TESTING HIGH FREQUENCY DATA COLLECTION IN ARMENIA, TUNISIA AND TURKEY October 2019 ACKNOWLEDGMENTS This report was prepared by a World Bank team composed of Emilie Perge (Senior Economist, Poverty & Equity GP) and Raisa Behal (Consultant, Poverty & Equity GP) with inputs from Jonathan Kastelic (Survey Specialist, Development Economics and Chief Economist), Shiva Shivakumaranand (Consultant, Poverty & Equity GP), Shinya Takamatsu (Consultant, Poverty & Equity GP) and Mohamed Mukhtar Qamar (Consultant, Macro Trade & Investment GP) with overall guidance from Sofia Ahlroth (Senior Economist, Environment GP), Werner Kornexl (Senior Natural Resources Management Specialist, Environment GP), Garo Batmanian (Lead Environment Specialist, Environment GP). The team would like to thank Craig Meisner (Senior Environmental Economist, Environment GP), Taoufiq Bennouna (Senior Natural Resources Management Specialist, Environment GP), Gillian Cerbu (Natural Resources Management Specialist, Environment GP), Cesar Cancho (Senior Economist, Poverty & Equity GP), Facundo Cuevas (Senior Economist, Poverty & Equity GP), Ani Avetisyan (Consultant, Environment GP), and Farah Hathout (Consultant, Environment GP) for their help to access data and their logistical support in the fieldworks. The team wants to thank all households who took part in the surveys in Turkey, Armenia and Tunisia. The team is very grateful to the participants of the FLARE conference in Stockholm in October 2017 and to the participants of a reference group for their comments and suggestions on the tool. This work was made possible thanks to a PROFOR grant (TF0A2918). i EXECUTIVE SUMMARY THREE TESTS OF FOREST-SWIFT •Consumption data from Household Budget Survey 2013 •Forest income data from Socioeconomic Household Survey 2016 Turkey •Poverty at 23.2% of forest village population in 2017 •Forest dependence at 19% in 2017 •Consumption data from Income and Living Conditions Survey 2015 •No baseline data on forest income Armenia •Armenia Forest, Energy and Poverty Survey 2018 collecting forest income •National poverty in 2018 at 39.5% of thepopulation. •Consumption data from Living Condition Survey from 2015 •No baseline data on forest income but descriptive information from an FAO study 2012 Tunisia •Tunisia Forest, Land Degradation and Poverty Survey collecting forest income •Poverty in 2019 at 16.6% of the population in rural areas of Center-West and North-West MAIN ADVANTAGES OF FOREST-SWIFT •LSMS-type surveys key to the implementation of Forest- Useful tool to SWIFT complement •Forest-SWIFT to fill the gap betwee two rounds of survey household surveys •Forest-SWIFT to provide more regular data on forest dependency and poverty •Predicted estimates of poverty are comparable to available Strong measures measures. of poverty and forest dependence •Forest dependence comparable to the ratios reported in other studies. Short, timely, and •As a standalone, Forest-SWIFT questionnaires administered working well as in less than 20 minutes part of a longer •Forest-SWIFT questionnaires easy to embed with the longer questionnaire questionnaires •To oversample forest areas to have good and reliable Innovative estimates for forest people. sampling design for data collection •Sample of forest areas only, forest versus non-forest, or forest and larger administrative areas. ii MAIN CHALLENGES OF FOREST-SWIFT •Baseline data for consumption are not representative of Baseline data not forest areas. available •Lack of forest income data limiting the full test of Forest- SWIFT. •Low participation in forest-related activities preventing from computing robust forest income and predicting forest Prediction of income using Forest-SWIFT forest income •Unable to model negative income values and the choice made towards modeling gross forest income. •No master frame, built from population census data to be Sample frames used to define forest stratum built using ad-hoc definitions of •Using satellite images on forest cover with population forest stratum. density data assuming that grids close to forest being in forests. iii 1 CONTENTS Acknowledgments ................................................................................................................................................. i Executive summary .............................................................................................................................................. ii 1 Introduction ................................................................................................................................................... 1 2 Poverty and forests...................................................................................................................................... 4 3 Forest-SWIFT methodology ................................................................................................................... 6 4 Data requirements ....................................................................................................................................... 9 4.1 Poverty data........................................................................................................................................... 9 4.2 Forest income data........................................................................................................................... 10 5 Piloting Forest-SWIFT ............................................................................................................................. 11 5.1 Turkey ................................................................................................................................................... 12 5.2 Armenia ................................................................................................................................................ 19 5.3 Tunisia .................................................................................................................................................. 23 6 Advantages and challenges using Forest-SWIFT .......................................................................... 28 7 Conclusion .................................................................................................................................................... 30 8 References .................................................................................................................................................... 31 Appendix ............................................................................................................................................................... 34 List of figures Figure 1 Kernel density distribution for consumption ....................................................................... 15 Figure 2 Kernel density distribution for forest income ...................................................................... 15 Figure 3 Kernel density distribution for consumption ....................................................................... 21 Figure 4 Kernel density distribution for consumption ....................................................................... 26 List of tables iv Table 1 Forest-SWIFT models ....................................................................................................................... 14 Table 2 Participation in forest and non-forest activities (percent)................................................ 17 Table 3 Results for imputed poverty rate and consumption per capita (Rural) ....................... 17 Table 4 Results for Imputed Forest Income and ratio below median forest income .............. 18 Table 5 Forest-SWIFT models ....................................................................................................................... 20 Table 6 Results for imputed poverty rate and consumption per capita ....................................... 23 Table 7 Forest-SWIFT models ....................................................................................................................... 25 Table 8 Results for imputed poverty rate and consumption per capita ....................................... 27 v 1 INTRODUCTION Forests can contribute significantly to the livelihoods of households living in their vicinity. Forest resources provide land and products destined for markets or self- consumption, and even materials for shelter (Byron and Arnold, 1999; Vira et al., 2015). Using the Poverty Environment Network (PEN) household-level data spanning 24 tropical developing countries on the role of environmental income in rural livelihoods (Wunder et al., 2014), Angelsen and colleagues (2014) estimated that around 20 of households’ income come from forest through extraction, processing forest products, wage activities, and other income. But strong measurement of forest income is still at its early stage. The Poverty and Environment Network1 data were a novelty in measuring forest income across multiple sites using quarterly surveys. For each forest products present in the site, households were asked how much they collected, they sold, and at what value. These survey tools were later used as a basis for the creation of the Forestry modules by the FAO, CIFOR, IFRI, and the World Bank. In these modules, one can measure forest income through the production values of forest products, forest wages, other payments for forest ecosystem services. These modules can be used as standalone or embedded in other surveys. And there is very little assessment of poverty within forests because of sampling design and budget constraints. While information on poverty levels and trends at the national level are still lacking (Serajuddin, et al. 2014), this information is quasi-inexistent at the forest level. National household surveys used to estimate poverty are time consuming and very expensive. Unlike population censuses, such surveys are conducted on representative samples stratified to increase the precision for consumption or income estimates, and other indicators of interest (Deaton 1997). These samples are always representative at the national, urban, and rural levels while forest areas are almost never 1 The PEN survey covers about 8,000 households in 24 countries across Sub-Saharan Africa, South and East Asia, and Latin America, and is representative of smallholder-dominated communities leaving close to forests (with access to forest resources). 