Adaptation to Natural Disasters through the Agricultural Land Rental Market: Evidence from Bangladesh

Analytical Framework

Available studies suggest that in areas subject to high risk but with poorly developed capital markets, land sales will likely be few and limited mainly to distress sales (Deininger and Jin 2008). In fact, most rural households consider land to be a source of regular income, a security against famine, and the foundation of their livelihood. Especially in the rural areas, land ownership is an important element of social, political, and economic empowerment. Therefore, most of the land transactions are rentals, and not sales, in a country like Bangladesh.

Land rentals can be a risk coping strategy if rental decisions are made in response to shocks resulting in income losses (Ward and Shively 2015). Farmers make livelihood decisions based on their owned land, and such decisions may be motivated by exposure to extreme climatic events. The key idea behind quantity adjustments through a land rental market is that larger farmers rent out their surplus lands to smaller farmers, who rent in to optimize their operational farm size. We hypothesize that this phenomenon is accelerated when such transactions take place in response to exposure to a natural disaster.

We develop a conceptual model that is similar to Skoufias (1995), Deininger and Jin (2008), and Deschênes and Greenstone (2007).⁴ For simplicity, we assume that the land rental market always clears irrespective of whether a disaster takes place. This assumption implies that in combination with high population density and low per capita arable land, any increased rental transaction in response to disaster exposure must be captured by observed heterogeneity in the socioeconomic and agricultural attributes associated with the groups of households involved in rent-in and rent-out transactions. Therefore, for a representative farmer, optimal rent-in and rent-out amounts, respectively, are

1

where Θ^I and Θ^O are the observed household attributes associated with a rent-in and rent-out farmer, respectively; τ = 1 represents exposure to a natural disaster; and τ = 0 indicates no such exposure. We suppress time and household subscripts to avoid confusions and simplify notations.

The representative farmer produces a mix of crops using its endowment of productive resources, including land. The farmer also obtains earnings from nonfarm economic activities at an exogenously determined wage rate.⁵ However, farmers normally supply nonfarm labor during lean seasons (Deininger and Jin 2008) and often experience credit constraint in doing so (Woldenhanna and Oskam 2001). Rural farm labor supply can be treated as inelastic in farm wage due to lack of labor mobility between sectors and locations.⁶ Considering these facts, we express output and cost as functions of operational land l only. Consequently, the representative farmer maximizes the following profit function by choosing land rental amounts l^I and l^O, which captures the effects of operational farm size adjustments on crop profits:

2

subject to equation [1]. p, q, l, c, and r, respectively, denote agricultural price, output, amount of owned land, cost of production, and the fixed rent per unit of land. Total operational farm size is l + l^I − l^O ∀ l^I ≥ 0, l^O ≥ 0. α ≥ 0 indicates the losses in crop profits due to disaster exposure that results in lowering the productivity of operated land. I is an indicator variable for rent in: 1 if rent in and 0 otherwise. Similarly, O is an indicator variable for rent out: 1 if rent out and 0 otherwise.

We assume that the land rental rates are predetermined either by the government (in case of share-cropping) or by locally practiced norms (especially in case of cash rentals) independently of the occurrence of a disaster, which is consistent with the Land Reform Act of 1984. Rental transactions also involve transaction costs, T^I > 0 for rent in and T^O > 0 for rent out, which are proportional to respective transaction amounts and not symmetric so that the net benefits from per unit of rent-in and rent-out transactions are different (e.g., Deininger and Jin 2008). This assumption follows the theoretical finding of Carter and Yao (1999), who suggest that rent-in transaction costs are higher than the simple rental rate due to associated search costs, while the rent-out receipts are lower than the pure rental rate because of redistribution costs.

We focus on total agricultural revenue instead of per acre yield. Because farmers maximize profits, predetermined rent allows us to normalize the productivity of each type of land, whereas the nonsymmetric transaction costs ensure the existence of institutional differences in returns from rent-in and rent-out transactions.

Prices of agricultural goods can be volatile, and an increase in crop profits may largely be due to increased prices resulting from postdisaster production shortages. We empirically tackle this issue by considering farm-gate and local market prices, whichever one is available, when calculating agricultural revenues. When farm-gate and local market prices are not available, we normalize prices to more aggregate levels (e.g., district or national level) by taking into account that potential relocation of agricultural operations through rental transactions to different plots of land might normalize prices over regions in a specific production year.

