Scholarly article on topic 'Weather shocks and cropland decisions in rural Mozambique'

Weather shocks and cropland decisions in rural Mozambique Academic research paper on "Agriculture, forestry, and fisheries"

CC BY-NC-ND
0
0
Share paper
Academic journal
Food Policy
OECD Field of science
Keywords
{"Land allocation" / Risk / "Household farms" / "Pooled Fractional Probit"}

Abstract of research paper on Agriculture, forestry, and fisheries, author of scientific article — César Salazar-Espinoza, Sam Jones, Finn Tarp

Abstract Economic development in low income settings is often associated with an expansion of higher-value agricultural activities. Since these activities often bring new risks, an understanding of cropland decisions and how these interact with shocks is valuable. This paper uses data from Mozambique to examine the effect of weather shocks on cropland decisions. We account for the bounded nature of land shares and estimate a Pooled Fractional Probit model for panel data. Our results show that crop choice is sensitive to past weather shocks. Farmers shift land use away from cash and permanent crops one year after a drought and from horticulture and permanent crop after a flood. However, this reallocation seems temporary as farmers devote less land to staples after two periods. This is consistent with the aim of maintaining a buffer stock of staples for home consumption.

Academic research paper on topic "Weather shocks and cropland decisions in rural Mozambique"

Contents lists available at ScienceDirect

Food Policy

journal homepage: www.elsevier.com/locate/foodpol

Review

Weather shocks and cropland decisions in rural Mozambique Crosse

César Salazar-Espinozaa'*, Sam Jones b'1, Finn Tarpc>1

a Department of Economics, University of Copenhagen, Denmark and Department of Economics and Finance, University of Bio-Bio, Chile b Department of Economics, University of Copenhagen, Denmark

c Department of Economics, University of Copenhagen, Denmark and UNU-WIDER, Helsinki, Finland

ARTICLE INFO

ABSTRACT

Article history:

Received 19 August 2014

Received in revised form 23 February 2015

Accepted 12 March 2015

Keywords: Land allocation Risk

Household farms Pooled Fractional Probit

Economic development in low income settings is often associated with an expansion of higher-value agricultural activities. Since these activities often bring new risks, an understanding of cropland decisions and how these interact with shocks is valuable. This paper uses data from Mozambique to examine the effect of weather shocks on cropland decisions. We account for the bounded nature of land shares and estimate a Pooled Fractional Probit model for panel data. Our results show that crop choice is sensitive to past weather shocks. Farmers shift land use away from cash and permanent crops one year after a drought and from horticulture and permanent crop after a flood. However, this reallocation seems temporary as farmers devote less land to staples after two periods. This is consistent with the aim of maintaining a buffer stock of staples for home consumption.

© 2015 Elsevier Ltd. All rights reserved.

Introduction

Development of the agricultural sector in Mozambique remains a pressing policy issue. Despite rapid rates of aggregate economic growth for almost two decades, headcount poverty rates and rural incomes appear to have remained broadly stagnant, particularly amongst the majority of households that rely on smallholder agriculture (Arndt et al., 2012; Jones and Tarp, 2013). Micro-survey evidence shows few signs of increased agricultural productivity via adoption of improved inputs and/or shifting into higher-return crops (World Bank, 2008; Mather et al., 2008; World Bank, 2012). At the same time, Mozambique faces increased risks from climate shocks. For example, estimates by UNISDR (2009) ranks Mozambique third among the African countries most exposed to risks from multiple weather-related hazards.

This study provides an empirical examination of the impact of weather shocks on crop portfolio choices of small-scale farmers in Mozambique. We address the following questions: are crop choices sensitive to weather shocks? If so, is there any pattern of reallocation in response to shocks? And, are there systematic patterns in response to shocks? For instance, farmers may be more sensitive to more severe shocks or farmers living in higher risk areas may be less responsive to weather shocks.

* Corresponding author at: 0ster Farimagsgade 5, Building 26, DK-1353 Copenhagen K, Denmark. Tel.: +45 35324401.

E-mail addresses: cesar.salazar@econ.ku.dk, csalazar@ubiobio.cl (C. Salazar-Espinoza), esamjones@gmail.com (S. Jones), Finn.Tarp@econ.ku.dk (F. Tarp).

1 Address: 0ster Farimagsgade 5, Building 26, DK-1353 Copenhagen K, Denmark.

The motivation for studying these questions relates to the impact of risks (and their realization in actual shocks) on the economic behavior of households. In the absence of functioning markets for credit, insurance and savings, rural households must largely rely on crop choice decisions to manage risk (Dercon, 2002; Kurukulasuriya et al., 2006). Furthermore, the incidence of shocks may shape farmers' perceptions of the general riskiness of their environment and influence crop portfolio choices. Following Gollier and Pratt (1996), farmers may be 'risk vulnerable' in the sense that the presence of an exogenous background risk (climate) raises their aversion to other risks (e.g., through crop choices).

Existing empirical evidence suggests that farmers react to weather risks by diversifying their cropping system, which acts as a form of self-insurance (Benin et al., 2004; Di Falco et al., 2010; Bezabih and Sarr, 2012; Bezabih and Di Falco, 2012). Rather than focusing on diversification per se, we explore changes in cropland allocation across different crop categories. In the case of Mozambique, some staples show risk-reducing properties in terms of drought tolerance and ease of storage. As such, it is an attractive choice for risk-averse farmers (Arndt and Tarp, 2000; Tarp et al., 2002). Equally, it is reasonable to assume that buffer stocks of staple foods, particularly grains, may be reduced in response to weather shocks to smooth consumption (Kazianga and Udry, 2006). Following a shock, households may prefer to devote a larger share of their land to staple foods in order to replace this buffer, implying income from higher value crops may be reduced. Accordingly, while diversification is of interest it is important to understand exactly how cultivation choices respond

http://dx.doi.org/10.1016/j.foodpol.2015.03.003 0306-9192/® 2015 Elsevier Ltd. All rights reserved.

to shocks (if at all) as well as the persistence of these portfolio changes.

The remainder of this study is organized as follows: Section 'Existing literature' reviews literature linking risk and crop choice. Section 'Agriculture and climate in Mozambique' describes key characteristics of the agriculture sector and climate patterns in Mozambique. Section 'Data' presents the data, including geospatial data on water availability, which we use to distinguish between drought and flood events. Reliance on external as opposed to self-reported data on shocks is helpful. It addresses concerns of systematic reporting bias since weather shocks are a function of geographical location (Cameron and Shah, 2013). Section 'Empirical strategy' describes our econometric model. We model cropland decisions as proportions; and, in order to address the fact that proportions are bounded between zero and one, we estimate a Pooled Fractional Probit (PFP) estimator due to Papke and Wooldridge (2008). We are unaware of existing studies that apply the PFP while controlling for unobserved characteristics. Section 'Results' discusses the main results; Section 'Robustness' considers a number of some robustness tests; and Section 'Conclusions' concludes.

Existing literature

Large fluctuations in weather conditions are generally associated with sizeable yield and price risk in agriculture. Moreover, since such shocks often affect an entire network, local mutual insurance schemes can break-down (Dercon, 2002). Consequently, in contexts of incomplete markets and limited asset holdings, ex post coping mechanisms cannot be relied upon to protect against exogenous shocks (Paxson, 1992; Townsed, 1994).2 Exposure to risk is therefore likely to affect ex ante crop choices (Fafchamps, 1992a; Chavas and Holt, 1996; Kurosaki and Fafchamps, 2002).

The concepts of 'risk' and 'shock' are often used to refer to situations characterized by uncertainty. Following Cohen et al. (2008), perceptions of context can be understood as being derived from a sequence of past events. The evaluation of risks by individuals can be expected to be dependent on past experiences. Under this process of adaptive expectation formation, weather risk can be proxied by past realizations of weather-related shocks. This means that droughts and floods occurring in the (recent) past are likely to shape farmers' perceptions of the current riskiness of their environment.

The incidence of a natural hazard is one element of background risk. If farmers are risk vulnerable, in the sense of Gollier and Pratt (1996), they may display more risk-averse behavior. The latter would be consistent with farmers preferring a crop-portfolio with a larger share of staples.3 Farmers may switch to staples after weather shocks for several reasons. First, some staples are relatively more drought resistant and less prone to crop failure during water shortage periods. Consequently, if the household consumes one of its crops, this provides self-insurance against production and consumption price risk (Fafchamps, 1992a). Second, some staples are less perishable and can be stored for future consumption. Food is likely to be expensive after weather shocks when the harvest is poor. In this case, households will use their stock of staples to smooth consumption in the current period and will expand staples production in the next period so as to replace the depleted stock. Even though general empirical evidence suggests that consumption smoothing

2 Credit constraints, commitment failure and imperfect flows of information among members of the community have been identified in the literature as potential causes of inefficiency of these institutions (Fafchamps, 1992b).

