishes smallholder and plantation operations is used to test the impact of district subdivision, which enhances local control of natural-resource revenues, on the location of smallholder forest conversion. Nonparametric analyses show that in subdivided districts, smallholders convert forests on steeper land that is farther from the nearest road and deeper into the forest. Smallholders are also less responsive to forest protection in subdivided districts. District subdivision imposes external costs of $1,629 to $4,941 per hectare due to increased carbon emissions associated with smallholder conversion deeper into the forest. (JEL C14, Q15)
1. Introduction
Reducing greenhouse gas emissions from tropical forest conversion is viewed as a critical component of plans to mitigate the effects of anthropogenic climate change. Emissions from forestry and other land use represented 10% of total anthropogenic greenhouse gas emissions in 2010 (Pachauri and Meyer 2014), a lower percentage than suggested in earlier reports. This reduced percentage is partly due to increased emissions from the fossil fuels sector but is also a result of more precise, spatially explicit estimates of the aboveground forest carbon stock (Baccini et al. 2012). Given this landscape heterogeneity, which affects the net benefits of forest conversion, effective forest management demands an understanding of the factors that determine the location of tropical forest conversion.
The management of tropical forests in Indonesia is of global concern, as this nation is in the top three in terms of extant stands of tropical forest and has forests that are home to countless endemic species. Rapid conversion to production systems has made Indonesia the top-ranked nation for deforestation (Margono et al. 2014) and consequently among the highest greenhouse gas emitters.
A key change in Indonesia with implications for forest conversion was the decentralization that occurred in the wake of Suharto’s 1998 resignation following the 1997 Asian financial crisis. To keep the nation intact, the post-Suharto regime decentralized the management of natural resources (Potter and Badcock 2001). As a result, many districts subdivided, ostensibly to secure as much control of their endowed natural resources as possible (Burgess et al. 2012).
Previous work has shown that this wave of district subdivision in Indonesia led to more forest conversion and plantation activity. Burgess et al. (2012) find that the increased number of political jurisdictions increased deforestation due to logging and reduced timber prices across Indonesia, consistent with districts engaged in Cournot competition as firms choose where to log. Suwarno, Hein, and Sumarga (2015) use surveys of government officials and stakeholders to evaluate forest governance quality. The authors find that deforestation has increased in Central Kalimantan as a result of district subdivision, because the quality of forest governance declined with district subdivision.
Decentralization of natural resource management can empower and democratize local communities and reduce poverty through more equitable access to resources. However, previous studies have shown that decentralization can also lead to further marginalization of poor and disadvantaged groups, with resource control accruing to the most powerful stakeholders (e.g., Larson and Soto 2008; Ribot, Agrawal, and Larson 2006). In Indonesia, district agencies grant concessions to plantation operators that convey secure property rights to the plantations and collect royalties from plantation operations. Smallholder production systems are not taxed and are not formally recognized by these agencies, meaning that smallholder tenure is insecure in the presence of plantations. The increased level of plantation activity in subdivided districts is expected to impact smallholder forest conversion due to the tensions between these two forces of forest conversion in Indonesia (see, e.g., Colchester et al. 2006; McCarthy 2004; Suyanto 2007; Suyanto et al. 2004).
To better understand how district subdivision affects the location of smallholder conversion, we explore the proximate determinants of smallholder forest conversion in subdivided and nonsubdivided districts between 2000 and 2008, using a unique GIS data set of 2008 land use and land cover that distinguishes between smallholder and plantation production at a fine spatial scale. We estimate these determinants on a randomly drawn set of pixels, as well as on a set of pixels that lies within 5 km of a district with the other subdivision status to ensure that the observed differences in the likelihood of smallholder conversion are due to district subdivision, and use a group unconfoundedness test to address endogeneity concerns.
We find that smallholders in nonsubdivided districts have a revealed preference for lowslope parcels that are near the edge of the forest and close to the nearest road. In subdivided districts, the location of smallholder production is not affected by slope or distance to the nearest road, and there is only a small reduction in conversion probability associated with increasing distance from the forest edge. These results suggest that the location of smallholder forest conversion, and therefore its private returns and external costs, is affected by district subdivision. We find that district subdivision leads to additional costs to society, stemming from increased carbon emissions due to the location of smallholder conversion farther into the forest in subdivided districts, with a present value of these external costs ranging from $1,629 to $4,941 per hectare, exceeding the range of net returns to smallholder palm oil production, the most valuable smallholder production system in the area (Grieg-Gran 2008).
Our findings suggest that policy action to encourage efficient forest management in Sumatra must address both the location and the amount of forest conversion. Doing so requires a comprehensive approach that acknowledges both smallholders and plantations as forces of forest conversion, unlike the two-year moratorium on plantation concessions enacted in 2011. We show that district subdivision, with a focus on the generation of natural-resource royalties, increases the external costs of smallholder forest conversion as it moves farther into the forest in response to district subdivision, suggesting that decentralization may have increased the challenges to effective forest management in Indonesia.
2. Study Area Background
The centralized distribution of locally generated tax revenues under President Suharto’s New Order regime was not popular among officials from particularly resource-rich regions of the nation, who felt that private-sector corporations with ties to decision-makers in Jakarta were too frequently granted access to these resources (Potter and Badcock 2001). Suharto was forced to resign his presidency due to fallout from the 1997 Asian financial crisis, which sent the rupiah into a major depreciation that lasted through 1998. During this time, there were calls by resource-rich provinces for independence, and Indonesia underwent an extensive, and immediate, wave of decentralization under the new regime to appease calls for more regional autonomy, while maintaining the extent of the nation (Potter and Badcock 2001). Several pieces of legislation, aiming to increase the generation of district royalties through increased natural-resource use, including increased issuance of concessions to plantation operators, facilitated the transfer of authority to provincial and district governments, allowing resource-rich regions to retain a greater share of the revenues generated within their jurisdictions (Potter and Badcock 2001).