Because disaster exposure affects rent-in and rent-out amounts as well as the output, we need to disentangle the direct and indirect effects of disaster exposure. Based on equation [2], changes in the representative farmer’s profits due to disaster for three alternative rental market participating decisions (autarky, rent in, rent out) are given in equations [3] and [4]. First, for an autarkic farmer who does not participate in the land rental market,

3

which implies that an autarkic farmer cannot overcome the losses in crop profits ∀ α > 0 and break even only when α = 0. Subscripts 0 and 1 denote τ = 0 and τ = 1, respectively. Next, for rent-in and rent-out farmers,

4

where for a rent-in farmer, ; ; and . Here, α ≥ 0 governs the direct effect of disaster exposure, whereas and govern the indirect effect of disaster on agriculture through the land quantity adjustments. The expression (r + T^I) refers to the per unit cost for rented in land; that is, the cash and in-kind payments made for the use of rented-in land. On the other hand, for a rent-out farmer, ; ; and . Similar to the case for a rent-in farmer, α ≥ 0 governs the direct effect of disaster exposure, whereas and govern the indirect effect of disaster. The expression (r − T^O) refers to the per unit cash and in-kind receipts from rented-out land. In addition, we must have Δl^I > 0 and Δl^O > 0 for a rent-in farmer, Δl^O > 0 and Δl^I > 0 for a rent-out farmer, and Δl^I = Δl^O = 0 for an autarkic farmer.

Rental market participating roles are affected by disaster exposure, and their corresponding transactions have indirect effects on crop profits. However, rental transactions are conditional on a number of socioeconomic factors. Rent-in and rent-out farmers must exhibit sufficient heterogeneity in their socioeconomic attributes (i.e., Θ^I ≠ Θ^O) for the indirect beneficial effects of land rental transactions to exist. That is, socioeconomic attributes vary across farmers’ rental market participating roles, and optimal adjustment of farm size through rental transactions must be conditioned on them.

Empirical Specifications

We examine the effects of disaster exposure on agricultural outcomes, controlling for land quantity adjustments through farmers’ participation and transaction decisions in the land rental market, using an econometric approach that accounts for extensive and intensive margins. The intensive margin measures the direct effects of disaster on crop profits, whereas the extensive margin considers the potentially mitigating effects of disaster-induced land quantity adjustments on the harms of disaster. Note that we restrict our estimation to agricultural plots to avoid any potential bias that might arise from multiple uses of land plots.

We estimate the effects of disaster exposure on land quantity adjustment through the rental market. As Figure 1 shows, the rent-in and rent-out amounts are left-censored because of farmers’ participation decisions: a positive amount of land brought into rental market for either renting in or renting out is observed only when a farmer decides to participate in the rental market. Moreover, Figure 2 confirms that the rent-in and rent-out amounts are left-censored for the subsamples of flood- and storm-affected and -unaffected households. Thus, the participating samples are nonrandom and are drawn from a wider population of farmers.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Distributions of Land Rental Transactions

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Land Rental Transactions by Year and Disaster

To capture their participation decisions, we employ a PPML model proposed by Silva and Tenreyro (2006) and Silva and Tenreyro (2010). The PPML model estimates Poisson regression by pseudo-maximum likelihood to identify and drop regressors that may cause the nonexistence of the (pseudo-) maximum likelihood estimates. Based on the maximum likelihood estimation method, PPML can be used for any kind of outcome variable provided that the mean function is correct (Wooldridge 1999). There are multiple benefits of using this approach instead of alternatives such as tobit models. First, PPML provides fully robust estimator of conditional mean parameters and overcomes many restrictive assumptions that the tobit estimates are based on. Second, it allows controlling for time and panel fixed effects—an important limitation that makes tobit estimates less attractive. It overcomes the heterogeneity problem pertinent to many panel data sets. Third, instead of basing the estimates only on the participants as in to-bit models, the estimating sample in PPML is representative of both the participants and nonparticipants in the whole sample. Therefore, estimated coefficients are more directly interpretable and can be compared with corresponding OLS estimates. Finally, in the statistical package Stata, this approach directly fits a semi-log specification where the outcome variable is automatically expressed in natural log form so the estimated coefficients are interpreted as semi-elasticity with respect to the explanatory variable.

