Impact of Weather Shocks on Food Security: How Effective are Forests as Natural Insurance?

Abstract Malnutrition and food insecurity affect nearly one billion people worldwide. In developing countries, adverse weather shocks exacerbate these challenges by reducing agricultural productivity. Rural households often rely on forests for food. We determine whether forest access is associated with a less severe effect of adverse weather shocks on food security in rural Malawi. Exploiting exogenous variation in weather shocks and predetermined forest access, we find that households without forest access experience drops in food security when confronted with shocks, while forest access is associated with insignificant changes in food security. This suggests that forests are used as natural insurance. For the period considered by the study, we find that most of the negative impact of shocks was driven by floods, which were more prevalent and severe than droughts. In addition, we find evidence that the role of forests as natural insurance improves with increased forest density (canopy cover). There is a minimum forest density threshold below which forests are not associated with natural insurance. These results suggest that efforts to protect forests should consider their natural insurance role, particularly in regions with weak social safety nets.


Introduction
Approximately 11 per cent of the global population, or over 828 million people, are food insecure, a 150 million increase since 2019 (FAO, IFAD, UNICEF, WFP, & WHO, 2022).Most of those affected are rural households directly relying on agriculture for food and income (Ruel-Bergeron et al., 2015), with low incomes, undiversified diets, and less reliance on food markets (Koppmair, Kassie, & Qaim, 2017).Increasing extreme weather events in sub-Saharan Africa may worsen food insecurity among these communities (Alderman, 2010;Creti, Delacote, & Leblois, 2021).Food insecurity, in turn, leads to malnutrition (both under-and over-nutrition), affecting the quality of life (Tanumihardjo et al., 2007).Early childhood malnutrition negatively affects education, income, and productivity (Muthuuri, Some, & Chege, 2020), causing economic costs of 2 per cent to 11 per cent of GDP in Asia and Africa (International Food Policy Research Institute, 2015).Developing countries, therefore, prioritise enhancing food systems' resilience to weather shocks.
Forests in many developing countries provide income and food through timber and non-timber forest products (NTFPs) (Tesfaye, Roos, Campbell, & Bohlin, 2011), which can supplement diets (Baudron et al., 2019).Rural communities mainly use timber for fuelwood and charcoal burning, providing energy for cooking and income (Ainembabazi, Shively, & Angelsen, 2013;Zulu, 2010).Common property forests accessed by communities can be used to support food security when other food sources, such as agriculture, fail due to climatic stressors, as forests are more resilient to weather shocks (Fentahun & Hager, 2009;Wunder, Angelsen, & Belcher, 2014).Households increase forest use to offset income drops resulting from shocks (Pattanayak & Sills, 2001;Smith, Hudson, & Schreckenberg, 2017).Specifically, as droughts and floods affect agriculture negatively (McCarthy, Kilic, Brubaker, Murray, & De La Fuente, 2021), households may resort to forests for food.However, because most NTFPs are non-market, they may not be accurately captured in the income effects, so their role in food security is underestimated (Barua, Boscolo, & Animon, 2020).Furthermore, not many studies (one exception is Hall et al. (2019)) focus on understanding how the natural insurance role of forests changes with forest densityan important aspect of forest conservation.Therefore, we focus on forest access, forest density, and food security outcomes that do not necessarily rely on market transactions to understand the role of forests in mitigating the effects of shocks on food security.
We define food security using two outcomesthe self-assessed food security (SAFS) data on whether the household had enough food for the whole year or not (Wegenast & Beck, 2020) and household dietary diversity score (HDDS) (Chegere & Stage, 2020).In addition to these, we also measure food security using the reduced coping strategy index (rCSI) (Vaitla et al., 2020) and the number of meals for children (Dureab et al., 2019).Using panel data, we explore whether households with access to common property forests experience smaller weather shock impacts on food security.We then examine forest access impact heterogeneity by forest density (measured by the proportion of total forest that is of higher canopy cover).Finally, we test the impact mechanisms by determining if households adjust labour away from agriculture to the forest in response to a shock.In summary, this study tests the effectiveness of forests in protecting food security in the event of a shock (i.e.forests as natural insurance that can mitigate the adverse effect of weather shocks on food security).
We contribute to the resource and development economics literature that focuses on shockcoping mechanisms or climate adaptation by highlighting the importance of forest access for household food security in the event of weather shocks.Specifically, our contribution is threefold.Though many studies have established a correlation between shocks and the use of common property resources (CPRs) as natural insurance (Agarwal, 1990;L opez-Feldman, 2014;Pattanayak & Sills, 2001), there is little evidence of the effectiveness of forests as a way to smooth or protect household food security against adverse weather shocks.Our main research question, therefore, attempts to bridge this gap in the literature: Does forest access protect food security in the event of weather shocks?If it does, this would mean that households with forest access should experience smaller (or zero) declines in food security in the event of weather shocks.The second research question we answer is: does the shock-impact mitigating role of forests improve with increased forest density?Specifically, we test whether the natural insurance role improves with forest density.Existing literature offers some insights into the correlation between forest density (or cover) and household welfare (Gergel et al., 2020;Hall et al., 2019) but does not focus on the ability to reduce the negative impacts of shocks on food security.
Our third contribution is to estimate the quantitative effect of forest access on mitigating the negative impact of weather shocks on food security.Thus, the current study adds to the literature that tests the effectiveness of shock-coping strategies (Janzen & Carter, 2019;Kazianga & Udry, 2006).Altogether, results from this study have policy-relevant implications for the conservation and use of forests as a safety net.

