This project was approved by the Animal Care and Use Committee of the Faculty of Veterinary Medicine (University de Montréal; protocol 17-Rech-1886). This paper was elaborated using the “STROBE-Vet statement” guidelines (12).

A list of farms using RMS was generated by contacting RMS equipment dealers and veterinarians and through social media. Straw farms were recruited with the help of Québec Dairy Herd Improvement Association (DHIA, Lactanet, Ste-Anne de Bellevue, QC, Canada). For both type of herds, to be eligible, farmers needed to be located within 250 km of the research facilities (St-Hyacinthe, QC), to have used the same bedding for >6 months prior to the farm visit, and for the straw farms, to be enrolled in a DHIA milk recording program. This latter condition was added for another part of the study on subclinical mastitis. We aimed at recruiting ~90 farms. This number was determined by an a priori power estimation. We estimated that, using 90 herds (20%: RMS bedding vs. 80%: straw bedding), milking an average of 50 cows, and a baseline probability of clinical mastitis of 20%, we would have >95% power to detect a difference of probability of mastitis corresponding to an odds ratio ≥1.4. For this calculation, we did not account for clustering of cows by herd. Thus, the real power is likely to be smaller than 95%.

All potential farms were contacted by telephone between July and December 2017 to verify their eligibility and willingness to participate. Basic demographic information such as the number of milking cows in the herd, type of bedding used, and for RMS farms, which equipment they used were also gathered from herds that were excluded, to study their similarities with participants and assess the presence of a selection bias.

Farm visits were described elsewhere (6, 11). Briefly, farms were visited once and producers had to answer a standardized questionnaire about their bedding management. Methods used to produce bedding were recorded and bedding samples as well as bulk tank milk samples were collected. Herd size was recorded as the number of lactating cows. We also recorded housing type (free or tie stall), time since the last renovations of the stalls (in years), and bedding thickness defined as shallow bedding (<10 cm of depth) or deep bedding (≥10 cm of depth). These latter covariables were pre-identified using directed acyclic graph as putative confounders of the association between bedding type and CM incidence.

During 1 year following the initial visit, producers were asked to sample aseptically each quarter of cows experiencing a CM. Farmers had to provide information regarding the identification of the cow, its parity, the position of the quarter affected, and the severity of the CM. For the latter, farmers had to categorize CM events as score 1 (abnormal milk only), score 2 (abnormal milk and udder, without systemic signs), or score 3 (systemic signs of illness such as fever, depression, and anorexia) as described by Sears and McCarthy (13). Two consecutives cases of CM in the same quarter of a cow were considered distinct if they were ≥8 days apart (14). Samples were sent on ice to the Faculty of Veterinary Medicine Laboratory (St-Hyacinthe, QC, Canada). The bacteriological analyses were realized following the National Mastitis Council guidelines (15). Briefly, 0.01 ml of milk was plated on blood agar and incubated for 24–48 h at 35°C. The sample was then classified as negative (no growth), pure intramammary infection (IMI) (one single phenotype of CFU), mixed IMI (two types of CFU), or contaminated (≥3 types of CFU). An IMI was defined as the isolation of ≥100 CFU/ml of a given phenotype. Pure and mixed IMI bacterial isolates where then identified by mass spectrometry (MALDI-TOF) using the database of the manufacturer (BDAL-8468) and a custom database validated specifically for staphylococci identification (16). Isolates needed to be identified to the species level (vs. genus level solely) to be retained for pathogen-specific statistical analyses.

The number of CM episodes was compiled for each farm, as well as the number of severe (score 3) CM episodes. Finally, we compiled the CM episodes by specific pathogens. To account for the varying herd size and the exact time period of follow-up, the number of milking cows in each herd and the length of the follow-up period were also compiled.

Most CM studies have to deal with different levels of the compliance of producers for reporting CM cases and/or submitting samples. To investigate this potential bias, we used two different approaches to estimate the herd animal-time denominator used to adjust CM incidence. First, we used a common approach which is the exact number of milking cows and the exact period of follow-up (17) to compute the number of animal-year at risk of the herd.

Then, as a sensitivity analysis, we also estimated the follow-up period using the interval between the first and last sampling dates as the definition. Thus, with this alternative method, farms who did not send any samples or that sent only one sample during the 1 year study period were excluded (i.e., they would contribute 0 animal-year at risk). Moreover, farms that may have sent >one sample but then stopped sending samples at some point in time would be included, but with a shorter time at risk period (i.e., only the time between the first and last sent samples would be compiled). Then, we computed the number of cow-year at risk of the herd for each farm by multiplying the number of milking cows of the herd by the follow-up time. Using this alternative method allowed to exclude producers who sent <2 samples during the study period and weighted down producers who possibly stopped sending samples during the study.

Statistical analyses were performed using SAS 9.4 (SAS Institute Inc., Cary, NC, USA). Descriptive statistics were used to explore relations between predictors. To compare the incidence of CM in RMS and straw farms, we used a binomial negative model with the number of CM cases on a given farm as the outcome (firstly total number of CM cases, then severe CM cases only, then CM cases by bacterial species), type of bedding used (RMS or straw) as the main predictor, and the natural logarithm of the number of cow-year at risk as an offset term. In this model, we also included a number of putative confounding variables as predictors: housing type (free or tie stall), time since the last renovations of the stalls (in years), bedding thickness (deep or shallow bedding), and herd size (number of milking cows). With such a model, we could thus compute the CM incidence ratio (IR) between RMS- and straw-bedded farms, after adjusting for these confounders, simply by exponentiating the bedding coefficient. Moreover, the mean estimated CM incidence (in cases/100 cow-year) for a given type of herd could be computed simply by adding the intercept and the coefficients corresponding to that farm description, then exponentiating the results and, finally, multiplying the results by 100 cows (to obtain an incidence per 100 cow-year). Finally, all models were ran twice, initially using the complete follow-up period to compute the animal-time at risk of the herd and, then, using the animal-time at risk computed using the alternative method.

