How costly is using livestock as a savings device?

Livestock is a major savings device in sub-Sahara Africa agriculture. I measure to what extent the value of livestock drops during food shortages. For this purpose I exploit Malawian prices of meat and maize for 72 markets from 1991 to 2009, a period with several food shortages. I show that large drops in the meat–maize terms of trade – our proxy for the value of livestock – are associated with food shortages. During food shortages the value of livestock decreases with 54% to 65%. The evidence is consistent with increased livestock sales during food shortages, but the drop in meat–maize terms of trade arises primarily due to increases of maize prices. Our results are robust to spatial spill-overs and various other threats. Similar drops in livestock value are shown to occur in other SSA countries. The value of livestock has decreased at the very moment livestock is sold on the market to purchase staple foods. Like produced staple foods, agricultural households systematically sell low. To bridge food shortage periods savings instruments are needed that do not lose value when liquidated. A few policy options are discussed. On-farm grain storage appears most promising.


Introduction
Agriculture in sub-Sahara Africa is mainly rainfed and the dominant risk in agriculture is a lack of rainfall or drought. Dependence on rainfall leads to occasional crop failures. Many farm households protect themselves against crop failures by saving through livestock. For several reasons the value of livestock drops during food shortages. The objective of the current paper is to measure the size of the drop of the value of livestock during food shortages, associated with increased sales of livestock and increased market prices of staple foods. On the basis of these measurements I discuss the welfare implications of using livestock as a savings device and lessons for potential alternative savings instruments to overcome periods of food shortage.
The literature on saving and risk is huge. Since our investigations are primarily empirical, I discuss a selected number of articles that focus particularly on the role of livestock for poor rural households in developing countries, rather than aiming at an exhaustive review of the literature on saving strategies. The key objectives of work in this area is to identify livestock as a major device for precautionary saving (Hänke & Barkmann, 2017;Kinsey et al., 1998;Mogues, 2011;Turner & Williams, 2002), to show that livestock sales are a primarily used to fund food purchases during food shortages (Hänke & Barkmann, 2017;Kinsey et al., 1998;Mogues, 2011;Turner & Williams, 2002), to assess the extent to which livestock savings are effective in establishing food security at the household level (Fafchamps et al., 1998;Hänke & Barkmann, 2017;Kazianga & Udry, 2006;Kinsey et al., 1998) and to what extent livestock sales are useful in smoothing consumption (Fafchamps et al., 1998;Kazianga & Udry, 2006).
Livestock sales occur often, and often exclusively, during food shortages (Hänke & Barkmann, 2017;Turner & Williams, 2002). More than two third of livestock sales in Niger are made to purchase food, under conditions with a high degree of urgency (Turner & Williams, 2002). Around 50% of sales of zebu and close to 80% of the sales of goats in semi-arid south-western Madagascar are driven by food shortages, while during crop failures on average 56% of food expenditures are funded with livestock sales (Hänke & Barkmann, 2017). Some authors find substantial and large contributions of livestock sales (Kinsey et al., 1998;Mogues, 2011). Livestock sales are the major resource that households use to fund food purchases during droughts in Zimbabwe (between 40 and 50%), a strategy followed by almost two-third of all households (Kinsey et al., 1998). Weather shocks in Ethiopia lead to asset drawdown by households and this is more pronounced for covariant than for idiosyncratic shocks. Precautionary motives of wealth holding are claimed to be more prevalent for liquid assets and for less productive forms of wealth, hence, larger for grain stocks than for livestock, and larger for small livestock than for large livestock (Kazianga & Udry, 2006;Mogues, 2011). Some authors are less confident about the effectiveness of livestock savings to protect against food shortages. Fafchamps et al. (1998) find that livestock transactions in Burkina Faso compensate for at most 30%, and probably close to 15% of income shortfalls due to village level shocks alone. Kazianga and Udry (2006) show that fluctuations in household consumption closely track fluctuations in household income associated with drought and subsequent recovery, and find no evidence that livestock sales or financial markets serve as an effective coping strategy against these income fluctuations.
In a useful survey on income risk and coping strategies Dercon (2002) highlights a major drawback of (assets like) livestock when used as a savings device: "…. A(nother) problem with holding assets to buffer consumption is that the terms of trade between goods for consumption and assets change as a result of a common shock. If a negative common shock occurs, households would like to sell some of their assets. However, if everyone wants to sell assets at the same time, asset prices will collapse and the amount of consumption that can be purchased will fall". The empirical work reported in the current paper aims to quantify the size of the fall in asset prices during food shortages in Malawi.
Most work on the impact of livestock sales during food shortages takes the household as the unit of research. This approach has a lot to recommend itself as the key objective is to assess the degree of food security, the degree to which households are capable of buffering fluctuations in income with savings like livestock and to preserve required levels of nutrition. I follow a different route in the current work. Instead of taking the household perspective, I look at markets. 1 I evaluate to what extent food shortages through their impact on market prices affect the value of livestock. Looking at markets rather than at households allows to assess the interaction between meat prices, staple food prices, and sales of livestock during food shortages at the level of the local market, and how these circumstances affect the value of livestock. Households make their decisions in the context of the market, taking prices as given. Food shortages affect large numbers of households simultaneously and so many households will make similar decisions, and market prices adjust accordingly. Availability of a large number of systematically recorded administrative monthly price data, covering an extensive number of geographical dispersed markets and a long period with several food shortages, makes this evidence an informative and indispensable complement to household survey based research.
In the remainder of this paper I supply background information on Malawi in Sect. 2. In Sect. 3 I propose a simple conceptual framework and I elaborate the empirical strategy. In Sect. 4 I show and discuss estimation results and robustness checks. In Sect. 5 I assess estimations and observed terms of trade. In Sect. 6 I discuss implications and policy alternatives, and in Sect. 7 I give a summary and conclusion.

