TY - JOUR AB - Abstract We investigate determinants of dairy producers’ risk exposure using a unique combination of foci on (i) downside risks, (ii) a holistic representation of revenues from milk and animal sales, (iii) climatic extremes and (iv) the role of animal health. A sample of German dairy farms reveals that animal health and heat stress indicators influence mean and semi-variance of revenues. For instance, heat stress exposure reduces expected milk revenues significantly. In the case of animal health-related indicators, our results show trade-offs between expected revenues and downside risks. Furthermore, variabilities in revenues from milk and animal sales are significantly interrelated. 1. Introduction The dairy sector is vitally important for agriculture, especially in Europe which is the largest milk producer worldwide (e.g. Bouamra-Mechemache, Jongeneel and Réquillart, 2008; Hirsch and Hartmann, 2014). Dairy producers face a number of risks, such as volatile quantities and qualities of outputs, institutional risks and fluctuating market prices (e.g. Chen, Roberts and Thraen, 2006; Wolf, Black and Hadrich, 2009; Valvekar et al., 2011; D’Antoni and Mishra, 2012; Henry et al., 2016). Production risks have been identified as crucial (Chen, Roberts and Thraen, 2006; Tveteras, Flaten, Lien and 2011; Berentsen, Kovacs and Van Asseldonk, 2012; Orea and Wall, 2012; El Benni and Finger, 2013) and are assumed to be driven by two main factors. Firstly, climatic variability and extreme climatic events affect the quantity and quality of animal products, such as milk and meat (e.g. Kadzere et al., 2002; Tubiello, Soussana and Howden, 2007; Key and Sneeringer, 2014). Secondly, animal health is essential for the production process and represents an important managerial component and source of income variability (e.g. Antle and Goodger, 1984; Lusk and Norwood, 2011; Hansson and Lagerkvist, 2014; Allendorf and Wettemann, 2015). Animal health can comprise various dimensions and is reflected in animal health-related indicators such as metabolic stress, the number of mastitis cases per cow or the frequency of animal losses (Oltenacu and Broom, 2010).1 Moreover, dairy production is characterised by a joint realisation of sales of milk and animals. In addition, revenues from animal sales are stochastic and contribute to adjustments of overall risks faced by the farm (e.g. Tveteras, Flaten and Lien, 2011). The identification of key risk determinants in dairy production and the quantification of their impacts, combined with a deeper understanding of farm-level responses to risks are of outmost importance when seeking to support farmers’ risk management strategies, develop improved risk management tools and better policies. We aim to identify determinants of downside risk exposure in dairy farms and quantify their effects. We provide the first analysis of risks in dairy production that covers a unique combination of foci on (i) downside risks, (ii) a holistic representation of revenues from milk and animal sales, (iii) climatic extremes and (iv) the role of animal health. This extends earlier research on risks in dairy production that has focused only on individual aspects (e.g. Tveteras, Flaten and Lien, 2011; Antle and Goodger, 1984; Key and Sneeringer, 2014). Our analysis tests the significance of different determinants of production risks in the dairy sector, quantifies marginal effects of various inputs on farm-level revenues and revenues variability and thus aids the development of better policies to overcome risks. Furthermore, we extend the current literature by identifying non-linearities and critical thresholds for risk exposure of dairy farms. To this end, we develop a multi-output stochastic production framework that is applied to a case study on dairy production in Germany, the largest European milk producer and processor. We use a rich panel dataset of 390 dairy farms over the time span from 2007 to 2014 that includes details on the economic performance of the farms, animal health indicators as well as production relevant indicators for weather extremes. Our econometric model accounts for the effects of various inputs and control variables on mean and semi-variance of revenues from milk and animal sales. Moreover, the joint production framework chosen in our analysis allows us to consider input and control variable effects on the semi-covariance between both outputs. The remainder of this paper is structured as follows. We start by developing the theoretical economic and econometric backgrounds. The econometric implementation and the dataset used are described in the next section. Subsequently, results are presented and discussed. The final section contains our concluding remarks. 2. Economic and econometric framework 2.1. Economic framework We use a stochastic production function framework to describe effects of different inputs on expected outputs as well as production risks (see Saastamoinen, 2015, for a recent survey). In line with Antle (1983), we assume that farmers maximise their utility with respect to a vector of inputs X taking into account moments (1,…, m) of different outputs produced (1,…, j) as follows: maxxE[U(π)]=h[μ11(X),μ21(X),…,μm1(X);…;μ1j(X),μ2j(X),…,μmj(X)] (1) where μ1j(X) represents the expected revenues of output j, and μk=E(Πj−μ1j)k,k=2,…,m denotes the mth moment of the distribution with Πj being the vector of realisations of revenues for output j. The analysis of risk in dairy production is often restricted to the variance of revenues (see Antle and Goodger, 1984, for an exception). The focus of this assumption is only valid if mean and variance are sufficient to describe the revenue distribution, i.e. no higher moments exist, and/or farmers have no preferences for higher moments. However, these assumptions are rarely fulfilled as agricultural returns are characterised by extreme loss events and farmers are often downside risk averse (e.g. Antle, 1983; Di Falco and Chavas, 2006; Koundouri et al., 2009). In fact, downside risk exposure can create liquidity problems for a dairy farm and potentially lead to bankruptcy (Hansson and Lagerkvist, 2014). Expanding the framework to higher moments such as skewness and kurtosis often faces an empirical challenge, namely that null hypotheses are not rejected (e.g. Kim et al., 2014).2 Therefore, we propose the use of a semi-variance approach (see also Miranda and Glauber, 1991; Vedenov and Barnett, 2004) to account for downside risk without expanding the number of equations to be estimated. Semi-variance only accounts for losses that are below a specific benchmark. Here we focus on revenue realisations below the expected value so that the semi-variance of output j can be defined as SVarj(X)=E{Πj(X)−E(Πj(X))}2∀Πj(X)0. Since milk production and animal sales are by nature closely linked, we assume inputs to be non-allocable (see also Tveteras, Flaten and Lien, 2011). Thus, input use not only influences the distribution of individual revenues but also the semi-covariance of revenues. With s1 and s2 representing the shares of milk and animal sales in total sales, mean and semi-variance of total revenues π arising from milk and sales of animals can be summarised as follows: E(π)=s1E(Π1(X))+s2E(Π2(X)) (2) for expected total revenues and SVar(π)=s12SVar(Π1(X))+s22SVar(Π2(X))+2s1s2SCov1,2(X) (3) for total semi-variance (e.g. Estrada, 2007). Note that we assume a (downside) risk averse farmer, so that the utility increases for higher levels of expected revenues E(π), but decreases for, ceteris paribus, higher levels of revenue semi-variance SVar(π). SCov1,2(X) is the semi-covariance between revenues from both outputs, defined as SCov1,2(X)=E[{Π1(X)−E(Π1(X))}·{Π2(X)−E(Π2(X))}]∀Πj(X)