Impact of weather shocks on food security 1761
The rest of this paper is organised as follows.Section 2 describes the context in which we empirically examine the linkages between shocks and food security and the data used.Section 3 describes our empirical strategy, while results are presented in section 4. Finally, section 5 concludes with a discussion of policy implications and avenues for future work.

Context
To explore if households can effectively use forests to protect food security, we examine the role of forest access in Malawi.Achieving good nutrition, which is a direct result of food consumption, is a challenge across the country.As of 2010, about 36 per cent of the population was categorised as food insecure (Aberman, Meerman, & Benson, 2018), and in 2016 (the final year for the data used in this study), about 3.3 million Malawians were food insecure (Mkusa & Hendriks, 2022).
In Malawi, about 80 per cent of rural households use forests (Fisher, Chaudhury, & McCusker, 2010) for either income or food.In our data, conditional on forest access, about 60 per cent of the households in each village use forests, with some villages having 100 per cent of the households using forests for food.Most households collect NTFPs such as firewood, insects, leafy vegetables, small animals, fruits, and insects.The main forest biome is the Miombo Woodlands, which covers the country's central, western, and eastern parts and is classified under the tropical and subtropical grasslands, savannas, and shrublands biome.The other main part is Lake Malawi and its surrounding rivers and streams.Forest cover is about 36 per cent of the country, of which 33 per cent is categorised as primary coverthe most diverse form of forest (Munthali & Murayama, 2015).Policymakers often underappreciate the role played by forests in rural livelihoods partly because quantitative estimates of the impact on nutrition and health are not available.With the Malawian national government committed to using forests for poverty alleviation (Government of Malawi, 2017), this study contributes to understanding forests' role in food security.
Malawi has two main seasons relevant to agricultural production: cold-dry and hot-wet.The dry season is from May to October, while November to April is the wet (rainfall) season.The country has experienced several climate shocks, including during the period considered by this study.Specifically, for the period under study, floods were more prevalent and severe compared to droughts (Figure A3 in the appendix).This makes Malawi a good context for understanding the impact of weather shocks on food security.

Data sources
We use the publicly available Living Standards Measurement Study-Integrated Surveys on Agriculture (LSMS-ISA) panel dataset that collects information on various aspects of livelihoods.The panel LSMS-ISA data we use were collected in Malawi (https://www.worldbank.org/en/programs/lsms/initiatives/lsms-ISA#3) in 2010, 2013, and 2016 as part of the Integrated Households Long Term Panel Survey (IHPS).The overall number of panel households was 1619 in 2010, the baseline year, and it was expanded in subsequent waves as households split up and new ones joined.The number of households increased from 1908 in 2013 to 2508 in 2016, according to the Malawi National Statistics Office (NSO).Between 2010 and 2013, the attrition rate was 3.78 per cent, and between 2013 and 2016, it was 4 per cent.Therefore, we use a panel of 1294 households from the 102 villages surveyed in 2010 and 2016.Even though the sample is nationally representative, it is restricted to these 102 villages which may not cover all forest types in Malawi.Therefore, results should be interpreted with context.
We then employ annual rainfall data to construct an objective measure of weather shocks using the Standardised Precipitation Index (SPI).Before making the LSMS-ISA data public and confidential (by jittering the GPS coordinates), the World Bank and country statistical offices use the GPS coordinates that are part of the household survey data to link the household data to the National Oceanic and Atmospheric Administration (NOAA) Climate Prediction Centre rainfall estimates captured at a resolution of about 120 km 2 (roughly a village).This is annual rainfall that captures even the non-growing months. 1 To test if using shocks defined from the growing months only changes the results, we also used monthly rainfall and potential evapotranspiration (PET) data from the Climatic Research Unit -University of East Anglia datasets (CRU-TS 4.05).The CRU-TS 4.05 data are available at 0.5 Â 0.5 grids (Harris, Osborn, Jones, & Lister, 2020).The CRU-TS data were merged with the household data using the GPS coordinates from the household data where each household was put in the grid in which they belong using their jittered coordinates (jittered for confidentiality).However, matching using jittered GPS coordinates does not likely affect our results as this does not change the grid in which a household is placed for the climate data (Michler et al., 2022).Finally, we use data on forest size and canopy cover (density) at the village level.For this, we use the Global Forest Watch data accessible for Malawi at the village level (Hansen et al., 2013).The data indicate the sizes of forests at different canopy-cover percentages at the village level (not pixels as in some satellite forest data) with village names for 2010.Therefore, we have one observation of forest size and density per village.The forest data are spatially merged with the LSMS data at the village level by matching the village names.More details on the data and summary statistics can be found in Appendix A.

