译者的话
抗生素（不包括疫苗）在国内规模化养殖场的生产成本中占比约为2.5-5%，不是很高，但头均绝对用量相对于欧美养殖业还有很大的降幅空间。荷兰汇总性数据显示：抗生素使用的减少不会对农场的技术或经济效益造成负面影响，对为改善动物健康而采取的措施进行了后续调查，这些措施使减少抗生素的使用成为一种可能。随着我国经济水平的进一步提升，养殖业的减抗、禁抗乃至无抗从由部分企业自发的行为逐渐演变为政府强制性的法规。无论是养殖业主还是养殖从业者，抗生素用量的降低不仅仅是养殖成本的关注，更多的是对美好生活的一种态度。打开南风窗，请赏《荷兰养猪场抗生素使用与盈利情况对比》。
荷兰养猪场抗生素使用与盈利情况对比（中）
Antibiotics Use Versus Profitability on Sow Farms in the Netherlands-Part 2
接上文...
2.方法Method
2.3.面板数据分析
Panel data analysis
第一步是分析荷兰过去几年抗生素的使用趋势。日使用剂量用线性回归函数近似来测定2008年前后的结构变化。随后，对固定效应模型和随机效应模型进行了估计。本质上，是对衡量生产水平的不同变量的函数进行了估计。我们能够检查抗生素的使用是否会影响任何表明农场生产水平的变量。模型定义如下(Verbeek,2012)：
The first step was to analyse the trend of antibiotics use in the Netherlands in the past years.NDD was approximated with a linear regression function to test for structural changes before and after 2008.Subsequently,both fixed effects and random effects models were estimated.In essence,functions of different variables measuring performance were estimated.We were able to examine whether the antibiotics use affects any of the variables indicating farm performance.The models were defined as follows(Verbeek,2012):
其中，yit是农场i在时间t的经济因变量，β0是所有观察中所有对象的截距项，x'it是农场i在时间t的独立变量的向量(包括抗生素使用)，β是描述x'it和yit之间的关系的系数向量，而和εit是观测比残差项。假设这些剩余项遵循一个正态分布，其平均值为0，且标准差为常数2，即σε2。在固定效应模型的情况下，αi是一个特定项的截距。在随机效应模型的情况下，αi被假设是一个特定项的残差，遵循一个标准偏差为σα2的分布(通常是正态分布)。随机效应模型假设在模型中残差(包括αi和εit)和所有自变量之间都没有相关性。在考虑所有其他协变量的情况下，通过观察抗生素使用系数的显著性来探讨抗生素使用与农场主收入之间的关系。本研究采用的显著性水平为0.05。此外，系数的标准误差是聚集在农场ID上的，使其对同一农场重复测量引起的异方差和自相关具有鲁棒性，降低了第二类误差发生的概率(Petersen,2009)。因此，由于公司效应，采用聚类标准误差来调整独立恒等分布的IID(0,σε2)的标准差(Abadie et al.,2017)。随机效应模型优于固定效应模型，因为随机效应模型假设这些农场是从分布更广的荷兰母猪场中抽取的。然而，因为特定项的残差(αi)以及模型中的任何自变量之间存在相关性(造成内生性和有偏差的估计)，并非总能使用随机效应模型。在这些情况下，本文使用了固定效果规范。Hausman检验可以用来确定是使用随机效应模型还是固定效应模型(Verbeek,2012)。零假设表明，固定和随机效应模型的系数估计是一致的，尽管随机效应模型的估计更有效。然而，在Hausman检验中，无法比较系数标准误差聚集的模型。因此，Hausman检验是在没有聚类标准误差的情况下进行的。Sargan-Hansen过度识别测试(以XTOVERID命令在Stata中使用(Schaffer and Stillman,2016))中允许聚类标准错误。因此，除了Hausman检验外，我们还使用了Sargan-Hanssen检验。
where yit is the dependent economic variable of farm i at time t,β0 is the common intercept term for all subjects over all observations,xit'is the vector of independent variables(including antibiotics use)of farm i at time t,βis the vector of coefficients describing the relationship between X'it and yit andεit is the observation specific residual term.These residual terms are assumed to follow a normal distribution with a mean of 0 and a constant standard deviation ofσε2.In the case of a fixed effect model,αi is a subject-specific intercept.In the case of a random effects model,it is assumed thatαi is a subject-specific residual that follows a(usually normal)distribution with a standard deviationσα2.The random effects model assumes that there is no correlation between the residuals(bothαi andεit)and all independent variables in the model.The relation between antibiotics use and the farmer income,given all other