A case-control study was conducted as described by Seyboldt et al.  and Jensen et al.  Cases were defined to fulfill at least three of the following five criteria: reduced milk yield (> 15% for at least three months compared to the milk yield of the year before), increased mortality (> 5% of the herd during the last year), increased culling rate (> 35% of the herd during the last year or an increase of > 10% compared to the year before), increased number of downer cows (> 10% of the herd during the last year) and farmers´ or veterinarians´ impression of herd health problems. The controls did not fulfill any of these criteria. All farms were located in the northwest of Germany (Lower Saxony, Schleswig-Holstein, and Northern part of North Rhine-Westphalia). In addition, all participating farms had a loose housing system for lactating cows, minimum herd size of 30 cows and were participating in Dairy Herd Improvement (DHI) milk tests.
Based on the sample size of 46 case and 46 control farms, an odds ratio of ≥4 was detectable (confidence 95%, power ≥ 80%, prevalence of controls 50%; calculated using NCSS Pass®).
All farms were visited once by a team of four research veterinarians who were trained with regard to the examination and data collection processes. During the farm visit, they scored the herd for body condition, hygiene, skin lesions and lameness; interviewed the farmers regarding herd health, management and diet composition; checked the housing conditions; assessed feedstuff; and examined five cows with obvious chronic conditions as well as five cows without obvious conditions. These ten cows were selected in accordance to defined eligibility criteria . If the five cows in a chronically sick condition showed lameness, they were examined in a claw trimming chute. In addition, silage, blood, feces, and bulk milk samples were taken. For all of these procedures, the four observers were trained prior to and during data collection. Standard operating procedures were used (SOPs; see Additional file 1: definition of risk factors). Different sections of data were collected by observers interchangeably. Inter-observer-reliability was not evaluated and observer effect was not considered during risk factor analyses. This was due to the a-priori training, usage of SOP’s and permanent training and supervision of the whole observer group by three different senior supervisors. Furthermore, a potential observer effect would not have affected data analyses due to the interchange between data collection parts and the fact that case and control farms were investigated by the same group of study vets, who visited every farm with a different composition of team members.
In addition to the evaluated risk factors, the three following confounders were studied: herd size (quantitative), season during which the farm visit took place (summer: May–October; winter: November–April), and access to a pasture (yes, at least seasonally; no, not at all). Descriptive statistical analyses, as well as single and multifactorial regression analyses, were utilized to assess the association of these confounding variables with case-control status.
Although the study region was chosen to reach a homogeneous study population with similar farm structures  and the definition of further eligibility criteria, structural differences were found: Slightly more case than control farms were visited during summer (Table 1). Case farms had fewer cows than control farms (Table 2) and cows from case farms more often had access to pastures (Table 1). These findings indicate a more extensive management system in case farms as compared to control farms. This is consistent with DHI data from Schleswig-Holstein, where larger farms had a lower culling rate and lower mortality . The confounders did not show a significant impact in the multi-factorial modelling.
The study veterinarians were asked, what they think which risk factors contribute to the fulfilment of the inclusion criteria on case farms. Based on their answers, four areas with a varying number of risk factors were identified, such as health management (including the sub-areas of infectious diseases and claw health), housing (including the sub-areas of stocking density, dimensions of cubicles, comfort of cubicles, and floors), hygiene, and nutrition (including the sub-areas of feeding management, silage quality, energy density, quantity of roughage, and crude fiber). Risk factors were aggregated at the farm level. An overview of each of the variables investigated is given in the following passages. More detailed definitions of the risk factors and references are provided in the Additional file 1 (definition of risk factors).
For the detection of liver flukes, lungworms and intestinal worms, feces samples from the ten cows that were examined clinically were tested for eggs via flotation, separately. In addition, a bulk milk sample was checked for antibodies against liver flukes (IDEXX©). For the detection of lungworms, serum samples of the ten examined cows were tested for antibodies. For the detection of MAP, feces samples from the five cows that were in a poor condition and five cows that were in good condition were pooled separately and examined via microbial culture. A farm was considered positive when at least one result from at least one sample was positive. The laboratory analyses were performed by different commercial service providers.
With regard to claw health, the frequency of herd claw trimming (quarterly or more often, every 6 months, longer than every 6 months or irregularly) was evaluated in the analyses. In addition, the number out of the ten examined cows with poor claw condition (no cows, one cow, more than one cow) was recorded, and whether high-grade dermatitis digitalis was found on at least one claw of the examined cows that showed lameness was also included in the statistical analyses (yes or no).
To evaluate the stocking density, the average ratio of the numbers of cows in the pen per cubicle (≤1 = no overcrowding; > 1 = overcrowding), feeding spaces and watering places (< 1 = no overcrowding, 1.01–1.5 = moderate overcrowding; > 1.5 = severe overcrowding), were calculated across all pens with lactating or dry cows on the farm (disregarding calving pens or pens for sick cows). In the case of absent feeding fences, one feeding space was defined as 0.75 m of the feed alley . To calculate the watering space, a cup drinker was assumed to be sufficient for eight cows. In the case of trough watering, a length of 8 cm was defined as one watering place .
To assess the comfort of cubicles, the number of pens with raised cubicles were counted (no pen, at least one pen but not all pens, all pens). It was also noted whether there was a pen without rubber mats or bedding material (no pen, at least one pen).
