Epidemiological factors and mitigation measures influencing production losses in cattle due to bovine viral diarrhoea virus infection: A meta‐analysis

Abstract Infection with bovine viral diarrhoea virus (BVDV) is associated with a loss in productivity in cattle farms. Determining which factors influence monetary losses due to BVDV could facilitate the implementation of mitigation measures to reduce the burden of BVDV. Mixed‐effect meta‐analysis models were performed to estimate the extent to which the costs of mean annual BVDV production losses per animal may be influenced by epidemiological factors such as BVDV introduction risk, initial prevalence, viral circulation intensity and circulation duration (trial 1). Additionally, changes in mean annual BVDV production losses per animal due to specific mitigation measures (i.e., biosecurity, vaccination, testing and culling, cattle introduction or contact with neighbouring cattle herds) were analysed (trial 2). In total, 19 studies were included in the meta‐analysis to assess mean annual BVDV production losses. The mean annual direct losses were determined to be €42.14 per animal (trial 1). The multivariate meta‐regression showed that four of the previously mentioned epidemiological factors significantly influenced the mean annual BVDV production losses per animal. Indeed, the per animal costs increased to €67.19 when these four factors (trial 1) were considered as “high or moderate” compared to “low”. The meta‐regression analysis revealed that implementation of vaccination and biosecurity measures were associated with an 8%–12% and 28%–29% decrease in BVDV production losses on average, respectively, when simulated herds were compared with or without such mitigation measures (trial 2). This reduction of mean annual BVDV production losses per animal due to mitigation measures was partially counteracted when farmers brought new cattle on to farm or allowed contact with neighbouring cattle herds. The influencing mitigation factors presented here could help to guide farmers in their decision to implement mitigation strategies for the control of BVDV at farm level.


Bovine viral diarrhoea virus (BVDV) is a Pestivirus related to both
border disease virus (BDV) and the causative agent of classical swine fever (CSF). BVDV infections have been detected in 88 countries worldwide (Richter et al., 2019) and represent an important infectious disease in the global cattle population Scharnböck et al., 2018). Infection causes substantial costs for farmers through increased production losses and mitigation expenditures. Worldwide BVDV production losses have been estimated to be up to 687.80 US dollars (USD) per animal .
Depending on the time and duration of infection, BVDV can cause a considerable number of direct losses, such as morbidity and mortality due to immunosuppression, reduced reproductive performance (e.g., first service conception, extended calving intervals), stillbirth and abortion, congenital deformities and malformations, growth retardation, reduced milk production and average daily weight gain (Burgstaller et al., 2016;Houe, 1999;Marschik et al., 2018;Richter et al., 2017). Mitigation measures may comprise (a) preventing BVDV transmission by control of cattle trade such as testing of cattle before movements and/or reduced replacement rate of cattle possibly carrying persistently infected (PI) foetuses (Houe, Lindberg, & Moennig, 2006), (b) application of a vaccine, (c) biosecurity strategies such as cleaning of equipment, protective clothing, double fencing and (d) general prevention of contact with potential PI animals (Evans et al., 2019) as well as (e) testing and culling to eradicate BVDV. The economic impacts of BVDV for cattle farms have prompted many countries to implement mitigation programmes and the success of these programmes on the reduction of BVDV prevalences in the global cattle population has been reported elsewhere (Scharnböck et al., 2018).
Determination of epidemiological and mitigation influencing factors on BVDV production losses can facilitate the implementation of control and prevention activities regarding BVDV at farm level.
The aim of this study was to analyse the extent to which epidemiological factors (e.g., BVDV introduction risk, initial prevalence, viral circulation intensity and BVDV circulation duration) and mitigation measures (e.g., biosecurity, vaccination, testing and culling, cattle introduction or contact with neighbouring cattle herds) may influence the ex-ante and ex-post estimated monetary level of production losses due to BVDV infections in the cattle population from the literature.

