An evidence and gap map of the environmental Kuznets curve and the forest transition hypothesis for estimating forest area development

Web of Science and Scopus were the bibliographic databases used in the literature search for this review. By selecting these two databases, we could cover almost all the relevant journals in the forestry sector. The systematic article search was limited exclusively to peer-reviewed articles in English. Search terms were limited to those in English due to project resource restrictions and time constraints. The search in these two databases was conducted in November 2020. The search strings used in each database are shown in table 1.

Using the advanced search function in the databases, we defined two main groups that are connected with 'AND.' These main groups consist of keywords related to forest-related terms and to the FTH or Kuznets curve. We searched for these terms in Web of Science in the titles, keywords, keywords plus^®, and abstracts. In Scopus, our search string was applied to the title, abstract, and keywords.

Articles eligible for analysis were identified using criteria that specify their suitability for inclusion. These inclusion criteria relate to the nature and content of the studies. We only considered peer-reviewed, English-language articles published between 1955 (emergence of the EKC theory) and 2020. Since we searched for internationally resilient and valid results, we only included cross-country studies with a continental or global scope (at least two countries had to have been included). Conversely, we excluded single-country studies. In terms of content and to answer our research questions, only studies that provided numerical estimates of forest development curves were included, while theoretical assumptions were excluded (see table 2).

After completing the literature research, the articles were stored in Citavi and duplicates removed. Each article was given an individual and unique number to be able to assign them exactly. The screening for eligibility of the studies was conducted in two stages according to the method established by ROSES. During the first stage, the titles, keywords, and abstracts of selected articles were screened. Based on the results of the first screening, full texts were screened in the second stage. Two authors performed the first screening of the articles, which were distributed randomly. To calibrate the convergence of the first screening, the authors processed ten percent of the articles in the first screening separately and compared the results afterwards. The collectively screened articles served as an exercise. Subsequent discussions were held until a unanimous opinion on the inclusion or exclusion of articles was reached. As the screening process continued, the authors held regular meetings to review consistency and clarify, for example, ambiguities in exclusion criteria, such as the delineation of regions or comprehensive synonyms for the emerging terms. The entire first screening process was completed in two months.

All articles that met the inclusion criteria after the first screening were transferred to the second screening level with a corresponding unique ID number. Articles that could not be unambiguously excluded after the first screening were also transferred to the second screening to check them again in full text.

The second screening process (full-text screening) followed the same procedure as the first screening process. Ten percent of the articles were screened in parallel by multiple authors to detect and eliminate different understandings on evaluating the eligibility of the articles. Afterward, the remaining articles were divided randomly between the authors. The inclusion results of the first screening were double-checked with the results of the second stage to eliminate inconsistencies between the first and second screenings. During the screening process, information on inclusion criteria was recorded.

The articles identified as relevant and eligible according to the inclusion criteria during the second screening often contained and analyzed more than one econometric model (in terms of, e.g. the estimation method used, the list of individual variables included, etc) for the estimation of the EKCd or FTH. These different estimation models are termed 'studies' in this article. In this context, the term 'model' refers to the estimation method used. The same model might be applied in an article in several studies, which differ, for instance, between the countries considered. From all included articles, the different studies estimating EKCd or FTH were identified, reviewed, and data was extracted. To avoid multiplicity of results, a study selection was performed. Multiplicity of results occurs in this case, when the same set of data is analyzed multiple times using different estimation methods or with overlapping sets of variables, time periods, or countries included (López-López et al 2018). López-López et al propose a framework for dealing with articles that report multiple effect sizes. In this case, the selection of a study for further analysis from a set of studies presented in an article is based on several criteria and involves a decision rule. We determined this rule of study selection before beginning the search. Here, we utilized the reductive approach of López-López et al for all effect sizes, e.g. time periods, regions, and dependent variables. If the same model setup was applied to analyze different regional areas (different sets of countries), all studies meeting the inclusion criteria were included, since in this case there is no multiplicity. If an article uses the same set of variables for short- and long-run models, we chose the long-run studies unless the authors stated that the short-run study delivered better results. If the studies only differed in their estimation methods or the set of variables tested, multiplicity might occur (López-López et al 2018). The main criteria used for study selection were (a) the significance and information of a model and, subsequently, (b) applied background knowledge. In other words, if the authors of an article stated the level of significance or information content of their different studies, we chose the study that the authors described as the best in terms of significance, consistency, and fit. If no statistical criteria were provided by the authors of a respective study on the most preferable models, we used background knowledge to guide model selection. In this sense, and according to the objective of obtaining the most comprehensive compilation of relevant dependent or independent variables for the estimation of EKCd and FTH, we chose the basic study (assuming it was not rejected by the author team of the article) and, in addition, the study that included the largest range of variables of a respective article. Here, we used the integrative approach (López-López et al 2018). For further analyses of the results (e.g. meta-analyses), a further reduction of effect sizes not only within but also between articles might be needed. According to our study selection, all results are displayed per article (n = 46), i.e. each article is only counted once. Only figure 10 displays the estimation methods used per study (n = 141). In cases of ambiguity, we relied on joint, unanimous team decision-making. Since all eligible articles were peer-reviewed and thus adequately fulfill scientific standards, an individual review of the quality of the articles was not conducted.