Livestock, staple food and food shortages in Malawi
Malawi is a landlocked country in the south of Africa, between -9.4° and -17.1° latitude and 32.7° and 35.9° longitude, around 800 km from north to south, and around 150 km from east to west, bordering Zambia, Tanzania and Mozambique (see Fig. 1 for a map of Malawi). A large lake, Lake Malawi, part of the Great Rift Valley, stretches from north to south, along the eastern border of the country. The Malawian population, which increased during the study period (1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)) from close to 9 million to 13-14 million, is mostly rural: only a small fraction (11% to 15%) lives in the cities Lilongwe, Blantyre, Mzuzu and Zomba. Per capita GDP, expressed in purchasing power parity US$, in 2009 is between 840 and 900 US$, making Malawi one of the poorest countries in the world, with a ranking in the bottom 14 of all countries. 2 More than 80% of the Malawi population depends for food and income on subsistence farming. The incidence of poverty is high: more than 50% of the population in Malawi is poor (various Integrated Household Surveys) and poverty is extremely high in remote rural districts (e.g. Chitipa in the north: 67.2%; Nsanje and Chikwawa in the south: respectively 76.0% and 65.8%, Integrated Household Survey 2004/2005. In the southern region poverty is at least 10%-points higher relative to the central regions. The key food crop in Malawi is maize. Cassava and rice have very modest market shares that are nevertheless increasing, and are important in a few districts (cassava in Nkhatabay and Nkhotakota, and rice in Karonga and Machinga). Other popular food crops are groundnuts and beans. Tobacco is by far the most important cash crop. Just like the other major cash crops, sugar and tea, tobacco cultivation dates back to the colonial period. Tobacco, however, has become nearly completely smallholder based in the course of the 1990s (Zant, 2020), while tea and sugar are still mainly produced on estates. Nearly every city, town or larger village has one or more markets for agricultural food crops on a regular basis, often daily or weekly: the market price data that I use in the empirical estimations are from 72 markets spread across the entire country (see Fig. 1 for the location of markets).
Maize in Malawi dominates both production and consumption of households. Maize is the major staple food in Malawi, accounting for 52% to 65% of the total per capita calorie intake. Due to high population density the largest market for maize is the southern part of the country. Also, nearly all households grow maize. Production of maize in Malawi is undertaken by households primarily for home consumption. As a consequence the quantity of marketed maize is limited: estimates of the marketed share of production range from 5 to 25% of domestic production (Jayne et al., 2008). The main maize crop in Malawi is planted from September to November and harvested from March to May. Agriculture in Malawi is rain-fed and rainfall risk is by far the dominant production risk in agriculture (Giné & Yang, 2009;Ahmed et al., 2020). Variation in rainfall and occasional droughts cause large fluctuations in production of maize. Apart from a distinct geographical variation, especially 1991/92 shows up as a year with an extreme drought, almost throughout Malawi (Table 4). Major crop failures with spatially varying intensity occurred in the crop seasons 1991/92, 1996/97, 2000/2002 and 2004/2005. 3 Periods with food shortages are identified by extremely high staple food prices, in particular maize prices (Fig. 9). 4 Just like prices of all agricultural products and common for sub-Sahara agriculture (Gilbert et al., 2017;Kaminski et al., 2016), Malawi maize prices follow a distinct seasonal pattern, with highs at the end of the marketing season from January to March, just before harvesting, and lows after harvesting from May to July. Within season price differences are large: median within season returns for maize are well above 100%. The first months of the calendar year, January to March, are the months where food shortages usually become apparent: during these months the highest maize prices are realized, and I also expect most extra livestock supply on the market, and thereby the lowest meat prices. This correspondence -large increases of maize prices jointly with modest, well below average increases or even decreases in meat prices -is clearly visible in the graphical evidence and seems to support increased supply of livestock during food shortages (Fig. 2). There is no systematic information on livestock dynamics and on livestock transactions at the household level. From this perspective, meat prices by month, for a substantial number of markets and for nearly two decades, are therefore the most systematic, frequent and granular source of information. There is, nevertheless, a census among farmers including a module on livestock and various other details (National Census of Agriculture and Livestock (NACAL), Malawi, 2006Malawi, /2007 1992 1993 1994 1995 1996 1997 1998 1999  6 Averages in the household survey data are conditioned on non-zero livestock in [2004][2005]. The share of households with zero livestock at ADD level varies from 1 to 19%, and averages 10%. 7 The number of chicken is largest and most wide-spread over households. However, their contribution to total livestock per household in tropical livestock units is limited. Also, unlike the other major types of livestock, I lack Malawian market prices data for chicken. Chicken are particularly popular because less lumpy, less valuable and therefore more practical: they can easily be sold on the market or consumed at home. Chicken are also attractive because of the short reproduction time. Asset lumpiness is known to limited access to livestock savings (Dercon, 2002). Note, however, that household livestock herds are mainly obtained by reproduction rather than by purchase. the data from different sources are reasonably close to each other, 8 which gives confidence about their reliability.
On the basis of the figure (Fig. 3), I argue that livestock rearing per household especially takes place in the northern and southern districts. The share of cattle per household in total livestock is the most important type of livestock, however with distinct regional variation: in the outer north and outer south (ADDs Karonga, Mzuzu and Shire Valley) cattle covers 65% to 84% of total livestock (see Fig. 7). In the central and urban part of the country (ADDs Lilongwe, Blantyre, Salima, Machinga) the smaller types of livestock (goats, sheep and pigs) are more prevalent, covering 49% to 79% of total livestock (see Appendix). Unfortunately, the core NACAL data on livestock and meat production, available annually from 2000-2001 onwards, did not allow meaningful observations about livestock and meat dynamics. 9 The size of livestock per household in the central ADDs (Lilongwe, Blantyre, Salima, Machinga) is low (compare, for example, Kinsey et al., 1998): it is difficult to attribute a serious role for livestock in terms of precautionary saving in these areas. Households in the outer north and outer south (Karonga, Mzuzu and Shire Valley) have larger herds of livestock, and may resort (more) to livestock savings because they have fewer alternative options to protect against crop failures. Conversely, households in the central ADDs, where the Malawian cities are located, have easier access to temporary urban wage jobs and other non-farm employment opportunities, a common technique to deal with income shortfalls arising from crop failure or drought. Also, easier access to food-aid in case emergency could play an important role. Dercon (1998Dercon ( , 2002 mentions off-farm activities as a rational choice for low-income households in case of credit constraints and agricultural risk. Finally, the correspondence between the regional pattern of poverty incidence and the per household number of livestock units is striking.  1 3 3 Conceptual framework, empirical strategy and data