| Standardization of collected data
In the present study, we analysed BVDV production losses and covariate data in the period from 1960 to 2015, provided by  and Richter et al., (2017). The literature research was extended to the period from January 2015 to June 2018 with the following predefined search terms ((bovine viral diarrh* OR bovine virus diarrh* OR bvd OR bvdv) AND (economic* OR financial OR cost*)) in order to identify the greatest possible number of recent publications concerned with the monetary level of BVDV production losses. A search for articles was performed in PubMed, ISI Web of Knowledge, and Scopus. Studies were included in the meta-analysis if the following criteria were met: (a) studies estimated production losses due to BVDV infections in monetary terms at farm, regional or national level; (b) studies analysed production losses in cattle, i.e., dairy and/or beef (without beef finisher); (c) studies should have published mean (average) annual production losses per animal which covered losses for all cattle in the associated population (even if the animals were not affected, as by Stott et al., (2012)). N.B. no restrictions were defined on the level of monetary losses due to BVDV infections. If studies reported initially losses of only infected cattle, as by Marschik et al., (2018), these losses should be transferable to all cattle in the associated population (infected and uninfected). Consequently, population characteristics such as number of cattle, herd size and in case that losses were published over multiple years, also the time periods of assessments, were essential for the standardization of the losses per animal and per year. Studies for which such standardization was not possible were excluded from further analysis; (d) additionally, studies were included which analysed at least one of the following mitigation strategy: biosecurity, vaccination, testing and culling, cattle introduction or contact with neighbouring cattle herds ( Figure 1). N.B. the term "animal" in the presented work covered dairy and/or beef production systems (without beef finisher). The total number of identified publications and the applied two-step selection process for eligible studies performed in accordance with the PRISMA guidelines (Preferred Reporting Items for Systematic Reviews and Meta-Analysis), are illustrated in Figure 1. All articles were screened in full by two reviewers (SG, DR) and eligible studies, i.e., which met the inclusion criteria, were then reviewed in full by one reviewer (SG) in accordance with the predefined variables shown in Table 1 Table S1 and in the Forest plots). N.B. we considered the fact that these two results from the same study are obtained in closer conditions than results from two separate studies by including "publication" as random factor in our meta-analysis (see following section: meta-analysis). The total number of publications included in the presented study is thus not identical to the total number of observations. Meta-analysis is used to detect effects across studies by analysing factors (covariate [i.e., independent variable of direct interest such as mitigation measures] or moderators [i.e., study-level covariates such as characteristics of the study or population]) that may influence the effect. All factors from Table 1 were analysed in the meta-analysis. However, some factors were only reported in part of the publication (e.g., approximately half of the studies provided PI prevalence values, see Table S1-S2) or the way in which the factors were presented were highly variable among the different publications (e.g., the infected status at the beginning of the studies may have covered seronegative, seropositive, transient infected, persistent infected or immune herds), which may cause bias in the F I G U R E 1 Flow chart of studies incorporated in the systematic review and meta-analysis TA B L E 1 Analysed influencing factors on the estimated mean annual BVDV production losses per animal and summarizing some of these variables into "new built" factors  Table 1) based on a categorical scale (low, moderate, high) were additionally created, as a combination of data originating from the incorporated studies. An assignment of factors incorporated in the "new build" factors is provided in Table 1, a detailed example of the variable construction is provided in Figure S1 and the associated data are provided in Table S1 and S2. Both the recorded raw data from the literature and the "new build" factors were analysed in the meta-analysis.
The production losses reported in the literature were standard- The index I_OCDEx includes the economic annual growth rate of the respective country and incorporates the inflation rate based on the consumer price index.

| Meta-analysis
The outcome variable in trial 1, which included the situation before any mitigation measure had been taken, is presented by the mean annual production losses per animal. The general and epidemiological factors (Table 1) were used as explanatory variables for the mean annual production losses per animal in the meta-regression analysis. In trial 2, the outcome variable is shown as the percentage decrease in the mean annual production losses per animal. Here, the effect of the recorded mitigation measures (Table 1) on the changes in BVDV production losses from the literature was investigated and the changes were expressed as the percentage difference of the production losses before and after implemented mitigation measures.
Trials 1 and 2 were performed independently in a random-effect meta-analysis model (without factors from Table 1 but with publication as random factor, i.e. if multiple data points were collected for a variable from one study (repeated measures), "publication" might be thought as random factor and is analogous to random effects in classical ANOVA) and mixed-effects model (with factors from Table 1 and with publication as random factor). In the first step, the heterogeneity of the incorporated studies in the meta-analysis was determined as follows: (a) calculating the percentage of total variation across the studies by estimation of the Higgins inverse variance (I 2 ) index (lay between 0% and 100%), whereby I 2 greater than 50% indicated substantial heterogeneity between studies and (b) calculating the degree of between study variance, i.e., the Cochran's Q-Test where p < .05 indicated heterogeneity. The limitation of I 2 and Cochran's Q-Test is that both provide only a value of the total heterogeneity between the considered studies in the meta-analysis but no information about the factors which causing the potentially heterogeneity. If evidence of high variability between studies was determined, the next step was to perform a meta-regression, i.e., quantification of heterogeneity in effect size among studies by including factors (covariate or moderators), referred to as mixed-effect model (Viechtbauer, 2010 Descriptive studies were also included because data about annual production losses per animal and epidemiological and/or mitigation covariates were provided. c Whether the factor was included in the meta-analysis is indicated with Y = Yes; N = No; and if the factor was included in the "build used" factor it is indicated with brackets and the abbreviation of the associated build used factor. TA B L E 1 (Continued) model (Bursac, Gauss, Williams, & Hosmer, 2008), followed by removing correlated and non-significant factors from the multivariate model, identification of significant factors combination that reduce the Akaike Information Criteria (AIC) and increasing the value of r square (R 2 ), as well as reducing the heterogeneity between the included studies in the meta-analysis. The τ 2 (residual heterogeneity variance) denoted the amount of the heterogeneity that may have not explained through the inclusion of the factors in the meta-analysis. A reference class for each factor was chosen to allow a comparison of the effect size ( Computing) using the Metafor package (Viechtbauer, 2010).