3 Some psychological studies suggest that individuals who are continually exposed to high risk environments may not care about the addition of a small independent risk (Kahneman and Tversky, 1979). This suggests that controlling for background risk may be important.

is limited in low income contexts, evidence does point to smoothing through the accumulation and depletion of staples stocks (Fafchamps et al., 1998; Kazianga and Udry, 2006). Indeed, Carter and Lybbert (2012) find that staples stocks play a more important role amongst very limited consumption smoothers.

A large literature studies the cropland decisions of small landholders in developing countries (see for example, Fafchamps, 1992a; Dercon, 1996; Kurosaki and Fafchamps, 2002; Masanjala, 2006; Damon, 2010; Chibwana and Fisher, 2012). One strand of the literature has investigated the potential advantages of multi-cropping as a risk management device (Adger et al., 2003; Benin et al., 2004; Di Falco and Chavas, 2009; Di Falco et al., 2010; Bezabih and Sarr, 2012; Bezabih and Di Falco, 2012). In addition, crop choice is identified as an adaptation strategy to climate change. For instance, Seo and Mendelsohn (2008) and Kurukulasuriya and Medelsohn (2008), using data of South-American and African farmers respectively, found that crop choices are highly sensitive to changes in precipitation and temperature under different climate change scenarios. Di Falco and Veronesi (2013) find that crop adaptation is more effective when it is implemented within a portfolio of actions rather than in isolation. For example, crop adaptation yields high net revenues when coupled with water conservation strategies or soil conservation strategies. We build on this literature, focussing on the Mozambican context, to which we now turn.

Agriculture and climate in Mozambique

Primary sector activities, which include agriculture and extractive industries, contribute around 30% of Mozambique's GDP; and agriculture alone employs 80% of the work force (Jones and Tarp, 2013). The agricultural sector remains relatively unproductive and consists mainly of smallholder farmers, who represent 85% of all rural households (World Bank, 2012). While rural agricultural markets are widespread, more than half of total household incomes correspond to the value of retained food. Major cash crops are sugar cane, coconuts, cotton, sesame, tobacco and cashews, and the main staple crops are maize, sorghum, millet, rice, beans, groundnuts, vegetables and cassava. More than 75% of small farms cultivate maize or cassava or both, which are also the main staples. Agriculture is predominantly rain-fed with less than 0.5% of total cropland under irrigation, almost all in sugar cane production (World Bank, 2010).

Mozambique has a rainy season lasting from October to April, with an annual average precipitation around 1000 mm. The rural population is frequently affected by extreme weather variations, where droughts and floods are the most common weather-related disasters (EM-DAT, 2013). Droughts are the most frequent natural phenomenon, occurring mainly in the southern and central districts, with a frequency of 7 in 10 and 4 in 10 years, respectively. Although less frequent, floods are more destructive and their effects can prevail for a longer time. They primarily occur in southern and central regions, along river basins, in low-lying areas, and in zones with poor drainage. They are caused by either heavy rainfall or increases in water levels in upstream neighbouring countries. Climate change will likely make weather fluctuations more frequent and extreme in the future. In particular, projections for Mozambique indicate that climate change is expected to increase the frequency and magnitude of droughts and floods, imposing important costs on Mozambique's economy and further complications for existing development challenges. Estimations for the worst case scenario suggest that GDP may fall between 4% and 14% relative to baseline growth in the 2040-50 decade in Mozambique if adaptation strategies are not implemented (World Bank, 2010; Arndt and Thurlow, 2013). Changes in cropland

are one of the key adaptation strategies to understand in order to assist planning by policymakers and quantify the impact of climate change (Seo and Mendelsohn, 2008).

Household data

We use a balanced panel of households from the 2002 and 2005 waves of the Trabalho de ¡nquérito Agrícola (TIA) survey collected by the Ministry of Agriculture of Mozambique in collaboration with Michigan State University (Ministério da Agricultura 2005; Ministério da Agricultura e Desenvolvimento Rural 2002).4 The TIAs are representative of small and medium-size farm households across rural areas of the 11 provinces in Mozambique (one province, Maputo City, is exclusively urban and not included here).5 The survey consists of a series of questions concerning household demographic characteristics, assets, farming techniques, access to services and community characteristics. Data also contain farmers' reports of amounts of hectares allocated to different crops. We use 3752 observations for which data on land shares are available.

Panel (a) of Table 1 reports descriptive statistics on changes in crop decisions from the dataset. It shows an increase of 2% in non-staple cropland share between 2002 and 2005. While cash crop and horticulture area increased during the period, permanent crop area decreased. The upward trend in non-staple crops was due to an overall increase in the cultivation of cash crops and horticulture. On average, farmers allocate around 50% of their land to cassava and maize. This percentage has remained unchanged during the study period. The uncultivated land share decreased around 2% between 2002 and 2005. The decrease in uncultivated land is more likely to reflect an expansion of cultivated area rather than changes in fallow land. This in line with the view that agricultural growth observed during that period was mainly driven by expansion in land use rather than productivity improvements (Mather et al., 2008).

Geospatial data

To identify which villages (locations) experienced weather shocks, we rely on information on villages' GPS coordinates recorded in the TIAs. To identify areas that have been flooded, we employ geospatial data recorded in the Global Active Archive of Large Flood Events from the Dartmouth Flood Observatory (Brakenridge, 2013). To identify drought areas, we use calculations of the Standardized Precipitation Index (SPI) by the National Centre for Environmental Predictions (NOAA) (McKee et al., 1993, 1995).6 Specifically, we use a SPI index constructed on 0.5° lat/lon grid

4 There are 8 TIA surveys conducted with interruptions during the period 19962012 (1996, 2002, 2003, 2005, 2006, 2007, 2008 and 2012). However, only the TIA 2002 wave contains a sample that was re-interviewed latter on in 2005, which makes it possible to make a panel solely using these two years. We exploit the panel structure of the TIAs since controlling for household heterogeneity is a critical issue when studying land allocation.

5 The sampling frame of the TIA survey was derived from the Census of Agriculture and Livestock 2000, and used a stratified, clustered sample design that is representative of small- and medium-scale farm households at the provincial and national levels, leaving out large commercial farms from the design. In particular, households cultivating more than 50 hectares of land or owning more than 20,000 fruit trees, more than 100 heads of cattle or more than 500 goats and pigs are classified as large-scale farmers and are not covered by the TIA surveys. Potentially, large-scale farmers may face very different trade-offs regarding crop choices than farmers in our sample. Heterogeneity between large and small landholders may be interesting to explore in future research.

6 The SPI is based on a long-term precipitation record of at least 50 years of

monthly values. This long-term record is fitted to a probability distribution, which is

then transformed into a normal distribution so that the mean SPI for the location and

desired period is zero.

monthly precipitations of 1949-2014 in Mozambique.7 We consider two time-scales. First, we compute the SPI over the main rainy season (November-April). When taking into account the rainy season, we assume that farmers respond to prospects of a good/bad season which is a function of how good/bad the general growing condition was in previous periods. Second, we compute a 3 months SPI index over the main planting/sowing period (October-December). That is relevant for most cash and staple crops. In a country dependent on rain-fed agriculture, erratic rains in the planting/sowing season will increase the probability of crop failure.

The SPI index includes both positive and negative values. Positive SPI values indicate that rainfall was above the median precipitation and negative values show that precipitation was below the median for that period. We define shock occurrence at the village level. Natural hazards are covariate shocks that are highly likely to have a common effect on the whole area of occurrence, and then over the entire village's population. We have a sufficient number of villages (525) to generate enough variation in our shock variables. We define two drought variables. First, we compute a drought indicator if the SPI value falls at or below minus 0.5. In addition, we exploit the continuity in the negative range of the index to explore drought intensity. We use the absolute values. Thus, a larger value would indicate a more severe dry cycle. Finally, we construct measures of the historical occurrence of natural shocks by counting the number of events in each village, going back 20 years to 1984. We use these measures to split the sample and study how weather shocks affect crop land decisions conditioned on background risk.8

We focus on whether a village was affected by a weather shock in t - 1 and/or t - 2. That is, shocks in 2000-2001 and 2003-2004 are used to explain the cropland allocations observed in 2002 and 2005, respectively. This lag is used because we are interested in how past events shape future behavior. Table 1, panel (b) summarises the weather shock data; and Fig. 1 maps flooded areas for the years of interest overlaid with the locations of surveyed villages. It shows that flooding predominantly affected villages in southern and central regions, although northern villages were also hit by the 2003 flood.