covariates,was explored by looking at the significance of the coefficient of the antibiotics use.This study used a significance level of 0.05.In addition,the standard errors of the coefficients are clustered around the farm ID,making them robust for heteroscedasticity and autocorrelation caused by repeated measures on the same farm(the firm effect)and reducing the probability of a type II error(Petersen,2009).Therefore,clustered standard errors were used to adjust the standard errors for deviations from an independent and identically distribution IID(0,σε2)due to the firm effect(Abadie et al.,2017).The random effects model was preferred over the fixed effects model as it was assumed that these farms are drawn from a larger distribution of Dutch sow farms.However,it may not always be possible due to a correlation between the subject-specific residual(αi)and any of the independent variables in the model(causing endogeneity and biased estimates).In these cases,this paper used a fixed effects specification.The Hausman test can be used to determine whether to use a random or a fixed effects model(Verbeek,2012).The null hypothesis stated that the coefficient estimates are consistent for both the fixed and the random effects model,although the random effects model is more efficient in its estimates.However,models with clustered standard errors of coefficients cannot be compared in the Hausman test.As a result,the Hausman test was done without clustered standard errors.The Sargan-Hansen overidentification test(implemented in Stata as XTOVERID(Schaffer and Stillman,2016))allows clustered standard errors.Therefore,the Sargan-Hanssen test is used in addition to the Hausman test.
分析表明，数据似乎不坚持Hausman检验的渐近性质。这可能是由于残差之间仍然存在自相关，因为Hausman检验不支持具有聚类标准误差的模型。在所提出的模型中，这种自相关在聚类标准误差中得到了修正，如前所述。Wooldridge(2001)指出“由此产生的测试其渐近水平可能大于或小于标称的大小”。在这种情况下，就需要Hausman检验中的sigmamore功能。这确保了Hausman检验符合所需的渐近性质。Sargan-Hansen过度识别测试是在测试结果不一致的情况下使用的，因为sigmamore是一个有点特别的临时解决方案。使用Stata 14(StataCorp,2015)建立并测试了这些模型。
The analysis showed that the data did not seem to adhere to the asymptotic properties of the Hausman test.This may be attributed to the fact that there is still autocorrelation between residuals,which is possible due to the Hausman test not supporting models with clustered standard errors.In the presented models,this autocorrelation was corrected for in the clustered standard errors,as indicated above.Wooldridge(2001)states that“the resulting test could have an asymptotic size larger or smaller than the nominal size.”In this case,the sigmamore option of the Hausman test was needed.This ensures that the Hausman test adheres to the needed asymptotic properties.The Sargan-Hansen overidentification test was chosen in case the results of the tests were in disagreement as the sigmamore is a somewhat ad hoc solution.These models were built and tested using Stata 14(StataCorp,2015).