| Trial 1
In total, 20 studies were included in the meta-analysis. Trials 1 and 2 included 19 and 6 publications with 83 and 87 observations, respectively. The influential case diagnostic of trial 1 indicated three observations and one study as sources of asymmetry ( Figure S2; Table S3). These observations were considerably higher regarding production losses (with a mean of €215) compared to other observations (with a mean of €40) and thus highly influence the results of the meta-regression. Consequently, three observations and one study from trial 1 were excluded in the present metaanalysis ( Figure 1). The funnel plot with 80 observations covering 18 studies (trial 1) did not show any asymmetry issues for the incorporated studies (Figure 2), despite the existence of many annual production losses per study is shown in Figure 3. In the univariate meta-regressions, BVDV production losses were associated with the factors production system, BVDV introduction risk, initial prevalence, viral circulation intensity and circulation duration. The other general and epidemiological variables in Table 1 were not statistically associated with mean annual BVDV production losses per animal. In the multivariate mixed-effect regression (trial 1), the mean annual BVDV production losses per animal were also significantly associated with the BVDV introduction risk, initial prevalence, viral circulation intensity and circulation duration (

| Trial 2
In trial 2, the funnel plot did not show any asymmetry issues for the incorporated studies (Figure 2), despite a few observations having  Figure) very high standard errors. Specifically, the influential case diagnostic indicated 2 outliers for both mitigation models (i.e., with cattle introduction or with contact with neighbouring cattle herds; Figure S3-S4) but exclusion did not change the significant associations or the coefficient observed in the final meta-regressions, which have been kept as the final ones. A publication bias was reported by the Egger's test (z = 6.4536, p < .0001). The heterogeneity of the dataset was high (I 2 = 98.55%; AIC = 2,179; Q-test: x 2 = 2,357; df = 80; p < .001).
In the univariate meta-regressions, the mitigation factors (i.e., biosecurity, vaccination, testing and culling and contact with neighbouring cattle herds) were statistically associated with BVDV production losses. In the multivariate mixed-effect meta-regressions, biosecurity, vaccination and cattle introduction or contact with neighbouring cattle herds were significantly associated with the change in the BVDV production losses (Table 2), whereas testing and culling was not identified as a significant factor. Implementation of vaccination and biosecurity measures were on average associated with an 8%-12% and 28%-29% decrease in BVDV production losses, respectively, when simulated herds from the literature were compared with or without such mitigation measures. This reduction of BVDV production losses per animal due to mitigation measures was partially counteracted on average by 18% when farmers brought new cattle on to farm (cattle introduction) or allowed contact with neighbouring cattle herds ( Table 2). The percentage difference of the production losses before and after implemented mitigation measures per study is shown in the Figure S5-S6.