1 sampled extensively to be representative, primarily because of the cost and effort of reaching forest communities. Forest-SWIFT is a cost-efficient and evidence-based tool designed to fill these gaps in the data on poverty and forest-dependence. An extension of the Survey of Well-being via Instant and Frequent Tracking (SWIFT), both tools build mini-surveys to estimate the welfare aggregate of choice using an econometric model developed using baseline data on the sample of interest. The ingenuity of SWIFT is that it does not collect any consumption or income data from households, only the 10-15 correlates identified in the baseline model. As such, Forest-SWIFT can measure poverty and forest dependence without multi-module questionnaires on household income and consumption allowing more frequent data collection on robust indicators that are comparable to national measures. Frequent and reliable data can help various actors, from local officials to development practitioners, identify effective policies, design more efficient programs, improve beneficiary targeting, and monitor and evaluate poverty targeting programs. The objective of this note is to present Forest-SWIFT, its applications and challenges. As of now, Forest-SWIFT has been piloted in three countries: Turkey, Armenia, and Tunisia, chronologically. Although SWIFT and its consumption estimation has been widely tested and adopted by various teams across the World Bank and the IFC, the forest-dependence component of Forest-SWIFT was still in a nascent stage of development and merited low- cost testing. As a result, the pilot sites were selected out of opportunity to team up with other projects, and availability of baseline data. Forest-SWIFT requires baseline data on both poverty and forest-dependence. To build the models and identify the necessary questions for the mini-survey, Forest-SWIFT requires data on poverty measured through consumption or income and on forest income. In the three countries where Forest-SWIFT was tested, the team had access to poverty data but only Turkey had forest income data. The data in Turkey come from two sources with poverty data coming from a nationally representative household budget survey and the forest income data from a socioeconomic survey in forest villages of Turkey. In the other cases, the team has had access to consumption data but no forest income data. The team had collected 2 a full-length forest income survey with a short module on SWIFT for poverty. The team has built the forest income model after the Forest-SWIFT survey. Forest-SWIFT present important advantages - Useful tool to complement exhaustive household surveys. LSMS-type survey to collect data to compute poverty or forest income are key to the implementation of Forest-SWIFT. Forest-SWIFT intends to fill the gap in monitoring poverty and forest dependence. - Strong measures of poverty and forest dependence. Predicted estimates of poverty are comparable to available measures and the examination of the trends (Turkey) and levels (Armenia and Tunisia) confirms that our predicted estimates are good. Forest dependence is also comparable to the rate of forest dependence reported in other studies (Angelsen et al. 2014). - Short, timely, and working well as part of a longer questionnaire. As a standalone, Forest-SWIFT questionnaires were administered in less than 20 minutes in Turkey. In the other countries, Forest-SWIFT questionnaires were easy to embed with the longer questionnaires created to collect data on energy (Armenia) or land degradation (Tunisia). - Innovative sampling design for data collection. In the tests provided here, the choice was to oversample forest areas to have good and reliable estimates for forest people. In some cases, the sample was restricted to forest households (Turkey), to forest and higher administrative levels (Armenia), or to forest and non-forest with higher administrative levels (Tunisia). But Forest-SWIFT faced some challenges. - Baseline data not available. Baseline data for consumption are not representative of forest areas. Forest income data were also missing which limited the full test of Forest-SWIFT. The use of forest income model will be part of a follow-up survey done by the respective teams in their project. - Prediction of forest income. The low participation in forest-related activities prevents one from computing robust forest income and predicting forest income 3 using Forest-SWIFT. The model is also unable to model negative income values and the choice was made towards modeling gross forest income. This choice was also justified by the fact that it was very hard to collect data on the input costs of forest collection for the last 12 months. - Sample frames built using ad-hoc definitions of forest stratum. To our knowledge, there are no master frame, built from population census data, that can be used to define forest stratum and decide which cluster is in a forest stratum and which cluster is not. Using satellite images on forest cover with population density data was the preferred way while making assumptions on which grid is forest and which grid is not (here the distance). The report is presented as follows. The following section 2 presents the importance of having data on poverty in forests. Section 3 explains Forest-SWIFT methodology while section 4 describes the data needed to implement Forest-SWIFT. Section 5 presents the results from three tests: Turkey (subsection 5.1), Armenia (subsection 5.2) and Tunisia (subsection 5.3). Section 6 describes the main advantages and challenges faced when developing and implementing Forest-SWIFT. Section 7 concludes. 2 POVERTY AND FORESTS Forest resources to household livelihood strategies and welfare. It is estimated that about 1.3 billion people—nearly 20 percent of humanity—rely on forests and forest products for their livelihoods (FAO et al., 2016). Forest people can either live in forest areas and live as hunter-gatherers or shifting cultivators; these people rely greatly on forest resources to fulfil their basic food needs. Another group of forest people are people living near forests whose primary activity is agriculture but use forest resources to fill their consumption and income gaps. The last category is people using forest resources to engage in the cash economy through timber logging, or the collection of minerals; these individuals mainly depend on income from forest-related activities rather than subsistence use activities (Byron and Arnold 1997). Households rely on a wide range of forest resources for their livelihoods. Households extract a variety of foodstuff, such as mushrooms, wild fruits, roots, honey, and medicinal 4 products from forests, some of which leads to commercialized and profitable harvesting. Forest resources also contribute to improved nutritional outcomes of poor households by supplying fuelwood for cooking. Trees on farms have also been proven to increase agricultural production by conserving soil nutrients, regulating water flows, and protecting against pests (Byron and Arnold 1997; Angelsen and Wunder 2003; Vedeld et al., 2007; Debela et al., 2012; Angelsen et al, 2014). Through a variety of means forest resources can be an important source of income. A global comparison using the Poverty-Environment Network (PEN) data found that income from forests represent between 20 percent (average Asian sites) to 28 percent (average Latin American sites), which is as important as agricultural income (crop and livestock), and wage income (Angelsen et al. 2014) in relation to average contributions to total income in these regions. Most of the contribution from forests to the livelihoods of the poor is not in the form of cash (Shepherd, 2012) but rather in products and services that are consumed by households. Nonetheless most of households living in or near forests are poor. Sunderlin and colleagues (2007) and Shepherd (2012) have observed that there was a huge overlap between areas with high forest cover and poverty pockets. As being remote, forest areas are poorly connected to markets to sell their agricultural and forest products. Also, forest households have little access to labor markets that could help them to diversify their livelihood strategies into off-farm off-forest activities (Sheperd 2012). While little is known regarding forest’s contribution to household’s movements out of poverty. Some case studies have reported that households can sustainably improve their welfare using forest resources (Godoy et al., 2009; Perge, 2010; Perge and McKay, 2016). The gap in dynamic forest-poverty studies is seen by the lack of panel data exploring factors explaining movements into and out of poverty and what role forests play in these dynamics. Existing studies on poverty dynamics in forest areas are mainly highly localized case studies using non-comparable measures. There exist multiple mechanisms through which forests can contribute to reducing poverty. A recent review of the literature identified five pathways to guide investments in 5 people, resources, and institutions so that forests can contribute to poverty reduction ( (Shyamsundar, et al. forthcoming). Namely these pathways are: (1) a) improvements in