Conceptual framework
I consider livestock the major input in the supply of meat, where livestock has a variety of uses (dairy production, draft animal, cattle raising for export) of which livestock raising for meat is one. 11 Fluctuations in livestock prices cause a shift of the meat supply curve: if meat demand is held constant, a decrease of the livestock price will shift the meat supply curve to the right and lead to a proportional fall in meat prices. For the meat markets I assume that standard demand and supply analysis applies. Hence, meat prices are adequately described by geographically specified local conditions and by time related developments that are similar to all markets. The exception that I elaborate on in the current paper is the exogenous shock that arises out of droughts and subsequent crop failures and food shortages. Droughts, crop failures and food shortages lead to a reduction in livestock value through several mechanisms. The first mechanism is that staple food prices increase extremely during food shortages, and thereby reduce the real value of livestock. As livestock is used to protect households in case of crop failures and maintain food security, I hypothesize that food shortages lead to increased sales of livestock: as many households will make similar decisions, the additional supply of livestock on the market will reduce livestock prices. Moreover, droughts may also force farmers to sell livestock because access to grazing becomes more limited. On both grounds the supply of livestock on the market increases and livestock prices decrease. With livestock the key input in the production of meat, these development exert a downward pressure on meat prices. Finally, drought and food shortages are likely to impact on the physical condition and body weight of livestock: a drop in the body weight of livestock will also affect the value of livestock negatively, 12 or under extreme conditions cause livestock mortality. Detailed data to measure the drop in value due to reduced bodyweight of livestock are lacking. However, in the approach set out below, I disentangle the reduction in livestock value due to a decrease in meat prices and the reduction in livestock value due to increased staple food prices, and elaborate on the likely additional impact on livestock value of drops in body weight.

