Why do forest products become less available?A pan-tropical comparison of drivers of forest-resource degradation

Our analysis draws on the PEN database, which was compiled by the Center for International Forestry Research (CIFOR)⁸ . The PEN data is a pan-tropical, standardized dataset of household incomes, assets, and livelihood systems. It allows for some global-comparative quantitative analyzes that go beyond case studies in terms of generalizable conclusions regarding factors affecting tropical forest resource base. The PEN data comprise social, ecological and economic information obtained from standardized interviews for 8301 households in 334 villages across 59 study sites in 24 tropical countries in Sub-Saharan Africa, Latin America and Asia. The data can be considered as a representative sample of smallholder-dominated tropical and sub-tropical landscapes with households having access to forest resources (Wunder et al 2014).

Our study draws on village surveys implemented between January 2005 and May 2010, providing information on villagers' perceptions of changing availabilities of main forest resources over the last five years, and the perceived causes of this change. In addition, the village focus groups provide information on land management, demographics (including migration) and other potential drivers of change. From the initial 334 villages, we restricted our analysis to those 233 villages for which the needed complete survey data existed (figure 1).

Zoom In Zoom Out Reset image size

The PEN data provides information on six forest resource types: firewood and charcoal, timber and other wood, food, medicine, forage, and others. Respondents ranked their relative importance, and stated what they perceived as their change in availability (stable, increase, or decline) over the last five years prior to the survey. Across the 233 villages the three most important forest products for local livelihood were firewood and charcoal, timber and other wood, and food. In two thirds of the villages, local people used all those three forest product types. Medicine and forage were less important for the village wellbeing, though 75% of the villages used one or other of these resources.

We then aggregated the changes in availability of the individual forest products across the six categories per trend (perceived increase and perceived decrease) to get a more general perception. Our aggregated indicator for perceived forest product decline was calculated as the percentage of the number of declining forest products from the total number of available products in each village as reported by the informants. Analogously, the percentage of perceived forest product increase was calculated as the percentage of the number of increasing forest products from the total number of available products in each village according to the informants.

The PEN household surveys provide reasons for changes of the individual forest product availability as perceived by the village population, specified for both increasing and decreasing availability. Those reasons were pre-structured, with binary responses (Does this suggested reason for change apply? Y/N). We then aggregated these reasons across the six resource types, indicating to what extent a particular reason, for example increased forest product use applied to the respective village or not. In this way, we identified the relevance of each reason for both perceived increase and perceived decrease of forest resources per village (see table 1; independent variables: self-reported reasons for forest product decline/increase). Finally, we juxtaposed the aggregated information on forest product availability to village-level context conditions: forest type, area, management regime, use regulations, demographics (see table 1; independent variables: framing conditions) in statistical analyzes. Appendices A and B show frequency distributions and correlations between the variables used, respectively.

We applied a random forest regression tree procedure to explore the drivers of both the perceived decline and perceived increase of forest products. Regression trees allow for revealing unknown nonlinear, multivariate patterns with high-order interactions, which typically occur in socio-ecological systems (De'ath 2002, Archibald et al 2009, DeFries et al 2010, Bonilla-Moheno et al 2012). Regression trees explain variation of a single dependent variable by recurrently splitting the data into more homogeneous subsets, referred to as 'nodes', by using combinations of categorical and/or numerical independent variables without making assumptions on the statistical distribution of the data (Breiman 2001). Each node is characterized by a typical value of the dependent variable, the number of observations in the node, and the values of the independent variables defining it. Random forest procedures apply bootstrapping to grow a large number of regression trees—hence, a forest of trees—to both improve the predictive power of regression tree models and reduce overfitting (Breiman 2001, Prasad et al 2006).

For the random forest analyzes, we used each percentage perceived forest product decline/increase as dependent variable. 500 trees were grown based on a randomized subset of independent variables using five independent variables at each split. All trees were grown without pruning, and the final result is the average over all 500 trees. Random forest regression tree procedures provide percent of the variance explained by the model and the mean-squared error (MSE) for each potential tree split. The increase in MSE (%IncMSE) of predictions is the result of a certain variable being permuted (i.e. randomly shuffled values) and thus is a measure of how much the predictive power of the statistical model increases by including this variable. As such, MSE was used to measure the relative importance of the independent variables within the model. The analysis was carried out using the randomForest package of the R software (R Development Core Team 2015). All variables included in the analyzes are presented in table 1.