All floods identified here were classified at least as class 1 or large flood events. This implies significant damage to structures or agriculture, fatalities, and/or a 1-2 decades-long reported interval since the last similar event (Brakenridge, 2013). However, floods vary in duration and extension. For example, floods in 2004 affected few cities or districts, covering around 4400 sq km and lasting for almost two weeks. In contrast, floods in 2000, 2001 and 2003 were national-scale disasters as effects extended to multiple provinces. To illustrate, these large scale floods covered areas between 200,000 and 440,000 sq km, and in some cases lasted for months (2000 and 2001). These large scale natural

7 Some other rainfall data sources may eventually be used. For example, remote sensing estimations developed by the Famine Early Warning Systems Network (FEWS NET) provide a higher resolution rainfall data at 0.1°, corresponding to around 10 x 10 km cells at the equator. However, this data is only available from 1995. Thus, the shorter temporal coverage makes it problematic to compute a reliable SPI index since it is highly recommended to have at least 50 years of historical rainfall data (McKee et al., 1993,1995). Alternatively, data with longer temporal coverage is also provided by the Climate Research Unit from the University of East Anglia at 0.5°. The data used here also has a resolution of 0.5° and goes back more than 50 years, fulfilling the criterion outlined above.

8 In order to guarantee sufficient observations, we use convenient thresholds to distinguish between low and high risk areas. We define a low flood risk village as that one has experienced between zero and one flood event in the last 20 years, and a high flood risk village as that one has been hit by a flood between 2 and 5 times. Similarly, we use the information on the number of droughts to distinguish low (between zero and 7 droughts] from high drought risk villages (between 8 and 11 events). We aim to identify the effect of recent weather shocks on cropland decisions, conditional on villages' background risk

Table 1

Descriptive statistics for 2002 and 2005 by crop category. Source: Authors' calculations based on TIAs 2002 and 2005, using the balanced panel (N = 3752).

Variables 2002 2005

Mean Dev Mean Dev

Panel (a)

Non-staple crop land share 0.12 0.16 0.14 0.19

Non-staple crop land (hect) 0.33 1.59 0.49 1.56

Ln (Non-staple crop land) 0.19 0.33 0.27 0.41

1 = Non-staple crop land > 0 0.44 0.50 0.44 0.49

Cash crop land share 0.04 0.11 0.05 0.12

Permanent crop land share 0.06 0.12 0.05 0.14

Horticulture land share 0.02 0.04 0.04 0.09

Maize-cassava 0.49 0.22 0.49 0.23

Sorghum-millet 0.08 0.16 0.08 0.15

Groundnut-beans 0.18 0.16 0.19 0.17

Sweet potatoes land share 0.02 0.06 0.02 0.05

Rice land share 0.07 0.16 0.05 0.15

Uncultivated land share 0.04 0.12 0.02 0.09

Panel (b)

1 = village was hit by a flood (t - 2) 0.21 0.41 0.21 0.41

1 = village was hit by a flood (t - 1) 0.22 0.41 0.01 0.12

# times a village has been affected 1.29 1.25 1.29 1.25

by a flood (last 20 years)

1 = village was hit by a drought 0.12 0.32 0.33 0.47

(t - 2) (rainy season)

1 = village was hit by a drought 0.05 0.22 0.48 0.50

(t - 2) (planting season)

1 = village was hit by a drought 0.05 0.22 0.00 0.00

(t - 1) (rainy season)

1 = village was hit by a drought 0.00 0.00 0.23 0.42

(t - 1) (planting season)

# times a village has been affected 8.10 1.6 8.10 1.6

by a drought (last 20 years)

Panel (c)

Total landholding (ha) 2.14 2.75 2.41 2.94

Ln (total landholding) 0.97 0.53 1.06 0.52

# plots 2.43 1.32 2.02 1.15

# family members 5.78 3.51 6.14 3.83

% family members with off farm jobs 0.07 0.16 0.13 0.23

% family members self-employment 0.15 0.23 0.21 0.27

Head's education level (years) 2.04 2.33 2.43 2.55

1 = HH received remittances 0.20 0.40 0.24 0.43

Wealth index 1.96 0.97 2.03 1.02

% plots with irrigation system 0.06 0.19 0.04 0.18

% plots with land title 0.01 0.11 0.03 0.15

1 = HH used animal traction 0.21 0.41 0.18 0.38

1 = HH used fertilizer 0.05 0.22 0.05 0.22

1 = HH received extension services 0.15 0.36 0.19 0.39

1 = HH belonged to farm 0.05 0.22 0.09 0.29

organizations

1 = HH received market price 0.31 0.46 0.39 0.49

information

Average regional retail maize price 2393.29 477.56 2378.46 481.14

(t - 1)

1 = village has electricity 0.08 0.26 0.13 0.33

% sick family members 0.01 0.06 0.02 0.09

1 = HH suffered a death (t - 1) 0.04 0.19 0.07 0.26

1 = HH suffered a divorce (t - 1) 0.01 0.09 0.03 0.16

Observations 3752 3752

Note: Panel (a) describes the dependent variables. Panel (b) shows the weather shock variables. Panel (c) displays the descriptive statistics for the controls.

hazards affected around 20% of households included in our sample (see Table 1). In particular, the 2000 flood is classified as a very large event (class 2) and is remembered as one of the worst natural disasters in 50 years in Mozambique (World Bank, 2010).9

Fig. 2 maps findings from the SPI for different years. For 2000 we see a dry cycle in central districts and wet cycles in the south. This extremely wet period is consistent with the flood identified in

9 The available data only allows us to distinguish intensity levels across different flood events but not within floods, which makes it difficult to formally test the effect of the duration/severity/magnitude of floods on cropland decisions.

Fig. 1 in the same year. In addition, drought events are detected in the south region in 2001 and in all regions in 2003.10 According to the SPI calculated over the rainy season, no droughts occurred in 2004. However, the SPI over the plating season does detect erratic rains at the beginning of the growing season in the south in 2004. Furthermore, it also shows a delay in precipitation in 2000, as illustrated in Appendix Fig. A1. Table 1 panel (b) shows that the percentage of households included in our sample affected by droughts ranges from 5% in 2001 to 33% in 2003.

Empirical strategy

Cropland decisions are commonly measured as proportions bounded between zero and one (Papke and Wooldridge, 2008). One challenge in modeling crop allocations in Mozambique is that there is a significant fraction of farmers that do not actually allocate land to non-staple crops (more than 50%, see Table 1), meaning many observations are corner solutions at zero. We address this statistical challenge by using the Pooled Fractional Probit (PFP) estimator. This relies on Bernoulli quasi-likelihood methods to ensure that estimates of predicted land shares vary between zero and one (Papke and Wooldridge, 1996). Furthermore, this model is appropriate for panel data that contains a large cross-sectional dimension and relatively few time periods (Papke and Wooldridge, 2008).

We consider a random sample of farmers i = 1.....N, repeated

across time period t =1.....T. The dependent variable yit corresponds to the land share allocated to a particular crop category (see below). Our empirical model is specified as:

E(yu\Xit, Zt, C) = U(fiXit + yzit_1 + c) (1 )

where xit is a vector of household and farm physical characteristics. zit_1 represents a vector of past weather shocks, i.e., flood and drought events. Coefficients b and y denote parameters to be estimated; ct refers to individual-specific unobserved characteristics; and U is the normal cumulative density function. In order to account for the unobserved effects ci, Papke and Wooldridge (2008) propose a conditional normality assumption to restrict the distribution of ci, given time averages of covariates:11

Ci = W + nzi + VA + ai (2)

where Xt = T-1^T=ixit and zt = T~1J2T=1zit-1 are vectors of time averages; and ai ~ N(0, aa) is a residual orthogonal term. With these assumptions, vectors b and y and associated average partial effects (APEs) can be identified up to a positive scaling factor. To see this, plugging (2) in (1) yields:

E(yit\Xit, Zt, ai) = U(W + bXit + yZit_1 + nXi + VZ + ai) (3)

Or equivalently:

E(yu\xu, Zit ) = E(U[W + bXit + yZit-1 + nz + Vzi + ai]\xu, zu ) (4)

Next, we employ a standard mixing property of the normal distribution (Wooldridge, 2010), yielding:

_ _ .,1 E(yit\Xit, Zit ) = U[(W + bXit + yZit-1 + nXi + VZ')/(1 + r2)2] (5)

10 We could have used flooded area data and the positive scale of SPI identifying wet scenarios to classify flood events according to the extent of seriousness, as in the drought case. However, a flood is a much more complex phenomenon that responds to other parameters than rainfall, which would invalidate any classification exclusively based on precipitation levels. For instance, the flood recorded in 2001 affecting mainly the central region originated from a very wet season in neighbouring Zambia and Zimbabwe that led to the opening of floodgates at the Kariba dam, and waters released from the Cahora Bassa Dam in Mozambique, flooding low-lying areas located further downstream.