In the wake of the national legislation supporting local oversight of natural resources, many districts chose to subdivide. By altering district borders, officials in district agencies would give themselves greater control over the natural resources found within their districts. Between 1998 and 2008, the total number of districts in Indonesia increased from 292 to 483, as local governments attempted to exert maximum control over the natural resource royalties originating in their district (Burgess et al. 2012).
The provinces of Riau, Jambi, and West Sumatra on the island of Sumatra have experienced substantial forest conversion since the mid-1990s, chiefly as a result of the rising price of palm oil that has made oil palm plantations quite profitable in the region (Figure 1). These provinces cover an area of approximately 17 million ha. From 1990 to 2008, 7 million ha of forest, or 58% of the total forested area in 1990, in these provinces were converted to production systems. While industrial plantations are responsible for the majority of forest conversion in this area, smallholder operations have been responsible for approximately 30% of forest conversion since 1990.
Study Area
Tensions exist between these two drivers of forest conversion because smallholders have only customary cultivation rights to forested parcels on the landscape, while government agencies offer concessions to plantation corporations to generate royalties (Potter and Badcock 2001). Smallholder production systems are frequently interrupted by corporations granted forest concessions by government agencies for plantation development (Suyanto et al. 2004), resulting in either incorporation of the communal producers into the corporate production system, or the forced, uncompensated relocation of smallholder production to lands not covered by the corporate lease rights.1 These incorporation agreements are perceived to be unfair, with promised payments occasionally undelivered and land-grabbing by plantations a frequent occurrence (Colchester et al. 2006). So, local smallholders rarely participate in these arrangements, and most plantations, including those in the nucleus system, in which smallholders effectively act as sharecroppers, rely on labor imported from other islands (e.g., Java) (McCarthy 2007). Smallholder production is occasionally cleared by plantations through the use of fire, even when the land was not included in the formal concession to the plantation operator (Suyanto 2007).
3. Model of Smallholder Location Choice
We present a single-period model in which a smallholder chooses where to locate a production system of fixed size to maximize the expected returns to conversion. The maximization of expected returns is a commonly applied framework in microeconomic models of land use conversion (e.g., Bockstael 1996; Irwin and Geoghegan 2001; Pfaff 1999).2 Let d measure the distance from the forest edge at which conversion occurs. The net returns to production conditional on maintaining control of the production system are given by π(d)=P×Y(d) – C(d), where P represents the market price for the good produced, Y(d) represents the yield of the product, and C(d) represents the cost of conversion and production, which is expected to increase at an increasing rate with distance from the forest edge (C′(d), C″(d) ≥ 0, C′(0) =0).3
Let S represent a district’s subdivision status, with S = 1 indicating a district that has been subdivided, and S = 0 indicating a district that has not. Let L(d, S) represent the probability that a forest patch is converted by a plantation. Assume that the probability of plantation interruption is decreasing in distance from the forest edge for the same reason that smallholder production costs increase with distance from the forest edge. Further, assume that L(0, S = 1) > L(0, S = 0) and that L′(d, 0) ≤ L′(d, 1)≤0 ∀ d, meaning that subdivision increases plantation activity at all distances from the forest edge. Such an outcome might obtain if subdivided districts attract plantations hoping to generate royalties from concessions and competition drives plantations farther into the forest in these districts. While this assumption is supported by the findings in previous studies that deforestation increased following district subdivision (e.g., Burgess et al. 2012; Suwarno, Hein, and Sumarga 2015), we test the validity of this assumption in our study area by examining the percentage of standing forest in 2000 included in plantation concessions following district subdivision.
To incorporate the probability of losing control of land to a plantation in the smallholder’s problem, consider the following progression of events. At the start of the period, the smallholder makes a choice about the distance from the forest edge at which to convert the forest to production. Before the end of the period, smallholder production is either interrupted by plantation activity or allowed to continue. At the end of the period, the smallholder receives the returns from production if her operation is not interrupted by plantations. If smallholder production is interrupted by plantation operations, then the smallholder receives compensation, if any, from the plantation manager at the end of the period.4 Then, the smallholder’s problem is given by
[1]
where W(d) represents the payment that the smallholder receives from the plantation operator for taking her land, which again decreases with increasing forest edge distance (i.e., W′(d) ≤ 0, W″(d) ≥ 0).5
Recall that L(0,S = 1) > L(0,S = 0) by assumption. It follows directly that the expected value of smallholder conversion at the forest edge in nonsubdivided districts is greater than that in subdivided districts. Namely,
Having seen that district subdivision lowers the value of conversion at the forest edge, we now turn to the optimal location of smallholder forest conversion.
From the first-order condition for the smallholder’s problem in [1], it is shown that the optimal location of smallholder forest conversion, d*, satisfies
[2]
The left-hand side of equation [2] represents the change in the expected net returns of smallholder production if she is able to maintain control of her production system with increasing forest edge distance, which is the marginal cost of moving production farther into the forest. The right-hand side of equation [2] represents the change in the expected net returns of smallholder production conditional on losing control of the land to the plantation operator, which is the marginal benefit of moving production deeper into the forest. Then, the optimal location of smallholder conversion satisfies
[3]
Note that
Also note that
which is strictly positive with insecure smallholder tenure.
Given these results, it is left to compare L(d*,0) and L(d*,1). Because L(0,0) < L(0,1) and L′(d,0)≤′(d,1) ∀ d, it follows from the above comparative statics that 
, where 
and 
are the optimal locations of conversion in nonsubdivided and subdivided districts, respectively.6
This model can offer insight into the differential impact of district subdivision across alternative smallholder production systems. Consider the difference between the returns to smallholder production, π(d), and payment offered by a plantation for the land, W(d), at a given location in the forest. From equation [3], it can be shown that the greater this difference, the farther into the forest the smallholder production system will move. Specifically, letting V(d) =π(d) – W(d), we see that
In the study area, where smallholder tenure is insecure, we expect V(d) to take on greater values for smallholder production systems that have higher returns (e.g., smallholder oil palm production). Then, high-return production systems will display a greater response to district subdivision, meaning a move farther into the forest interior in subdivided districts. This issue highlights the difference between plasma oil palm operations and smallholder systems at risk of expropriation (mixed agriculture, oil palm, and rubber).7 In this model, plasma operations, which are directly linked to plantation operations and face no risk of expropriation by other plantations, can be thought of as a system for which L(d) = 0 ∀d, meaning that the location of these operations should not be affected by district subdivision.