Finally, we fitted a single regression tree for each trend—i.e. one for perceived product increase, and one for perceived decline—to classify homogeneous groups of the original observed data on forest-product changes. The two regression trees were grown and pruned by cross-validation to remove splits that do not substantially decrease the mean square error. We applied a minimum node size of ten; hence all final nodes contain at least 10 data points, i.e. villages. The obtained splits and nodes were used to quantify the combination of drivers resulting in a gradient of little-to-strong change in forest-product availability. For building the single regression trees we used the rpart package of the R software (R Development Core Team 2015).

Perceived decline of resource availability was the dominating trend, which—except for firewood, medicine and forage in South America—applies to all six product types across the three continents (figure 2 and appendix C). More specifically, the two wood-product categories (timber and fuelwood) perceived the clearest decline across our villages, followed by the forest food category—although these trends differ somewhat across continents. 209 of the 233 villages perceived a decline of at least one forest product. Of these, 87 villages perceived a decline of all their available forest product types. In contrast, availability of at least one forest product increased in 89 villages during the same period according to the informants. This implies that on average 66% of the available products per village were perceived to decline while 16% of the available products were perceived to increase during the same period.

Zoom In Zoom Out Reset image size

Our most important driver of perceived forest product decline is the increased collection of forest resources by local people; it increases the MSE by 30% (figure 3). Forest clearing, due to both small-scale clearing for agriculture and large-scale projects including plantations and new settlements, is the second-most explanatory factor for the perceived forest-product decline. Among the four different forest-ownership types, community forest area is the most important predictor for decreasing forest-product availability. Required permissions, climate change and harvesting (timber) and burning (charcoal, brick and bush) are least important in explaining perceived resource availability decline (figure 3). The random forest regression model, which averages the results of the 500 individual trees, explains just over half (52%) of the variance of forest-product decline.

Zoom In Zoom Out Reset image size

We further identified the combination of factors that determine the intensity of the decline in product availability by growing a single regression tree. Based on the derived split thresholds, indications regarding the processes perceived to fuel the forest-product decline across the pan-tropics can be retrieved. Different combinations of drivers can result in a range of intensities of decline (figure 3 center). Increased forest product use is the first split in the regression tree, meaning that it is the most powerful discriminator between villages with relatively high and low forest-resource decline. In general, increased forest-product use applies to 175 out of 233 villages. The strongest decline is perceived in villages characterized by an increased usage of forest products, a share of community forest area of maximum one third, and immigration. The numerous villages characterized by this combination of drivers (102) have perceived a decline of the available forest products of on average 87% in the past five years. The second-largest group of villages is characterized by an increased forest-product use, a share of community forest area of at least one third, and forest clearing. These characteristics cause on average a decline of 71% of the available forest products. The least decline (with less than one percent) occurs in cases of no increased forest product use by local people, no forest clearing, combined with the absence of legal restrictions with respect to forest use.

Compared to the random forest, the predictive power of the single regression tree is somewhat lower (R² = 0.49), yet provides explicit splitting conditions pointing to nested factors conjointly driving forest-resource decline.

The predictions of forest-product decline obtained from the single regression tree show varying certainties. The predictions are remarkably high, i.e. having little variance, for the villages with the smallest perceived forest-product decline (figure 4). Apparently, the combination of factors determined by the model explains small product declines well. In contrast, predictions of medium-to-strong product decline show a much lower certainty. Probably, in those cases a variety of factor combinations lead to forest-product decline, which were not fully captured by the regression tree.