Empirical strategy
In line with the conceptual framework, I postulate that meat prices are empirically determined by local supply and demand conditions, and by food shortages. The key objective of the empirical estimations is to find evidence that food shortages are negatively correlated with meat prices. I assume that the influence of local supply and demand conditions on spatial meat prices will be fully captured by market and time fixed effects. I start with the following specification: where p meat jt is the price of meat in market j at time t , and food shortage jt is a variable that indicates if market j at time t experiences a food shortage. Parameters j represent market fixed effects and t time fixed effects, and jt is an error term with zero mean. The equation is a standard Two Way Fixed Effect specification (TWFE). The important feature of this specification is that it estimates the extent that a food shortage in a specific market raises prices in that market by more than it does in other markets in the same year. Note that the year fixed effects also absorb occasional net imports and ADMARC stock releases at the national level. If households sell livestock during food shortages, I expect that a food shortage has a negative impact of meat prices and hence 1 <0.
Prices of consumer goods and budget compositions of households fluctuate between years and across markets and, most importantly, within the season. For these reasons using monthly market prices over a long period and for a large number of markets in empirical estimations, makes it necessary to adopt a technique to make meat prices comparable over time and across markets. However, to find adequate consumer price indices for converting local prices into real prices in developing country agricultural settings is notoriously difficult. The available national consumer price index simply fails to take differences between prices across markets into account and is biased towards urban households. I propose the following solution. As subsistence households will normally value their livestock in terms of the quantity of staple food that can be purchased, the natural way to make meat prices comparable over time and across markets is by using staples food prices as reference prices, in the Malawi case maize prices. Since I have maize prices for a large number of markets (the same number of markets as in case of meat prices), using the terms of trade -the price of meat relative to the price of maize -is an elegant way to avoid the need for consumer price indices at the district or market level, price indices that are typically not available. Using the meat-maize terms of trade also generates a more complete and more accurate picture. The specification changes into: (1) p meat jt = 0 + 1 food shortage jt + j + t + jt , 11 Note that livestock markets are not well developed and tend to be thin leading to extreme and unreliable market prices and missing observations in the data, while meat markets are more regular with sufficient volume on both sides of the market and market price data which are reasonably complete. 12 Barrett et al. (2003) for an elaboration of this mechanism. I further propose to approximate food shortages with a physical measure of food shortages: I exploit the dominance of maize in the diet of Malawi population, and construct the requirement of maize by district, as a linear transformation of population and average per capita requirements. 13 I assume that actual previous season maize production relative to current season maize requirements is an adequate approximation of food shortages: values lower than 1 characterize a food shortage. I propose the following adjustment to Eq. (2): Due to its definition, I now expect 1 > 0 : a shortage of maize -a deficit of local maize production vis-à-vis local maize requirements -will have a negative impact on the meat-maize terms of trade, and likewise, a relative abundance of maize -a surplus of locally produced maize vis-àvis local maize requirements -will have a negative impact on the meat-maize terms of trade.
Next, I express our dependent variable -the meat-maize terms of trade -as a between-year seasonal change: I replace the terms of trade variable by the seasonal gap, the (log) ratio of the terms of trade in the three pre-harvest months to the same ratio in the three post-harvest months in the previous year. 14 Equation (3) changes into: This transformation allows to disentangle effects on meat prices and maize prices: as I can write the log of a ratio as the difference between the logs ( ln(p meat ∕p staplefood = ln(p meat ) − ln(p staple food) ), the decomposition into a 'meat,' and a 'maize' gap equation is straightforward: (2) p meat jt ∕p maize jt = 0 + 1 food shortage jt + j + t + jt . (3) Obviously, the coefficients of the food shortage proxy will be different for the meat equation and the maize equation.
To illustrate the seasonal gap in meat prices vis-à-vis the seasonal gap in maize prices, Fig. 4a and b show these gaps for the different identified meats and maize in two specific markets. The figures confirm a consistently opposite movement of meat price gaps and maize price gaps during food shortages. The size of the maize price increases is, however, of a much a larger scale than the meat price decreases.
The figure also highlights that the meat price gaps move fairly independently from each other outside years of food shortage. In line with this observation, I hypothesize that the abundance-scarcity variable is likely to impact meat prices in an asymmetric way: no impact under normal circumstances and a significant impact in case of food shortages. It is straightforward to split the explanatory variable into abundance-scarcity with food shortage, and abundance-scarcity without food shortages ( maize production jt−1 ≤ or > maize requirement jt ).
These changes are incorporated in Eqs. (7), (8), and (9): where abundance-scarcity absc jt = [maize production jt−1 ∕ maize requirement jt ] . The indicator shortage conditions the abundance-scarcity variable to be equal to measured abundance-scarcity if maize production jt−1 ≤ maize requirement jt , and zero elsewhere, while the indicator no shortage does the reverse. I expect 1 , 1 and 1 to be significant and 2 , 2 and 2 to be insignificant. Equations (7), (8), and (9) are the basic specifications for the empirical estimations.
With price data for many markets in a geographically limited space, spatial spillovers are very likely. Spatial spillovers potentially affect the estimated relationship and need to be controlled for. I therefore employ fixed effects spatial regression. Estimations allow for spillovers from other gap p m jt ∕p mz 13 See Sect. 3.3 for the construction of the maize requirement variable. 14 This transformation will relate the analysis more directly to the seasonality literature (Gilbert et al., 2017;Kaminski et al., 2016). . Note: The figure shows the maize and meat price gap for specific markets. Price gaps are calculated as the average price in January, February and March in a specific year (the lean season prices), relative to the average price of the lowest three prices from May to November of the previous year, for a specific market. Source: Calculations based on price data from the Agro-Economic Survey, Ministry of Agriculture and Food Security, Government of Malawi markets, through the dependent variable, through the explanatory variable and in the residual. To specify the influence of spatial spillovers I make use of an inverse distance matrix. The inverse distance matrix has zeros on the diagonal, and, by construction, attributes large weights to nearby markets and small weights to far-away markets. Spatial spillovers are incorporated in the estimation by including as regressors the product of the inverse distance matrix and the dependent variable, the independent variable and the residual. I estimate this adjusted equation with fixed-effects spatial regression for panel data, which is available in STATA. Because of high transaction costs and the characteristics of domestic trade (Fafchamps et al., 2005) I have truncated the spatial effects to 50 km distance between markets. 15 Several other concerns are addressed in additional estimations (regional variation, serial and spatial correlation in standard errors, potential endogeneity). I consider regional variation by re-estimating the equations separately for the northern, central and southern region. I expect our data to be serially and spatially correlated and consequently estimated standard errors will be inconsistent. Clustering standard errors with the panel identifier, in our case market, results in standard errors that are robust to serial and spatial correlation (Driscoll & Kraay, 1998). Finally, there are potential concerns about endogeneity of the abundance-scarcity variable: I have addressed this by estimating with IV-2SLS and instrumenting the abundance-scarcity variable with rainfall by district.

Data, data sources, data availability and variable construction
The core data for the empirical estimations are monthly market prices for goat meat, steak & bone, pork and maize, for a total of 72 markets, all sourced from the Agro-Economic Survey, Ministry of Agriculture and Food Security, Government of Malawi (see Fig. 1 for the location of markets). I have chosen the price of goat meat, steak & bone and pork since these prices correspond with the major livestock categories reared by Malawi households (See Figs. 7 and 8). For some predominantly Islamic districts (Machinga) there is no (or only a limited) market for pork. I have chosen the price of maize as staple food price, as maize is the major staple food grown and consumed by most Malawi households. From the background section on of Malawi agriculture it is clear that maize dominates both in production and consumption, in all districts of Malawi (and more so relative to many other countries). Cassava is a popular crop in the northern districts that border lake Malawi (Nkhatabay, Nkhotakota, and Karonga), but becoming increasingly popular throughout Malawi. Rice is especially grown in Karonga in the north, and Machinga and Zomba in the south, and primarily consumed in the cities. The availability of price data increases drastically after January 2004: from this date onwards observations increase, for all series, from less than 30% to nearly 80% of all markets (See Fig. 9). The background of this increase is a change in the coverage of markets by the administrative organisation that collects the market price data (Agro-Economic Survey, Ministry of Agriculture and Food Security, Government of Malawi). Additionally, the data suffer from missing observations reflecting lack of supply in the market and associated with seasonality. This is common to agricultural price series. Since I am particularly interested in prices during lean season, it may possibly affect the estimations. However, missing lean season observations fortunately do not jeopardise the estimations, partly due to the aggregation of monthly data into annual observations.
For the construction of a variable reflecting maize abundance-scarcity by district I have used census based population data by district (Rural Development Project) from the National Statistical Office -dated 1987, 1998 and 2008 -that are interpolated for intermediate months, and annual maize production data, also sourced from Agro-Economic Survey, Ministry of Agriculture and Food Security, Government of Malawi. Rainfall data from 31 weather stations are from the Department of Climate Change and Meteorological Service in Zomba. Monthly rainfall observations, aggregated by season -where the season runs from April to March -are first attributed to Extension Planning Areas (EPAs, 102 in total) and subsequently averaged by district (Rural Development Project). For descriptive purposes I show the development of real prices of meat and maize (See Fig. 9). I use the Malawi consumer price index, sourced from the IMF International Financial Statistics database, to deflate the price series. Malawi has a total of 26 districts (Rural Development Project (RDPs)), eight Agricultural Development Divisions (ADDs) and three regions (north, central and south): Table 5 shows how the 72 markets for which price data are available, are related to RDPs, ADDs and regions. The latitude-longitude coordinates of markets needed for the spatial estimations are taken from Google maps.
The abundance-scarcity variable is constructed by exploiting the dominance of maize in the diet of Malawi population: it is defined as actual previous year maize production versus the households requirements of maize for consumption, both by district. Values lower than 1 of the abundance-scarcity variable characterize a food shortage. Households requirements of maize for consumption are a simple transformation of district population. Maize required for basic nutrition per person is equal to maize kcal share in the diet times the total kcal needs per person, divided by kcal content per kg of maize. Unfortunately maize diet shares and kcal needs are based on case studies, and are not region (and time) specific. Hence, I calculate maize required for basic nutrition, in district j and date t as follows: For the average maize kcal share I use 0.5 and 0.65 and for average kcal needs per person (per day) I use 2100 kcal and 2300 kcal. 16 The per kg kcal content of maize is 3570, which completes the construction of maize requirements (Zant, 2012). Applied to the data the abundance-scarcity variable confirms documented food shortages and also shows substantial geographical variation (See Fig. 10). Documented food shortages are confirmed by the development of the abundance-scarcity variable over the years. The maize abundance-scarcity variable also supports pronounced variation across districts. The plots of the maize abundancescarcity variable further indicate a larger intensity of food shortages during 1992-93 and 1994-95, relative to 2005-06. This contrasts with observed food price increases which are req jt = (pop jt * share in diet * needs per person) ∕maize kcal content per kg larger during the 2005-2006 food shortage. The use of the abundance-scarcity variable implicitly assumes that the geographical variation in crop outcomes and population density, jointly with high transaction costs 17 create sufficient variation across markets.