For data extraction and data coding, templates were developed in Microsoft Excel to systematically record study metadata, analysis outcomes, and contextual information. Data extraction and coding were conducted during both screening stages, but in different templates. Each time, the screening templates were tested and piloted for ten percent of the articles, subsequently discussed, and adopted until a uniform template structure was created to ensure consistent data gathering and entries between the four reviewers. For articles written by members of the reviewer team, we ensured that the respective authors did not screen or code their own work.

The main categories for data extraction and coding included article ID, reviewer, and bibliographic information, and in the first screening, the type of evidence, the regions and periods studied, the forest development theories considered, and the estimation method.

In the second screening stage, the results of the first screening stage were reviewed and adjusted if necessary. In addition, data were extracted on the dependent and independent variables applied in the studies, the assumed curve shapes, the statistical estimation method, and the indication of estimation coefficients. Whether the estimates were applied to a specific forest type was also examined.

The systematic literature search yielded 1317 articles from Scopus, 1167 articles from Web of Science (a total of 2484 articles), and nine supplementing articles from the bibliographies of screened articles. After removing duplicates, we screened the title, keywords, and abstracts of 1633 articles. A total of 1552 articles were excluded after the first screening as these did not meet the respective inclusion criteria. Some articles did not match several inclusion criteria. Five articles did not fulfill the peer-review requirement, two articles were not written in English, and 780 articles did not deal with forests, wood, or trees. A total of 874 articles did not meet our inclusion criteria in terms of regions to be studied, i.e. they studied national, sub-national, or local areas. The estimation condition was not met by 257 articles, and 654 articles did not refer to the concepts of EKCd or FTH. In total, 73 articles could not be clearly identified as either eligible or ineligible during the first screening stage. Their inclusion or exclusion was subsequently decided based on a joint discussion by the author team. Ultimately, 25% of the articles in question were included. After the exclusion of non-relevant articles, 82 articles remained, which were moved to the full-text screening stage. During the second screening process, 36 articles had to be excluded due to missing numerical estimation models, missing output of coefficients, likelihood estimations of forest transition in the FTH, or lack of consideration of a forest-related dependent variable. After the second screening, 46 articles fulfilling all the inclusion criteria remained. The following evaluations of the results and the charts refer to these 46 articles. From these 46 articles, 141 studies were selected as relevant for the analysis of this EGM. Figure 3 shows an adapted flow diagram of ROSES highlighting our results.

Zoom In Zoom Out Reset image size

The 46 articles included in the analysis were published in 32 different journals. Most of the articles were published in the Ecological Economics, Forest Policy and Economics, and the Journal of Development Economics journals. Table 3 in the appendix provides an overview of all 46 selected articles and information on the data sources applied as well as the study specifications.

3.1.1. Extent of publications on different forest development concepts and appearance over time

We found that the majority of articles (76%, n = 35 articles with 90 relevant studies) estimated EKCd, five articles (with 20 relevant studies) estimated FTH, and six articles (with 30 relevant studies) estimated both concepts of forest development (see figure 4(a)).