productivity (P) of forest land and labor; b) governance reform to strengthen community, household and women’s rights (R) over forests and land; c) investments (I) in institutions, infrastructure and public services that facilitate forest-based entrepreneurship; d) increased access to markets (M) for timber or non-timber forest products; and e) mechanisms that enhance and enable the flow of benefits from forest ecosystem services (E) to the poor (Shyamsundar, et al. forthcoming). This framework sheds light on the importance of synergetic investments and pre- requisite conditions. The pathways enounced above are intertwined and non-exclusive. For instance, if one wants to improve market access for the sales of non-timber forest products (NTFPs), one needs to ensure that forest households have the rights to extract and sell such resources. If so, the second pathway become a pre-requisite. If not, investments have to be made in this pathway. Multiple interactions can be identified which complicates the result chains and measurements of these investments. Monitoring whether the impact on poverty is achieved through these interventions requires more frequent data on poverty. Forest investment projects often are focused on measuring outputs without a true reflection on how to make sure these outputs will achieve the poverty impacts mandated by the World Bank Group. Such lack of strong monitoring of poverty results, as explained in introduction, from the prohibitive costs (financial, human, logistics) of doing a poverty survey even at the project level. However, only with a strong measurement of poverty within the projects one can affirm that forest investments can contribute to poverty reduction. The methodology tested here intends to fill this gap and bring forward forests into the national development strategies. 3 FOREST-SWIFT METHODOLOGY Forest-SWIFT is a data collection method developed to provide timely, quick, and accurate data on poverty and forest dependence through a small set of country-specific questions. As an extension of the Survey of Wellbeing via Instant and Frequent Tracking (SWIFT), a methodology developed by the World Bank to estimate poverty incidence between 6 consecutive comprehensive household surveys (Ahmed et al., 2014; Yoshida et al., 2015), Forest-SWIFT additionally tracks forest dependence, defined as the forest income share of permanent household income. Although the model can use either consumption or income as the welfare measure, the country preference takes precedence to maintain comparability. Forest-SWIFT develops country-specific models for each indicator, which often requires different variables per model. Each model assumes a linear relationship between household total consumption/income, or forest income (ℎ ) and their correlates (ℎ ) with a projection error ℎ The inclusion of this error term differentiates this model from other predictive tools ℎ = ℎ ′ + ℎ (1) Forest-SWIFT estimates the log transformation of the dependent variable to smooth asymmetries and normalize the distribution of the variable, making it easier to estimate. Forest-SWIFT controls for issues linked to over-fitting – when a model performs well within the sample but poorly outside the dataset – by cross-validating the model (Kuhn and Johnson, 2013). The purpose of cross-validation in Forest-SWIFT is to identify the optimal level of significance- or p-value- in the model, which would balance the number of determinants and the goodness of fit across the sample. Cross-validation consists of two steps: (a) splitting the sample in n-folds and running the model in n-1 folds and testing it on the nth fold 2 and (b) running multiple models per fold, testing various thresholds of significance for model variables. This process of ‘stepwise’ selection entails adding variables to the Ordinary Least Square (OLS) model sequentially if they bring enough information, and simultaneously removing them if they do not. Each fold has a chance to be the testing data and this process is repeated n times by changing the nth fold each time. The optimal p-value performs best in terms of mean-squared errors3 between actual and projected welfare, and the absolute value of the difference between the actual and projected poverty (or forest- dependence) rates. This concludes the cross-validation process and the stepwise OLS 2 We pick ten folds, but it could be any number of folds. 3 The average of the sum of squared differences between ℎ and ′ ^ℎ = ℎ ^ ∗ 7 regression is run a final time on the full sample of data using the selected p-value. The resulting regression is the SWIFT poverty model. To ensure the quality and robustness of the models, Forest-SWIFT carries out two tests (if data are available): backward imputation and validity test. The former applies the final model to a previous round of data to check the stability of the model over time. The latter tests whether the error term follows a normal distribution using a simulation method developed by Elbers, Lanjouw and Lanjouw (2002, 2003). The result is a small set of questions that considerably simplifies data collection and encourages teams to collect data quicker and more frequently than traditional lengthy household surveys. Forest-SWIFT data are collected through CAPI (Computer Assisted Personal Interviews) to ensure quick analysis and results. After creation of the questionnaire, the last phase of Forest-SWIFT is to predict consumption and forest income based on the coefficients from the models. In this final prediction phase, Forest-SWIFT utilizes multiple imputation estimations to apply the coefficients from the respective models to the variables in the new dataset. Random error is simultaneously introduced by adding 1000 imputations with error per household estimate. SWIFT does not provide a single point estimate for consumption or forest income per household but flags the consumption estimates in each fold that are below or above a threshold. While SWIFT cannot identify a household as uniquely poor or non-poor, it allows one to create an indicator that flags whether each consumption estimate, of the 1000 per household, is above or below the poverty line. To calculate the poverty incidence in the sample area, SWIFT uses all 1000 imputations per household, across all households (1000 * NH) and estimates sample poverty using the multiple imputation estimate4 program. Similarly, Forest-SWIFT estimates forest dependence as the ratio of forest income and consumption, per imputation. 4 https://www.ssc.wisc.edu/sscc/pubs/stata_mi_estimate.htm 8 4 DATA REQUIREMENTS As explained above, Forest-SWIFT requires baseline data in terms of poverty and forest dependence. The following subsections explain how these measurements can be obtained. 4.1 Poverty data Poverty can be estimated using a monetary aggregate of either consumption or income. National poverty rates are calculated either with income (most countries in the LAC region) or with consumption. Using either aggregate requires long questionnaire with more than 80 questions. In the poorest countries, preference is given towards consumption since income fluctuates substantially introducing a bias in poverty estimation (Dercon and Krishnan 2000). A consumption aggregate is less volatile than income as households tend to smooth consumption to fulfill basic needs (Dercon and Krishnan 2000). In all the countries where Forest-SWIFT has been tested, poverty is measured using a consumption aggregate normalized to be per capita. Building a per capita consumption aggregate is a four-step process: • The first step is to calculate total annual household consumption. Consumption aggregate is composed of food and non-food consumption. Food consumption is disaggregated into i) the value of purchased food; and ii) the implicit value of non- purchased food, including household production, food aid, and other donations. Non- food expenditures comprise current and occasional expenditures, including (imputed) rent and spending on durable goods and services, such as utilities, health and education. • The second step is to remove the effect of inflation from household consumption. Households can be surveyed at different time periods, and during this period the consumer prices (CPI) including price levels of food can fluctuate. This adjustment does not often extend beyond food expenditure, since the reference period for non- food expenditure of annual and price levels of durable goods are stickier and do not fluctuate across months. • The third step consists of normalizing household’s total consumption to represent per-capita, or per-adult equivalent, household consumption instead. The per-capita 9 adjustment grants equal value to each member of the household i.e. assumes a 5-year- old child consumes as much as a 30-year-old adult, whereas per-adult equivalent adjustment accounts for differences in the household composition and weights consumption by age and gender. The method of normalization varies by country. • The fourth step is to divide per capita (per adult equivalent) consumption by a spatial deflator, which adjusts for differences in the cost of living between regions and, when possible, across rural and urban geographies, which affect various transaction costs such as supply and transport. Using a poverty line (national or international), one can build poverty headcount, gap, and depth. A national poverty line, which is preferred for within country comparisons, is built using a basket of goods consumed by surveyed households, including a selection of food items that provide a minimum calorie requirement. The poverty headcount is the fraction of the population whose per-capita consumption aggregate is below the (per-capita) poverty line. The poverty gap, which represents the population’s average shortfall between consumption and the poverty line, is a very useful measure to track the depth of poverty. 