To evaluate the dimension of the cubicles, the width of cubicles (> 120 cm; yes or no), average height of neck rails (> 115 cm; yes or no), and average distance from neck rail to curb (> 195 cm; yes or no) were measured at four randomly chosen cubicles in every pen with lactating or dry cows . Normally, the fourth and the fourth-to-last of the cubicles of the row next to the wall, the fourth-to-last cubicle of the middle row and the fourth cubicle of the row next to the feeding fence were measured. Firstly, the mean of cubicle sizes was calculated at pen level. Secondly, the mean of all pens with lactating or dry cows was calculated to aggregate the data at the farm level and was compared to recommendations mentioned above.
In addition, the percentage of pens with slippery floors was assessed (no pen, 1–50% of the pens, more than 50% of the pens) as well as whether or not at least one pen had damaged floors (no pen, at least one pen with damaged floors).
The percentage of pens with dirty or very dirty floors (< 50% of the pens, 50–99% of the pens, 100% of the pens) and dirty or very dirty lying areas (no pen, at least one pen, but not all pens, all pens) was calculated and included in the analyses.
To assess feeding management, the frequency of daily feed delivery and frequency of pushing the feed back to the fence for early lactating cows (first 100 days after parturition) were included in the analyses based on farmers´ statements (see Additional file 2).
Silage quality was investigated whether or not at least one silage fed to lactating or dry cows was considered to be below current recommendations for sensory status (decomposition, loss of structure or high-grade mildewed; yes or no) assessed by the study veterinarians, crude ash content in grass silages (> 8% of dry matter; yes or no), true protein content (grass silage < 50% true protein of crude protein content; yes or no), dry matter content (grass silage: < 30% or > 40% or corn silage: < 28% or > 35%; yes or no), pH-value (grass silage: > 4.7 or corn silage: > 4.2; yes or no), and microbiological deviations (assessment based on recommendations by VDLUFA ; at least one silage with profound variation; yes or no). The analyses of the silages concerning the ingredients and the microbiological status were performed by an accredited service provider.
During the interview the farmer was asked for the composition of the diet for fresh lactating cows. Diets were calculated based on farmers´ statements using Futter R® (dsp agrosoft). For the silages, the results of the laboratory analyses of the sample taken at the farm visit were used. The declaration of concentrates and supplements was assumed as stated at the product or its delivery receipt . The energy density in the roughage diets (silage, hay, straw) was calculated as composite in the diet for early lactating cows. In addition, the energy density in the whole diet (with concentrates and other feedstuff) for early lactating cows was calculated. Both variables were measured as net energy content for lactation (MJ NEL) per kilogram of dry matter (DM). Additionally, the quantity of fed roughage (kilogram of DM per cow per day; quantitative) for early lactating cows was included in the analysis.
With regard to the potential lack of crude fiber, the ratio of crude fiber within the diet [< 16% for TMR (total mixed ration), < 18% for PMR (partial mixed ration; crude fiber was regarded in the fed ration without individual concentrate supply); yes or no] and ratio of roughage to the whole diet (%; quantitative) were calculated for early lactating cows. Additionally, the percentage of cows in the herd with a fat content < 3% in milk (< 3%; 3–5% or > 5% of the herd) and a fat-protein-quotient < 1 (%; quantitative) of the last DHI milk recording before the farm visit were evaluated.
Statistical analyses were performed as described in detail by Jensen et al. . After entry into a relational SQL online study database, all analyses were conducted using SAS 9.3® (SAS Institute Inc., Cary, NC, USA). Data were checked for plausibility and missing values. Variables were aggregated at the farm level (statistical unit) as described above and in the Additional file 1 (definition of risk factors). Overall, only nine data points were missing, indicating excellent data quality.
First, a descriptive analysis was performed stratified by case- and control-status. Then, the linearity of the relationship between the quantitative variables and the logit of the case control status was evaluated. Linearity was confirmed graphically using R®, version 3.1.1 (R Foundation for Statistical Computing, Vienna, Austria). Two variables (ratio of roughage to the whole ration for early lactating cows and quantity of fed roughage) had a quadratic relationship to the logit of the health status. The quadratic terms of these two variables were included in the statistical analyses. If no quadratic or linear relationship was found, the variables were categorized. Associations among risk factors were investigated using Cramer’s V (cut-off: 0.7), Spearman’s rank correlation coefficient (cut-off: |0.8|) or analyses of variance (cut-off for coefficient of determination: 0.64). No association between risk factors was beyond these cut-off values. Therefore, no risk factor was excluded from further analyses. After the tests for association among the risk factors, a single-factorial logistic regression was performed. Variables with P < 0.2 were included in a multifactorial logistic regression analysis. To achieve an informative model, variables in the multifactorial model were excluded using stepwise backward selection, if the corresponding P value was greater than 0.05. The correlation matrix of the predictors was investigated to review the associations in the final statistical models. Two-way interactions among the risk factors were included in the backward-selected model and checked for statistical significance with P < 0.1. After backward selection of the interactions, no interactions with P < 0.1 remained in the model.
ROC curves were computed for the multifactorial model assessing the performance of the model. Due to the explorative nature of this study, a multiplicity correction was omitted .