| D ISCUSS I ON
In order to analyse previously published studies with a specific emphasis on production losses incurred by BVDV infection, we re- Over all studies from the literature (see Table S4), we calculated a mean annual BVDV production loss of €42.14 per animal (without factors; random-effect meta-analysis model). This estimated mean annual production loss per animal covered infected and uninfected animals and is probably lower than if only losses from infected cattle would be taken into account. Our meta-analysis demonstrated that studies which assumed a high BVDV introduction risk, initial prevalence, viral circulation intensity and circulation duration, compared to low, significantly increased the mean annual BVDV production losses to €67.19 per animal (Table 2). For instance, in the present work, it was shown that across all studies incorporated in the metaanalysis, the mean annual BVDV production losses per animal were €34.33 higher in cattle herds with a high simulated introduction risk compared to herds with a low risk (  et al., (2012), identified twice higher mean annual BVDV production losses per animal in dairy than in beef herds. Nonetheless, over all incorporated studies in the meta-analysis, the factor "production system" was not determined to be a significant factor on mean annual BVDV production losses per animal. One reason could be that some of the considered studies in this meta-analysis estimate production losses for beef herds based on observations in dairy herds, when no observation for beef herds were available (e.g., in the study by Valle et al., (2005)).
The present study confirms the success of mitigation activities with regard to reduction in mean annual BVDV production losses per animal. Mean annual BVDV direct losses per animal were 8%-12% and 28%-29% lower in studies including vaccination and biosecurity as compared to studies omitting these mitigation measures, respectively (Table 2). This result is in agreement with the meta-analysis by Newcomer, Walz, Givens, and Wilson (2015). The study reveals that abortion decreased by 45% and the foetal infection rate decreased by approximately 85% in cattle herds vaccinated against BVDV compared with non-vaccinated herds (Newcomer et al., 2015). In contrast to vaccination, biosecurity reduces BVDV production losses more effectively (Table 2). This may be related to the fact that farmers often fail to apply the vaccine correctly, vaccines are not proven to be fully protective (Evans et al., 2019), e.g., in the prevention of in-utero transmission of the virus (Moennig & Brownlie, 2001), the BVDV vaccine does not provide life-long immunity and hence periodic vaccination is required (Weldegebriel, Gunn, & Stott, 2009), live BVDV vaccine could be contaminated with other viruses (Lindberg, 2003), and/or a critical vaccination coverage rate should be reached to prevent new PI animals (Scharnböck et al., 2018). In the present multivariate-regression, testing and culling was not identified as a significant factor in changing the mean annual production losses per animal due to BVDV infection. The fact that a low number of studies incorporated in the present study have analysed the effect of culling strategies on BVDV production losses (n = 26 out of 87 observations; see Table 1) may have contributed to culling measures being not statistically significant. Another reason for this is that Pasman, Dijkhuizen, and Wentink (1994) showed that a testing The difficulty faced in reducing heterogeneity in the present work suggests that other covariates not taken into account may contribute to the heterogeneity and identified outliers. For instance, the herd immunity, improvements of the breeding performance over the time, period of gestation, management practices, age of animals, duration of mitigation activities, different BVDV status, level of herd production, stocking density, community pasturing activities, case-selection procedure, virulence of the infecting BVDV genotype or strain (Hessman et al., 2009;Houe, 1999;Scharnböck et al., 2018). Different modelling approaches, study assumptions, input parameters and the unbalanced number of studies identified for some factors may have also contributed to the high heterogeneity between the studies presented here. A range of different methods were applied in the literature, such as stochastic simulation models (in 12 of the 19 studies), deterministic models (n = 4) such as decision trees, and other methods (n = 3). In total, five studies (Stott & Gunn, 2008;Stott, Humphry, & Gunn, 2010;Stott et al., 2012;Stott, Lloyd, Humphry, & Gunn, 2003;Weldegebriel et al., 2009) used the same stochastic simulation approach, developed by Gunn, Stott, and Humphry (2004). All of these studies considered four disease states (susceptible, transiently infected, immune and/or persistent infection), constant herd size with animal movements (replacement or death), naïve herds at the beginning of the simulation, initial source of BVDV introduction due to contact with neighbouring herd or introduction of BVDV infected animals, and different annual transmission rates, such as the probability assumed of infectious contact between a susceptible and a PI animal.
Although the modelling approach and the estimated input parameters used in these five studies are closely followed that of Gunn et al., (2004), the following model modifications were incorporated which may influence the apparent spread of the animal disease and the economic impact of BVDV: constant herd size varied between all studies (ranged from 14 to 230 head) which may influence the basis reproduction number (R 0 ); Stott et al., (2003) and Stott and Gunn (2008) considered not only naïve herds at the beginning of the simulation but also herds with unknown BVDV status. This modification lead to the simplification in which animals were allocated F I G U R E 3 Forest plot of the meta-analysis models including the significant epidemiological factors ("build used" factors of Table 1). The column on the right refers to the mean annual BVDV production losses per animal with the corresponding confidence intervals (shown in