Zoom In Zoom Out Reset image size

Compared to the perceived resource decline, perceived increases of resource availability exist for at least one resource in 89 out of 233 villages. The most important factor perceived to support an increase is changes in forest management. Besides, less clearing of forests for agriculture and reduced use of products—mainly because fewer local people collect less, and the usage from large-scale commercial users is reduced—are considered to significantly increase the availability of forest products. These three factors individually increase the MSE by 22%–28% (figure 5). Forest clearing, the second-most important factor for a loss of forest products, also provides opportunities for forest-product increase. Post-clearing product supply, such as fuelwood, is a key predictor of forest-product increase. This reminds us that increased availability is not automatically a sign of more sustainable forest management. Improved forest access, due to both more illegal access of protected areas and improved access rights, in-migration and open access forests are the least important variables, hence contributing least to the perceived increase in availability of tropical forest products. The random forest regression model explains 63% variance of perceived forest product increase. Hence, predicting the minority scenarios of increases in forest-product availability is more reliable than predicting declines.

Zoom In Zoom Out Reset image size

Again, we grew a single regression tree to scrutinize combination of factors determining the intensity of perceived increases in product availability. Figure 5 shows that decreased forest-product use is the first split, i.e. the most powerful discriminator between villages with high versus low forest-resource increase. Yet, decreased forest-product use applied to only 40 villages. The steepest increase was perceived where a decreased use of forest products conjointly occurs with a population growth below 30%. Villages characterized by this combination of factors have perceived an increase of on average 67% of the available products over the past five years. However, this occurs in only 26 of the 233 villages. 78% of villages are characterized by a negligible increase of forest-product availability (less than 0.1%). The single regression tree explains 62% of the variance of forest product increase (R² of single regression tree), hence the predictive power is similar to the random forest model.

Once more, model certainties vary considerably (figure 6). For villages where on average two thirds of the resources are perceived to increase (node 4), variance is also large, and prediction certainty is low. Hence, these villages are associated with a lower certainty with respect to their predicted resource increase. Analogous to the resource decrease results, in these cases the variety of different combinations of driving factors was not fully captured in the regression tree.

Zoom In Zoom Out Reset image size

Comparisons of previous case studies—e.g. systematic quantitative meta-analyzes—have been hampered by different study designs, methods and scales (Rudel 2008). Our comparative results, in turn, have the advantage to build on a standardized data collection approach. Our approach does not, however, allow us to infer causal relationships between forest product availability and the independent variables. In addition, we need to critically reflect on the fact that the reported dynamics of forest resources are not objective measurements, but perceptions of local people expressed in village meetings. First, these perceptions may be subject to recall errors over a certain period (five years in our study): the most recently occurring changes and drivers are often better recalled than events farther back in time, which is a well-known disadvantage of retrospective studies. Cross-checks with independent measurements of forest-product changes could in more detailed case studies be important meaningful supplements for correcting possible biases. However, in our case of a large standardized, global-comparative survey, quite naturally they were not available. We therefore lacked the means to validate such stated perceptions. Second, responses may also reflect cultural biases regarding how sensitive local people are to the perceived dynamics, and how they express them to a visiting research team. Sometimes responses should be considered a vested discourse, for example when local communities bemoan the decline of firewood resources, allegedly due to external resource pressures, because they would like the outside world to restrict external access. Perceptions may also introduce biases regarding the drivers of resource change. For example, in certain cases local people might confuse their knowledge that population has increased with their notion that resource use has also increased, without being able to prove this based on independent measurements. Furthermore, sampling of opinions of community members may have introduced potential biases regarding the role of the community with respect to resource use, expressed for example in differences regarding forest access and management. Local community members who respond to the survey most likely do not aim at indicting the local community but may tend to be more willing to accuse the outside community. Considering these aspects recording perceptions of resource change and its drivers might introduce biases in the results of the survey. In certain cases such biases may be manifested in shifted variable importance. However, the most important driving factors of resource change (for both decrease and increase of resources) show a distinct importance and we therefore presume that possible biases do not affect our analysis. Since the village surveys provide no means for validating or correcting possible biases our analysis had to consider the survey data being an accurate representation of actual field conditions. In fact, a few studies have tested for biases of self-perceived dynamics of environmental resources. They concluded that reasons for divergence between measured and perceived changes are indeed often associated with (changes in) people's vested interests in certain resources (Meze-Hausken 2004, Amsalu et al 2007, Adimassu et al 2014). In contrast, if over-exploitation of natural resources has been mostly locally self-inflicted, or even caused by illegal extraction e.g. from a protected area, responses may under-report emerging resource shortages; an aspect that was considered in the PEN survey design (Babigumira et al 2014).