The above model of a single smallholder’s choice of conversion location can be linked to a discrete choice model in which the aggregate probability of forest conversion to a smallholder production system is a function of parcel characteristics (see, e.g., Bell and Irwin 2002; Bockstael 1996). Specifically, let Rjk represent the net expected return from converting parcel j from forest to smallholder production system k. Following [1] above, define Rjk as follows:
In practice, only some of the factors affecting the net expected returns of forest conversion are observed by the researcher, so we can consider that the net expected returns of conversion to system k include a random portion, η, which is unobserved by the researcher. Then, the probability that forest parcel j is converted to smallholder production system k is given by
where h indexes the alternative smallholder production systems, including leaving the forest unconverted.
In this framework, the above model results can be used to generate empirically testable predictions. The finding that the returns to conversion at the forest edge are greater in the nonsubdivided districts relative to the subdivided districts leads to the following prediction:
Prediction 1. The probability of smallholder conversion at the forest edge will be greater in nonsubdivided districts than in subdivided districts.
The finding that the optimal location of smallholder conversion in subdivided districts lies farther into the forest than the optimal location in nonsubdivided districts leads to the following prediction:
Prediction 2. At some positive distance into the forest, the probability of smallholder conversion in subdivided districts will be greater than that in nonsubdivided districts.
Finally, the result that the optimal location of smallholder conversion will lie farther into the forest as the returns to the smallholder production system increase leads to the following prediction:
Prediction 3. District subdivision will have the greatest impact on the location of smallholder forest conversion for high-return production systems (e.g., oil palm).
Section 5 describes in more detail the assumptions about the distribution of η and the relevant set of land uses made in the parametric and nonparametric models estimated to test the above predictions.8 Before turning to those descriptions, we introduce the data used to explore the impact of district subdivision on the location of smallholder forest conversion.
4. Data
This analysis explores how district subdivision affects whether forested land is converted to smallholder production systems. Satellite data provide the best opportunity to account for both legal and illegal forest conversion, which is necessary to obtain accurate estimates of conversion rates and the determinants of conversion (Burgess et al. 2012).
Data Sets
The key data for this analysis are maps of land use and land cover (LULC) from 2000 and 2008. The 2000 LULC map is based on Landsat imagery and identifies pixels as either forest or nonforest. The map of LULC in 2008 is derived from manual classification of Landsat and IRS-P6 imagery with validation through ground checks (Setiabudi 2008). The 2008 LULC map was commissioned by World Wildlife Fund Indonesia and has been used in recent policy and research efforts related to the Sumatran Tiger (e.g., Bhagabati et al. 2014). The combination of high-resolution satellite imagery and ground-truthing allows for detailed LULC classification in the 2008 map that includes identification of both smallholder and plantation operations.
The 2000 LULC map is used in conjunction with data on plantation concessions in 2006 and 2012 to explore whether district subdivision is associated with greater plantation activity. The percentage of extant forest in 2000 covered by plantation concessions within each district is compared across subdivided and nonsubdivided districts. Table 1 displays these mean values and the results of a Wilcoxon rank-sum test of their equality. The tests show that subdivision is correlated with greater plantation activity across both data sets, supporting results from the literature and the assumptions of the theoretical model.
Percentage of 2000 Forested Area Allocated to Plantations
The dependent variables in our analyses measure whether a pixel converts from forest to smallholder production between 2000 and 2008. The binary indicator variable takes on a value of one if the pixel is engaged in any of the following smallholder production systems in 2008: smallholder oil palm, smallholder rubber, or mixed agriculture. As mentioned above, the nucleus estate system of oil palm production essentially involves smallholders working as sharecroppers on land controlled by oil palm plantation operators (referred to as plasma production).
The presence of plasma smallholder operation could complicate our exploration of the effects of district subdivision on smallholder production if it were mistakenly included with other smallholder oil palm operations. Fortunately, the data include a separate category, smallholder oil palm plantation, that describes the plasma system. The main results of our binary analysis (smallholder or not) exclude this category of production from our classification of smallholder systems. We include this production as a separate category in our multinomial regressions, meaning that there are four categories of smallholder production (smallholder oil palm, smallholder rubber, mixed agriculture, and plasma oil palm).
Several different biophysical and infrastructural data layers were combined to generate a vector of pixel characteristics that would be expected to determine the probability of a pixel being converted from forest to smallholder production, including elevation, precipitation, slope, soil depth, distance to the forest edge, and distance to the nearest road, settlement, and town. The details of the development of these data layers are provided by Bhagabati et al. (2014).
Data Selection
The unit of observation in the following econometric analyses is a 30 m by 30 m pixel. The forest conversion decisions made on the Sumatran landscape by smallholders take place at different spatial scales than our unit of analysis. The data selection process and the choice of regression models were undertaken to ensure the independence of the outcome variable across pixels in the following parametric and nonparametric analyses.
We randomly drew 13,025 pixels from the set of pixels that were forested in 2000. Of the full set of pixels, 9,767 pixels lie within districts that were subdivided, while the remaining 3,258 pixels lie in districts whose borders were unchanged.9 Table 2 presents summary statistics of the physiogeographic characteristics of the 13,025 randomly drawn pixels.
Characteristics of Randomly Selected Pixels
The location of smallholder operations seems to vary based on district subdivision, with conversion of pixels farther into the forest and farther from the nearest settlement in subdivided districts. However, it also appears that smallholder production occurs on lower-elevation and lower-sloped pixels in subdivided districts. In general, the subdivided districts seem to have more favorable pixel characteristics (e.g., elevation, slope, nearest road and town distances) than the nonsubdivided districts. This outcome makes it less likely that we would observe smallholders in more remote locations in these districts simply due to increased competition with plantations, allaying endogeneity concerns. To explore the causal relationship between pixel characteristics, district subdivision, and smallholder conversion, we turn to the econometric analysis.