brackets). The different single numbers attached before the authors' names (left column) represent the "build used" factors of Table 1. The numbers represent the following scores, i.e., 1 = low; 2 = moderate and 3 = high. The order of the numbers can be classified as follows: first number covered BVDV introduction risk, followed by initial prevalence, circulation intensity and duration (see also Table S1). N.B. the forest plot may include the same combination of numbers more than one before the authors' names within a study because different sets of input parameters were used, resulting in different estimated mean annual production losses per animal. The grey diamonds represent the effect size adjusted for the "build used" epidemiological factors. Forest plots of mitigation measures are provided in Supplementary Figure S5-S6. N.B. Full references of authors shown in the forest plots are available in Supplementary Table S4 randomly to the disease states to fulfil the assumed fixed antibody (positive) prevalence at a herd level of 0.95 and antigen (positive) prevalence of 0.50. Furthermore, the five studies considered the probability of successfully avoiding contact instead of the probability of infectious contact between a susceptible and a PI animal.
The rate of avoiding infection taken into account in the literature differs slightly, e.g. Stott et al., (2010) incorporated a larger range of probabilities regarding the contact aversion between the animals as compared to Stott et al., (2003). Additionally, the mitigation options, i.e. biosecurity and/or vaccination (vaccination efficacy ranged from 60% to 90%) were considered by Gunn (2008) andStott et al., (2012), while Gunn et al., (2004) assume no mitigation and no re-infection with the disease. The latter differs also from the work by Stott et al., (2010), who analysed the potential impact of re-infection on the BVDV production losses. Weldegebriel et al., (2009) andStott et al., (2012) adapted the model by Gunn et al., (2004) to suit dairy herds rather than only beef suckler herds and therefore the input parameters varied regarding the assumed replacement rate (changed from 15% to 30%), model time steps (changed from yearly to quarterly period in order to reflect the seasonal milk production cycle), prevalences, economic parameters and the slightly different probability of biosecurity breakdown in any year of the 10-year simulated epidemic. All these differences between the original study by Gunn et al., (2004) and the other five studies may explain the wide range of mean annual production losses per animal (ranging from €2.50 to €69.00; mean: €29.19) reported.
However, when these six studies were compared with the other incorporated studies (n = 13, see Table S4) which do not use the stochastic model by Gunn et al., (2004), the following differences were identified: the mean annual production losses per animal was approximately €7.00 higher; different ex-ante or ex-post methods were used, while the studies based on the model by Gunn et al., (2004) taken largely constant epidemiology and economic values into account. Further, the majority of the studies assume different levels of discounting rates and the higher the discounting rate, the lower the level of the current production losses in the studies. The six studies based on the model by Gunn et al., (2004) took immunosuppression into account, whereas the other studies (such as the study by Thomann et al., (2017) and Marschik et al., (2018)) did not use the BVDV model by Gunn et al., (2004) neglected it. Further difference between applied BVDV models are given in the review article by Viet, Fourichon, and Seegers (2007) data from numerous studies and countries is helpful as it provides a more general overview of the influencing factors and is more powerful and less biased than any individual study or conventional methods (Gurevitch et al., 2018;Scharnböck et al., 2018). Thus, the results of the presented study could be used to increase awareness of factors influencing mean annual BVDV production losses per animal and to support decision-making by farmers and veterinary authorities implementing mitigation measures such as biosecurity measures or control of cattle introduction against BVDV. The latter is particularly essential for cattle owners located in BVDV-free regions to guarantee their freedom from an animal disease that is not globally regulated. BVDV-free countries with a fully susceptible population will have a higher BVDV introduction risk and thus higher impact on production compared to countries with a high proportion of seropositive cattle. Nonetheless, the monetary benefit of mitigation strategies, such as biosecurity, will be highly variable due to (a) wide range of determined annual production losses per animal (

| CON CLUS IONS
The mean annual production losses due to BVDV infection was found to be €42.14 per animal. The costs increased to €67.19 when the BVDV introduction risk, initial prevalence, viral circulation intensity and circulation duration were "high or moderate" compared to "low." Our results reveal that the implementation of vaccination and biosecurity measures was associated with an 8%-12% and 28%-29% decrease in mean annual BVDV production losses on average, respectively, when simulated herds with or without such mitigation measures were compared. This reduction of BVDV production losses per animal due to mitigation measures was partially counteracted on average by 18% when farmers introduced new cattle into a farm or allowed contact with neighbouring cattle herds.

ACK N OWLED G EM ENTS
This work was supported by the Project VET-Austria, a cooperation between the Austrian Federal Ministry of Health, the Austrian Agency for Health and Food Safety and the University of Veterinary Medicine Vienna.

CO N FLI C T O F I NTE R E S T
None of the authors of this paper has a financial or personal relationship with other people or organizations that could inappropriately influence or bias the content of the paper.

E TH I C A L A PPROVA L
Ethical Statement is not applicable because the manuscript is a systematic review of the literature.