4.2 Forest income data Forest income is typically overlooked in traditional multi-topic household surveys that are used to compute poverty. Traditional household surveys in countries with large forest cover (Brazil, Democratic Republic of Congo, …) do not include questions on forest activities. Consequently, specificities of forest households’ livelihoods are neglected when designing and implementing national policies to pull them out of poverty. Forest dependence is often computed using case studies, which select different methodologies and yield incomparable results. For example, studies usually include some combination of the following forest related incomes: value from harvested products, value from planted products, forest related wages, payments for ecosystem services. The PEN initiative, discussed earlier, is one of the first to collect comparable data on forest activities across 24 countries (Wunder et al., 2014). To fill this data gap, the Forestry modules were launched in November 2016 to be used either as a standalone or as part of a larger questionnaire. The Food and Agricultural 10 Organization (FAO), the Center on International Forestry Research (CIFOR), the International Forestry Resources and Institutions Research Network (IFRI), and the World Bank designed a comprehensive set of questions to allow governments, practitioners, and researchers to robustly assess the livelihoods of forest households across various value generating activities, forest tenure, and access to forest resources (FAO et al., 2016). Thanks to this module, national policies and programs can consider the specificities of forest households when designing their policies. The Forestry modules are composed of community and household questionnaires that either make up the core standard module or are included in optional extensions. The standard community questionnaires identify the most important forest products for the community, annual harvesting schedules, quantities collected and pricing, among other additional community benefits from forests (FAO et al., 2016). The extended community questionnaires consider governance and forest institutions, community environmental services. The standard household questionnaires encompass questions on forest income, forest resources used for energy, health, and construction, and food shortages. The extended household questionnaires are about forest clearance and changes. The flexibility and customizability of the Forestry Modules make them a valuable addition to the forest data collection tool-kit. Information on forest income comes from multiple sources of income. Forest income includes benefits from extraction of timber and non-timber forest products that are sold on the market or consumed by the household, forest-related wage and non-wage activities. Such data can be disaggregated by cash and non-cash sources of income and be used to estimate the value of forest extraction. To robustly estimate forest income, information is also collected on inputs and costs. 5 PILOTING FOREST-SWIFT Forest-SWIFT can complement traditional household surveys collecting consumption and forest income. Data to compute a consumption aggregate, usually collected by National Statistic Offices, are not collected more than every few years. The Forestry modules, despite being a much-needed tool to collect forest data, has had limited uptake, and has been 11 implemented only once in a handful of countries (Turkey, Sao Tome e Principe…). Forest- SWIFT fills this additional gap in the frequency of tracking, by estimating poverty as well as forest income, between rounds of more comprehensive household surveys. To pilot Forest-SWIFT, three countries were selected after discussions with task-team leaders. During early stages of the project, the team had chosen to work in Turkey, Argentina, and Mozambique based on preliminary discussions with task-team leaders (TTLs) and existence of investment projects in the countries. However, Argentina and Mozambique dropped out and did not pursue Forest-SWIFT. Argentina had been collecting detailed household data as part of an ongoing investment project and could not pursue additional data collection. In the end, the pilots were conducted in Turkey (2017), Armenia (2018), and Tunisia (2019). Data requirements were not met in two out of three countries. Unfortunately, forest income data were non-existent in Tunisia and Armenia. To overcome this issue, both forestry module and Forest-SWIFT were implemented at the same time. With this newly collected forest income data, we built models ex-post that can be used by the teams in the Environment Global Practice (ENR GP) for future projects. 5.1 Turkey Turkey has all data requirements to conduct the first test of Forest-SWIFT. Not only is the national Household Budget Survey (HBS) collected annually, but also a fellow team in the ENR GP collected forest Socio-Economic Household Survey (SEHS) in 2016, thus providing us with recent data on both consumption and forest income. Although Turkey discontinued urban and rural disaggregation in 2014, the 2013 survey was still available and could be used to build the consumption model. The SEHS 2016 had extensive questions on harvested forest products, forest wage income, and payments for environmental services, and more which allowed the ENR GP to compute a robust forest income aggregate. Sampling designs in the two surveys were different. While the 2016 SEHS is representative of forest villages with a sample of 2,037 households and collects data on forest participation, income, and forest products, the 2013 HBS is nationally representative 12 and has 10,058 households with 36,812 individuals.5 The 2013 HBS collects information about household composition, consumption, assets, employment categories, and education. Although the data includes urban-rural classification, regions remain unidentified. In the 2016 SEHS, forest income is measured as the total monetary value of forest product extractions, wages from forest employment, and payments from forest services. Although the survey asked about a wide range of forest products, only those that were collected by at least 10 households in the sample were included in the analysis. As such, only income from the following 16 forest products are included: firewood, mushroom, herbs, thyme, sage, hazelnut, linden, sting nettle, walnut, rosehip, pinecone, chestnut, industrial wood, blackberry, trefoil, opium. Despite only 15 observations, forest payments were included in the forest income aggregate. On average, households diversify their livelihoods with 60% of households having three or more sources of revenue (PROFOR, 2017). At the same time, 15 percent of households specialize, in agriculture or livestock, 10 percent solely in forest-related activities, while 11 percent of households combine forest-related activities and agriculture or livestock (PROFOR, 2017). Researchers in this report find that when using a relative poverty line equal to 60 percent of the median income in the villages (TL 480 per capita per month in 2016 about US$12 per capita per day in 2011PPP), 40 percent of sample households fall below this line (PROFOR, 2017). Household characteristics differ across rural households in 2013 HBS and forest households in 2016 SEHS. First, forest households tend to be larger than rural households, due to a larger share of prime-age adults (i.e. fewer dependents such as children and elders). Second, forest household heads are mostly men and better educated than their rural counterparts. Third, labor force participation of household heads or adults is lower in forest villages (table 1 in appendix). However, these differences are not significant and the observations in these two datasets are comparable. We can therefore confidently develop the model with these two surveys. 5 Here we are using World Bank’s 2013 ECAPOV Turkey dataset. ECAPOV- the World Bank’s harmonized poverty database for Eastern and Central Asia- harmonized Turkey’s 2013 Household Budget Survey (HBS) to create a number of standard indicators that can be compared across countries in the region. These indicators have been then published in the World Bank open data portal and World Bank World Development Indicators (WDI) http://databank.worldbank.org/data/reports.aspx?source=2&Topic=11 13 Consumption model with HBS 2013 The consumption model was designed to estimate the log per-capita expenditures for the rural sample. The estimation was a regression of this dependent variable on household characteristics, such as household composition, assets, employment categories, and education. The poverty rate among rural households was estimated at 34.9 percent using a poverty line of US$7 (in terms of 2011 PPP). Turkey being an upper-middle country, the poverty line had to be set at US$7 a day to ensure that the poverty rate is high enough to calibrate the models with precision. Lower lines (US$1.90 or US$3.20 a day in 2011 PPP) would give poverty rates that are too small for such a precision. Table 1 Forest-SWIFT models Consumption Forest income Age of HH head 0.006 (0.001) HH head is married (=1; 0 if not) -0.133 (0.046) HH head education is primary (=1; 0 if not) 0.137 (0.048) HH head education is secondary (=1; 0 if not) 0.231 (0.056) HH head education is tertiary or higher (=1; 0 if not) 0.393 (0.075) Head is employed (=1; 0 if not) 0.312 (0.047) Number of retired 0.347 (0.04) HH head is an employer (=1; 0 if not) 0.264 (0.091) HH head employed in agriculture (=1; 0 if not) -0.139 (0.029) HH with members working as unpaid worker in family -0.995 (0.296) business HH receiving forest wages 2.617 (0.184) Household size -0.071 (0.01) -0.261 (0.034) Dependency ratio -0.423 (0.044) Household having central heating system (=1; 0 if not) 0.305 (0.051) Non-overcrowding (=1; 0 if not) 0.23 (0.041) HH owns computer (=1; 0 if not) 0.272 (0.028) HH owns tractor 0.359 (0.112) HH owns chainsaw 0.292 (0.114) HH owns car/truck -0.229 (0.110) Village electrical network 0.249 (3.47) Village water network 0.143 (-4.14) Village with net migration and high poverty 0.279 (-3.96) Village with no net migration and high poverty 0.127 (-4.85) Constant 8.15 (0.11) 5.304 (0.285) Number of obs 3006 1155 R-squared 0.439 0.352 Root MSE 0.524 1.535 Note: non-overcrowding is defined as less than 1.5 persons per room. Standard errors in parenthesis. All variables are significant at 0.005 in consumption model and at 0.01 in the forest income model. 