5. Econometric Methods
Chomitz and Gray (1996) provide the econometric template for most ensuing studies of land-cover change in the tropics (e.g., Nelson and Hellerstein 1997; Pfaff 1999). Our exploration of district subdivision and the location of smallholder production is guided by this existing work and also employs nonparametric estimation techniques, which allow for unrestricted interaction between explanatory variables and consider flexible functional forms. We use the nonparametric techniques to explore the reliability of commonly applied parametric models, including the spatial autoregressive linear probability model, in the context of forest conversion.
Latent Variable Motivation
Assume that the land rent on parcel j engaged in production system k, 
, is defined as the difference between the value of outputs and inputs, Qjk, and Yjk, at their respective location-specific prices, Pjk and Cjk:
[4]
Spatially disaggregated price and yield data, namely, P, C, and Q, are not available for every parcel j on the landscape. Assume that prices vary spatially as a function of a vector of parcel-specific variables, Zjk, related to the distance to markets, including distance to the nearest road or town. Further assume that parcel characteristics will impact the yield of production on parcel j, so let Mjk represent a vector of productivity shifters allowing for spatial heterogeneity in production, as it might be expected for crop yields to vary across the landscape based on parcel elevation and slope. To explore the impact of district subdivision, include a district-level indicator of subdivision, Sj. We include district indicator variables to absorb the group-level shocks induced by the inclusion of a variable measured at the district level and return ηjk to homoskedastic, idiosyncratic error terms. We add province indicator variables to control for district- and province-level characteristics that have been explored as determinants of subdivision and deforestation in previous studies, including, notably, corruption (Burgess et al. 2012; Suwarno, Hein, and Sumarga 2015). We describe additional steps taken to address potential endogeneity of district subdivision in the next subsection. Standard errors are clustered at a level of 2 km by 2 km cells to acknowledge the difference in scale between pixels and smallholder conversion decisions.10 Then, the rent on parcel j engaged in smallholder production system k, 
, becomes
[5]
where Dj is a vector of district indicator variables and Pj is a vector of province indicator variables.
Motivated by the predictions from the theoretical model regarding different locations of smallholder conversion across district types, we can split the sample of forested parcels into those located in subdivided districts and those located in nonsubdivided districts. This approach leads to
[6]
where Sj is an indicator function taking on a value of one if parcel j lies in a district that was subdivided following Suharto’s resignation and is zero otherwise.
The choice of production system is described using logit and multinomial logit models, in which the observed outcome, smallholder conversion (Rjk), takes on values based on the latent variable’s magnitude relative to a threshold value (McFadden 1973). To account for the difference between the scale of observation and the scale of forest conversion by both plantations and smallholders, we also estimate a spatial autoregressive linear probability model, in which the W matrix is row-normalized and based on k nearest neighbors between pixels, through two-stage least squares, with WX as the instrument for WR* (Drukker, Egger, and Prucha 2013).
Nonparametric Estimation
To ensure that our results are not driven purely by parametric assumptions, we also employ a nonparametric approach to estimating the probability of smallholder forest conversion. The nonparametric approach is robust to misspecification in the functional form and in the distribution of the unobservables. To describe the nonparametric methodology, start with equation [6]:
In the nonparametric approach, f(Zjk, Mjk, Dj, Pj) and g(Zjk, Mjk, Dj, Pj) are allowed to be completely unknown. The nonparametric approach further makes no assumptions regarding the joint distribution of the unobservables ψ(ω0, ω1, …, ωK).
The goal of this analysis is to estimate the probability of smallholder conversion, R, which is based on the latent returns to production, R*, across different sets of pixels. The specific unknown probability of interest is
[7]
where ϕ(k,x) is some unknown function. A feasible estimator for equation [7] can be obtained using the generalized product kernel function of Racine, Li, and Zhu (2004):
[8]
where Kγ2 (Xj,x) has the following representation:
[9]
[10]
[11]
where Xc are continuous variables such as elevation or precipitation, while Xd are discrete variables such as the Jambi or Riau province indicator variables. The dimensions of Xc and Xd are r1 and r2, respectively. γ1 and γ2 = [hp, λp] are the bandwidth parameters for the R and X variables. The kernel function for the discrete dependent variable has the following representation:
[12]
The analysis is designed to detect differences in the preferred location for smallholder production as a result of a district’s subdivision status. To do so, the probability of smallholder conversion, 
, is estimated separately for pixels located in subdivided and nonsubdivided districts. To limit the possibility that differences in average marginal effects are due to evaluation of these densities at different values of the explanatory variables (due to different ranges of values for the explanatory variables in each of the district types, as suggested by Table 2), the separately estimated functions 
are evaluated at all of the 13,025 points to determine the average marginal effect of changes in each explanatory variable.
The districts that were subdivided following the end of Suharto’s regime are relatively concentrated spatially.11 To ensure the robustness of our results, we consider the full sample and conduct the same analysis using only parcels that lie within 5 km of a district with a different subdivision status. In addition to this simple analysis, we also employ more formal approaches to addressing concerns regarding the endogeneity of the subdivision indicator used in the analysis.12
Taking the approach related to group unconfoundedness presented by Rosenbaum (1987), we assess whether our subdivision indicator satisfies the unconfoundedness (exogeneity) assumption required for the reliable interpretation of our results. Recall that Sj is an indicator function equal to one if a district was subdivided and zero otherwise. For unconfoundedness to hold, we require that Sj ⊥ R(0), R(1)|X. To apply the group unconfoundedness approach to our problem, consider taking two subgroups of observations in the nonsubdivided districts we denote by G = c1, c2. By comparing the response variable across these two subgroups, we can assess whether one of the groups is comparable to the subdivided districts. As described by Imbens and Rubin (2015), we can formally test this by estimating the following difference:
[13]
Rosenbaum (1984) proves that a failure to reject H0 is equivalent to a failure to reject the equality of the distributions of unobservables across each group under consideration. A failure to reject H0 provides evidence in favor of unconfoundedness, thus supporting the validity of our results.