Construction of key variables
We use household dietary diversity score (HDDS) and a binary-defined self-assessed food security indicator as the main dependent variables measuring food security.Household dietary diversity score (HDDS) was obtained from a one-week recall of the foods eaten by the household (for details, see Appendix A).HDDS strongly correlates with per capita food consumption, food consumption score (FCS), and energy availability (Ruel, 2003).Results are robust to the use of either HDDS or FCS.The choice of HDDS over FCS is because HDDS is often used in studies linking forests to diets (Baudron et al., 2019;Hall et al., 2019) and we intend to situate this study within this literature.
In addition to HDDS and SAFS, we also construct rCSI as a weighted sum of the frequency of the different food coping strategies used by the household (Vaitla et al., 2020).We also included food expenditure in Malawian Kwacha adjusted for inflation as a dependent variable to capture households who may depend on market-bought foods though these are a minority in the data.
We then employ annual rainfall data to objectively measure weather shocks using a 12-month Standardised Precipitation Index (SPI).We focus on a 12-month SPI instead of just the growing season that mainly focuses on maize, as shocks to other crops in other months may affect food security.For example, some households in Malawi grow vegetables and crops such as millet within seasonal lakes and rivers that may benefit from rainfall in the dry season (Mungai, Messina, & Snapp, 2020).Furthermore, water available during growing season may depend in part on soil moisture that accumulates from rainfall outside the growing months.A focus on rainfall only in the growing season could miss this effect.However, rainfall in the non-growing months is negligible nationally (Figure A1 in the appendix).All SPI values below -1.6 (droughts) and above 1.6 (floods) were taken as weather shocks because this is when major crop/pasture losses occur 2 (Leng & Hall, 2019).The choice of 1.6 as a cut-off for defining shocks is also supported by the empirical analysis (Figure S2 in the Supplementary Material) which shows that shock defined above 1.6 are very rare while those below 1.4 have no significant negative effect on food security.Using the same SPI approach, we also construct seasonal weather shocks using the CRU-TS dataset for the growing months' rainfall (November to Impact of weather shocks on food security 1763 April).We calculate a two-monthly SPI for the planting months (November and December), growing months (January and February) and harvesting months (March and April) and construct weather shocks to understand if rainfall shocks in different stages of the agriculture calendar have different effects.We also define shocks using rainfall per annum, a non-relative measure, and results are robust to such a definition (Table S1 in the Supplementary Material).
Forest access is defined at the village level using a community focus group discussion question.The community-level question asked the community members (including leaders) through focus group discussions if they had access to a community resource, such as a forest, where they could extract NTFPs.Forest density is defined using three thresholds: the proportion of forests or tree cover in the village with >35 per cent , >50 per cent , and >75 per cent canopy cover.More details on data and the construction of key variables are in Appendix A.