14 Source: authors’ estimates using HBS 2013 and SEHS 2016 Following the SWIFT methodology, an evaluation of the mean-squared errors of the consumption estimates, and the absolute differences in the poverty estimates yielded an optimal p-value of 0.005 for the consumption model. The final stepwise regression selected 14 out of 23 potential explanatory variables. Table 2 lists the final regression for the poverty model. The kernel density plots for both original and measured values confirm that the estimated consumption density across the population is like the original (Figure 1). Figure 1 Kernel density distribution for consumption .6 kdensity logwelfare .4 .2 0 4 6 8 10 12 x kdensity logwelfare Original kdensity logwelfare imputed Source: authors’ estimates with 2013 HBS Forest income model with SEHS 2016 The forest income model was designed to estimate the log of per capita forest income on a similar set of household characteristics. For cross-validation purposes, the median per capita forest income serves as the threshold in place of a poverty line. The model only includes households who report non-zero and positive forest-related income from any source, which represents around 60 percent of the original sample. The final model selects 10 variables out of 25 using a p-value 0.01 (table 2). The following kernel density plots confirm the matching distribution of the income estimates. Figure 2 Kernel density distribution for forest income 15 .2 .15 kdensity lnpc_fbinc .1 .05 0 -5 0 5 10 15 x kdensity lnpc_fbinc kdensity lnpc_fbinc _mi_m=7 Source: authors’ estimates with 2016 SEHS 2017 Forest-SWIFT survey and results These two models led to the creation of 20-minute questionnaire. The questionnaire focused on the determinants identified in the two models discussed above and included 20 questions about forest collection and wage, dwellings, and assets. The 2017 Forest-SWIFT survey was administered to a sub-sample of the 2016 SEHS (1000 households across 100 out of the 202 villages) and thus also form an unbalanced panel dataset. The sample remains representative of national forest villages. The survey took place over a three-week period. GPS coordinates, cell phone numbers from respondents were recorded to monitor survey progresses and to check quality of answers on collection of forest products. Participation in forest related activities remains high in 2017. Nearly all households in the sample do forest-related activities with 83 percent of households generating income values from these activities and 77 percent extracting forest products for their home- consumption (table 2). The most important source of forest income remains forest-wage activities. The forest products collected the most were walnut, hazelnut, linden, and herbs, while hazelnut and walnut were the most likely to be commercialized (table 2 in appendix). On average, 37 percent of households collected between 4 and 9 different forest products but only 14 percent of households sold at least 1 forest product; households mainly collect forest products for their home-consumption (table 3 in appendix). 16 Table 2 Participation in forest and non-forest activities (percent) Participation (% of households) FOREST-RELATED 94.78 Income from Forestry and/or NTFP Production 83.20 Wage 82.35 Market Sales from Collections 14.06 Other, Unidentified 0.34 Subsistence Value from Collections 76.85 NON-FOREST RELATED INCOME 96.57 Wage Agricultural Income 59.12 Pensions 70.68 Other, Unidentified 47.99 Source: authors’ estimations using 2017 Forest-SWIFT. Note: Results are weighted at the household level Forest households in 2017 are better-off than rural households in 2013, but one out of four households are poor. SWIFT consumption estimates in 2017 are higher than consumption values in 2013 (table 3).6 Over time poverty appears to have decreased, which is also reflected in the national poverty rate (Cuevas and Rodriguez-Chamussy 2016). However, absolute poverty levels are not comparable between PROFOR (2017) and this analysis due to a difference in welfare aggregate (Forest SWIFT estimated consumption, while PROFOR 2017 measured income) and poverty lines. Table 3 Results for imputed poverty rate and consumption per capita (Rural) HBS 2013 (original) SWIFT 2017 Poverty rate (%) 34.9 23.2 Mean per capita Consumption (TL) 5,906 6513.206 Note: Consumption values in HBS 2013 and in Forest-SWIFT 2017 are all in Turkish Lira 2013. The poverty line equal of US$ 7 per capita per day was converted to TL 2013 using 2011 PPP. Source: authors’ estimations using HBS and Forest-SWIFT data. Forest income is greater in 2017 than in 2016. In 2017, forest income is on average equal to TL 1223.82 in 2016 prices (table 4). Households appear more likely to participate in forest-related activities given higher response rates in 2017 than in 2016, this can also be 6 Since values in the original model are in Turkish Lira of 2013, the resulting projected consumption aggregate is also in 2013 values. Deflating these values to PPP of 2011, poverty rate for forest households is lower than poverty rate in the original dataset. 17 attributed to the change in questionnaire design and length. While the original questionnaire (2016 SEHS) collected data on 90 forest products, the follow up survey focused on a subsample of forest items, potentially reducing respondents’ fatigue. The increased response and accuracy is also reflected in higher forest incomes across households. Table 4 Results for Imputed Forest Income and ratio below median forest income SEHS 2016 Forest-SWIFT 2017 Ratio below median forest income (%) 50.02 46.70 Mean log per capita forest income 4.631 5.003 Mean per capita income 893.49 1223.82 Forest dependence (percent) n.a. 14.71% Note: Forest incomes with SEHS 2016 and Forest-SWIFT 2017 are in Turkish Lira 2016 Suggested formula for inverse natural log = Exp(m+ sigma (m)^2/2) Forest dependence is equal to the average household forest income over household average consumption that is a proxy of permanent income. Source: authors’ estimations using HBS and Forest-SWIFT data. Forest income contributes to 19 percent of households’ total income. Using the predicted values of consumption and of forest income, the average relative contribution of forests to total consumption is approximately 15 percent. This result is comparable to the levels of forest dependence reported by Angelsen and colleagues (2014) in their global study. It is evident that even though forests may not be their main source of income, households depend on this source for their livelihoods. Poor forest households are more dependent on forest income for their livelihoods than non-poor forest households. On average, poor forest household generate 30% of their income through forest-related activities. Similar to Angelsen et al (2014), we found that poor households’ relative benefit from forest income was higher than non-poor households in terms of forest-income share of household consumption, although their absolute benefit was lower i.e. they received lower monetary benefits This first pilot in Turkey was very instructive for the next two pilots. The team used the takeaways from the Turkey pilot to inform and improve the next pilots, especially limiting respondents’ fatigue when asking forest questions by asking fewer questions about fewer forest products. The pilot provided insights on how to build and use the model to predict consumption and forest income. Finally, this pilot demonstrated the need to have data on 18 households living outside forests to be able to illustrate differences between forest and non- forest households. 5.2 Armenia Armenia has very recent poverty data but lacks forest income data. The poverty data comes from the 2015 Income and Living Condition Survey (ILCS). These data were shared by the Poverty GP team through the ECAPOV platform. In this data, poverty is measured using a consumption aggregate per adult equivalent resident household members, who had spent at least 1 day in the survey month in the household. This aggregate accounts for both scale effects within the household and the differential in needs between the different demographic within the household. The ILCS is a multi-topic survey with questions on household composition, jobs, migration, dwellings, income, and consumption. The absence of forest data was filled with an original household survey. The Forest- SWIFT team worked with the Armenia ENR team to collect primary data on forest uses and forest income, in addition questions required for SWIFT estimation of consumption. The Forest-SWIFT constructed the forest module to ensure that a detailed forest income aggregate could be built after data collection. This forest module has questions on non- timber forest and timber forest products harvested, source of harvest, quantities sold and values, women participation in harvesting NTFPs, wage forest income, and payments for forest services. Consumption model with ILCS 2015 This model was designed to estimate the log per adult equivalent consumption for the national sample using household characteristics, such as household composition, assets, employment, and education. The poverty rate for the full sample was 29.60 percent using the national poverty line. The first stage of SWIFT estimation to cross-validate an optimal p-value identified a value of 0.075 pre-survey, and 0.005 post-survey. The SWIFT model was updated post-survey to reflect changes in survey responses. The pre-survey stepwise regression selected 23 out of 88 potential explanatory variables, and the post-survey selected 36 out of 84. Table 5 lists the final regression for the poverty model. 