Specification Testing
To investigate whether misspecification is present within the parametric models described above, Fan, Li, and Min’s (2006) nonparametric bootstrap test of conditional distributions is employed for each of the parametric models estimated.13 This test looks for evidence that the true conditional distribution is different from that implied by an arbitrary parametric specification. The null hypothesis of Fan, Li, and Min (2006) is that the true population distribution is equal to some parametric conditional distribution given by ϕ(R = k|X = x, β). Formally,
[14]
For a logit model, ϕ(R = k | X = x, β) = RΛ(βx) + (1 – R)(1 – Λ (βx)), where Λ is the standard logistic distribution function. Rejection of the null hypothesis suggests misspecification in the parametric model, which casts doubt on the reliability of the parametric results, due to either an incorrect distributional assumption on the unobservables or function specification.
6. Results
We first present the results of the specification testing of the parametric models of forest conversion. The results of the nonparametric analysis testing the predictions about smallholder conversion in the subdivided and nonsubdivided districts follow.
Specification Testing
The results from the specification test show that the functional form and error distribution assumptions of each of the proposed parametric models are rejected via the nonparametric bootstrap test of conditional distributions based on subdivided (N = 9,192) and nonsubdivided (N = 3,258) subsamples.14 These results imply that the coefficient estimates and resulting marginal effects provided by each of these parametric models are inconsistent.15 There are numerous explanations for the rejection of the proposed models. The logit model might be rejected because it aggregates production systems and fails to acknowledge spatial autocorrelation. The multinomial logit model might be rejected for its assumption of the independence of irrelevant alternatives and the omission of spatial autocorrelation. The spatial autoregressive linear probability model may be rejected due to a functional form misspecification or a misspecified weight matrix. Any of these models might be rejected based on their assumed distribution of the error terms. The specification test proposed by Fan, Li, and Min (2006) is able to reject only the assumptions of the model, not provide insight into the particular assumption that fails to hold. Due to the rejection of each of the proposed parametric specifications of smallholder conversion, the paper will proceed with a focus on the nonparametric results.
Smallholder Conversion
Table 3 contains the average marginal effects from the binary analysis of the determinants of smallholder forest conversion. These results show that the probability of smallholder conversion decreases with increasing slope, distance from the forest edge, and distance from the nearest road for pixels located in nonsubdivided districts. Additionally, pixels in these districts are less likely to be converted to smallholder production if they are located in protected areas.16 Notably, elevation has no impact on the probability of smallholder conversion for the pixels in these districts, which is a different result from that in the subdivided districts, where the probability of smallholder conversion decreases with increasing elevation. The difference between the marginal effect of each of these variables across subdivided and nonsubdivided districts is statistically significant, except for distance from the nearest road. These results suggest that smallholder conversion in subdivided districts occurs in locations that would not be favored in nonsubdivided districts.
Nonparametric Results for Determinants of Smallholder Forest Conversion
We explore the change in conversion probability for a one standard deviation change in the explanatory variables relative to a one standard deviation change in the dependent variable to gauge the salience of the impacts.17 One standard deviation changes in forest edge distance, road distance, and protected area status generate impacts in nonsubdivided districts that are 5.68, 4.82, and 17.14 times larger than the impacts in subdivided districts, respectively. A one standard deviation change in forest edge distance, road distance, and protected area in the nonsubdivided districts leads to a 74%, 48%, and 16% increase in conversion probability over the unconditional probability of conversion. These results suggest that moderate changes in forest edge distance, road distance, and a change in protection status have very substantial economic impacts when compared to the unconditional probability of conversion.
To pinpoint the source of the average marginal effects and to directly address Predictions 1 and 2 from the model of smallholder conversion, Figure 2a and b show the pointwise conversion probabilities for forest edge and road distance for the binary analysis.18 The probability of conversion is highest at the forest edge in both the subdivided and nonsubdivided districts, with the conversion probability greater in the nonsubdivided districts than in the subdivided districts over the first 1 km into the forest (Figure 2a). This result is consistent with Prediction 1. Regarding Prediction 2, that the probability of smallholder conversion in the forest interior would be greater in subdivided districts than nonsubdivided districts, we see that the point estimate of conversion probability is in fact greater in subdivided districts from 1.5 to 5 km into the forest, though the difference is not statistically significant. The marginal effects in the nonsubdivided districts are significantly less than those in the subdivided districts between 0.5 and 1.75 km into the forest, which is additional support for Prediction 2.
Conversion Probability: (a) Binomial Smallholder, by Forest Edge Distance; (b) Binomial Smallholder, by Road Distance; and (c) Smallholder Oil Palm, by Forest Edge Distance
The probability of smallholder conversion is greater at the road’s edge in nonsubdivided districts than in subdivided districts, which is consistent with Prediction 1, when applied to road distance. While the gap in estimated conversion probability narrows substantially between 2 and 5 km from the nearest road, it is never greater in the subdivided districts, which does not support Prediction 2 in this context (Figure 2b). The marginal effect for the subdivided districts is effectively zero at most distances, except for the 2.5 to 3.25 km range, where the marginal effect is positive, offering some support for Prediction 2 in this context. For the nonsubdivided districts, the marginal effect is negative and statistically different from zero over the first 3 km from the nearest road.
The inclusion of district indicator variables in the models ensures that unobserved district characteristics (e.g., corruption) are not responsible for the observed results, though it is possible that an omitted variable that might determine smallholder conversion decisions (e.g., local climatic conditions) could be driving the results. We conduct the same analysis whose results are displayed in Table 3 on a subset of pixels that lie within 5 km of the border of a district whose subdivision status differs from that of the district in which the pixel is located. The results described above are robust across this subset of pixels, confirming our confidence that the change in smallholder conversion-location preferences is due district subdivision.19
We conduct a test of the null hypothesis presented in equation [13] regarding group unconfoundedness to explore exogeneity of district subdivision status. Recall that a failure to reject the null is evidence in favor of the exogeneity of district subdivision status. To form valid subsets on which to test unconfoundedness, we consider groupings based on the distance of a nonsubdivided pixel from a subdivided district border starting at 1 km and ending at 20 km. We fail to reject the null in 18 of the 20 groupings considered.20 In the two cases for which the null is rejected, the upper bound in the 95% confidence interval contains a near-zero number. The results suggest that group unconfoundedness, and therefore exogeneity, is likely to hold in the analysis.