Although the same group of farmers can be involved in both rent-in and rent-out transactions (Rahman 2010),⁷ only around 5% of farmers from our estimating sample make simultaneous rent-in and rent-out decisions on different plots of agricultural land. Moreover, related literature generally treats rent-in and rent-out farmers as two separate groups of people with distinct attributes (Kung 2002; Deininger, Zegarra, and Lavadenz 2003; Teklu and Lemi 2004; Deininger and Jin 2005; Vranken and Swinnen 2006; Masterson 2007; Holden, Deininger, and Hosaena 2007). By assuming that rent in and rent out are two independent decisions, our econometric investigation involves separate estimations for different rental market roles.

At any point in time, the decision to participate in the land rental market and the optimal rent-in and rent-out amounts by each farmer can be estimated as a two-step process. First, a farmer participates in the land rental market in time according to the following linear probability model (LPM) with two-way fixed effects:

5

where the binary outcome variables representing farmer’s willingness to participate in the land rental market, I_it and O_it, are defined as I_it = 1 if the farmer rents in land and 0 if not and O_it = 1 if the farmer rents out land and 0 if not. Despite the binary nature of the dependent variable, LPMs provide good estimates of the partial effects for average values of the explanatory variables, suffer less from measurement errors, and the coefficients allow for a straightforward interpretation of the effects (Wooldridge 2010). We report robust standard errors, clustered at the union level, to overcome the heteroskedasticity problem.

The dummy variable flood_i denotes flood exposure: 1 if household i is affected by any flood between the BIHS rounds 1 and 2 and 0 if not. Similarly, storm_i denotes storm exposure: 1 if household i is affected by any storm between the BIHS rounds 1 and 2 and 0 if not. post_t denotes postdisaster year: 1 if BIHS round 2 and 0 if BIHS round 1. Δ_i and ρ_t are the household and survey round fixed effects to control for any potential omitted variable bias. Due to the inclusion of Δ_i and ρ_t, collinear variables flood_i, storm_i, and post_t drop out of our estimating regressions.

We are interested in the coefficients of the interaction terms. In particular, β_I and β_O1, the coefficient of (flood_i × post_t), estimate the effects of flood exposure on rental participations, whereas β_I2 and β_O2, the coefficients of (strom_i × post_t), estimates that for storm exposure. Because disaster-exposed farmers either experience floods or storms, we do not consider the possibility of multiple disaster types in our empirical specification.

Farmers’ adaptation is also associated with characteristics of households and local communities (Skoufias 1995). Therefore, our empirical approach to estimating equation [5] involves specifying the components of the vector x_i based on the information available in the BIHS data set. We follow existing literature to specify generic determinants, x_i, of agricultural land rental decisions, which commonly include household- and farm-level characteristics (Taslim and Ahmed 1992; Skoufias 1995; Deininger, Zegarra, and Lavadenz 2003; Teklu and Lemi 2004; Deininger and Jin 2005; Rahman 2010). A household is a group of people who live together and take food from the same pot. The BIHS considered a household member as someone who has lived in the household at least six months and at least half of the week in those months (Ahmed 2013). Household characteristics include the age of the household head, years of schooling of the highest educated family member, and household size. Farm-level characteristics include ownership of a tractor or plow-yoke and access to agricultural facilities; agricultural facilities include agricultural extension services (defined as 1 if the household has access to agricultural extension services and 0 if not) and subsidy (defined as 1 if the household has received agricultural subsidy and 0 if not).⁸

We control for infrastructural variables: proximity to public transportation (defined as 1 if the household is located within 1 km of a bus stop, main road, or train station; 0 otherwise) and public finance (1 if the household is located within 2 km of a bank or a microfinance nongovernmental organization; 0 otherwise). Typically, proximity to public transportation measures the access to market and access to nonagricultural employment, which might also have mitigating effects on the harms of disaster exposure. Controlling for access to nonagricultural employment is important. For example, Kung (2002) found that Chinese households with active participation in off-farm labor markets have rented less land. On the other hand, proximity to public transportation and public finances indirectly controls for the nonagricultural and commercial use of a plot of land. Generally, better access to such infrastructural facilities lowers the dependency on agriculture and may affect rental market participation and transactions. In the absence of a direct measure of migration in response to disaster exposure, infrastructural variables also control for a farmer’s likeliness to migrate to unaffected or urban areas. Finally, we control for exposure to idiosyncratic shocks and long-term variations in monsoon rainfall.