19 Table 5 Forest-SWIFT models Consumption Urban HH (Excluding Yerevan) -0.35 (0.02) Rural HH -0.28 (0.04) HH: HH cannot afford to own car/truck -0.24 (0.02) Pensioner share of Real members -0.22 (0.05) Road to town/mkt status= poor -0.20 (0.04) Real hhsize (present at least 1 day) -0.17 (0.02) No washing machine in HH -0.17 (0.04) No internet connection in HH -0.16 (0.02) Not registered in pov scheme b/c well off -0.11 (0.02) No hot running water in HH -0.10 (0.02) No central gas supply in HH -0.09 (0.02) No local sanitation compound/hole in HH -0.09 (0.02) Head is a dependent (>65yrs) -0.09 (0.03) Max educ in HH: High School -0.06 (0.02) Head Education Level: Middle School (Basic) -0.06 (0.03) Max educ in HH: Vocational -0.05 (0.02) HH Type: Pvt. House -0.05 (0.02) Real HH Size squared 0.01 (0.00) HH Remittance received 0.05 (0.02) Head Migrated (attempted/successfully) 0.09 (0.02) HH Remittance sent 0.09 (0.02) Head is pensioner 0.10 (0.03) Head: Other self-employed worker 0.12 (0.03) Marz: Sjunik 0.14 (0.03) Marz: Gegharkunik 0.14 (0.03) Hired labor share in HH 0.15 (0.05) Empl Activity = Agriculture, forestry and fishing 0.16(0.03) Marz: Lori 0.16 (0.02) Marz: Armavir 0.17 (0.03) Marz: Aragatsotn 0.21 (0.03) Disabled share of Real members 0.21 (0.06) Dependent share of Real members 0.25 (0.04) Share of real members:Employee with a written 0.27 (0.04) contract Marz: Ararat 0.29 (0.03) Marz: Kotayk 0.32 (0.03) Head is student 1.09 (0.47) Constant 11.86 (0.06) Number of obs 3006 R-squared 0.439 Root MSE 0.524 Source: authors’ estimates using ILCS 2015 The kernel density plot of original and estimated consumption shows a non-normal 2-peak distribution of consumption, such that the proportion of middle-income households dips. 20 Non-normal distributions require an alternative method of multiple imputations for the consumption estimates i.e. predictive mean matching (PMM). Unlike linear regression estimation, which calculates a consumption figure per household only using the regression coefficients and random error i.e. a true estimate, PMM only identifies a consumption figure in the pool of baseline data that is closest to the true estimate. As such a common limitation of using PMM is the inability to use this method when the time gap between the baseline and survey data is too large. Figure 3 Kernel density distribution for consumption Source: authors’ estimations with ILCS 2015 2018 Armenia Forest, Energy and Poverty Survey and results The 2018 Armenia Forest, Energy and Poverty Survey (AFEPS) uses a comprehensive questionnaire to collect information on forest uses and energy. As there were no detailed data on forest income, a detailed questionnaire was created with the environment team working in Armenia. This questionnaire was done following the general guidelines of the Forestry modules. It also included modules on energy and energy uses to assess households’ dependence on fuelwood. Sample was designed for the data to be representative of forest and poor households. Explicit stratification for the 2018 AFEPS is determined by the primary domains of inference, which include Yerevan, Urban (excluding Yerevan), and Rural. Poor households, specifically those utilizing forestry resources, are of interest in the AFEPS to design the sample. Based 21 on available data, each community in Armenia can be classified based on estimated poverty incidence and proximity to forested areas. The sample frame was constructed using satellite images and all grids were classified in one of the following strata: Yerevan, Urban Poor in Forest, Urban Non-poor in Non-forest, Rural Poor in Forest, Rural Non-poor in Non-forest. Data collection took place from October 2018 to December 2018 across the country. The sample size for the survey was about 748 households in 180 grids. There is an oversampling of poor grids in forests in both rural and urban areas to make sure that data are representative at that level. Over-sampling of poor households in forests allows us to capture more variation in household natural resource use. Although the non-poor and non- forested strata are included in the sample design, they are not intended to produce usable estimates at this level. Listing was conducted on a subset of rural and urban grids by the team supervisors. Participation in forest related activities is low in 2018. Overall, less than seven percent of households in Armenia collect forest products. In rural areas, nearly one out of five households collect any type of forest products (table 6). Outside households living in rural areas, households do not perceive forest areas as being useful. Table 6 Forest participation and perception of usefulness Armenia Yerevan Overall Rural Harvest forest 7 0 4 18 products (%) Perception of usefulness of forests (%) Essential 4 5 7 Not essential 23 28 45 Do not use forest 73 100 67 48 Source: Authors’ estimates using AFEPS 2018 The small number of observations for forest income prevents the team from calibrating a Forest-SWIFT modeling. Participation in relative terms at the national level is low with a small number of observations (about 50 households). Doing multiple imputations, SWIFT methodology requires that the baseline data have at least 500 observations to be able to compute the model. A small number of observations introduces some biases in the estimation techniques. 22 Poverty has not changed between 2015 and 2018 although it increased slightly in urban areas outside Yerevan. SWIFT consumption estimates in 2018 are higher than consumption values in 2015 (Table 7). However, this trend is mostly driven by Yerevan while in rural areas, consumption only increased slightly. In urban areas, excluding Yerevan, consumption per capita has decreased leading to an increase in poverty. This geographical disparity in poverty is in line with evidence from the latest Poverty and Equity Brief for Armenia.7 However, one can note from this brief that the estimated poverty rate in 2018 is higher than the measured poverty rate in 2017 reported in the poverty brief. The brief reports that in 2015 there was a surge in poverty and the consumption model relies on 2015 data, the model might be picking up this different poverty pattern. Table 7 Results for imputed poverty rate and consumption per capita ILCS 2015 (original) AFEPS 2018 Poverty rate (%) Armenia 29.60 29.46 Yerevan 24.85 20.56 Other Urban 33.64 38.33 Rural 30.73 29.86 Consumption per capita (drams) Armenia 69207.32 66579.91 Yerevan 80593.67 75270.9 Other Urban 60802.72 59090.46 Rural 65456.21 65294.76 Source: Authors’ estimates using ILCS 2015 and AFEPS 2018 5.3 Tunisia Tunisia has a comprehensive household survey dataset used to measure poverty. The National Statistics Office of Tunisia collected a Household Budget, Consumption, and Living Conditions Survey (LCS – Enquête nationale sur le budget la consommation et le niveau de vie des ménages) in 2015. Poverty in Tunisia is measured using consumption per capita and a national poverty line. Survey data on forest households were limited. The forest data on the other hand, were collected by the United Nations Food and Agriculture Organization (FAO) in 2012 using the 7https://databank.worldbank.org/data/download/poverty/33EF03BB-9722-4AE2-ABC7- AA2972D68AFE/Global_POVEQ_ARM.pdf 23 study on characterizing forest population (“Étude Sur La Caractérisation De La Population Forestière) from the National association on sustainable development and wildlife conservation (Association Nationale de Développement Durable et de Conservation de la Vie Sauvage). About 2040 households were surveyed for this study with questions related to household composition and individual education, uses of forest resources, and labor participation. The sampling strategies are different between the two surveys. Twenty-five thousand households were surveyed for the LCS 2015, and the sample was built to be representative at the urban, rural, and regional level. Due to lack of metadata on the FAO study, there was no clear indication on the sampling strategy within this survey and all the results appear to be unweighted. The forest income is not comparable to the one described in the Forestry Module and is not used to develop a Forest-SWIFT model. The forest income includes gross revenues (market and subsistence) from collections of (20) forest products, and gross revenue from sales of coal. There are no questions on payments for environmental services, income from forest wage activities, or any other forest-related activities. Only 340 households have reported income from one of the sources above, which substantially reduces the number of observations. Since we were not able to use Forest-SWIFT to model forest income, we used linear regressions to identify the main correlates to forest income in the data. Consumption model with LCS 2015 The consumption model is run for households living in the rural areas of the two lagging regions. Since the number of observations is high, the team was able to compute a model using the observations on households living in rural areas of the two regions of the interest: Center-West and North-West. The consumption model was designed to estimate the log per-capita