We utilize a multinomial analysis to explore the possibility that the binary results might be dampened by aggregating different smallholder production systems and to test Prediction 3 from the theoretical model. These results are depicted in Table 4.21 The average marginal effects depicted in this table again indicate that smallholder conversion occurs in different locations based on district subdivision status, with the differences suggesting that smallholder conversion in subdivided districts occurs at locations that would not be preferred in nonsubdivided districts. This result is consistent with the predictions of the model of smallholder forest conversion. For oil palm production, the probability of conversion decreases with increasing elevation and forest edge distance and is significantly lower in protected areas in nonsubdivided districts; these effects are significantly dampened in subdivided districts, with the magnitude of the average marginal effect for these variables decreasing by roughly 85, 17, and 17 times, respectively, in the subdivided districts.
Nonparametric Multivariate Results for Determinants of Smallholder Forest Conversion
Figure 2c illustrates how the probability of conversion to smallholder oil palm production varies at different distances from the forest edge across the subdivided and nonsubdivided districts. These results show that smallholder oil palm production pushes into the forest interior from the forest edge in subdivided districts in contrast to outcomes in nonsubdivided districts, where the forest edge is the preferred location for oil palm production. The results in this figure offer support for Predictions 1 and 2 from the theoretical model. The increased probability of smallholder oil palm conversion in the forest interior in subdivided districts relative to nonsubdivided districts emphasizes that the results in Tables 3 and 4 are not merely due to the fact that there is less smallholder forest conversion overall in subdivided districts.
The results for plasma oil palm systems support the prediction that smallholder location choice is affected by district subdivision. These systems do not face a risk of expropriation from plantations, and we expect the determinants of plasma oil palm system location choice to be independent of subdivision status and that these systems would locate in areas with different characteristics than independent smallholder oil palm production. The results seem to support these expectations. First, there are only two statistically significant differences in the average marginal effects across our explanatory variables for this production system (precipitation and distance from the nearest town), and one of these differences is of marginal significance (town distance). Neither of the significant differences are for variables that had differential impact on the probability of independent smallholder oil palm production across district subdivision status. Furthermore, the difference in average marginal effect between subdivided and nonsubdivided districts, across all variables with statistically significant differences, with the exception of precipitation and protected area, is of a different sign for independent smallholder oil palm production relative to plasma systems.
The results for mixed-agriculture smallholder systems again offer some support for the displacing effect of district subdivision on smallholder forest conversion in Sumatra. Specifically, these systems are 7.26 times more responsive to forest edge distance in nonsubdivided districts than in subdivided districts, meaning that these systems might be expected to be located farther from the forest edge in subdivided districts, which is consistent with the predictions from the theoretical model. Furthermore, there is no significant decrease in the likelihood of forest conversion to smallholder mixed-agriculture production on pixels located within protected areas in subdivided districts, while the deterrence is significant at the 5% level in nonsubdivided districts. There is less evidence provided by the results for smallholder rubber production. One explanation for the lack of differences in conversion location for rubber production is that these trees grow in the wild and adult trees might be tapped where they grow naturally, avoiding the need to wait for seedlings to mature. The observed variation in the impact of district subdivision on the location of the different smallholder production systems is consistent with Prediction 3 from the theoretical model, with the observed effect greatest for the oil palm and mixed-agriculture systems, which have higher returns (oil palm) and higher investment (oil palm and mixed agriculture) than smallholder rubber production.
7. Welfare Impacts of the Location of Smallholder Forest Conversion
The concern about conversion of more-remote forests in subdivided districts is that these areas may not have been converted in nonsubdivided districts. Given the vast amount of endemic biodiversity contained in the forests of Sumatra and the suite of ecosystem services provided by those same forests that generate tangible benefits both globally (e.g., carbon sequestration and storage) and locally (e.g., avoided sedimentation, storm-peak reduction), it seems clear that the external costs of forest conversion vary spatially and that these costs will increase as conversion occurs in more remote locations.
The private and social benefits and costs of forest conversion vary spatially, making the location of conversion a key determinant of the welfare effects of such action. The literature demonstrating the negative impacts of edge effects on biodiversity and forest health is quite robust (see, e.g., Gascon, Williamson, and da Fonseca 2000). Recent work has highlighted that the carbon stored in forest biomass increases substantially with increasing distance from the forest edge, meaning that forest conversion to agricultural production in the interior of the forest can significantly increase carbon emissions (Chaplin-Kramer et al. 2015). These ecological changes negatively affect social welfare through reductions in the value of ecosystem services provided by the extant forest, as well as potential losses of biodiversity.
Our welfare analysis of district subdivision’s impact on the location of smallholder conversion is based on the increased external costs due to carbon emissions from smallholder forest conversion deeper into the forest in subdivided districts. This is a lower bound on the actual external costs of more-remote smallholder conversion, as it omits the value of other lost ecosystem services and negative biodiversity impacts. We use the results of our nonparametric analysis to generate estimates of conversion probability in nonsubdivided districts based on the model from subdivided districts in order to quantify how the location of forest conversion in nonsubdivided districts would change under conditions of subdivision. Each pixel is classified as predicted to convert to smallholder production under subdivision if 
, where 0.1913 is the conversion probability value that balances the specificity and sensitivity of the nonparametric model. At this threshold, the model accurately predicts 98% of both positive and negative outcomes.22
We calculate the biomass of the forest in the nonsubdivided districts using an equation derived from data from Chaplin-Kramer et al. (2015) that links aboveground forest biomass with forest edge distance for lowland forests in Sumatra. We find that the average edge distance for pixels predicted to convert under subdivision is 0.955 km, while the average edge distance is only 0.266 km under nonsubdivided conditions, meaning that pixels converted under subdivision have 44.14 additional metric tons of biomass per hectare, on average, with a bootstrapped 95% confidence interval of 40.03–45.23 metric tons of biomass per hectare.