The PEN cross-sectional data does not support a study of explicit inter-temporal transitions in the sample villages. Yet, it provides some clues as to potential factors driving current perceptions of resource change. Whether or not natural resource change will continue depends largely on the extent of local adaptation (Cumming et al 2014). Our most important driver of declining availability is the increased local use of forest products. This suggests per se a low degree of local adaptation to increasing scarcity—at least so far. Conversely, the smallest forest product declines were perceived in villages without increased forest product use by local people, and in absence of both forest clearing and legal restrictions with respect to forest use. This may indicate, among this minority of cases, a capacity to formulate effective village responses to emerging scarcities in natural resources through the adaptation of use patterns.

Potentially, villages could respond by regulating forest-product access through various institutions and strategies (Armitage 2005, Ostrom and Nagendra 2006). Access to forest products proved to be an important factor for explaining forest product dynamics. Our results show that when at least one third of forest area is owned by local communities, villages are less likely to perceive resource declines than where non-community forest ownership prevails (64% versus 84% perceived product decline).The transition from open access to actively managed community forests often occurs when resource degradation has become substantial, and village institutions are strong enough to enforce rules. Again, our analysis points to a cluster of driving factors here, but we cannot clearly distinguish cause and effects.

Our analysis singled out population growth as a key factor in explaining forest resource trends. In fact, 91% of our villages faced a growing population from natural population growth and net-immigration combined. The results show that relative population change is an important factor mainly for perceived increases in availability of forest resources; excluding it from the model reduces the predictive ability of the model by 12%. We identified 40 villages where perceived forest resource increase is responsive to population change, with the strongest rise (67% of the available forest resources) occurring in villages with either population declines or little growth. The majority of villages (85%) experienced immigration within the past ten years; 20% of these villages have an immigrant share of at least 25% of village population. Immigration is moderately important for explaining tropical forest-product decrease. The results show that immigration, controlling for increased product use and community forest area, raises the perceived decline from 59% to 87% of the available forest products. This is in line with findings of several regional and local case studies that illustrate the impact of immigration on forest cover (Geist and Lambin 2002, Carr 2009).

Change in product use by the local population is a key driver of resource availability; it is the most discriminating factor in both single regression tree analyzes. Increased product use leads, in various combinations with other factors, to a perceived decline of between 51% and 87% of the available forest resources. For the respective villages, we conclude that resource decline has not led to an adaptation of use patterns to counteract trends. Similarly perceived resource decline (64% of the available forest resources) in the absence of increased product use only occurs in case of forest clearing. Overall, forest clearing poses a significant threat for forest products in the villages included in our analysis. As such, forest clearing is the second most important factor for determining forest product changes, in either direction.

Obviously, availability of forest products is linked to the existence of forest. Yet, the incremental effect of deforestation on forest products had been little studied so far, especially in comparative terms. In our study, forest clearing includes both small-scale clearing for agriculture (125 villages) and larger-scale clearing for project development (46 villages). The former is often related to subsistence and cannot be separated from demographic factors (Geist and Lambin 2002). The latter cannot be explained by locally occurring processes alone. In our study, both internally and externally driven forest clearing contribute to the decline of forest product availability. Yet, forest removal also creates opportunities, leveraging fresh supplies of clearance-related products (e.g. fuelwood), and enabling the growth of secondary forests. Post-clearing product supply is the fourth most important contributor to perceived increased forest product availability. Forest management is applied very differently across all villages. While 28% of the villages do not practice any form of active and deliberate forest management, almost half of all villages apply at least three types of management. Sustainable forest management strategies such as selective logging have the potential to directly provide forest products, and to maintain the forest sustainably. Forest management changes, including tree planting, is the most important factor for determining perceived forest product increase, which can even compensate for continuing resource usage.