2.方法Method
2.4.与抗生素使用有关的行为因素
Behavioural factors attributing to antibiotics use
为了探究减少抗生素使用决定的行为因素，我们使用了与计划行为理论(TPB)相关的结构(Ajzen,1991)。理论指出，个人执行某种行为的意愿，例如减少母猪使用抗生素，会收到三个方面的影响，即：其自身对这种应为的态度、对此人很重要的其他人对该行为的意见(社会规范)、以及此人对自己控制行为的能力的感知(感知行为控制)(Ajzen,1991)。态度可以用行为信念和对结果的评价来解释。行为信念是对拟议行动的后果的信念(例如，试图限制抗生素的使用将导致较低的经济表现)，而结果的评估是个人对结果的赋予的价值，社会规范可以与规范信念和遵守的动机相联系。规范性信念可以定义为“重要的参照物个人或团体同意或不同意执行某一特定行为的可能性”，遵从动机可以认为是对遵循规范信念的倾向，它依赖于代表该信念的人(例如：兽医对你决定限制使用抗生素的影响有多大？)。感知行为控制(PBC)可以定义为一个人对其自身执行行为技能的信心。它可以分为PBC能力和PBC控制力，并与控制信念相关，控制信念可以被定义为“存在或缺乏必要的资源和机会”(Ajzen,1991)。
In order to explore the behavioural factors that explain the decision to decrease the antibiotics use,constructs related to the Theory of Planned Behaviour(TPB)are used(Ajzen,1991).The theory states that the intention of an individual to perform a certain behaviour–for example to use less antibiotics for sows–is influenced by his or her attitude towards the behaviour,the opinion of important others(social norm)about the behaviour and the individual’s perception of the control he or she has over the behaviour(perceived behavioural control)(Ajzen,1991).Attitude can be explained by behavioural beliefs and evaluation of the outcome.Behavioural beliefs are the believed consequences of the proposed action(e.g.attempting to limit antibiotics use will result into lower economic performance),and the evaluation of outcome is the value that individuals place on an outcome social norms can be linked to the normative beliefs and the motivation to comply.Normative beliefs can be defined as“the likelihood that important referent individuals or groups approve or disapprove performing a given behaviour”,and motivation to comply can be referred to as the tendency to follow a normative belief,dependent on the person who represents the belief(i.e.what is the influence of the veterinarian on your decision to limit antibiotics use?).Perceived behavioural control(PBC)can be defined as the confidence an individual has in his or her skills to perform the behaviour.It can be divided in PBC-capability and PBC-controllability,and is linked to control beliefs,which can be defined as the believed“presence or absence of requisite resources and opportunities”(Ajzen,1991).
除了与计划行为理论相关的构念外，还包括与感知风险、不确定性和相对风险知觉(环境知觉)相关的构念。这一点很重要，因为认为风险较小的农场主可能不太愿意采取预防措施(Kunreuther et al.,2001;Ogurtsov et al.,2008)。感知的不确定性可以被描述为行为成功的感知概率(即，我不确定是否可以限制我农场的抗生素使用)。
In addition to the TPB-related constructs,constructs related to perceived risk and uncertainty and relative risk perception(perception of the environment)are included.This is important because farmers who perceive less risks are probably less willing to adopt preventive measures(Kunreuther et al.,2001;Ogurtsov et al.,2008).Perceived uncertainty can be described as the perceived probability to succeed in the behaviour(i.e.I am uncertain whether I can limit my antibiotics use).
概念变量(构念)是由FADN后续调查中单独的项目(问题)组合而成的，因为这些概念变量很难直接测量。单独的项目大多是按7分制来衡量的，其中，1是最否定的答案(如完全不同意，非常不可能)，7是最肯定的答案(如完全同意，非常可能)。用否定词的条目被反编码。接下来，通过Cronbach’s alpha(Tavakol和Dennick(2011)所阐述的方法)检验将被组合成一个概念变量(构念)的单独条目的内部一致性。Tavakol和Dennick(2011)指出，这些值应该在0.70到0.95之间。达到满意的结果后，对这些项目的所有测量取平均值，以获得一个单一概念结构的值。
The concept variables(constructs)are created by combining separate items(questions)of the FADN follow-up survey as these concept variables are hard to measure directly.The separate items were mostly measured on a 7-point scale where 1 was the most negative answer(e.g.totally disagree,very unlikely),and 7 the most positive answer(e.g.totally agree,very likely).The negatively worded items are inversely recoded.Next,the internal consistency of the separate items to-be-combined into one concept variable(construct)are tested by Cronbach’s alpha(as explained by Tavakol and Dennick(2011)).Tavakol and Dennick(2011)state that these values should range from 0.70 to 0.95.After the result is deemed satisfactory,the mean is taken for all the measurements of these items to obtain a value for a single conceptual construct.