Zoom In Zoom Out Reset image size

Figure 4(b). shows the number of relevant publications per year. An increasing, yet not constant, trend can be seen in figure 4(b) Most of the articles were published between 2017 and 2020 with a peak in 2017. The joint consideration of both concepts in one paper was only applied from 2011 onwards. Remarkably, in 2017, 50% of the published articles were on EKCd and 50% of the articles examined both forest development theories together.

3.2.1. Temporal coverage

The period length per article is the number of years on which the analyses are based (see figure 5). Taking all estimation approaches together, the articles use a period length of 24.7 yr on average. The individual analysis revealed that the applied period lengths of the various concepts are, on average, close to each other (EKCd: 24.6 yr, FTH: 24.8 yr, both: 25.2 yr). No trend over time is apparent.

Zoom In Zoom Out Reset image size

3.2.2. Geographical coverage

The regions specified by the authors of the selected articles (e.g. global, developing, or tropical cross-country selection) can be assigned to nine country groups (see figure 6(a)). Global data sets are used most often, followed by articles that refer only to developing countries and tropical countries. Few articles select countries from sub-Saharan Africa (SSA) or Asia. In addition, few articles considered either tropical and developing countries together or countries that form an intersection of these groups. In individual cases, articles consider countries that have primary forests, non-OECD countries, and countries located in continental Europe (see figure 6(a)). The world map shown in figure 6(b) illustrates the frequency distribution of the estimated countries. Here, all individual countries considered in the articles are plotted and shaded by their frequency (global estimations are excluded from the presentation as well as six articles without an exact country description). The countries most commonly studied are Bolivia, Brazil, Colombia, Ecuador, India, Ivory Coast, Kenya, Malawi, Malaysia, Nigeria, Peru, Philippines, and Thailand. We also distinguished the country clusters by development concepts and found that FTH articles most often examined global and Asian datasets, EKCd mostly examined global, developing, or tropical countries, and the articles that used both development concepts predominantly looked at developing country datasets. All articles except Turner et al use regional categories to describe their database. However, there are no uniform definitions for the regional categories to which authors assigned their data sets. Further, the regional categories often overlap. In addition, the authors who assigned their database to the global group sometimes examined less than 100 countries. For example, Andrée et al (2019) examine 95 countries, Ehrhardt-Martinez et al (2002) examined 74 countries, and Kahuthu (2006) examined 84 countries.

Zoom In Zoom Out Reset image size

3.3.1. Dependent variable

In addition to the variety of dependent variables used in the articles, the terms are also defined differently. The rate of deforestation is described, e.g. as a change in natural forest area (Bhattarai and Hammig 2004) and as a multiplier between forest cover and deforestation ratios (Leblois et al 2017). Since the definition of forest area is wide-ranging and there is no unified underlying agreement for this, we rely on the definitions in the individual articles. We have assigned comparable variables to six forest-related (figure 7) and two agriculture-related groups. For example, the differences in data set, logarithms, time lag, and forest types remain in the application of the variables. For better clarity of the dependent variables, we assign the variables listed in table 3 in the appendix to figures 7 and 8. To estimate forest area development, the articles use and test eight different dependent variables as representatives of forest development (see figure 8). In figure 7, we distinguish absolute forest area ('forest area'), absolute forest change ('forest area change'), and relative annual change ('annual rate of forest change'). 'Annual rate of forest change' was also expressed by some article authors as 'forest area change rate' or 'forest cover change rate.' Further variables applied are 'forest cover' (relative share of forest area), and change in relative share ('forest cover change'). The same formula applies to the 'rate of deforestation' and 'annual rate of forest change' but with different input data. A distinction is made between gross deforestation and net deforestation. The latter considers the forest development at a certain point in time at which the net forest development is negative and the forest area is decreasing (figure 7).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The results of table 3 in the appendix show that across all articles (n = 46), the dependent variable 'rate of deforestation' occurred most frequently (n = 23). Other dependent variables repeatedly used are 'forest cover' (n = 8), 'annual rate of forest change' (n = 5), 'forest area' (n = 5) and 'arable land' (n = 3).