Empirical models
To understand the mitigating role of forest access on food security in the event of a weather shock, we begin with Equation ( 1): where Y h ivt is the outcome variable for individual i in village v and time t, and h ¼ HDDS, SAFS are household dietary diversity and self-assessed food security, respectively.a h i is the household fixed effects.S vt is an indicator of whether village v experienced weather shocks in period t (S vt ¼ 1 if jSPIj>1.6 in village v and year t), Acc v is forest access, X ivt is a vector of household characteristics (household size, age of household head, number of other households in the village allocating labour to the forest, income, remittances, temperature, and livestock owned), c j t is a year fixed effect, and e j ivt is an idiosyncratic error term.d h 2 , which tests for the effect of access on food security is the parameter of interest in each equation.
The expectations operator allows for nonlinear specifications (i.e.count models).Forest access, Acc v , is defined at the village level using the community-level key-informant focus group questionnaire that asked communities if they have access to a communal forest resource.
The starting point for the estimation is the two-way fixed effects (TWFE) models.We include the household fixed effect that controls for unobserved time-invariant heterogeneity specific to each household that would otherwise compromise causal inference due to omitted variable bias (Gormley & Matsa, 2014).Year-fixed effects capture all shocks common to all households in each year, such as a policy change.
When h ¼ HDDS, Equation ( 1) is estimated using the fixed effects Poisson (FEP) model to consider the count nature of HDDS. 3 In this case, E Y HDDS ivt Â Ã is the expected log of the household dietary diversity count.The coefficients of the FEP are interpreted as the proportional change in the expected value of the dependent variable if the regressor changes by one unit.
When h ¼ SAFS, Equation ( 1) is estimated using a fixed effect (FE) linear probability model (LPM) with a linear error term.Nonlinear probability models would be a logical choice for this analysis but since our identification strategy relies on FE, we have resorted to an LPM that can be estimated with FE.Heteroscedasticity, a problem in LPMs, is corrected by clustering the standard errors at the village level (Fern andez-Val, 2009) the level at which shocks are measured.
Note that the variable for access does not appear on its own in the model because it is timeinvariant, and as such, it is absorbed by the household fixed effects.The parameter, d h 1 , tests for the impact of a weather shock on food security for households living in villages without forest access, while d h 2 tests for the differential impact for those with forest access.The impact of the weather shocks on those with access to forests is To test if forest access offsets the impact of weather shocks for those with access, we test the null hypothesis that To address the effect of forest density on the natural insurance role of forests, we estimate Equation ( 1) with an additional independent variable: an interaction between forest access, weather shocks, and a measure of forest density.This augmented regression is used to test if the shock-mitigating role of forests is increasing in forest density.We present results using alternative definitions of forest density (described in section 2.2).A further robustness check is conducted using a kernel estimator to account for the nonlinear interaction effect (details of the kernel estimator and main results are in Supplementary Material B).
Lastly, we explore the impact mechanisms by determining how households adjust labour among different activities when faced with weather shocks, considering any differences between those with and without forest access.Using Equation (2), we examine if labour allocated to the collection of forest resources, agriculture, and other activities (L ivt j ) responds to weather shocks.
Where L ivt j is labour allocated by household i belonging to village v in period t to activity j ¼ agriculture, CPR forests, off-farm work, and non-agricultural work.e j ivt is an idiosyncratic error term and the rest of the variables remain as defined before.Similar to Equation ( 1), this regression does not include the forest access variable on its own since it is time-invariant.
Labour allocated to agriculture, off-farm work, and non-agricultural work is measured in total household person-hours per week.Due to data limitations and the desire to determine if there is an increased reliance on NTFPs, we use the value of forest collection estimated from what households report.Using value as a proxy for labour allocation to the forest requires an assumption that NTFPs output value (at constant prices) is proportional to the amount of labour used for extraction (Delacote, 2009).This is a reasonable assumption for two reasons.First, NTFPs collection relies almost exclusively on labour and no capital inputs (Neumann & Hirsch, 2000).Second, it is difficult to change the types of NTFPs collected and move from low-value to high-value ones since the mixture of NTFPs depends mostly on forest type and landscape in each village.Hence, any increase in the value of NTFPs collected likely reflects increased effort (labour).The real value of output is also a better measure instead of the total output quantity (i.e. in kilograms) as it allows for aggregation across different forest products.

Identification and estimation issues
In the absence of random forest access, our study faces several identification challenges.One potential threat relates to the identification of the impact (d h 2 ) in the estimation of Equation ( 1).Specifically, household forest access may be endogenous to food security.However, we argue that access is predetermined as almost no villages report a change in access between 2010 and 2016.Less than 3 per cent of villages report a change, and we do not use this variation for identification.Second, our main focus, for which most households access the forest, is for collection of NTFPs, which we argue does not deplete the forest.We test if the use of forests for NTFPs collection is correlated with forest loss or density and find that there is no significant correlation (Supplementary Material C)suggesting that forest use for NTFP collection does not cause loss or depletion.Therefore, access is predetermined and not affected by use (for NTFPs).In this case, to identify the impact of the interaction of access and shocks, we leverage an important feature of exogenous treatments often exploited in the literature (Dreher, Minasyan, & Nunnenkamp, 2015;Nunn & Qian, 2014) and recently demonstrated formally by Nizalova and Murtazashvili (2016) and Bun and Harrison (2019).Specifically, the coefficient estimate for the interaction of an exogenous variable (weather shocks) and an endogenous but predetermined variable (forest access) can provide a causal impact estimate.