11 This strategy was first suggested by Chamberlin (1980).

Fig. 1. Polygons of flooded areas and villages' locations. Source: Authors' elaboration using TIA data and information from Dartmouth Flood Observatory.

which can be estimated via maximum likelihood methods treating ra as a parameter to be estimated.

To verify estimates from the PFP approach, we also estimate a Correlated Random Effects (CRE) Tobit model for panel data, which assumes crop land decisions are simply censored at zero. However, if household decisions regarding crop participation and land amounts are determined by different underlying decision processes, this approach may be restrictive. Thus, we also estimate the Double-Hurdle (D-B) model due to Cragg (1971).12 Finally, for comparison, we show results of a simple linear fixed effect (FE) model. Note that in all estimations we control for a large number of covariates. Descriptive statistics for these covariates are shown in Table 1, panel (c). Further details can be obtained on request.

With respect to the dependent variable(s), we begin by classifying the household production portfolio into staple and non-staple crops; we then study changes in the land share allocated to non-staples. Subsequently, we consider a more disaggregated classification covering nine non-overlapping categories: cash crops, permanent crops, horticulture, cassava-maize, sorghum-millet, groundnut-beans, rice, sweet potatoes, and uncultivated land.13 This disaggregation is important. First, while annual crops are

12 We follow the same strategy as in the PFP model to account for household heterogeneity.

13 Uncultivated land is defined as land that has been ploughed and harrowed

previously but has been left without being sown, typically because of lack of means to work it, to restore its fertility or to avoid surplus production.

produced from plants which last one season, permanent crops are perennial and not replanted after each harvest. Thus, it is not as easy to adjust permanent crop land in the short-run. Second, the distinction between cash and food crops is important. Whilst all crops have potential to be sold, cash crops are those that are non-edible and which cannot serve as (food) self-insurance. Third, similar to staple crops, horticulture has a short farming cycle, needs minimal capital investment, and part of its production can be used to satisfy food needs.14 However, horticulture is generally irrigated and is found more extensively near main urban areas.

Fourth, we distinguish maize and cassava from other staples. These two crops are the main staples in rural diets and are also important cash generating source. Fifth, we aggregate sorghum and millet. They can be considered general substitutes for maize, but are more drought resistant, and have roughly the same growing season. Sixth, groundnuts and beans are studied together. They are often used in rotation with the main cereal. Seventh, we distinguish rice from other staples since rice is not sensitive to flooding and is mostly sold as a cash crop to urban areas. Finally, we study land allocations to sweet potatoes, a classic crop for food security. This crop has a shorter and flexible farming cycle and has the capacity to grow in poor growing conditions.

14 In 2002, there were some missing observations for the horticulture category. For this year, we computed estimates of horticulture land shares by subtracting all the rest of crop categories from total land. We then replaced the missing information with these estimates.

Fig. 2. Drought identification based on a 6-month SPI (November-April). Note: Red color identifies droughts (SPI lower than -0.5); yellow show normal climate conditions (SPI between -0.5 and 0.5); and green areas identify wet periods (SPI greater than 0.5). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Results

Weather shocks and non-staple cropland share

Columns 1-3 of Table 2 report our main results. They are derived from the PFP estimator, from which average partial effects are calculated. Column 1 includes only the flood shock variables; column 2 replaces the flood shocks with drought shocks; and column 3 includes both flood and drought shocks simultaneously, which is our preferred specification.15 All specifications include a full set of control covariates (shown) as well as the average of covariates to control for unobserved household fixed effects (not shown). The remaining columns of Table 2 report results for the same specification using alternative estimators. Column 4 is a simple fixed effects panel estimator; column 5 reports APEs of the CRE model; and columns 6-7 report the participation and quantity equations from the D-H model.

Across all specifications and estimators, we note that cropland decisions are sensitive to recent weather shocks. Whilst, there are some differences in the magnitude of estimated coefficients, they are similar. Results from the D-H model are not directly comparable to the other columns. However, they continue to indicate a significant effect of weather shocks on both participation and quantity allocated to non-staple crop farming.

Taken together, the estimates show that farmers switch away from higher-value non-staple crops in response to prior flooding. On average, farmers reduce the land share allocated to non-staple crops by 4.2% and 2.5% following a flood in t - 1 and t - 2, respectively (Table 2, column 1). The marginal effect due to a flood in t - 2 is reduced while the marginal effects associated with a flood in t - 1 slightly increase to 4.7% after controlling for recent drought shocks (see column 3). In comparison, the average farmer reduces the land share allocated to non-staple crops by 8% after a recent drought event (t - 1). In sum, the evidence indicates that farmers are more responsive to droughts and that responses to shocks are strongest in the short run.16

Weather shocks and crop portfolio changes

Table 3 reports results for the effect of past weather shocks on disaggregated crop categories. In keeping with the results discussed above, floods drive a switch away from permanent crops and horticulture toward both maize-cassava crop farming and uncultivated land. Changes in uncultivated land are also driven by the effect of a flood in t - 2. While we investigate this latter result further in Section 'Robustness'; we note here that this may be due to the extreme devastation of flooding in 2000 (World Bank, 2010). Substantial losses in terms of arable land, equipment and livestock, as well as actual displacement of households, may account for the increase in uncultivated land share since many farmers were left with limited means to work their land.17 We also note that farmers respond to flood events in t - 1 by reducing the sorghum-millet land share at time t, which is consistent with farmers substituting sorghum-millet for other staples.

Recent droughts produce a similar pattern of reallocation. Farmers move away from cash and permanent crop farming to staple crops. While groundnuts-beans increase after a drought in t - 1, farmers respond by increasing sorghum-millet and sweet potatoes land shares after a drought in t - 2. However, we did not find statistically significant changes in maize-cassava land shares

15 Standard errors for the APEs were obtained using 500 bootstrap replications clustered at the household level.

16 This difference is significant at 1% (t = 124).

17 On the other hand, more land left without being sown may simply reflect a farmer's decision to avoid surplus production or more time required to restore land fertility after this devastating flood.

after a drought. The negative effect of droughts on the permanent crop land share must, as already noted, be interpreted with caution. However, the most plausible explanation of this result refers to how permanent crop farming is carried out and estimated. In most cases, farmers practice inter-cropping meaning that it is unusual for tree crops to be the only cultivated crop in an area. Consequently, a negative change in permanent crop land share probably indicates that farmers are simply intensifying intercropping practises.

We also note that farmers respond to drought events in t - 1 by increasing the horticulture land share at time t. Horticulture in Mozambique is predominantly carried out in peri-urban areas, and therefore is more likely to have access to reliable (e.g., piped) water sources. Additionally, the rice land share is reduced after a drought in t - 1 and t - 2. This persistent effect responds to the fact that the flooded condition of rice fields is necessary for rice growth, implying that drought events are an important source of production risk for rice. Finally, we note that farmers also tend to increase the land share that is uncultivated after a drought in t - 2. This may appear to contradict the need to restore household food stock. However, although arguably less devastating than floods, droughts can generate important material and human losses. Also, depending on their severity, they can exhaust soil quality (FAO, 2005). The implication is that farmers may consider it optimal to work intensively on a smaller cropped land area following a drought, thereby allowing land to recover.

Robustness

Timing of drought events

Poor and erratic rains in the planting/sowing season may lead to a reduction in potential yields and overall crop production. In turn, this may induce farmers to alter their crop portfolio. Thus, rather than defining drought events with respect to rain shortages during the rainy season, we re-estimate the model and define drought shocks with reference to the main planting/sowing period (October-December). Since this period is most relevant for cash and staple crops, we focus on these categories for clarity.