Estimations and robustness checks
I start with estimating the basic specification (Eqs. (7), (8), and (9)). Apart from indicator variables, both the dependent variable -the meat-maize terms of trade gap, the meat gap and the maize gap -and the abundancescarcity variable are transformed into natural logarithms. This transformation allows the interpretation of coefficients in terms of elasticities. Estimation results, reported in Table 1, confirm that the impact of food shortages on meat-maize terms of trade is statistically significant -nearly all at the 1% level -with the expected sign for all meat prices. More importantly: food shortages widen the maize price gap, as expected, but lead to an opposite movement in the meat price gap. The positive and significant coefficients in the meat price gap estimations (column 2) are clear support for increased livestock sales Price gaps are calculated as the average price in January, February and March in a specific year (the lean season prices), relative to the average price of the lowest three prices from May to November of the previous year, for a specific market. The source data are monthly market price observations for 72 locations (markets, villages and towns), from January 1991 to October 2009, taken from Agro-Economic Survey, Ministry of Agriculture and Food Security, Government of Malawi (see Sect. 3.3 for further details). Equations are estimated with OLS. All estimations include market and year fixed effects. Standard errors clustered by market are reported in brackets below the coefficient * p < 0.10,**p < 0.05,***p < 0.01 during food shortages. The estimations further support an asymmetric impact of the abundance-scarcity variable: coefficients during no shortage periods are without exception insignificant. Robustness checks, documented in Appendix (Tables 6,  7 , 8, 9, 10, 11, 12, 13, and 14), by and large support the results reported in Table 1. Estimation by region point to regional heterogeneity: the relationship is supported in estimations for the south and the north, but much less (not) in estimations for the central region. I also find stronger support (higher and more significant coefficients) for rural areas vis-à-vis urban areas. Controls (serial correlation in errors) and alternative estimation techniques (Fixed Effects (FE), Instrumental Variables -Two Stage Least Squares) generate similar coefficients.
Prior to considering the size of impacts and the economic implications of the estimations of the meat-maize terms of trade, I report estimations of another robustness check. Since spatial spillovers across markets are likely to occur, I applied an estimation technique to control for spatial correlation. Estimation results also documenting direct and total impacts are reported in Table 2.  Table 1 for data details. The fixed effect spatial regression (spxtregress) specifies spatial correlation of the dependent variable, the explanatory variable and the error term. The inverse-distance matrix has zero weights for markets more than 50 km apart. The table reports the average impacts (estat impact). Standard errors are reported in brackets next to the coefficient. * p < 0.10,**p < 0.05,***p < 0.01 In terms of the sign and significance of the coefficients, Tables 1 and 2 estimations are similar. All direct and total impacts on either the meat price gap, the maize price gap of the meat-maize term of trade gap are statistically significant, mostly at the 1% level and have the expected sign. However, spatial spillovers are statistically well supported, most convincingly in the maize price gap estimations. The difference between direct and total impact is large in the maize price gap (with indirect impact 69% of total), negligible in the meat price gap and in between in the meatmaize term of trade gap (indirect impact 11%-57% of total). Controlling for spatial spillovers has major impact on the size of the coefficients. Overall, Table 2 coefficients, both direct and total, are much higher than Table 1 OLS coefficients, except for the coefficients in meat price gap estimation (direct impact in case of the meat price gap, ranging from 0.041 to 0.047 (T2), compared to 0.032 to 0.059 (T1)). In summary, controlling for spatial spillovers further supports the impact of food shortages on the meatmaize terms of trade, reveals a substantial contribution of surrounding markets and generates strongly increased impacts of food shortages. Particularly the coefficients in maize price gap and meat-maize term of trade gap estimation increase.
The evidence, reported in Tables 1 and 2, indicates that the meat-maize terms of trade has decreased during food shortages. The decomposition into meat and maize price gaps shows that the bulk of the decrease originates from the increase in the maize price gap (T1: 77%-86%; T2: 79%-91%), while a modest part (T1: 14%-23%; T2: 9%-21%) originates from the decrease of the meat price gap. The estimates are consistent with increased livestock sales and support the hypotheses that the value of livestock have decreased during food shortages. With the double logarithmic specification coefficient can be interpreted as elasticities: considering the total impact with spatial spillovers, a 100% decrease of the abundance-scarcity variable leads to a decrease in meat-maize terms of trade varying from 54 to 65%. During food shortages observed decreases of the district abundance-scarcity variable ranged from 50 to 100%.
For several reasons, however, even this large drop in livestock value severely underestimates the actual drop. The time fixed effects absorb the country-wide covariant part of impact: coefficients further increase without time fixed effects. Next, the estimations report averages and employ district aggregates in the key explanatory variable, which both conceal distributions while food shortages occur in the tail of distributions. For these reasons, it is more informative to take a closer look at the change in meat-maize terms of trade during food shortages per se. From a food-security perspective a typical household is, in the end, indifferent where the drop in livestock value comes from: What matters for households is to have sufficient food to survive a food shortage. Nevertheless, the estimation results do provide sound support for the negative correlation between meat-maize terms of trade and food shortages.