Impact of weather shocks on food security 1765
The other challenge, potentially, is the co-location of villages with access to forests and lower exposure to weather shocks.If villages with forest access are located in areas that are humid and experience less severe effects from a weather shock, it may be that the coefficient we find in the empirical analysis, compensating the negative impact of weather shocks, is simply because these locations are more humid and less impacted by negative rainfall shocks.Though we cannot completely resolve this, we address the challenge in two ways.First, we determine if there are systematic differences in weather shocks, long-term rainfall, temperature, potential evapotranspiration (30-year period from the CRU-TS dataset) and outcome variables between villages with forest access and without.We also extend this analysis to forest density.We estimate a pooled OLS model since forest access and density are time-invariant.Results in Table 1 show no significant correlations between weather/climate patterns and forest access.This suggests there is no co-location of forest access and our measure of weather shocksa plausible finding for a small country like Malawi with a relatively uniform climate (Warnatzsch & Reay, 2019).The only significant correlation is between dietary diversity, forest access, and densityshowing that villages with high density forests have low dietary diversity, possibly because they are more remote.Our empirical analysis is not about comparing dietary diversity levels across villages, but how the levels change within a village across time because of shocks between those with and without forest access.
In addition to the TWFE, we implement three different specifications of panel data models that seek to control for unobserved factors to test the robustness of the results.We begin with a matching estimator.We use propensity score weighting to compare households in villages that are otherwise similar except for forest access to further reduce potential bias due to time-varying differences (Rajkhowa & Qaim, 2022).Propensity scores for forest access are recovered and used as weights in estimating equation (1).Propensity score weighting results in improved covariate balance and adjusts for potential bias due to non-random selection into villages with or without forest access (Li, Morgan, & Zaslavsky, 2018).To generate propensity scores, we utilise a rich set of village-level covariates, including rainfall, distance to the main road, distance to town, village socio-demographics and temperature.We employ inverse propensity score weighting to estimate the effect of forest access on food security, restricting the estimation to observations with common support.Second, we include an interaction between long-term average rainfall and weather shocks.Long-term average rainfall captures the differences in climate between villages, and this interaction captures any differences in weather shocks across the different climates.If climate is correlated with forest access, then this could be driving the heterogeneity we estimate by forest access.The interaction term captures such effects.
Lastly, we implement an interactive fixed effects model (Bai, 2009) by adding a vector of common factors with loadingsthe interactive fixed effects.The interactive fixed effects allow household unobservables to change over time by including the interaction of time-specific factors with individual-level fixed effects through principal component analysis.Unobserved, timevarying heterogeneity may stem from omitted common variables or global shocks that affect each unit differently (Coakley, Fuertes, & Smith, 2006).Interactive fixed effects models have been employed in other contexts such as the evaluation of enterprise zone policies (Gobillon & Magnac, 2016).For the interactive fixed effects, Equation ( 1) is estimated using a linear model for both HDDS and FS because currently it can not handle non-linear models.
To determine whether an increase in forest density is associated with an improvement in the effectiveness of forests as natural insurance, we rely on the previously stated assumption that the impact of local households allocating labour to the collection of NTFPs has little impact on forest depletion (Gbetnkom, 2008;Sassen, Sheil, & Giller, 2015).We test the impact of total labour allocated to NTFPs collection on forest density (in Supplementary Material C, Table S8 column 2) and find that NTFPs collection effort is not significantly correlated with forest density.Timber harvesting, mainly done by urban-based and non-Malawian companies (Kafakoma & Mataya, 2009) depletes forests and is largely exogenous to local community decisions.Community participation in management is low in commercialised forests (mostly for timber) (Jumbe & Angelsen, 2007), supporting the larger role of outside agents.
Finally, despite these efforts to reduce the influence of both observed and unobserved confounders, we cannot fully eliminate endogeneity concerns.Potentially, unobserved differences in villages with and without forest access could be driving the heterogeneity in shock impacts that we estimate.While we present a wide range of specifications meant to alleviate this concern, it is not feasible to introduce truly random variation in forest access.Therefore, we interpret the results in this study as robust associations and provide avenues for future research to further investigate the natural insurance role of forests.