Table 4 reports our results now using the modified drought indicator. The results suggest that the timing of rain shortage is relevant. Specifically, farmers respond to a drought in t - 1 by reducing land shares to maize-cassava and groundnuts-beans and increasing land allocated to sorghum-millet and sweet potatoes. This reduction in the maize-cassava land share may seem inconsistent with food security concerns. However, sorghum and millet resist drought better than maize. Also, evidence indicates that sorghum, although mostly substituted by maize in the 1940s, is now being promoted to provide greater resilience to drought. Furthermore, sweet potatoes are well known for being a classic crop for food security. This crop provides, on average, more micro-nutrients per hectare and day than maize and cassava, has a shorter and flexible farming cycle and has the capacity to grow in poor growing conditions and during post-disaster periods. These characteristics also make sweet potatoes one of the preferred crops when maize and cassava fail. Finally, we also note that, farmers respond to drought events in t - 2 by reducing the sorghum-millet land share at time t. This is consistent with a re-adjustment of their buffer stocks of food staples. That is, in t - 1 farmers may have deliberately over-produced sorghum-millet to replace a diminished buffer stock. At time t, farmers then lower the sorghum-millet land share in line with normal consumption needs.

Drought intensity

A further concern with our definition of drought shocks is that it relies on a binary distinction between events. To explore whether

Table 2

Average partial effects for land allocated to non-staple crops.

Variables (1) PFP (2) PFP (3) PFP (4) FE (5) CRET (6) (7) Double hurdle model Probit Tobit

Flood (t - 2) -0.025*** -0.012* -0.013* -0.010 0.033** -0.046***

(0.005) (0.007) (0.009) (0.010) (0.017) (0.012)

Flood (t - 1) -0.042*** -0.047*** -0.048*** -0.060*** -0.077*** -0.066***

(0.007) (0.007) (0.008) (0.011) (0.026) (0.020)

Drought (t - 2) 0.001 0.001 0.002 0.013 0.059*** -0.032*

(0.008) (0.01) (0.009) (0.013) (0.023) (0.017)

Drought (t - 1) -0.085*** -0.083*** -0.120*** -0.142*** -0.258*** -0.184**

(0.01) (0.012) (0.022) (0.021) (0.068) (0.087)

Ln (landholding) 0.054*** 0.055*** 0.057*** 0.062*** 0.267*** 0.117*** 0.290***

(0.008) (0.008) (0.008) (0.009) (0.015) (0.021) (0.021)

# plots -0.010*** -0.009*** -0.011*** -0.012*** -0.016*** -0.009** -0.009*

(0.003) (0.003) (0.003) (0.003) (0.005) (0.005) (0.005)

# family members -0.004** -0.003** -0.004** -0.004** -0.005* -0.002 -0.006*

(0.002) (0.002) (0.002) (0.002) (0.003) (0.005) (0.003)

% family members with off farm jobs 0.037** 0.037** 0.036** 0.036*** 0.031* 0.028 0.063**

(0.014) (0.015) (0.015) (0.014) (0.019) (0.037) (0.032)

% family members self-employment 0.034*** 0.029** 0.033*** 0.033*** 0.039** 0.024 0.062**

(0.012) (0.012) (0.012) (0.012) (0.015) (0.029) (0.024)

Head's education level (years) 0.004 0.004 0.004 0.004 0.005 0.004 0.003

(0.003) (0.003) (0.003) (0.003) (0.004) (0.007) (0.005)

1 = HH received remittances 0.009 0.009 0.010 0.009 0.013 -0.002 0.021

(0.007) (0.008) (0.008) (0.008) (0.010) (0.018) (0.015)

Wealth index 0.009* 0.006 0.007 0.008 0.011 0.003 0.015

(0.005) (0.005) (0.005) (0.005) (0.007) (0.014) (0.009)

% plots with irrigation system 0.029* 0.027 0.028 0.031* 0.060*** 0.079* 0.048

(0.018) (0.017) (0.017) (0.018) (0.023) (0.046) (0.032)

% plots with land title -0.021 -0.018 -0.019 -0.020 -0.032 -0.031 -0.064

(0.021) (0.020) (0.020) (0.021) (0.033) (0.059) (0.056)

1 = HH used animal traction 0.007 0.004 0.0030 0.004 -0.003 0.027 0.001

(0.012) (0.012) (0.011) (0.012) (0.017) (0.025) (0.022)

1 = HH used fertilizer 0.099*** 0.100*** 0.106*** 0.113*** 0.180*** 0.096*** 0.119***

(0.021) (0.021) (0.021) (0.019) (0.036) (0.029) (0.026)

1 = HH received extension services 0.004 0.005 0.005 0.006 0.005 0.001 -0.004

(0.007) (0.008) (0.008) (0.008) (0.011) (0.019) (0.015)

1 = HH belonged to farm organizations 0.013 0.013 0.014 0.016 0.031 0.042 0.010

(0.012) (0.012) (0.012) (0.013) (0.019) (0.028) (0.021)

1 = HH received price information -0.007 -0.003 -0.004 -0.004 0.001 0.012 -0.017

(0.006) (0.006) (0.006) (0.006) (0.008) (0.015) (0.011)

Average regional retail maize price (t - 1) 0.000 0.000 0.000 0.000 0.000 0.000 0.000

(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)

% sick family members 0.006 0.002 0.004 0.004 -0.011 0.024 -0.071

(0.042) (0.041) (0.042) (0.044) (0.039) (0.084) (0.078)

1 = HH suffered a death (t - 1) 0.003 0.004 0.004 0.006 0.026 0.037 0.002

(0.012) (0.012) (0.012) (0.012) (0.018) (0.028) (0.023)

1 = HH suffered a divorce (t - 1) 0.010 0.009 0.010 0.009 0.015 0.089*** -0.040

(0.022) (0.021) (0.022) (0.018) (0.025) (0.035) (0.034)

1 = village has electricity 0.075*** 0.069*** 0.074*** 0.055*** 0.075*** 0.085*** 0.081*

(0.021) (0.021) (0.022) (0.016) (0.029) (0.031) (0.042)

Year dummy Yes Yes Yes Yes Yes Yes Yes

Observations 7504 7504 7504 7504 7504 7504 7504

Note: Columns (1)-(3) display APEs of the PFP estimator Column (4) presents marginal effects of the FE model. The dependent variable in these estimations is the land share allocated to non-staples crops. Column (5) shows APEs of the CRE Tobit model. The dependent variable in this model is the logarithm of the amount of land allocated to non-staples. Column (6) shows APEs of the Probit model corresponding to the first equation of the D-H model. The dependent variable in this model is the probability of farming non-staple crops. Column (6) shows APEs of the Tobit model corresponding to the second equation of the D-H model. The dependent variable in this model is the logarithm of the amount of land allocated to non-staples crops. All specifications include a full set of control covariates (shown) as well as the average of covariates to control for unobserved household fixed effects (not shown). Bootstrapped standard errors for PFP, CRE Tobit and D-H models (Replications = 500), and clustered standard errors for the FE model are shown in parentheses. *** p <0.01. ** p <0.05. * p <0.1.

this is material, we re-estimate the models in Table 3 replacing the binary drought variable with the underlying continuous SPI metric, where a larger number indicates a more severe dry cycle. These results are reported in Table 5. As before, we find a negative and significant effect of rain shortages on the share of land allocated to non-staple crops. The results also show a similar pattern of reallocation - farmers move to sorghum-millet cultivation from cash, rice and permanent crop farming. Moreover, impacts are larger in

zones affected by more severe drought events. In line with previous results, we also find that uncultivated land increases in the face of a more severe drought.

Background risl<s

Our main results assumed that responses to shocks are homogenous. However, it may be the case that individuals who

Table 3

Average partial effect estimates of the Pooled Fractional Probit (PFP) model for the land share allocated to different crop categories.