We convert the excess biomass content into an external cost of smallholder forest conversion farther into the forest by converting this biomass measure into metric tons of CO2 and multiplying this amount by the social cost of carbon (SCC).23 We use three estimates of the SCC to account for a range of emissions scenarios and discount rates in the year 2010: $21.40 (average emissions, 3% discount rate), $35.10 (average emissions, 2% discount rate), and $64.90 (95th percentile emissions, 3% discount rate) (Interagency Working Group on Social Cost of Carbon 2010). The present value of the per-hectare costs of the altered location of smallholder forest conversion following district subdivision due to carbon emissions ranges from $1,629 to $4,941, with an average of $2,672.
Comparing the per-hectare costs calculated above to the returns to smallholder agricultural production generates estimates of the net benefits of smallholder forest conversion. Grieg-Gran (2008) provides estimates of the opportunity cost of avoided forest conversion in numerous tropical nations based on the foregone system of production. According to Grieg-Gran (2008), per-hectare returns to smallholder oil palm production range from $960 to $3,439 (in 2007 dollars), depending on whether a high or low yield is assumed for palm oil production and whether one-time timber sales during forest conversion are included in the calculations. The returns to these production systems are at least one order of magnitude greater than those for other systems (e.g., smallholder rubber production is estimated to yield $72 per hectare), and we see that they are comparable to, though generally lower than, the external costs of excess carbon emissions alone due to conversion of more remote forest patches by smallholders in subdivided districts. Such conversion is shown to be potentially welfare reducing to society, emphasizing the potential gains of more efficient forest management in Sumatra.24
The counterfactual generated for the welfare analysis also allows us to explore how the extent of smallholder forest conversion is impacted by district subdivision. The number of pixels predicted to convert under subdivision, namely, those whose predicted probabilities exceed 0.1913 using the model from the subdivided districts, is greater than the observed number of converted pixels in the nonsubdivided districts (a 95% confidence interval of 192–379 pixels relative to 175 observed conversions). This result suggests that the aggregate external costs of smallholder forest conversion would have been greater had these pixels been in subdivided districts, due to both more remote location and a greater amount of conversion.25
These comparisons illustrate how substantial the external costs of more remote forest conversion are and suggest an opportunity to address these unintended consequences of district subdivision. Our analysis emphasizes that forest hectares are not fungible, underscoring the point made by Vincent (2016) that impact evaluation of forest management programs based on biophysical terms (e.g., hectares of avoided forest conversion) is difficult to link to economic measures of program performance, because the biophysical units typically ignore heterogeneity in the costs and benefits of these programs. Such points call for monetizing the benefits and costs of forest conversion to aid in the development of successful forest policy.
8. Discussion
We take advantage of a uniquely detailed GIS data set to explore how district subdivision in Indonesia impacts the location choice and associated external costs of smallholder forest conversion. Our analysis shows that smallholders convert forest at different locations within districts that have been subdivided relative to those that have not; this finding holds when the models are estimated on the full set of randomly drawn pixels that were forested in 2000, as well as a subset of pixels that lie within 5 km of the nearest district with a different subdivision status. Smallholders are located on more marginal lands (steeply sloped, farther from the forest edge and nearest road) within subdivided districts, supporting predictions from a simple model of smallholder conversion. Our analysis also shows that the likelihood of conversion to smallholder mixed agriculture or oil palm systems increases for forested pixels within protected areas within subdivided districts, ceteris paribus. This response increases the external costs of conversion, measured only by carbon emissions, with the estimated costs slightly greater than the present value of per hectare net returns to smallholder palm oil production, the most valuable smallholder production system in the region.
The results allow for some commentary on the impact of decentralization on welfare outcomes. There is theoretical justification for both improved and worsened natural resource management due to decentralization, stemming from either a race to the top among jurisdictions in competition for constituents who might vote with their feet (see, e.g., Tiebout 1956), or a race to the bottom among jurisdictions aiming to appease firms that are sensitive to the costs of compliance with environmental and social regulations (see, e.g., Musgrave 1997). As emphasized by Ostrom (1990), institutional context is a key determinant of decentralization outcomes, making it difficult to come up with consensus support for either of the above theorized outcomes. McCarthy (2004) finds that in Kalimantan, with uncertainty about property-rights security and governance responsibilities, decentralization led to a race to exploit forest resources.
Our analysis suggests that the focus on increasing local government control of naturalresource extraction royalties may offset some of the benefits of these increased royalties by failing to protect the livelihoods of smallholders. These results are consistent with those presented by Suwarno, Hein, and Sumarga (2015). Increased attention to the plight and behavior of smallholders might be achieved with more formalized smallholder tenure, which would make smallholders another royalty source for district agencies. More secure smallholder property rights could alleviate some of the displacement predicted by the above model and demonstrated in our empirical analysis.
The results of our welfare analysis suggest meaningful carbon benefits associated with prevention of smallholder forest conversion deeper into the forest. Carbon payments for reducing emissions from deforestation and degradation (REDD+) might achieve greater emission reductions by acknowledging smallholders as agents of forest conversion and including this stakeholder group more directly in the forest-management process. In examining the formation of comanagement agreements between resource users and the state, in which the state achieves lower management costs by allowing increased resource use, Engel, Palmer, and Pfaff (2013) suggest that the inclusion of carbon rights in forest comanagement agreements would further motivate communities to increase forest carbon stocks. Our finding of the similarity between the present values of per hectare external carbon costs and net agricultural returns emphasizes that smallholders in these areas would have incentives to conserve forests with access to these carbon rights. Our findings suggest that inclusion of smallholder behavior in natural resource management decisions might improve outcomes under decentralization.