运用回归分析来探讨框架中所描述的概念之间的联系。将日剂量数的偏斜数据进行对数变换以符合正态性。然后，分析平均年使用量和从调查数据中收集的信息之间的联系。在回归分析中，意图、态度、积极和消极的行为信念、结果评价、社会规范、标准信念、遵从动机、感知行为控制、控制信念以及感知风险和不确定性作为预测因子。在多元回归分析中，如果两种构念之间的皮尔逊相关系数为0.50，则它们不用于同一个模型中。如果变量不是正态分布，则将变量分为三分位数，分值从1到3。
Regression analysis is applied to explore the linkage between the concepts described in the framework.The skewed NDD data is log transformed to conform to normality.Next,the association between the average yearly use and the information gathered from the survey data are analysed.In the regression analyses intention,attitude,positive and negative behavioural beliefs,evaluation of outcome,social norm,normative beliefs,motivation to comply,perceived behaviour control,control beliefs and perceived risk and uncertainty are used as predictors.In the multivariable regression analyses,two constructs are not included in the same model if the Pearson’s correlation coefficient between them>0.50.If variables were not normally distributed,the variables were divided into tertiles,with scores ranging from 1 to 3.
3.结果Results
3.1.描述与趋势分析
Descriptive and trend analysis
农场数据集由每个变量最多620和最少371个观察值组成。农场的平均年数约为6年(表1)。在农场之间和农场内部(在时间上)观察到显著的异质性。
The farm dataset consisted of maximally 620 and minimally 371 observations per variable.The mean number of years a farm in the panel was approximately 6 years(Table 1).Substantial heterogeneity was observed between farms as well as within farms(in time).
表1：2004-2016年荷兰母猪场的描述性统计数据。
Table 1.Descriptive statistics Dutch sow farms for time period 2004–2016.
图2描绘了每年每头猪抗生素日剂量(NDD)的趋势和分布。2008年之前的趋势不显著(0.97 NDD,p=0.36,R2adj=0.01)，而2008年之后的线性回归模型中有显著的负系数(−2.44 NDD,p<0.01,R2adj=0.15)。并且，据荷兰兽医局报告称，2009年至2016年，养猪生产中的抗生素每日使用剂量下降了57%(Stichting Diergeneesmiddelen Autoriteit,2018)。
Fig.2 depicts the trend and the spread of the number of daily doses of antibiotics per animal per year(NDD).Trends were insignificant up to 2008(0.97 NDD,p=0.36,R2adj=0.01),while there was a significant negative coefficient in the linear regression model after 2008(−2.44 NDD,p<0.01,R2adj=0.15).Also the Dutch Veterinary medicine authority reported a 57%decline of NDD in pig production from 2009 to 2016(Stichting Diergeneesmiddelen Autoriteit,2018).
图2：2004-2016年荷兰农场每头猪每日抗生素剂量(NDD)的平均值和25%-75%间隔。
Fig.2.Mean and 25-75%percentile interval of the number of daily doses of antibiotics(NDD)per animal year per Dutch farm from 2004-2016.
案件回顾
3.2.面板数据分析
Panel data analysis
通过Sargan-Hansen过分识别检验和Hausman检验，五种回归模型均包含固定效应规范(表2)。此外，使用固定效应模型时，场内R2大于场间R2和总体R2。这种差异是由于固定效应模型侧重于解释对象内部的差异，而不是解释对象之间或整体的差异(Wooldridge,2001)。
All five regression models included fixed effects specifications as indicated by the Sargan-Hansen overidentification test and the Hausman test(Table 2).Moreover,the within R2 tended to be higher than the between R2 or overall R2 with fixed effects models.This difference was caused by the focus of the fixed effects model on explaining within-subject variation rather than between or overall variation(Wooldridge,2001).