The dependent variable 'rate of deforestation' was applied 21 times for estimating EKCd, zero times for FTH, and two times for both (see figure 8). Thus, EKCd articles are dominated by the estimation of 'rate of deforestation.' FTH articles (n = 5) distinguish four different variables, with 'forest cover' being estimated in two articles as well as 'forest area' (see figure 8(b)). Four dependent variables are examined in the articles that consider both concepts of forest development. The most frequent specification for articles that estimate both concepts is the 'rate of deforestation,' which refers only to decreasing forest development and the 'annual rate of forest change,' which considers net forest development (each n = 2) (see figure 8(c)).

3.3.2. Independent variables

All selected articles apply sets of independent variables in their estimation models. Even though some independent variables are similar across the different articles, they are specified in slightly different ways. Therefore, we assigned them to 11 thematic classes (see figure 9). The independent variables most frequently examined are economic income variables (applied in 41 articles). The demographic variables contain variables such as gross domestic products (GDP) growth rate, GDP, GDP², and GDP per capita. Demographic variables were applied in the articles just as often; these include population density, rural population as a percentage of total population, and population size. Half of the articles use governmental or other institutional variables (e.g. level of democracy, political rights, and political stability) as well as forest-related (e.g. timber price and labor in forest areas), agriculture-related (e.g. agricultural expansion and cereal yield), and trade-related (e.g. trade openness and export price index) variables in their models (see figure 9). All independent variables tested are shown in figure 9, regardless of whether they were significant or not.

Zoom In Zoom Out Reset image size

3.3.3. Curve shapes

3.3.4. Statistical estimation methods

Overall, 14 different estimation methods occurred across all articles (see table 3). Eight articles used different estimation methods for the different studies included in their analysis. The fixed effects (FE) and ordinary least square (OLS) estimators appeared most frequently (17 times) (see figure 10). In the publications before 2015, 75.8% of the studies used FE and OLS estimates. In the more recent articles from 2015 onwards, the methods became more diverse, and the application of OLS and FE estimation methods decreased to 35.3% of the articles. Seven new estimation methods were added in recent years.

Zoom In Zoom Out Reset image size

The first research question addressed in this EGM targets the evidence base of empirical estimations of forest area development based on EKCd and FTH. By carefully screening the relevant literature since 1955, this article finds that empirical estimations of EKCd were performed more frequently (n = 35 articles), while empirical estimations of the FTH and of both EKCd and FTH were published infrequently (n_FTH = 5, n_both = 6). The joint consideration of both concepts in one publication only began in 2011 and has been examined regularly since then. This aligns with discussions on the similarities and synergies of the two concepts as mentioned, e.g. by Caravaggio (2020b). Investigations on these two forest development concepts have increased in recent years, demonstrating the actuality and ongoing debate relating to these concepts.

The second research question analyzed in the EGM considers the geographical and temporal coverage of articles on the estimation of EKCd and FTH. We observed similarities regarding the regional distribution of the areas studied and the period lengths used for analysis in the different studies. No difference in data usage period lengths could be detected between the two concepts. This is surprising, since the concepts focus on different parts of forest development. Whereas EKCd articles require the analyses to be limited to periods of deforestation, the FTH concept is based on analyzing the entire forest development, even beyond deforestation. The period length is partly determined by the underlying, globally available data sets. But no increase in period length is observed in more recent publication years, although data availability would have allowed this.

The articles considered analyze different groups of countries, where almost all parts of the world were covered. However, certain regions (especially tropical developing countries) were repeatedly studied. Rather than covering a complete region, most articles focused on a subset of countries within one region only (which is mostly located in developing countries). Nevertheless, 35.5% of the articles reviewed referred to global datasets. This is interesting, as only global datasets can avoid a country selection bias. However, although global studies do not necessarily contain all countries of the world, they at least combine a subset of developing and developed country data. For these mixed samples, it might be even more difficult to detect an inverted U-shape for regressions due to greater heterogeneity (as pointed out by, e.g. Choumert et al (2013).

In this context, research gaps occur. For example, developed countries are less frequently studied on an individual basis than developing countries tend to be. In cases where they are considered, these countries are often differentiated by their economic units or by continental boundaries (in global studies). In addition, climate regions or regions unified by trade agreements were not considered in the analyzed studies. Specifications of the concepts for empirical estimates in this direction could be a vital approach for future research.