Variables (1) (2) (3) (4) (5) (6) (7) (8) (9)

Cash Permanent Horticulture Maize- Sorghum- Groundnut- Sweet- Rice Uncultivated

crop crop Cassava millet beans Potatoes land

Flood (t - 2) -0.006 0.002 -0.003 0.006 0.011** -0.012** 0.002 -0.021** 0.019***

(0.004) (0.005) (0.003) (0.009) (0.005) (0.006) (0.002) (0.004) (0.006)

Flood (t - 1) -0.007 -0.016** -0.008** 0.030** -0.020** 0.002 -0.002 0.003 0.0160*

(0.005) (0.006) (0.003) (0.012) (0.004) (0.009) (0.0024) (0.005) (0.009)

Drought -0.008 0.012 0.002 0.002 0.015** -0.039*** 0.011*** -0.014** 0.0243**

(t - 2) (0.005) (0.008) (0.004) (0.013) (0.006) (0.008) (0.004) (0.006) (0.010)

Drought -0.029** -0.048*** 0.033* 0.039 0.010 0.039* -0.004 -0.058** 0.043

(t-1) (0.011) (0.005) (0.019) (0.025) (0.267) (0.020) (0.005) (0.003) (0.027)

Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes

Year dummy Yes Yes Yes Yes Yes Yes Yes Yes Yes

Observations 7504 7504 7504 7504 7504 7504 7504 7504 7504

Note: Dependent variables are the land share allocated to crops as indicated in the column headers and which vary between zero and one. Column (1) displays APEs for the cash crop category. Column (2) shows APEs for the permanent crop group. Column (3) presents APEs for horticulture farming. Column (4) shows APEs for cassava-maize farming. Column (5) displays APEs for sorghum-millet. Column (6) shows APEs for groundnut-beans. Column 7 presents APEs for sweet potatoes. Column 8 displays APEs for rice farming. Column (9) shows APEs for the uncultivated land category. APEs were calculated after the estimation of the PFP model. All specifications include a full set of control covariates as well as the average of covariates to control for unobserved household fixed effects (not shown). Bootstrapped standard errors are shown in parentheses (Replications = 500). *** p <0.01. ** p <0.05. * p <0.1.

Table 4

Average partial effect estimates of the Pooled Fractional Probit (PFP) model for the land share allocated to different crop categories.

Variables (1) (2) (3) (4) (5)

Cash Maize- Sorghum- Groundnut- Sweet-

crop Cassava millet beans Potatoes

Drought 0.010* -0.023* -0.026*** 0.025*** 0.002

(t-2) (0.006) (0.013) (0.006) (0.009) (0.003)

Drought -0.011 -0.053*** 0.031*** -0.022** 0.015***

(t-1) (0.007) (0.014) (0.008) (0.009) (0.005)

Flood Yes Yes Yes Yes Yes

controls

Control Yes Yes Yes Yes Yes

variables

Year dummy Yes Yes Yes Yes Yes

Observations 7504 7504 7504 7504 7504

Note: The table replicates selected columns of Table 3. The unique difference is that the drought covariate has been modified to reflect rain shortages during the planting/sowing season. All specifications include a full set of control covariates as well as the average of covariates to control for unobserved household fixed effects (not shown). Bootstrapped standard errors are shown in parentheses (Replications = 500). *** p <0.01. ** p <0.05. * p <0.1.

live in higher (background) risk environments react differently to those living in lower risk areas. This is important because the decreasing trend in precipitations observed in the last years in Mozambique suggest a higher incidence of natural disasters. This may have shaped adaptation - i.e., a shock in high risk areas may have a lower impact since farmers are more prepared for it.

In Table 6, we test if responses to recent weather shocks vary according to the magnitude of background risk. To do so, we interact dummies for low and high risk areas with the drought and flood event variables. We define a high flood risk village as one that has been hit by a large flood more than once in the last 20 years. High drought risk villages are those that have experienced more than 7 droughts over the same period. The table focuses on the effect of recent weather shocks on the land share allocated to cassava-maize crops. This is to ease interpretation and minimize chances results are driven by agro-ecological conditions.

We find that farmers living in higher drought-risk villages are most sensitive to floods, but are not more/less sensitive to

droughts. The latter suggests a reinforcement effect rather than adaptation in high risk areas. Since droughts are more frequent in Mozambique than floods (on average), farmers in high drought risk areas may be more aware of the losses from these natural hazards, making them more resistant to adoption of a riskier production portfolio.

Other input choices

A further concern with our model is that we implicitly ignore how production decisions other than crop allocation may adjust to weather shocks. Put differently, interpretation of the estimated APEs for the shock variables requires that all other aspects of production remain fixed. However, it is reasonable to suppose that fertilizer use, livestock activities, off-farm employment and remittances (among others) may respond to shocks and that changes in these factors may indirectly affect crop allocations. If so, then their presence in the model as covariates effectively over-controls for the impact of shocks on crop allocation decisions, ruling out indirect effects. To address this, we first remove all 'suspect' covariates and re-run the baseline model. These results are reported in column 1 of Appendix Table A1. The results remain fundamentally unchanged, implying that the direct effect of shocks on crop allocations is significant and dominant.

As an alternative approach, which also extends our analysis, we consider models for alternative outcomes. For instance, previously we noted that the increase in uncultivated land after a weather shock may be due to displacement of households from their farm (or part of it). It would also be consistent with household members seeking alternative, off-farm income sources. Thus we run the reduced form model presented above using the following outcome variables: the share of family members in off-farm jobs (i.e., who have wage labor outside the farm); the proportion of family members who are self-employed (i.e., undertake activities other than farming); use of fertilizer; and receipt of remittances. We find that the occurrence of flood shocks increases the proportion of off-farm labor as well as the probability of receiving remittances, supporting the notion that these act as coping mechanisms for flood events, but not for drought shocks. Moreover, we find that the probability of using fertilizer increases after a weather shock. As argued above, this result may be in line with a decline in soil

Table 5

Average partial effect estimates of the Pooled Fractional Probit (PFP) using drought intensity.

Variables (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

Non- Cash crop Permanent Horticulture Cassava- Sorghum- Groundnut- Sweet- Rice Uncultivated

staples crop maize millet beans Potatoes land

Drought Int 0.009 -0.013*** -0.001 0.020*** -0.023** 0.037*** -0.039*** 0.001 -0.007 0.025***

(t - 2) (0.008) (0.004) (0.008) (0.003) (0.010) (0.006) (0.007) (0.002) (0.008) (0.006)

Drought Int -0.192*** -0.052* -0.146*** 0.054*** -0.0002 0.140*** -0.016 -0.012 -0.056* 0.106***

(t - 1) (0.036) (0.029) (0.027) (0.017) (0.032) (0.051) (0.025) (0.008) (0.031) (0.022)

Flood controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes

Control Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes

variables

Year dummy Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes

Observations 7504 7504 7504 7504 7504 7504 7504 7504 7504 7504

Note: The table replicates Table 3 replacing the binary drought indicator with a continuous version. All specifications include a full set of control covariates as well as the average of covariates to control for unobserved household fixed effects (not shown). Bootstrapped standard errors are shown in parentheses (Replications = 500). *** p <0.01. ** p <0.05. * p <0.1.

Table 6

Average partial effect estimates by background risk level.

Variables (1) (2)

Flood (t - 1) 0.010 0.019

(0.016) (0.015)

Drought (t - 1) -0.068 0.059

(0.127) (0.108)

Flood (t - 1)*high risk flood area 0.0220 (0.0198)

Flood (t - 1)*high risk drought area 0.0350* (0.0205)

Drought (t - 1)*high risk flood area -0.0224 (0.111)

Drought (t - 1) *high risk drought area 0.108 (0.128)

Flood (t - 2) Yes Yes

Drought (t - 2) Yes Yes

Control variables Yes Yes

Year dummy Yes Yes

Number of observations 7504 7504

Note: Dependent variable is the land share allocated to cassava-maize crops. Column (1) displays APEs for the model interacting recent weather shocks with high drought risk indicators (see text). Column (2) shows APEs for the model interacting recent weather shocks with high flood risk indicators (see text). APEs were calculated after the estimation of the PFP model. All specifications include a full set of control covariates as well as the average of covariates to control for unobserved household fixed effects (not shown). Bootstrapped standard errors are shown in parentheses (Replications = 500). ***p <0.01. **p < 0.05. * p <0.1.

quality after a flood/drought, leading farmers to purchase inputs to recover productivity soil.

Crop rotation

Rotation of crops may be an important driver of land allocation changes.18 Again, this was not captured (controlled for) in our main specification. To address this, a history of crop allocation patterns would be required for each household. However, this is not available in our data. Minimally, however, we do have basic information on whether or not the household employed rotation practices. According to this, which is only available in 2005, about 35% of farmers pursue some form of rotation. To examine whether our results may be confounded with crop rotation, we simply re-estimate our

18 For example, beans are normally planted in rotation with the main cereal and cultivations without fertilizers may benefit from the input remains of the preceding year from intensive productions, mainly cash crops.

full model excluding households that rotate. Overall our main findings are unchanged (results are available on request). Finally, we include crop rotation as an outcome variable and re-estimate the reduced form model discussed in the previous sub-section. These results are shown in the final column of Appendix Table A1. They show that rotation is lower among farmers after a weather shock. Since food insecurity substantially increases after a natural disaster, agricultural practices whose productivity benefits are ambiguous (at least during/after a shock) may be of reduced concern during such periods. In sum, we conclude that crop rotation is unlikely to be a key driver of our results.