Alternative explanations
I discuss two alternative explanations of a drop in the value of livestock. First, the same weather conditions that lead to a crop failure may also result in underfed cattle which than command lower proceeds in the market. A drop in body weight and health of livestock during a food shortage due to underfeeding is clearly a channel that directly affects the value of livestock to the household (Barrett et al., 2003). However, such a drop in value is likely to be concentrated in the volume or quantity rather than the price of meat: 1 kg of meat from well-fed or 1 kg of meat form underfed cattle will fetch more or less similar prices. It is especially a drop in the quantity, the volume of meat, that arises due to underfeeding. Consequently, the potential additional reduction in livestock value arising out of underfeeding, makes the measured drop in value a lower bound of the actual loss in value. Secondly, the price inelasticity in maize demand may result in a strong negative income effect on the demand for food, which also reduces meat prices. The evidence tells a different story: Market participation of farm households in sub-Sahara Africa agriculture is limited. Many households are subsistence farmers and do not purchase or sell on the market. However, with a crop failure households are forced to purchase food on the market with their available savings, partly livestock. Also, households that normally would sell maize on the market, have to purchase maize on the market in case of a crop failure. Hence, a food shortage -a shift of the supply curve -also triggers increased staple food demand, complemented with increased livestock supply from most households.  Figure 5 illustrates graphically the development of the goat meat-maize terms-of-trade for several markets, supporting an extreme drop of the relative value of goat meat during the climax of this food shortage.

Taking a closer look at the meat-maize terms of trade
As already hinted at in the previous section, a more direct measurement of the size of the reduction of the wealth value of livestock will be more informative and can be obtained by simply summarizing the change in meat-maize terms of trade during specific periods of food shortage. The major food shortage episodes in Malawi in our sample occur in 1997-1998, 2001-2002 and 2005-2006. Table 3 reports the relative change in meat-maize terms of trade during food shortages, calculated as the average terms of trade from January to March -the lean season months in Malawi -relative to these terms of trade in the post-harvest months during the previous year. The relative change is calculated for all villages, towns and cities recorded in the data, and subsequently averaged over all these markets. The table indicates that the average meat-maize terms of trade decreased during these episodes from 56 to 73%. The estimated decrease in livestock value is likely to be an underestimate since our data do not allow to measure the reduction in value due to a lower body weight of livestock during food shortages (see also Sect. 5.1).
Another interesting period is the period from early 2008 to late 2009. During this period staple food prices in Malawi increased substantially (See Fig. 9). However, this price increase originated from the world market and had no relation with the domestic agricultural outcome in Malawi. During this period staple foods were available in   (Fig. 2), indicating normal and not excess supply of livestock.

Meat-staple food terms of trade during food shortages in other SSA countries
Similar developments in meat-staple food terms of trade are observed in many sub-Sahara African countries. Figure 6 shows the meat-maize terms of trade in Uganda during the 2011 food shortage. Figures 11, 12, 13, 14, 15, and 16 show additional and similar evidence for several countries (Uganda, Kenya and Somalia), for several meats, for several staple foods and during several periods of food shortage. 19 In short, the experience in many sub-Sahara African countries indicate a systematic pattern of large drops in meat-staple food terms of trade during food shortages, reflecting drops in the value of livestock.

Implications for households
How does the drop in livestock value affect the average household? With a simple back-of-the-envelope calculation   (Fafchamps et al., 1998;Kazianga & Udry, 2006). The observed preference for using grain stocks to bridge periods of staple food shortages is clearly not an issue of liquidity (Kazianga & Udry, 2006;Mogues, 2011): in contrast with livestock, grain stocks do not lose value during a food shortage. The extreme drop in livestock value during food shortages highlights a major drawback of using livestock sales to overcome adverse impacts of crop failures. A savings device that drastically decreases in value when needed is a poor savings device. 21 An adequate savings device should not lose value at the very moment when it is liquidated. The lack of adequate savings devices creates obvious risks: it makes it more difficult to protect against food shortages and maintain food security, and it aggravates the risk of agricultural production. In case of credit constraints and risky agriculture, Dercon (1998Dercon ( , 2002 shows that low income households tend to choose low return-low risk activities. Conversely, with savings instruments that adequately protect against food shortages, households have better opportunities to take up high return-high risk activities (like cash crops or cattle rearing). The long run effects of eliminating ex-ante risk on welfare and growth are shown to be large (Elbers et al., 2007).