表2：荷兰母猪场的回归分析模型，在控制2004-2016年期间其他重要因素的同时，评估生产性能指标和NDD之间的关系。
Table 2.Regression models for Dutch sow farms,estimating the relation between a performance indicator and NDD while controlling for other significant factors for time period 2004-2016.
注：括号里是(聚类)标准误差和P值。该表还包括观察的数量、每个模型中对象的数量、Hausman检验结果，Sargan-Hansen过度识别测试的结果、所选的模型、场内R2、场间R2和总体R2。场间R2给出了模型所解释的农场内部变异的份额。FE=固定效应模型，RE=随机效应模型。(1)表示使用了sigmamore法。
The(clustered)standard errors and P-values are between brackets.The table also includes the number of observations and number of subjects used per model,together with the Hausman test outcome,the outcome of the Sargan-Hansen overidentification test,the chosen model,the within R2,the between R2 and the overall R2.The within R2 gives the share of within-farm variation that is explained by the model.The between R2 gives the share of variation between farms that is explained by the model.FE=fixed effects model,RE=random effects model.(1)Means it used the sigmamore option.
在0.10的显著性水平上，NDD在母猪年提供仔猪数量、母猪动保成本、母猪总成本、母猪总收益和农场主每头母猪的收入模型中的系数均不显著。总的来说，没有发现NDD(降低成本或增加收入、生产或收入)与生产性能增长的显著关联。
The coefficients of NDD in the delivered piglets per sow,animal health costs per sow,the total costs per sow,total revenue per sow and farmer income per sow models were all not significant at a significance level of 0.10.Overall,no significant performance-increasing association of NDD(decreasing in costs or increasing in revenues,production or income)were found.
所有模型的Wald检验或f检验均拒绝了模型中所有系数的零假设(显著水平为0.05)，与仅有截距的模型相比，该模型没有额外的解释力。正如预期的那样，输入和输出混杂变量对总成本、总收益和收入(即、每头仔猪所需饲料、仔猪价格、每头母猪所需饲料、每头母猪饲料价格、对每头母猪的工作时间)的影响非常大。此外，相比每头母猪提供的仔猪数量和动保成本，农场现代化程度和农场生产业绩之间的关联具有统计学意义(显著水平为0.01)。猪舍建筑的现代化程度越高，相应的动保成本就越低。
The Wald or F-tests of all the models rejected the null hypothesis(with a significant level of 0.05)of all coefficients in the model jointly having no additional explanatory power compared to an intercept-only model.As expected,input and output confounding variables had a large influence on overall costs,revenues and income(i.e.,piglet feed per piglet,piglet price,sow feed per sow,sow feed price and labour hour per sow).Furthermore,the association between modernity of the farm and farm performance was statistically significant with respect to delivered piglets per sow and animal health costs(at a significance level of 0.01).More modern buildings were associated with less animal health costs.
其他混淆变量在任何模型中都不显著或几乎不显著。这些变量包括农场主的最大年龄和和母猪的数量等。用这些不显著变量重新估计模型几乎不会影响NDD的系数估计。
Other confounding variables were not or hardly significant in any model.These variables included for example the age of the oldest farmer present and number of sows.Re-estimating models with these insignificant variables hardly affected coefficient estimates for NDD.
炎炎夏日，送上猪舍省钱、降温、效
提高仔猪存活率不难，就做好这5点！
猪场驱虫如何实现简单、更高效？
教你如何给产房母猪降温！
母猪产仔率提高了，仔猪的成活率却
脚下有泥，心中有光 | 牧原农艺师
携手共赢，友谊长存：江西正邦作物保
国新办举行7月份国民经济运行情况新
农业农村部就2022年“三夏”生产形势