Results
In this section, we present our econometric results.We begin with estimating the effect of weather shocks on food security, the role of forest access, and forest density, and then provide some tests for the impact mechanism.

Weather shocks, forest access, and nutrition
Results on the mitigating role of forest access on the effect of weather shocks on food security are displayed in Table 2.We begin with the standard TWFE, propensity score weighting, longterm rainfall and weather shocks interaction, and an interactive fixed effects models.Overall, the findings are robust to different estimation strategies.Except for results on food security using propensity score, coefficients from these different specifications are similar.The coefficients from propensity score weighting on SAFS suggest that when we do not match households with and without forest access, the estimates may be biased towards zero, indicating that some omitted variable(s) (e.g.population density, access to other coping strategies, living in more Impact of weather shocks on food security 1767 Having access to a forest is associated with a reduction in the severity of the impact of weather shocks on dietary diversity and self-assessed food security.The effect of weather shocks on HDDS and food security is negligible for those with access to forests.When we compute the marginal impact of a weather shock on those with forest access, we find that the effect is not significantly different from zero for both HDDS and SAFS for most models.Even for models with negative and significant effects, it is much lower than for those without forest access.While studies such as Hall et al. (2019) fail to find a significant relationship between forest use and dietary diversity, we demonstrate that in the event of shocks when more labour is allocated to the forest, there is a significant positive relationship between forest access and dietary diversity and that forests protect food security.
In further robustness checks below, we show that there are differences on the impact by whether the shock is a drought or flood.Generally, results seem to hold for floods and not droughts and this is likely that more severe floods were experienced for the period under review.This shows that for severe shocks which affect food security, forest access reduces the impact.Given the robustness of the standard TWFE when compared to other more complex specifications (results that include an interaction of 30-year rainfall and shocks and interactive fixed effects produce qualitatively similar results), we estimate the remaining regressions with a TWFE model.
In Table 3, for robustness, we show the impact of weather shocks (using the original categorical SPI) on food security using four different measuresfood expenditure in real Malawian Kwacha, rCSI, and the number of meals eaten by children per day.Food expenditure captures market-bought foods, rCSI captures the intensity of reliance on food consumption reduction as a coping strategy, while the number of meals for children captures food security for a different household demographic.In all regressions except for rCSI, the main results are confirmed.For Impact of weather shocks on food security 1769 rCSI, even though the results are not statistically significant, they are qualitative similar to those from other measures of food security.

Role of forest density on the use of the forest as natural insurance against food insecurity
To determine if dense forests are associated with better mitigation against the negative impact of weather shocks, we estimate a model that includes interaction between shocks, access to forests, and forest density.We use three different thresholds to define dense forests as described earlier.The results are in Table 4.We find that an increase in the proportion of dense forests is associated with an increase in the natural insurance role of forests.This relationship is much stronger when using a higher threshold for defining dense forests.It is noteworthy to state that high forest density (i.e.>50% canopy cover) is very limited across the country (see Figure S1 in the Supplementary Material) while forests with <50 per cent canopy cover do not appear to provide enough NTFPs to act as natural insurance.
Our results are consistent with the hypothesis that dense forests contain a greater diversity of products (Baudron et al., 2019;Hall et al., 2019).It is also possible that dense forests improve the average product of labour (i.e., the number of NTFPs collected for each hour a household allocates to the collection) in the resource.