Dampening supply shocks and tilting the demand curve
What possibilities are there to improve the quality of livestock savings? And what alternative savings devices are available? The trivial but unsatisfactory answer is to avoid crop failures in agriculture. Nevertheless, making agriculture less sensitive to the vagaries of the weather through wider application of irrigation, water harvesting and conservational cultivation techniques will dampen peaks in staple food prices and thereby also strengthen the adequacy of using livestock as a savings device. An alternative way to dampen supply shocks is to make staple food imports less costly. Improved transport links, in the first place with South Africa, Mozambique (Nacala) and Tanzania (Dar es Salaam), but also domestically, have the potential to make Malawi less self-reliant with the result that staple food prices would be less sensitive to domestic production. A more modest increase in staple food prices will automatically also improve the usefulness of livestock savings. Another way to improve the adequacy of livestock as a savings device is to diversify staple food demand. The highly inelastic maize demand in Malawi is driven by traditionally strong consumer preferences (Smale, 1995). These preferences translate into a high price inelasticity 20 A typical household -6 persons: 2 adults, 2 teen-agers, and 2 children below 10 years of age -that grows maize for home consumption requires around 700 kg of maize for a whole year, equivalent to around 54 kg per month, to adequately feed all household members. The quantity of maize required to feed a household is calculated as follows: [2 (adults) × 2300 (daily kcal requirement of adults) + 2 (teenagers) × 1800 + 2 (children) × 1200)] × 60% (maize calorie share in household diet) × 365 (number of days per year) / 3570 (per kg kcal content of maize). I calculate the market value of selling goats by this household in terms of maize. I propose that the household can sell two goats. Goats have a weight of 30 kg and the meat content of a goat is 25 kg. With a complete crop failure and assuming that households fully and perfectly anticipate maize prices and household maize needs in the remaining marketing season, the household sells its two goats directly after harvesting time and -also directly after harvesting time -purchases maize. With these assumptions households can cover approximately 33% of maize needed till the next harvest with livestock sales. With an average reproduction rate of around 60% per year and depending on the size of the household herd before the foodshortage, it might take a few years to recover the previous household herd of livestock, and to re-establish pre-food-shortage levels of food security. 21 The more general message is that agricultural households in developing countries suffer rather than benefit from seasonality: households sell agricultural output when prices are low: livestock during food shortage, but also, staple food crops directly after harvest. Most households also purchase staple food when prices are high, in the lean season or during food shortages. There are substantial and largely unexploited arbitrage opportunities (Burke et al., 2019). 1 3 of maize demand. A larger response to prices or a flatter maize demand curve would significantly reduce extreme maize price increases during food shortages. Substitution into other staple foods will increase responses to maize price increases. Rice and cassava are obvious candidates, sorghum is also likely to work: diversification of the consumption diet into these staples will contribute to make the maize demand curve flatter, mitigate maize price increases during food shortages and, jointly with this, reduce the drop in livestock value.

Formal savings and insurance instruments
Putting an alternative savings device in place, like a formal bank or savings account, or a formal insurance, is another technique to protect against crop failures. However, formal bank accounts or savings accounts in sub-Saharan countries are shown not to be attractive for most households, even if subsidized (Dupas et al., 2018). Major reasons for limited attractiveness and low active usage are extreme poverty ("too poor to save"), high transaction costs and coexistence with several alternative types of informal savings (Dupas et al., 2018). Monetary savings also suffer from the same drop in purchasing power with extreme food price increases.
Protection against the risk of crop of crop failure is possibly better addressed with formal insurance. Conventional insurance in agriculture has a bad record and proved infeasible due to severe problems of moral hazard and adverse selection, high monitoring costs and high administrative costs. Index insurance resolves these fundamental problems and is thereby considered an attractive and promising alternative for developing economy agricultural settings. During the last decades several practical applications of rainfall index insurance have seen the light, often on a pilot basis. The potentially attractive properties of index insurance has also triggered a large research effort on impact and pitfalls of rainfall index insurance (Ahmed et al., 2020;Cole et al., 2013;Dercon et al., 2014;Giné & Yang, 2009;Giné et al., 2008;Jensen et al., 2016;Karlan et al., 2014). Unfortunately index insurance is not the silver bullet that is hoped for: willingness-to-pay and take-up appear to be notoriously low (Ahmed et al., 2020;Giné et al., 2008). Low take-up is attributed to a variety of causes like unfamiliarity with formal insurance, lack of understanding and poor information dissemination (Ahmed et al., 2020), the extent of basis risk (Dercon et al., 2014;Giné et al., 2008;Jensen et al., 2016), and the interaction with informal insurance arrangements (Dercon et al., 2014) or credit (Ahmed et al., 2020;Giné & Yang, 2009;Karlan et al., 2014). In this context livestock saving, another type of informal insurance, can also be expected to affect household demand for rainfall index insurance. Much research focuses on the impact of credit contingent with rainfall index insurance on technology adoption and investment in agriculture, rather than on food security (Ahmed et al., 2020;Karlan et al., 2014). Our measurements of staple food price increases during food shortages point at an essential requirement of indemnity claims of rainfall index insurance schemes (but also rainfall-insurance-backed social safety nets): these indemnity payments need to be defined in real terms, for example in kg of maize, corresponding with the loss of maize harvest, or its money equivalent at the price level of the moment the claims are made. To achieve this will be a nasty hurdle for insurance contracts.