Conclusions

Agricultural growth and development typically involves transformation in the form and structure of rural activities. In particular (some) farmers reallocate resources away from food self-sufficiency toward higher value, higher risk agricultural activities. However, farmers may be reluctant to exit food crop cultivation as it helps insure them against food shortages. This suggests that an understanding of cropland decisions and how they interact with weather shocks is an important policy relevant challenge. It is an even more crucial issue in light of the expected higher frequency of natural disasters due to climate change.

In this study we combined panel data and geospatial information for Mozambique to analyze the impact of weather shocks on cropping activity. We took into account the bounded nature of land allocation decisions and used the Pooled Fractional Probit model due to Papke and Wooldridge (2008). We found that crop choice is sensitive to recent weather shocks and farmers are more responsive to more severe droughts. Farmers tend to reallocate land from high risk to low risk cropping activities after a natural hazard. While farmers mainly move out of horticulture and permanent crops after a flood, they reallocate resources away from cash and permanent crops when hit by a drought. We also found that crop reallocation seems to follow a short-term pattern, which is consistent with the maintenance of a buffer stock of food staples within the household.

These findings were found to be robust to alternative definitions of shocks as well as to the exclusion of variables that also may be affected by weather events. This indicated that these shocks primarily have a direct effect on cropland decisions. In addition, we noted that the share of land that is uncultivated tends to rise as a consequence of some weather shocks; and that farmers living in higher drought risk areas appear more responsive to flood shocks.

Fig. A1. Drought identification based on a 3-month SPI (October-December). Note: Red color identifies droughts (SPI lower than -0.5); yellow show normal climate conditions (SPI between -0.5 and 0.5); and green areas identify wet periods (SPI greater than 0.5). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Some caveats with respect to our results merit comment. First, in our examination of the disaggregated production portfolio, we do not account for simultaneity and correlation among crop

categories. Second, our framework implicitly assumes that one type of crop is a substitute for others, which does not take land suitability constraints into account. Nonetheless, switching to

Table A1

Average partial effect estimates for reduced form model, with alternative dependent variables.

Variables (1) (2) (3) (4) (5) (6)

Flood (t - 2) 0.002 0.023** 0.029** -0.018*** 0.015 0.135***

(0.008) (0.011) (0.013) (0.006) (0.019) (0.026)

Flood (t - 1) -0.054*** -0.004 0.026* 0.035*** 0.037* -0.142**

(0.007) (0.011) (0.014) (0.009) (0.019) (0.065)

Drought intensity (t - 2) 0.007 -0.001 0.004 -0.012* 0.019 -0.042*

(0.008) (0.010) (0.013) (0.007) (0.019) (0.022)

Drought intensity (t - 1) -0.191*** 0.005 0.003 0.040* 0.031 -0.668**

(0.036) (0.031) (0.041) (0.0237) (0.064) (0.283)

Ln (landholding) 0.058*** -0.003 0.008 -0.011 0.014 0.050***

(0.008) (0.008) (0.011) (0.009) (0.018) (0.018)

# plots -0.008*** 0.007** 0.018*** 0.014*** -0.002 0.021**

(0.003) (0.003) (0.004) (0.003) (0.007) (0.008)

# family members -0.003* -0.001 -0.007*** 0.004* -0.002 -0.008***

(0.002) (0.002) (0.003) (0.002) (0.004) (0.002)

Head's education level (years) 0.00418 0.007** -0.004 0.003 -0.0005 0.005

(0.00258) (0.003) (0.003) (0.003) (0.007) (0.003)

Wealth index 0.00851* -0.003 0.005 0.010* 0.003 0.015

(0.00484) (0.006) (0.007) (0.005) (0.012) (0.009)

% plots with land title -0.0224 -0.032 -0.006 -0.006 -0.062 0.108*

(0.0199) (0.022) (0.029) (0.034) (0.052) (0.056)

1 = HH received extension services 0.008 0.018* 0.029** 0.017* 0.016 0.075***

(0.008) (0.009) (0.011) (0.009) (0.017) (0.021)

1 = HH belonged to farm organizations 0.0172 -0.009 0.025 0.032** 0.008 0.057**

(0.012) (0.011) (0.017) (0.016) (0.028) (0.029)

1 = HH received price information -0.0004 0.029*** 0.036*** -0.008 0.018 0.057***

(0.006) (0.007) (0.008) (0.006) (0.013) (0.017)

Average regional retail maize price (t - 1) 0.000 0.000 0.000 0.000 0.000 0.0001***

(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)

% sick family members 0.005 0.013 0.031 -0.013 0.138* 0.030

(0.042) (0.051) (0.058) (0.028) (0.077) (0.077)

1 = HH suffered a death (t - 1) 0.006 -0.005 0.032** 0.001 -0.005 -0.009

(0.012) (0.012) (0.016) (0.009) (0.027) (0.030)

1 = HH suffered a divorce (t - 1) 0.012 0.021 0.035 0.003 -0.009 0.044

(0.021) (0.020) (0.033) (0.022) (0.045) (0.046)

1 = village has electricity 0.061*** -0.029** -0.027 -0.002 0.005 0.024

(0.020) (0.012) (0.018) (0.015) (0.038) (0.027)

Year dummy Yes Yes Yes Yes Yes No

R square - - - 0.026 0.010 0.070

Observations 7504 7504 7504 7504 7504 3752

Note: Column 1 shows the APEs for the land share allocated to non-staple crops without potential endogenous variables (irrigation, fertilizer) and controls that may potentially respond to shocks (off-farm activities, remittances and animal traction). Column 2 shows the APEs for the proportion of family members with off-farm jobs. Column 3 reports the APEs for the proportion of family members self-employed. These models are estimated by the PFP and include a full set of control covariates and the average of covariates to control for unobserved household fixed effects (not shown). Column 3 shows the marginal effects for fertilizer use (1 if farmer uses fertilizer). Column 4 reports the marginal effects for the receipt of remittance (1 if farmer receives remittances).These models are estimated by the FE and include a full set of control covariates. Column 5 present the marginal effects for crop rotation (1 if farmer practices rotation).This model is estimated by the OLS for 2005 and include a full set of control covariates, regional and agro-ecological dummies. Bootstrapped standard errors are shown in parentheses (Replications = 500). *** p <0.01. ** p <0.05. * p <0.1.

staples is less likely to be constrained by agro-ecological conditions - e.g., the distribution of maize-cassava production is less dependent on geographical factors. Third, given the limited temporal dimension of our panel, we are not able to fully explore longer-run dynamics in cropland decisions. These may be important, especially in explaining changes in permanent crops, use of annual rotation, and in exploring the role of prices on crop choices. Moreover, while we have uncovered clear evidence of short-term farmer responses to weather risks, future development of the sector will depend in fundamental ways on structural changes in the wider economy, including the articulation between industry and agriculture.

Despite these considerations, an important implication of our empirical findings is that climate change, which is expected to increase the frequency of extreme weather events, is likely to have a material effect on crop choices in developing countries such as Mozambique. More specifically, it may slow the adoption of new commercial crops (or technologies) by smallholder farmers, especially where these expose households to food security risks.

Additionally, climate change may accelerate movement out of agriculture into off-farm activities, potentially spurring an increase in rural-urban migration.

Acknowledgments

We wish to acknowledge the data collection efforts of the Ministry of Agriculture and Rural Development of Mozambique without which this study would not be possible. The same goes for collaboration with staff at the Ministry of Planning and Development, now integrated in the new Ministry of Economics and Finance. We also wish to express our gratitude for comments received on various occasions at Department of Economics in Copenhagen, at the University of Nottingham and at The European Summer School: The Economics of Adaptation to Climate Change. Special thanks go to John Rand, Oliver Morrissey, Robert Mendelsohn, Brent Sohngen and Marcela Jaime for their valuable comments on earlier versions. The usual caveats apply.

Appendix A

See Fig. A1 and Table A1.

References

Adger, W.M., Huq, S., Brown, K., Conway, D., Hulme, M., 2003. Adaptation to climate change in the developing world. Prog. Dev. Stud. 3 (3), 179-195.