Robustness checks
We conduct a range of robustness checks to test if our results are robust to different estimation strategies, alternative definitions of weather shocks, and the inclusion of other controls.Table 5 summarises each robustness check.The main focus is on the results of forest access, but we also include robustness checks on forest density.
Generally, most robustness checks confirm the main results.The results are confirmed when we use rainfall levels to define weather shocks (as annual rainfall below 500 mm, and above 1000 mm, based on maize crop water requirements; Wiyo, Kasomekera, & Feyen, 2000).Furthermore, we define shocks using rainfall from the growing season only (November-April) using the CRU-TS dataset.Results are confirmed in terms of the signs of the coefficients but only significant for SAFS.The non-significant results on the effect of growing season-defined shocks on HDDS may mean that seasonal rainfall does not adequately capture the effect on all crops grown or that soil moisture from out-of-season rainfall can be important.This justifies our preference for a 12-month SPI-defined shock.The next robustness check uses bimonthly shocks for November-December, January-February, and March-April, which we categorise as planting, growing, and harvesting months, respectively, based on the maize calendar in Malawi.Results are generally mixed in terms of signs and significance.This is likely because of collinearity among the bimonthly shocks (all three had significant pairwise correlations).Therefore, given our data, we are not able to separately isolate the impact of shocks that occur at different times of the season.The other robustness check is to define shocks based on different SPI cut-off points.Generally, mild shocks have no impact on food security, while fewer households experience very severe shocks (e.g.only about 11% experience shocks defined as jSPIj>1.8,and those with access who experience this shock only make up 7% of the sample), making the estimates more imprecise.Shocks defined by droughts and floods separately generally confirm the observed results.Furthermore, including potential evapotranspiration among the controls does not affect the results.To control for possible co-location of weather shocks and forest access, we interact long-term rainfall and shocks, and results are confirmed for the moderating role of forest density.These results on the moderating role of forest density are also confirmed when we estimate the effect using a kernel estimator.Impact of weather shocks on food security 1771 4.4.Impact mechanism: are shocks associated with increased labour to forests?
Results in Table 6 show that households decrease labour to agriculture in the event of a weather shock, and this decrease is much more substantial (and significant at the 10% level) for those with forest access. 4Shocks are associated with a decrease in the value of NTFPs collected for those without access within the village (but who may access forests in other villages) and an increase in the value of NTFPs collected for those with access, though not statistically significant.This is suggestive evidence that labour to forests (for those with access to forests) increases while the labour to agriculture reduces in the event of a shock.On off-farm (ganyu) labour, there is no statistically significant effect of shocks though the signs of the coefficients suggest  S3 Results from all three bimonthly shocks have the right signs but not the interaction of shocks and forest access.Significance levels are also different.We suspect the correlation between these variables makes the estimates imprecise (the correlation for all three bi-monthly shocks was significant at 1% level).Use different SPI cut-off points for defining shocks.

Figure S2
Shocks defined as jSPI| < 1.6 do not affect HDDS or SAFS negatively.Shocks defined as jSPIj>1.6,or 1.8, significantly negatively affect HDDS and SAFS, but as the proportion of households that experienced such a shock reduces, the estimates become imprecise (e.g.jSPIj>1.8).

Flood and drought shocks separately
Table S4 Generally (except for drought on HDDS), flood and drought shock results are similar to the main results.Flood shock results are more consistent with the main results.This likely occurs because floods are more prevalent and severe for the period we consider (see Figure A3 in the appendix).Control for the potential evapotranspiration (PET) Table S5 The results are confirmed even when we control for differences in potential evapotranspiration that can affect the moisture content.Include the interaction of long-term rainfall and weather shocks on the regression to understand the role of forest density Table S6 Results are robust to the inclusion of this interaction.
Using the kernel estimator for forest density heterogeneity to account for the nonlinear interaction effect

Figure S3
Results from the fixed effects model are generally confirmed, and this estimator provides a more nuanced view of the role of forest density in food security.
that there is an increase in labour allocated to these activities for those without forest access.
Those with forest access significantly increase labour allocation to non-agricultural enterprises.This is plausible as forest products-based activities (including collection of NTFPs) are categorised as non-agricultural.This result is consistent with other studies (Fisher et al., 2010;Pattanayak & Sills, 2001;Wunder, B€ orner, Shively, & Wyman, 2014) and suggests that households allocate labour to forest resources when returns from agriculture fall because of adverse shocks.This explains the natural insurance role of forests observed in the main results above.