expenditures (including all frequent food and non-food expenditures, including health expenditures, imputed rents and estimated depreciation of durable goods) on household characteristics, such as household composition, assets, employment categories, and education for the rural sample. The poverty rate among rural households across the two lagging regions was estimated at 35.6 percent using the national poverty line. 24 Table 8 Forest-SWIFT models Consumption HH Head: Age 0.00 (0.00) Total HH Members -0.11 (0.00) Share of women in HH -0.12 (0.03) Dependency Rate (Share of HH Size) -0.29 (0.02) HH has at least 1 pensioner 0.08 (0.02) Max education level in HH: Tertiary 0.08 (0.02) Max education level in HH: Primary -0.04 (0.01) Head Health Coverage: Free Health Card -0.12 (0.02) Head Health Coverage: Discounted Health Card -0.08 (0.01) Head's current employment: No Contract 0.08 (0.01) Head Employment: Public Admin/Company 0.12 (0.02) Head Employment: Employee -0.11 (0.01) Head Participates in Ag for self-consumption 0.10 (0.01) HH Type: Semi-detached or paired housing 0.05 (0.02) HH Type: Villa (Full or single floor 0.09 (0.02) HH has at least 2 rooms -0.09 (0.01) HH Drinking Water Source: Public 0.05 (0.01) HH Water Heating Energy Source: Electricity -0.07 (0.02) HH Type: Residential/Collective Housing 0.08 (0.02) HH owns: furniture living room 0.03 (0.01) HH owns: Kitchen with sink 0.18 (0.03) HH owns: DVD player 0.12 (0.02) HH owns: Toilet without flush 0.04 (0.01) HH Owns: Air Conditioner 0.20 (0.02) HH does not have a bathroom -0.08 (0.01) HH owns: Kitchen without sink 0.09 (0.02) HH Owns: Fan 0.06 (0.01) HH owns: Toilets with flush 0.08 (0.02) HH Owns: Computer 0.12 (0.02) HH Owns: Washing Machine 0.06 (0.01) Constant 14.67 (0.04) Number of obs 4607 R-squared 0.5112 Root MSE 0.3478 Source: authors’ estimates using LCS 2015 Following the SWIFT methodology, an evaluation of the mean-squared errors of the consumption estimates, and the absolute differences in the poverty estimates yields an optimal p-value of 0.005 for the consumption model. The final stepwise regression selects 30 out of 74 potential explanatory variables. The kernel density plots for both original and measured values confirm that the estimated consumption density across the population is like the original (Figure 4). 25 Figure 4 Kernel density distribution for consumption Source: authors’ estimates using LCS 2015 Forest correlates with FAO 2012 A simple OLS regression identified simple correlates that explain forest income levels in Tunisia. Correlates for forest income include characteristics of household head such as occupation, education, and age, as well as number of household size. Dwelling characteristics and access to basic services seem to matter as well to explain differences in forest income. The types of NTFP harvested explain differences in forest income with households collecting more valuable forest products (honey, Aleppo pine, or pinecones) having more income than others. 2019 Forest, Land Degradation and Poverty survey and results The 2019 Tunisia Forest, Land Degradation and Poverty Survey (TFLPS) uses a comprehensive questionnaire to collect information on forest uses and land degradation. As there were no detailed data on forest income, a detailed questionnaire was created with the environment team working in Tunisia. This questionnaire was done following the general guidelines of the Forestry modules. It also included modules on land degradation from LSMS-type survey, and questions on forest and land restoration. Sample was designed for the data to be representative of rural households in forest and non-forest areas of North-West and Center-West Tunisia. Explicit stratification for 26 the 2019 TFLPS is determined by the primary domains of inference, which include rural areas of North-West and Center-West, and for forest and non-forest areas. We consider as forest all areas less than 5 km away from a forest body. The sample frame was constructed using satellite images and all grids were classified in one of the following strata: rural North- West in forest, rural North-West in non-forest, rural Center-West in forest, and rural Center- West in non-forest. Data collection took place from June to July 2019 across the two provinces. The total proposed household count is 960 over 115 grids. Over-sampling of the forested further provides increased opportunity to capture more variation in household natural resource use. Non-forested areas are also included in the sample design and are intended to produce usable estimates at this level. Listing was conducted on all grids by the team supervisors and eight households are selected in each grid. Forest participation in Center-West and North-West is low. In 2019, about 6.3 percent of households collect forest products including products for subsistence and sales, as well as payments for ecosystem services. Of the 6.3 percent households who collected forest products, 89 percent consumed these products at home, and 81 percent sold the products in markets. Most of these households engaged in both self-consumption and sales. Most households collecting forest products extract only one product. On average, eight out of ten household collect a single product in the last 12 months, with 11 percent collecting 2 products, and 5 percent collecting 3 products. Households selling forest products report collecting more products than households using forest products for subsistence purposes only. Honey and medicinal plants are the products most consumed and sold (table 4 in appendix). Households report only a single collector for all products, of whom 79 percent are male, 86 percent are the household head, and 12 percent are either children (below 15 years old) or elder people (above 65 years old). Households in 2019 are better-off than rural households in 2013, but one out of four households are poor. SWIFT consumption estimates in 2019 are higher than consumption values for Center-West and North-West provinces in 2019. Table 9 Results for imputed poverty rate and consumption per capita LCS 2015 (original) TFLPS 2019 27 Poverty rate Rural North-West & Center-West 35.61% 15.63% Rural North-West 35.26% 16.64% Rural Center-West 35.88% 14.75% Center-West, Forest n.a. 24.54% Center-West, Non-Forest n.a. 11.31% North-West, Forest n.a. 20.03% North-West, Non-Forest n.a. 15.26% Consumption per capita (Tunisian dinar) Rural North-West & Center-West 2,039,740 2,946,391 Rural North-West 2,098,560 2,957,152 Rural Center-West 1,996,397 3,036,000 Center-West, Forest n.a. 2,947,600 Center-West, Non-Forest n.a. 3,067,136 North-West, Forest n.a. 2,626,878 North-West, Non-Forest n.a. 3,091,228 6 ADVANTAGES AND CHALLENGES USING FOREST-SWIFT Forest-SWIFT is a useful tool to complement exhaustive household surveys. LSMS-type survey to collect data to compute poverty or forest income are key to the implementation of Forest-SWIFT. Forest-SWIFT intends to fill the gap in monitoring poverty and forest dependence, hence complementing exhaustive household survey between two data collections. Forest-SWIFT provides strong measures of poverty and forest dependence. The cases presented above report trustworthy predicted estimates of poverty that are comparable to available measures. It is true that in none of the cases we had truly comparable measures of poverty but the examination of the trends (Turkey) and levels (Armenia and Tunisia) confirms that our predicted estimates are good. Forest dependence is also comparable to the rate of forest dependence reported in other studies (Angelsen et al. 2014). Forest-SWIFT questionnaires are short, timely, and work well as part of a longer questionnaire. As a standalone, Forest-SWIFT questionnaires were administered in less than 20 minutes in Turkey. In the other countries, Forest-SWIFT questionnaires were easy to embed with the longer questionnaires created to collect data on energy (Armenia) or land degradation (Tunisia). Forest-SWIFT questions were often easy to plug in the questionnaire 28 and to some extent of interest for the wider analysis of the data. Forest-SWIFT can help to assess the characteristics of poor forest households. Forest-SWIFT allows one to innovate in designing sample for data collection. Being a short questionnaire, people using Forest-SWIFT can design a sample with multiple strata and domains of inference. In the tests provided here, the choice was to oversample forest areas to have good and reliable estimates for forest people. In some cases, the sample was restricted to forest households (Turkey), to forest and higher administrative levels (Armenia), or to forest and non-forest with higher administrative levels (Tunisia). Differences in sampling design and sample size were driven by survey objectives and budget. Implementing Forest-SWIFT faced some challenges. Baseline data not often available • Baseline data for consumption are not representative of forest areas. The observed values of the consumption correlates from Forest-SWIFT can be quite different from those of the baseline potentially affecting the variance of the estimates. In addition, baseline data for forest income are scarcely available. • The Forest-SWIFT tool could not be fully tested in cases without baseline data for forest income. The use of forest income model will be part of a follow-up survey done by the respective teams in their project. Prediction of forest income • Forest participation with a small number of households collecting of forest products for sales. This prevents the computation of a robust forest income aggregate and the estimation using Forest-SWIFT of forest income for future rounds. • The model is unable to model negative income values. When calculating income, one could put inputs and obtain negative values if revenues are smaller than costs. The modeling cannot predict negative values and will give estimates differing greatly from the observed ones. It is then preferable to model gross forest income. This choice was also justified by the fact that it was very hard to collect data on the input costs of forest collection for the last 12 months. 