Arndt, C., Tarp, F., 2000. Agricultural technology, risk, and gender: a CGE analysis of Mozambique. World Dev. 28 (7), 1307-1326.

Arndt, C., Thurlow, J., 2013. Climate uncertainty and economic development: evaluating the case of Mozambique to 2050. WIDER Working Paper, No. 2013/ 042. UNU-WIDER, Helsinki.

Arndt, C., Hussain, M.A., Jones, E.S., Nhate, V., Tarp, F., Thurlow, J., 2012. Explaining the evolution of poverty. The case of Mozambique. Am. J. Agric. Econ. 94 (4), 854-872.

Benin, S., Smale, S., Pender, J., Gebremedhin, B., Ehui, S., 2004. The economic determinants of cereal crop diversity on farms in the Ethiopian highlands. Agric. Econ. 31 (2-3), 197-208.

Bezabih, M., Di Falco, S., 2012. Rainfall variability and food crop portfolio choice. Food Sec. 4 (4), 557-567.

Bezabih, M., Sarr, M., 2012. Risk preferences and environmental uncertainty: implications for crop diversification decisions in Ethiopia. Environ. Resource Econ. 53 (4), 483-505.

Brakenridge, G.R., 2013. <http://floodobservatory.colorado.edu/Archives/index. html>.

Cameron, L., Shah, M., 2013. Risk Taking Behaviour in the Wake of Natural Disasters. NBER Working Paper Series, No. 19534. Cambridge, MA.

Carter, M.C., Lybbert, T.J., 2012. Consumption versus asset smoothing: testing the implications of poverty trap theory in Burkina Faso. J. Dev. Econ. 99 (2), 255264.

Chamberlin, G., 1980. Analysis of variance with qualitative data. Rev. Econ. Stud. 47 (1), 225-238.

Chavas, J., Holt, M., 1996. Economic behaviour under uncertainty: a joint analysis of risk preferences and technology. Rev. Econ. Stat. 78 (2), 329-335.

Chibwana, C., Fisher, M., 2012. Cropland allocation effects of agricultural input subsidies in Malawi. World Dev. 40 (1), 124-133.

Cohen, M., Etner, J., Jeleva, M., 2008. Dynamic decision making when risk perception depends on past experience. Theory Decis. 64 (2-3), 173-192.

Cragg, J.G., 1971. Some statistical models for limited dependent variables with application to the demand for durable goods. Econometrica 39 (5), 829-844.

Damon, A.L., 2010. Agricultural land use and asset accumulation in migrant households: the case of El Salvador. J. Dev. Stud. 46 (1), 162-189.

Dercon, S., 1996. Risk, crop choice and saving: evidence from Tanzania. Econ. Dev. Cult. Change 44 (3), 485-513.

Dercon, S., 2002. Income risk, coping strategies and safety nets. World Bank Res. Observer 17 (2), 141-166.

Di Falco, S., Chavas, J., 2009. On crop biodiversity, risk exposure, and food security in the highlands of Ethiopia. Am. J. Agric. Econ. 91 (3), 599-611.

Di Falco, S., Veronesi, N., 2013. How can African agriculture adapt to climate change? A counterfactual analysis from Ethiopia. Land Econ. 89 (4), 743-766.

Di Falco, S., Bezabih, M., Yesuf, M., 2010. Seeds for livelihoods: crop biodiversity and food production in Ethiopia. Ecol. Econ. 69 (8), 1695-1702.

Fafchamps, M., 1992a. Cash crop production, food prices volatility, rural market integration in the third world. Am. J. Agric. Econ. 74 (1), 90-99.

Fafchamps, M., 1992b. Solidarity networks in pre-industrial societies: rational peasants with a moral economy. Econ. Dev. Cult. Change 41 (1), 147-174.

Fafchamps, M., Udry, C., Czucas, K., 1998. Drought and saving in West Africa: are livestock a buffer stock? J. Dev. Econ. 55 (2), 273-305.

FAO, 2005. Drought-Resistant Soils. Optimization of Soil Moisture for Sustainable Plant Production. FAO Land and Water Bulletin 11, Rome.

Gollier, C., Pratt, J., 1996. Risk vulnerability and the tempering effect of background risk. Econometrica 64 (5), 1109-1123.

Jones, S., Tarp, F., 2013. Jobs and welfare in Mozambique. WIDER Working Paper, No. 2013/045. UNU-WIDER, Helsinki.

Kahneman, D., Tversky, A., 1979. Prospect theory: an analysis of decision under risk. Econometrica 47 (2), 263-292.

Kazianga, H., Udry, C., 2006. Consumption smoothing? Livestock, insurance and drought in rural Burkina Faso. J. Dev. Econ. 79 (2), 413-446.

Kurosaki, T., Fafchamps, M., 2002. Insurance market efficiency and crop choices in Pakistan. J. Dev. Econ. 67 (2), 419-453.

Kurukulasuriya, P., Medelsohn, R., 2008. Crop switching as a strategy for adapting to climate change. Afr. J. Agric. Econ. 2 (1), 105-126.

Kurukulasuriya, P.M., Mendelsohn, R., Hassan, R., Benhin, J., Deressa, T., Diop, M., Eid, H.M., Fosu, K.Y., Gbetibouo, G., Jain, S., Mahamadou, A., Mano, R., Kabubo-Mariara, J., El-Marsafawy, S., Molua, E., Ouda, S., Ouedraogo, M., Sene, I., Maddison, D., Seo, S.N., Dinar, A., 2006. Will African agriculture survive climate change? World Bank Econ. Rev. 20 (3), 367-388.

Masanjala, W., 2006. Cash crop liberalization and poverty alleviation in Africa: evidence from Malawi. Agric. Econ. 35 (2), 231-240.

Mather, D., Cunguara, B., Boughton, D., 2008. Household income and assets in rural Mozambique, 2002-2005: Can pro-poor growth be sustained? Research Paper Series, Ministry of Agriculture, Republic of Mozambique.

McKee, T., Doesken, N., Kleist, J., 1993. The relationship of drought frequency and duration to time scale. In: Proceedings of the Eighth Conference on Applied Climatology. American Meteorological Society, Anaheim: CA/Boston:MA, pp. 179-184.

McKee, T., Doesken, N., Kleist, J., 1995. Drought monitoring with multiple timescales. In: Proceedings of the Ninth Conference on Applied Climatology. American Meteorological Society, Dallas: TX/ Boston: MA, pp. 233-236.

Ministério da Agricultura (Direcgao de Economia: Departamento de Estatística), 2005. Trabalho de Inquérito Agrícola 2005. Maputo, Republic of Mozambique.

Ministério da Agricultura e Desenvolvimento Rural (Directo de Economia: Departamento de Estatística), 2002. Trabalho de Inquérito Agrícola 2002. Maputo, Republic of Mozambique.

Papke, L.E., Wooldridge, J.M., 1996. Econometric methods for fractional response variables with an application to 401(k) plan participation rates. J. Appl. Econ. 11 (6), 619-631.

Papke, L.E., Wooldridge, J.M., 2008. Panel data methods for fractional response variables with an application to test pass rates. J. Econ. 145 (1-2), 121-133.

Paxson, C., 1992. Using weather variability to estimate the response of savings to transitory income in Thailand. Am. Econ. Rev. 82 (1), 15-33.

Seo, S., Mendelsohn, R., 2008. An analysis of crop choice: adapting to climate change in south American farms. Ecol. Econ. 67 (1), 109-116.

Tarp, F., Arndt, C., Tarp Jensen, H., Robinson, S., Helberg, R., 2002. Facing the Development Challenge in Mozambique. An Economywide Perspective. Research report 26, International Food Policy Research Institute, Washington, D.C.

Townsed, R., 1994. Risk and insurance in village India. Econometrica 62 (3), 539591.

UNISDR (United Nations International Strategy for Disaster Reduction), 2009. Global Assessment Report on Disaster Risk Reduction. UNISDR: Geneva.

Wooldridge, J., 2010, . Econometric Analysis of Cross Section and Panel Data, second ed., vol. 1. MIT Press Books.

World Bank, 2008. Beating the Odds: Sustaining Inclusion in a Growing Economy: a MOZAMBIQUE POVERTY, Gender and Social Assessment. Washington D.C.

World Bank, 2010. Economics of Adaptation to Climate Change - Mozambique. Washington D.C.

World Bank, 2012. Country Partnership Strategy FY12-15 for the Republic of Mozambique. Washington D.C.