Conclusion
In Malawi, we investigated the impact of forest access on food security during adverse weather shocks.We used different estimations approaches that control for both observed and unobserved heterogeneity.Our panel data analysis reveals that forests mitigate the negative effects of weather shocks on food security and dietary diversity.This protective effect is stronger in areas with higher forest density.Forests can be considered an important strategy to safeguard food security amidst climate change-induced weather shocks.However, increased deforestation could undermine this strategy.Households adjust their labour from agriculture to forests during shocks, highlighting the need for policies that promote resilient agriculture without encouraging deforestation.While forests play a significant role in coping strategies, further research is required to assess whether dependence on forests traps households in poverty.Our study is limited by observational data and potential non-random forest access, suggesting caution in interpreting causal relationships.Future research should seek natural experiments to isolate the mechanisms through which forest access improves well-being during weather shocks and collect data on non-timber forest product collection and quantities.Impact of weather shocks on food security 1773 The distribution of the HDDS calculated from the listed food groups is shown in Figure A2.In the current study, we observe a mode of 9 food groups consumed out of 10 possible groups, a mean of 8, and a variance of 2.83.
In Figure A3, the distribution of the SPI is shown as a continuous variable.The figure shows that total rainfall was above the long-term average in 2013 and below the long-term average in 2016.Malawi experienced flooding in the 2012/2013 season, which affected 16,000 households. 5In 2016, there were widespread droughts that affected Malawi and most of Southern Africa.The El Niño-induced drought left more than 6.5 million people in need of humanitarian aid across Malawi, Mozambique, and Zimbabwe because agricultural production was below national food demand. 6In our data, more floods were experienced (11% of the household-years) than droughts (3% of household-years).
In addition to the key outcome and independent variables, we also summarise the control variables.Control variables included in the models are time-variant household characteristics such as family size, age of household head, education level of the household head, and household income.As a village level control, we include the number of households allocating labour to forest collection, less household i.We include the number of households allocating labour to the forest within the village with the understanding that as more labour is allocated, households have to compete for the limited forest products, which might reduce the effectiveness of forests as natural insurance.Results are robust to the exclusion of all village and household controls.
Forest access is defined at the village level using a community focused group discussion question.The community-level question asked the community members (including leaders) through focus group discussions if they had access to a community resource such as a forest where they were able to extract NTFPs.We use these responses to define forest access at the village level.This variable is mostly time-invariant as 97 per cent of the villages do not have within-variation over time (i.e. they have access or no access throughout the period under consideration).The remaining 3 per cent 7 are categorised as having access if they had access in at least one of the three years.The key variables are summarised for each year in Table A2.Note: The overall mean is 7.93, and the variance is 2.83.
Impact of weather shocks on food security 1777  Impact of weather shocks on food security 1779

Figure A2 .
Figure A2.Distribution of household dietary diversity scores (N ¼ 1294).Note: The overall mean is 7.93, and the variance is 2.83.

Figure A3 .
Figure A3.Standardised Precipitation Index distribution by year for Malawi.SPI < À1.6 or SPI > 1.6 is categorised as a shock.

Table 1 .
Correlation between forest access, forest density, weather shocks, and food security Note: Robust standard errors clustered at the village level are in parentheses.ÃÃÃ p < .01,ÃÃ p < .05,Ã p < .1.

Table 2 .
Mitigating role of forest access on the effect of weather shocks on food security Standard errors are clustered at the village level are in parentheses.All models include household fixed effects, and year fixed effects.Other controls include family size (total labour endowment), age of household head, the number of households in the village allocating labour to the forest, income, total livestock owned (in tropical livestock units), remittances received, and temperature.Note that the sample size for propensity score regressions is lower as we restrict the sample to observations for which the weights are not missing.The marginal impact of weather shocks for those with and without forest access is calculated at the long-term rainfall mean for each group.villageswhichareless likely to have forest access and more likely to be recipients of NGO relief programs) are negatively correlated with forest access.Results on SAFS from the interactive fixed effects are not significant.Results in Table2(columns 1-6) indicate that experiencing a weather shock negatively affects food security.Weather shocks reduce the HDDS by 4-6 per cent and decrease the probability of being food secure (SAFS) by 18-68 percentage points for households without forest access. accessible

Table 3 .
Impact of weather shocks on food borrowing and food expenditure Notes: Standard errors clustered at the village level are in parentheses.Others controls are similar to those included in Table 2. Other controls are similar to Table 2. ÃÃÃ p < .01,Ã p < .1.

Table 4 .
The moderating role of forest density on forests as natural insurance Standard errors clustered at the village level are in parentheses.All models include village and household time-varying controls, household fixed effects, and year fixed effects.HDDS: household dietary diversity score; SAFS: self-assessed food security.Models with HDDS as the dependent variable use FEP model while those with SAFS use LPM.Other controls are similar to Table2.

Table 5 .
Summary of robustness checks

Table 6 .
Effect of weather shocks on labour adjustment to different activities Robust standard errors are in parentheses clustered at the village level.

Table A1 .
List of food groups used to calculate dietary diversity score (Tea; coffee/cocoa/milo; salt; spices; yeast/baking powder; tomato/hot sauce; fish powder/sauce; other condimentincluding small amounts of milk for tea/coffee)

Table A2 .
Summary of key variables used in the analysis MK is Malawi Kwacha, the currency for Malawi.All variables measured in MK are in real terms with 2010 as the base year.Monetary variables are converted to 2010 Malawian Kwacha (real values) by adjusting for inflation using consumer price index data from the World Bank.In 2010 $1¼ MK150.5.HH ¼ Household.For all binary variables, the mean is the average proportion for the category labelled 1.The numbers in parenthesis for forest density represent the proportion of forests with a part of forest having the indicated canopy cover.