The third research question examined within the scope of this EGM refers to the specifications of the concepts for empirical testing relating to the choice of dependent and independent variables, curve shapes, and estimation methods. The great heterogeneity in the study specifications is striking, and can be attributed to a lack of theoretical specification in the underlying concepts. Thus, different authors interpret the concepts differently. The variation in independent variables tested as well as in the statistical methods applied is an additional benefit for gaining knowledge about the concepts being considered. These variations (independent variables and estimation methods) are attributed partly to varying data availability and access, as well as the presumed varying statistical skills of the different authors, which are also determined by data and method development over time. Econometric methods and statistical programs have evolved, which can be clearly seen in the more sophisticated methods that have being applied in recent publications. In some publications, basic studies are calculated using the FE or OLS estimators and were then compared to the results of more innovative estimators.

The forest-related dependent variables are variously defined by the authors of the selected articles. Forest definitions are often determined by the datasets applied or are not specified in detail by the authors. This was also observed by Choumert et al (2013), who explained the variation in variable specification by the different data sources and the various official values for forest-related variables. Although the underlying forest development concepts themselves already often lack a precise forest definition, this is, however, indispensable for empirical estimates for practical reasons of data application. Therefore, forest definitions are often practice- rather than theory-based. The greater heterogeneity of study specifications makes comparisons of estimates more complicated. This includes the specification of the shape of the curve, which depends on the chosen dependent variable. For example, when estimating the dependent variable 'rate of deforestation' for EKCd, the assumption of this concept results in an inverted U-shape of the curve, whereas for the dependent variable 'forest area,' a U-shaped curve is obtained. Thus, depending on the different dependent variables, the curve shapes also differ.

By disclosing all decisions and process steps, and by observing the approved protocols, the EGM method attempts to keep selection bias as low as possible. However, the present study omitted gray literature or non-English language articles. By considering peer-reviewed studies only, a potential publication bias exists. This publication bias occurs because peer-reviewed articles usually include statistically significant results (Rosenthal 1979). Therefore, peer-reviewed literature is considered to be only a subset of all the work done in the field and is not a perfect representation. Although we do not anticipate statistically insignificant results on this topic to be published increasingly in gray literature, the risk is more that statistically insignificant studies are likely to be underrepresented in peer-reviewed articles. The bias of ignoring non-English literature is considered to be low since supra-regional articles in particular that are relevant to this study are most likely to be published in English.

The methodological approach of a systematic map shows the research landscape of the topic. The method collects and maps the results but does not further analyze them. Conclusions about the quality of the studies, the dependent and independent variables selected, and the estimation methods cannot be drawn. The method serves as a first step by providing an overview of the comprehensive implementation of EKCd and FTH. For applying further research built on this EGM, e.g. meta-analyses or systematic reviews, it is relevant to additionally consider the potential multiplicity of studies. In this EGM, we reduced the number of effect sizes in the articles, however, multiplicity between articles needs to be considered in further studies.

The literature on quantitative assessments of the EKCd and FTH is extensive. At the same time, the results presented here show that the existence of both concepts and the de-facto policy implications failed to generate consensus. However, it is undisputed that political measures potentially facilitate or slow down forest area change. Thus, effective policy interventions for halting deforestation are important in shaping future forest development.

By displaying whether the hypotheses hold between different socio-economic variables and forest development, this EGM sheds light on research gaps relating to factors that actually affect forest area change. The number of published articles, as well as the broad geographical coverage, are promising approaches for gaining further insights into forest area development research. However, it also indicates that the FTH estimations are employed less often, even though the FTH concept is more far-reaching than EKCd and considers forest area development as a whole. There is a need for a more holistic approach to the concepts. The limited regional focus and period length of less than 30 yr covered by the studies seems very short when studying forest development. Although this EGM is a first step in the further analysis of which factors affect forest development, it already reveals gaps with regard to the limitation of the numerical estimations. The results of this study need to be incorporated in more in-depth analyses, e.g. a systematic review.

Future studies should extend the present work by pointing out the most relevant variables for future forest agenda setting and those policies that could be effective in halting deforestation by targeting certain baseline variables.