Storage and storage policies
In short, formal devices like bank or savings accounts, or index insurance schemes cannot effectively protect against the extreme staple food price increases that occur after droughts and crop failures. So what than offers adequate protection against food shortages? Savings in the form of storage of grain is a savings device that maintains its real value, even under extreme food shortages or food price increases. Evidence suggests that farmers often do not have access to appropriate storage technologies for staple foods (Aggarwal et al., 2018). Lack of adequate storage devices also could be a major reason for farmers to sell maize shortly after harvest. Alternatively, availability and use of good grain storage technologies is shown to have behavioral impacts on farmers (Aggarwal et al., 2018), 22 to increase storage duration and to reduce storage losses (Luo et al., 2022). If farm households are able to commit to store a quantity of grain equivalent to three months of staple food consumption after a normal or bumper harvest -around 7 to 8 times out of 10 harvests -this would create sufficient protection against the following season possible crop failure. 23 Even storage for shorter periods is shown to generate major arbitrage gains (Burke et al., 2019). Implementing storage of individual farmer output at the village level appears to be effective as it enhances commitment and organization, and keeps incentives close to beneficiaries, the farmers. Both shifting grain stocks to high price periods -either for home consumption after a crop failure, or to sell on the market -and sales of livestock to normal periods would both generate arbitrage gains for farm households (Burke et al., 2019). 24 For an individual household, selling livestock, and subsequently purchasing and storing maize, both directly after the maize harvest, appears a strategy that benefits to the utmost from potential arbitrage opportunities and simultaneously creates food security. A policy which promotes grain stocks at the household level by making crop storage bags available -hermetically sealed storage bags (like the so-called Purdue Improved Crop Storage bags (PICS)) -and by subsidizing farmers' group level, village level or household level granaries recommends itself (Aggarwal et al., 2018). There may also be a role for governments to (partially) subsidize replenishment of maize after a food shortage. A combination of credit contingent with grain storage, implemented by commercial banks, could further enhance food security jointly with securing funds for risky investment (Aggarwal et al., 2018;Basu & Wong, 2015;Channa et al., 2022;Omotilewa et al., 2018).

Summary and conclusion
I measure to what extent food shortages affect the value of livestock savings. For this purpose I exploit monthly market prices of meat and maize in Malawi, for 72 locations (towns, villages and markets), for the period from 1991 to 2009. The empirical estimations offer convincing support for the claim that decreases in real meat prices, and thereby in the value of livestock are associated with food shortages. The value of livestock -measured with the meat-maize terms of trade -decreases substantially during food shortages, with reductions averaged over villages, cities and towns, ranging from 54 to 65%. The evidence is consistent with increased livestock sales during food shortages, but the drop in meat-maize terms of trade arises primarily due to maize price increases. Results are robust for serial and spatial correlation, for potential endogeneity of food shortages, and for various other threats. Since the data do not allow to estimate the decrease in livestock value due to a lower body weight of livestock during food shortages, our estimates need to be interpreted as lower-bound estimates: the actual drop in livestock value is likely to be even larger. The reduction in the value of livestock occurs at the very moment livestock is sold on the market to purchase staple foods. Similar to produced staple foods, households systematically tend to sell low. Dampening price peaks in staple food prices during food shortages through diet diversification, and irrigation and conservational cultivation techniques in agriculture will help to mitigate the drop in value of livestock savings. However, large price fluctuations in staple foods will remain. Therefore savings devices are needed that do not lose value during food shortages and thereby offer adequate protection. Such savings devices will generate large welfare gains and enhance economic growth. Most formal instruments are not suitable for this purpose. Storage of staple food at the household level appears a strategy with the best perspectives.     -90 1990-91 1991-92 1992-93 1993-94 1994-95 1995-96 1996-97 1997-98 1998-99 1999-2000 2000-01 2001-02 2002-03 2003-04 2004-05 2005- The darker shades indicate increasingly lower rainfall levels relative to minimum rainfall levels required for vegetative growth (around 700 mm per season) Additional estimation output      Two-way fixed effects

Instrumental variable estimation
Strictly there is no need to instrument the abundancescarcity variable or apply any other adjustment (for omitted variables, reverse causality, endogeneity, etc.): production of maize relates to the previous season and is hence predetermined, and population by district develops only gradually over time, without major shifts and shocks, and is also largely predetermined. However, to avoid potential issues on this account between maize prices and maize production (versus maize requirements), I instrument the abundancescarcity variable with seasonal rainfall by district and rainfall levels (low-medium-high) by ADD.
Hence, I have: and Equation (10) is estimated to address potential biases due to endogeneity and estimation results are reported in Table 13.
(10) gap x jt = 0 + 1m aize production jt−1 ∕maize requirement jt maize production jt−1 ∕maize requirement jt = 0 + 1 seasonal rainfall jt + ∑ 2,km rainfall level tk * ADD m  Price gaps are calculated as the average price in January, February and March in a specific year (the lean season prices), relative to the average price of the lowest three prices from May to November of the previous year, for a specific market. The source data are monthly market price observations for 72 locations (markets, villages and towns), from January 1991 to October 2009, taken from Agro-Economic Survey, Ministry of Agriculture and Food Security, Government of Malawi (see Sect. 3.3 for further details). Equations are estimated with IV-2SLS. Instruments for the abundance-scarcity variable are last season rainfall by district and rainfall by indicator (low-medium-high) interacted with ADD. All estimations include location (market) and time fixed effects. Standard errors clustered by market are reported in brackets below the coefficient.

Driscoll-Kraay standard errors
Kleibergen-Paap rk LM statistic under-identification test, Cragg-Donald Wald F statistic and Kleibergen-Paap rk Wald F statistic weak identification test; Hansen J statistic over-identification test of all instruments * p < 0.10,**p < 0.05,***p < 0.01 (1)  Price gaps are calculated as the average price in January, February and March in a specific year (the lean season prices), relative to the average price of the lowest three prices from May to November of the previous year, for a specific market. The source data are monthly market price observations for 72 locations (markets, villages and towns), from January 1991 to October 2009, taken from Agro-Economic Survey, Ministry of Agriculture and Food Security, Government of Malawi (see Sect. 3.3 for further details). The fixed effect spatial regression (spxtreress) specifies spatial correlation of the dependent variable, the explanatory variable and the error term. The inverse-distance matrix has zero weights for locations more than 50 km apart.. Standard errors are reported in brackets next to the coefficient * p < 0.10,**p < 0.05,***p < 0.01