In our empirical analysis, a household’s endowments of family labor, tractor, and access to facilities such as subsidy, extension services and transportation are treated as exogenous. However, these resources are often endogenous because households may determine their optimal levels through different means. For example, family labor can be adjusted through migration and gestation, and a tractor can be traded in the market. Such adjustments can take a longer planning horizon. To be effective, migration or gestation can only adjust labor endowment after a certain time. Therefore, it can be assumed to be exogenous in the short run. Similarly, in agricultural society, selling tractors or plow-yokes is extremely rare and done mostly during extreme liquidity crises. Finally, in Bangladesh, access to above-mentioned facilities becomes publicly available only when the facilities are built by the government. Although this will remain a limitation, since appropriate instruments for these potentially endogenous variables are either unavailable or difficult to conceive, we follow the tradition of Skoufias (1995) and treat them to be determined outside of the model.

We investigate rental transactions using the PPML regression models:

6

where and . Vector x_i is as described for equation [5]. We define the outcome variables, ln LI_it and ln LO_it, as the natural logs of hectares of agricultural land rented in and rented out by farmer i in time t. The PPML approach directly fits a semi-log specification where the outcome variable is automatically expressed in natural log form, so that the estimated coefficients are interpreted as semi-elasticity with respect to the explanatory variable.

Effects of disasters on crop profits are conditional on rent-in and rent-out amounts, which are determined by equation [6], according to farmers’ corresponding participating roles in the rental market. We use two-way fixed-effect panel regression models to estimate the effects of disaster on farmer i in time t by participating roles I (i.e., rent in) or O (i.e., rent out):

7

where and are predicted rental transactions amounts from equation [6]. Y_it represents the measure of agricultural outcomes: crop profits. We include all harvested crops and their local market prices reported by farmers when calculating profits, which are then expressed in thousands of US$ in purchasing power parity (PPP) terms. We deduct the cost of production and the monetary value of all the cash and in-kind payments made for rented-in land from the market value of total harvested crops and then add the monetary value of all the cash and in-kind receipts from rented-out land to calculate crop profits.

In fact, we adopt a modified Ricardian model in equation [7] where we use total crop profits as our outcome variable instead of land value to capture the effects of disaster exposure in agriculture. The use of profits is particularly appropriate in this setup because land markets are often imperfect in Bangladesh, like many other developing areas (Di Falco, Veronesi, and Yesuf 2011), and the use of land values requires fully functioning land markets so that land prices reflect the present discounted value of land rents into the infinite future (Deschênes and Greenstone 2007).

Predicted rent-in and rent-out amounts in equation [7] connect the coefficients of the components of z_i in equation [6] with the outcome variable in equation [7] and yield the indirect effects or extensive margins of disaster exposure through land rental transactions. On the other hand, ϖ₁ and ϖ₂; that is, the coefficients of (flood_i × post_t) and (strom_i × post_i) in equation [7], yield the direct effects or intensive margins of disaster exposure on crop profits. Following Moore, Gollehon, and Carey (1994), the total margins, or total marginal effects, of disaster exposure is the sum of the effects along the intensive and extensive margins for the land rental market participants:

8

where ϖ_I1 and ϖ_I2 are the intensive margins from rent-in transactions in response to flood and storm exposures, respectively. Similarly, ϖ_O1 and ϖ_O2 are the intensive margins from rent-out transactions in response to flood and storm exposures.