29 Sample frames built using ad-hoc definitions of forest stratum • The sample frames were built using satellite images and grid sampling. To our knowledge, there are no master frame, built from census data, that can be used to define forest stratum and decide which cluster is in a forest stratum and which cluster is not. Using satellite images on forest cover with population density data was the preferred way. • When building the grids, one had to define forest grids and non-forest grids for cases that were not fully covered by forests. In such cases, the preference was geared towards using a distance from the centroid of the grid to the forest with all grid less than 5 km away from a forest being considered in the forest. However, such choices can have implications on estimates and more can be done to decide how to stratify forest and non-forest areas. 7 CONCLUSION Contribution of forest resources and areas is often undermined in national development policies as data explaining this contribution are missing. Forest-SWIFT is one instrument that tries to fill this gap by giving strong poverty estimates for households living in, and relying on, forests. Forest-SWIFT provides in a timely and regular manner forest dependency ratio and can be used to identify characteristics of poor households living in forest areas. It is important to recognize that the tests of Forest-SWIFT have suffered delays due to data availability and constraints in sampling design. It was hard for the team to get access to baseline data and documentation when available. In addition, sampling strategies have required the use of GIS data and multiple tests to define grid sizes. The lack of up-to- date sample frame has meant that listing of all grids was required. A listing being a census of all dwellings and households in a cluster is long and endearing task that is necessary to compute weights. Further research is undertaken to assess the differences between cluster size from listing and from most recent satellite images. If the differences are not too important, one could restrict listing to areas with high population mobility and use population number from satellite images. 30 In addition, the lack of forest income data be a deterrent for future application of Forest-SWIFT. However, the SWIFT team in the Poverty GP is currently developing a new way to collect the data; this would consist of collecting a full-length questionnaire (consumption and forest income) on a subsample of households (sample A) while collecting SWIFT-type questions (only the determinants for consumption and for forest income) on the remainder (sample B). The model would be built during data collection on sample A and tested on sample B. 8 REFERENCES Ahmed F., C. Dorji, S. Takamatsu, and N. Yoshida. 2014. “Hybrid Survey to Improve the Reliability of Poverty Statistics in a Cost-Effective Manner” World Bank Policy and Research Paper 6909. World Bank, Washington D.C. Angelsen, A., P. Jagger, R. Babigumira, B. Belcher, N.J. Hogarth, S. Bauch, J. Borner, C. Smith- Hall, and S. Wunder. 2014. "Environmental Income and Rural Livelihoods: A Global- Comparative Analysis." World Development 64. Angelsen, A. and S. Wunder. 2003. “Exploring the forest-poverty link: Key concepts, issues and research implications. CIFOR Occasional Paper No. 40. Bogor, Indonesia: Center for International Forestry Research. Byron, N., and M. Arnold. 1997. What Futures for the People of the Tropical Forests? Working Paper, Jakarta, Indonesia: CENTER FOR INTERNATIONAL FORESTRY RESEARCH. Cuevas, P. F., and L. Rodriguez-Chamussy. 2016. Exploring the determinants of the positive trends in poverty and shared prosperity in Turkey. Working Paper, Washington, D.C.: World Bank. Deaton, A. 1997. The analysis of household surveys : a microeconometric approach to development policy. Washington DC, USA: The International Bank for Reconstruction and Development. Debela, B. L., G. Shively, A. Angelsen, and M. Wik. 2012. “Economic Shocks, Diversification, and Forest Use in Uganda.” Land Economics 88, no. 1 (2012): 139–54 31 Dercon, S., and P. Krishnan. 2000. "Vulnerability, seasonality and poverty in Ethiopia." The Journal of Development Studies 36 (6): 25-53. doi:10.1080/00220380008422653. Elbers C., J.O. Lanjouw and P. Lanjouw, 2002. “Micro-Level Estimation of Welfare,” Policy Research Department Working Paper No. WPS2911, The World Bank. Elbers C., J.O. Lanjouw and P. Lanjouw, 2003. “Micro-level estimation of poverty and inequality.” Econometrica, Vol. 71, No. 1 (January, 2003), 355–364 FAO, CIFOR, IFRI and World Bank. 2016. National socioeconomic surveys in forestry: guidance and survey modules for measuring the multiple roles of forests in household welfare and livelihoods, by R.K. Bakkegaard, A. Agrawal, I. Animon, N. Hogarth, D. Miller, L. Persha, E. Rametsteiner, S. Wunder and A. Zezza. FAO Forestry Paper 179. Godoy, R., V. Reyes-García, V. Vadez, W. Leonard, S. Tanner, T. Huanca, D. Wilkie, and TAPS Bolivia Study Team. 2009. “The Relation between Forest Clearance and Household Income among Native Amazonians: Results from the Tsimane Amazonian Panel Study, Bolivia.” Ecological Economics 68 (2009): 1864–71. Kuhn, M. and K., Johnson. 2013. “Applied predictive modeling” Springer Perge, E. 2010. “Testing a Poverty Trap Mechanism with Tsimane’ Panel Data.” Working Paper: Chronic Poverty Research Center CPRC, Manchester, UK. Perge, E. and A. McKay. 2016. “Forest clearing, livelihood strategies and welfare: Evidence from the Tsimane' in Bolivia.” Ecological Economics 126:112-124. Sahn, D. E., and D. C. Stifel. 2000. "Poverty Comparisons Over Time and Across Countries in Africa." World Development 28 (12): 2123-2155. Serajuddin, U, H Uematsu, C Wieser, Nobuo Yoshida, and Andrew Dabalen. 2014. Data Deprivation. Another Deprivation to End. World Bank Policy Research Working Paper, Washington DC, USA: The World Bank, 1-22. Shepherd, G. 2012. Rethinking Forest Reliance: Findings about poverty, livelihood resilience and forests from IUCN’s ‘Livelihoods and Landscapes’ strategy. Switzerland: International Union for the Conservation Nature. 32 Shyamsundar, P., S. Ahlroth, P. Kristjanson, and S. Onder. forthcoming. "Supporting pathways to prosperity in forest landscapes – A PRIME framework." World Development. doi:https://doi.org/10.1016/j.worlddev.2019.104622. Vedeld P., A. Angelsen, J. Bojö, E. Sjaastad, and G. Kobugabe Berg. 2007. “Forest Environmental Incomes and the Rural Poor.” Economic Perspectives and Analyses of Multiple Forest Values and Sustainable Forest Management 9 (7): 869–79. https://doi.org/10.1016/j.forpol.2006.05.008. Vira B., C. Wildburger and S. Mansourian. 2015. “Forests, Trees and Landscapes for Food Security and Nutrition. A Global Assessment Report.” IUFRO World Series. Volume 33. Vienna Wunder S., A. Angelsen, and B. Belcher. 2014. Forests, Livelihoods, and Conservation: Broadening the Empirical Base. World Development, 64, Supplement 1, S1–S11. https://doi.org/10.1016/j.worlddev.2014.03.007 Yildirim H. T. and Kose M. 2013. “Non-Wood Forest Products Production Policy and Marketing in Turkey.” Conference Paper. Forest Products Society's 67th International Convention Forest Products in a Global Sustainable Economy. Yoshida, N., R. Munoz, A. Skinner, C. Kyung-eun Lee, M. Brataj, W. Durbin and D. Sharma. 2015. “SWIFT Data Collection Guidelines version 2”, World Bank. Washington D.C. 33 APPENDIX Table 1 Summary statistics for SEHS 2016 and HBS 2013 HBS 2013 SEHS 2016 (rural) Household size 3.88 4.52 Dependence ratio 0.607 0.532 Head Characteristics Age 52.08 53.35 Male 87.33% 96.48% No school 19.70% 9.06% Primary school 80.30% 90.06% Employed 71.81% 65.48% Prime aged adults’ characteristics No school 18.49% 9.85% Primary school 81.51% 90.04% Employed 61.01% 49.60% Neither student, nor employed (15-29 yrs) 29.45% 21.16% Unemployed 2.74% 48.89% Labor force participation 61.01% 56.71% Female labor force participation 38.67% 26.30% HH Characteristics a Household with access to gas 1.82% 0.00% Household with access to piped water inside or outside house 98.43% 89.42% HH Assets HH Assets: Total out of 5 3.27 3.94 Household with access to refrigerator 97.84% 97.50% Household with fixed or at least one cell phone 96.66% 98.53% Household with access to internet at home 16.44% 7.64% Household with access to washing machine 92.29% 94.27% Observations 3,006 1,256 Source: authors’ computation using SEHS 2016 and HBS 2013. Weights applied. Note: All statistics are at the HH level. a SEHS 2016 collected this data at the community level Table 2 Most important forest products and sales Products Extraction (percent) Sales (percent) Hazelnut 50.19% 12.93% Industrial wood 41.95% 0.32% Linden 58.13% 0.03% Opium 0.95% 0.33% Pinecone 15.91% 0.31% Sting nettle 37.14% 0.00% Trefoil 21.68% 0.11% Walnut 76.84% 4.12% Herbs 56.87% 0.07% Note: Weighted at the household level 34 Percent of households extracting and selling forest products measured on sample of households extracting at least one forest product (970 of 1256 HH). Table 3 Total forest products collected in Turkey Total (of 9) Collected For Sale Only 0 22.58 85.94 1 16.54 14.01 2 10.76 0.05 3 12.54 4 9.16 5 12.12 6 9.31 7 6.24 8 0.72 9 0.03 Table 4 Main forest products collected in Tunisia (% of households) Forest Product Collection rate (%) Honey 44.61 Medicinal Plants 29.91 Wood products 13.72 Aleppo Pine seeds 8.50 Essential Oils 5.73 Nuts 2.94 Other 0.26 Carob (Fruit) 0.21 Source: Authors’ estimates using TFLPS 35