Although we are able to identify differences in smallholder behavior across districts based on subdivision status, the mechanism behind this result cannot be determined in our analysis. The two most likely pathways are that subdivided districts are characterized by more plantation activity, leaving smallholders with less desirable land, or that smallholders locate their operations on marginal lands in areas with high threat of plantation operations due to insecure smallholder tenure. The first alternative would be predicted via a von Thunen or Alonso bid-rent model in a landscape with a functioning market for land. The latter alternative seems more relevant for outcomes in Indonesia, where smallholder land tenure is insecure, and is supported by the existing literature.
Our results suggest that insecure tenure may affect both the location and extent of smallholder forest conversion, which has implications for the private and social net benefits of such action. Future work that identifies the direct impact of insecure smallholder tenure on the location and extent of forest conversion in the tropics would be a meaningful contribution to the literature that has considered the interaction between heterogeneous agents of deforestation (e.g., Angelsen 2001) and could be of great use to policy makers in tropical nations.
The conditions that exist in forested regions of tropical nations often prevent optimal management of this important source of local and global social benefits, stemming from their production of valuable ecosystem services. Our research effort sheds light on how the presence of heterogeneous drivers of forest conversion leads to particular inefficiencies in the management of these valuable global resources through inefficient spatial distribution of conversion in areas with more localized resource control. Additional research into how governments in these regions might best allocate their limited resources to monitoring and enforcement, development of missing institutions, and anticorruption measures is necessary to ensure that these areas are as well equipped as possible to mitigate and adapt to anthropogenic climate change.
Acknowledgments
The authors wish to thank Bradley Eichelberger, Oki Hadian, and Nasser Olwero for their help in assembling and organizing the GIS data used in the analysis.
Footnotes
↵1 Under the nuclear estate scheme, oil palm plantations gain access to land that was under the de facto control of local landholders and allow these individuals to manage small areas of oil palm that had been planted by the plantation manager with an understanding that the smallholders will sell their output solely to their affiliated plantation (referred to as plasma operations) (Levang 2003). These arrangements subject smallholders to monopsony conditions, and the plantations tend to operate with limited transparency, leading to smaller than expected payments and a lack of clarity about when, if ever, the land will be returned to smallholder control (McCarthy 2007).
↵2 In this model, land use change is the result of many individual smallholders making conversion decisions simultaneously, a common approach in the literature (e.g., Bockstael 1996; Parks and Hardie 1995; Wu and Segerson 1995).
↵3 There is no evidence to motivate a specific relationship between yield and forest-edge distance. We proceed by assuming that the relationship between profit and forestedge distance is driven by the relationship between cost and distance from the forest edge (namely, π′(d)≤ 0).
↵4 As noted by Pfaff (1999), the choice of forest conversion is a dynamic problem; however, we can explore the issue in a static context by making an assumption about the decision rule used to determine the timing of forest conversion. If we assume that the deforestation choice (i.e., whether or not to deforest a patch at time t conditional on the plot being forested at this point) can be reduced to comparing the benefits from having the plot remain in forest relative to those of deforestation in time t + 1, then the conversion choice can be studied as a static problem.
↵5 Insecure tenure, which raises the possibility of land expropriation, meaning W(d) = 0 ∀ d, has been shown theoretically and empirically to increase deforestation rates (Deacon 1994; Mendelsohn 1994) and to suppress investment in agricultural production (Jacoby, Li, and Rozelle 2002).
↵6 The higher cost of conversion farther into the forest means that smallholder profits conditional on maintaining parcel control will be lower inside the forest than at its edge. Achieving the same level of profits in the forest as at the edge, conditional on maintaining control, would require converting a larger forest area to production in the forest interior than at the edge. We explore the impact of district subdivision on conversion extent in our welfare analysis.
↵7 The issue of expropriation risk is another impact of plantation conversion in addition to the infrastructure and land-scarcity effects explored by Angelsen (2001).
↵8 See Appendix A for additional detail regarding the nonparametric approach we employed.
↵9 See Appendix Figure B1 for a map of the pixels included in our analysis.
↵10 These cells include 4,444 30 m by 30 m pixels and cover an area of 400 ha, which is 20 times larger than the maximum smallholder production system of 20 ha.
↵11 See Appendix Figure B1 for a map of district subdivision status.
↵12 See Appendix Table B6 for a comparison of the normalized differences in the means of our observable characteristics across both district types.
↵13 All code was written in Matlab and is available upon request.
↵14 See Appendix Table B1 for detailed results.
↵15 See Appendix Tables B7–B9 for the results of each of the inconsistent parametric models.
↵16 The possibility that protected areas serve as a refuge from plantations when facing a high threat of plantation conversion could explain the heterogeneous effectiveness of protected areas in addressing issues of conservation and poverty that have been reported in the literature (e.g., Ferraro, Hanauer, and Sims 2011).
↵17 See Appendix Tables B2 and B3 for further details on the methodology.
↵18 See Appendix Figures B2 and B3 for the pointwise derivatives for forest edge and road distance, respectively.
↵19 See Appendix Table B4 for these results.
↵20 See Appendix Table B5 for the detailed results.
↵21 See Appendix Table B10 for the optimal bandwidth for the various explanatory variables in the binary and multinomial analyses.
↵22 See Appendix Figures B4–B6 for graphs of model sensitivity and specificity for the nonparametric model, spatial lpm, and logit model.
↵23 The additional metric tons of biomass per hectare are first converted to metric tons of carbon using a multiplier of 0.47 recommended by the Intergovernmental Panel on Climate Change (2006) for tropical forests and then converted to metric tons of CO2.
↵24 The SCC is designed to be a comprehensive estimate of climate change damages and thus typically exceeds the price paid in carbon payment schemes (Hamrick and Gallant 2017).
↵25 Given the pervasiveness of corruption in Indonesia (Fisman 2001; Olken 2007), quantifying the aggregate net benefits of district subdivision would require a credible means of estimating the cumulative welfare impacts of the increased royalties from greater plantation activity that accounts for the impact of corruption on the collection and use of these tax revenues, in addition to aggregating the perhectare returns and external costs of smallholder conversion described above.
References
This article requires a subscription to view the full text. If you have a subscription you may use the login form below to view the article. Access to this article can also be purchased.
Log in using your username and password
Purchase access
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.