The corresponding extensive margins for flood and storm are and for rent-in transactions, and and for rent-out transactions. Respective coefficients of the interaction terms in equation [6] are expressed as semi-elasticity and transformed to marginal effects according to

9

Data and Variables

Data for our analysis are from two rounds of the BIHS. The USAID-funded survey was designed and supervised by the International Food Policy Research Institute, administered by Data Analysis and Technical Assistance, Dhaka, Bangladesh, and approved for publication by the government of Bangladesh (Ahmed 2013). The first round of the BIHS data set was collected between October 2011 and March 2012. Statistically, BIHS is nationally representative of the rural areas of the seven administrative divisions of Bangladesh, with a sample size of 6,500 rural households from 325 primary sampling units. The second round of the survey was collected from January to June 2015 (Ahmed 2016), which was administered on the same sample of households surveyed in the baseline, creating a two-round panel, when 6,260 households from the baseline survey were reinterviewed. We exclude 1,127 households that have not reported their agricultural revenues and rental transactions, and 5 households that were affected by both floods and storms. Our estimating sample consists of 5,128 rural agricultural households. Table 1 describes and summarizes the variables we use in the empirical analysis.

The BIHS data set reports information on a household’s exposure to any negative shock—both idiosyncratic shocks (e.g., death of main earner, loss of a regular job, loss of assets, crop loss, and loss or decrease of remittances) and covariate shocks (e.g., natural disasters). We are particularly interested in household-specific reporting of exposure to natural disasters such as floods and storms. These disasters affect 3% of the surveyed households in between BIHS rounds 1 and 2 (Table 1). We use the self-reported household-level exposure to disaster from the BIHS in our subsequent analysis, therefore overcoming the limitations of regional level disaster exposure data. Most of the small-scale disasters affect specific regions of Bangladesh. Certain regions experience recurring events of natural disasters, making it difficult to identify random treatment and control groups at the regional level. Moreover, the EM-DAT database only reports a disaster if one of these four criteria is fulfilled: (1) 10 or more people are reported killed, (2) 100 or more people are reported affected, (3) there is a declaration of a state of emergency, and (4) there is a call for international assistance. In many cases, this restricts identification of the number of affected people and therefore underestimates the potential effects of disaster exposure at the household level.

The BIHS data set contains information on farm and nonfarm incomes in addition to detailed reporting on revenues and costs associated to crop cultivation, which we used to derive our outcome variables measuring welfare. All monetary values are expressed in thousands of U.S. dollars in PPP terms at the rates of 26.61 and 21.87 Bangladeshi taka per US$ for BIHS rounds 1 and 2, respectively. Table 1 reports that on average, crop profits are evaluated at $1,695 and $887 in 2011 and 2015, respectively. Crop profits are adjusted for the cost of production, cash and in-kind (imputed) receipts from rented-out land, and cash and in-kind (imputed) payments made for rented in land.

Table 1 also reports land management variables, including farmers’ land rental market participation and transaction decisions. All land measures are expressed in hectares, where 1 decimal = 0.00405 ha. In 2011, a total of 40% of farmers participated in the land rental market to rent in 0.13 ha lands on average, and 22% of farmers participated in the land rental market to rent out 0.08 ha land on average. These rental transactions increased operational farm size from 0.13 ha (which is a farmer’s owned-operated land) to 0.26 ha (which includes rented-in land and excludes rented-out land). Similarly, in 2015, a total of 37% and 26% of farmers participated in the land rental market to rent in 0.12 ha and rent out 0.09 ha of agricultural land. Operational farm size remained at similar levels, 0.26 ha in 2011 and 0.25 ha in 2015, although these were rental transactions.

Such rental transactions are conditional on a number of socioeconomic factors, such as household, farm, and regional attributes. Table 1 reports the baseline summary statistics of all the explanatory variables used in the empirical analysis. All data are from the BIHS data set round 1. On average, household size is 4.24 members; household heads are 45 years old, with 85% of them being male. The highest educated member in the family has 6.6 years of schooling on average. Eleven percent of farmers own tractors or plow-yokes, which is an important technology for cultivation in rural Bangladesh. Seven percent and 10% of farmers, respectively, benefit from agricultural extension services and agricultural input subsidy. On average, 62% of households are located within 1 km of a bus stop, a main road, or train station; whereas 26% are located within 2 km of a bank or a source of microcredit.

We use rainfall data from the Bangladesh Department of Meteorology (BMD), which is an organization under the Ministry of Defense responsible for maintaining the weather stations and collecting daily weather data, including rainfall and temperature, at the station level. The BMD website provides monthly average rainfall data from 1980 to 2013. We take the difference between the average rainfall over the monsoon months of June–September for 2006–2010 and 1980–2010 for BIHS round 1, and 2009–2013 and 1980–2010 for BIHS round 2.