The assumption of linearity of the relation between quantitative predictors (time since the last renovation of the stalls and herd size) and the outcome (logarithmic transformation of the incidence ratio) was verified with the addition of polynomial terms (square and cubic terms) after centering the predictor. If the polynomial terms were significant (p < 0.05), the polynomial presentation of the variable was retained in the final model. If overdispersion was observed in the data (Pearson chi-square > 1.2), robust variance was used. Significance level was fixed at p ≤0.05. Data and the SAS code used to construct the models are publicly available at https://doi.org/10.5683/SP2/KIEMHY.

We obtained a list of 49 RMS and 139 straw farms and recruited 27 and 61 RMS and straw farms, respectively. Reasons for exclusion of RMS farms were as follows: four had recently changed their bedding type to non-RMS bedding, 11 were outside the defined geographic location, one did not use RMS under the milking cows, and six could not be reached despite several attempts. From the 139 straw farms interested in the project, 61 were selected on their ability to provide computerized health records.

We visited the farms between January 15th and July 10th, 2018. The farms recruited have been described in Lasprilla-Mantilla et al. (6) and Gagnon et al. (11). Briefly, recruited RMS farms had 55–900 lactating cows (median 111) and straw farms had 43–229 lactating cows (median 65). An automatic milking system was in used in 37% of RMS farms and in 3% of straw farms. Furthermore, 59% of RMS farms and 98% of straw farms participated regularly in a DHI program. Even though participating in a DHI program was an inclusion criterion in straw herds, one farm did not record any data (i.e., dropped out of DHI) during the follow-up period. In the 27 RMS farms, 26 used a separation process as first step for producing RMS bedding. One RMS farm used an anaerobic digester as the first step and a separation process as the second and last step. From the 26 farms who did a separation first, one used the solid fraction immediately after separation, two used a rotative drum to turn into compost the solid fraction, 10 allowed the solid fraction to mature in a heap, and 13 allowed the solid fraction to mature in an enclosed container. During the monitoring year, two farms (one RMS and one straw) burned and five others (one RMS and four straw) dropped out of the project. Two of these farms (one RMS and one straw) dropped off very early and without sending any milk samples. They were, therefore, excluded from the analyses. For the three other farms, data collected until they left the project were used in the analyses. When we used the alternative follow-up period (from the first to the last sampling dates) to compute the animal-time at risk, 28% of RMS farms and 35% of straw-bedded farms had a follow-up time of 6 months or less. The characteristics of the general herds are reported in Table 1.

We received 1,247 samples during the study period (Figure 1). We excluded 11 samples because they were collected on the same mammary gland quarter <8 days since the last CM episode. From the 1,236 remaining samples, there were no information about CM severity in 69 (5.6%) of the samples, 92 (7.4%) were submitted as score 0 (no clinical mastitis), 492 (39.8%) as score 1, 426 (34.5%) as score 2, and 157 (12.7%) as score 3. We excluded the 92 samples for which the producers explicitly reported a severity score of 0. Samples with no reported severity were, however, retained. We observed a pure IMI in 624 (54.5%) of the 1,144 remaining samples, a mixed IMI in 192 (16.8%), and no growth in 165 (14.4%). Of the collected samples, 163 (14.2%) were considered contaminated. The proportion of contaminated samples was not associated with bedding type (chi-square test; p = 0.89).

The most frequent recovered pathogens (in pure or mixed IMI) by bedding type and severity are reported in Table 2. Briefly, in RMS farms, the most frequent pathogens were Streptococcus uberis (16.0%), Escherichia coli (13.8%), K. pneumoniae (13.2%), Streptococcus dysgalactiae (6.2%), and Staphylococcus aureus (3.4%). In straw-bedded farms, S. aureus (16.6%), S. uberis (11.0%), E. coli (9.1%), S. dysgalactiae (8.0%), and K. pneumoniae (1.1%) were the most frequent pathogens. Clinical mastitis episodes due to coliforms (K. pneumoniae and E. coli) were more often severe. Clinical mastitis episodes due to S. uberis, S. dysgalactiae, or S. aureus were, in general, mainly mild or moderate (Table 2).

Clinical mastitis (all cases) incidence distribution estimated using a period at risk extending from start to end of the study is illustrated in Figure 2. Unconditional least square means (LSM) estimates (i.e., not adjusted for putative confounding variables) are presented as Supplementary Table 1. After adjusting for potential confounders, there was no statistical difference in the general CM incidence between the two farm types. A LSM estimate of 14.0 cases/100 cow-year (95% CI: 8.3, 23.7) was obtained for RMS farms, whereas a LSM of 16.3 cases/100 cow-year (95% CI: 9.0, 29.6) was obtained for straw farms (Table 3). Moreover, there was no difference in the incidence of severe CM episodes (severity score 3) between the two types of farm. A LSM of 2.1 severe cases/100 cow-year (95% CI: 1.1, 4.1) was obtained for RMS farms, and straw farms had a LSM of 1.6 severe cases/100 cow-year (95% CI: 0.8, 3.4).

The estimated incidence of K. pneumoniae CM, however, was higher in RMS farms with a LSM of 1.6 cases/100 cow-year (95% CI: 0.8, 3.4) compared with 0.2 cases/100 cow-year (95% CI: 0.1, 0.6) for straw farms. For this comparison, a 7.0 times (95% CI: 2.0, 24.6) higher incidence was observed in RMS farms. There was no significant difference between the two types of farms regarding the incidence of CM due to S. uberis, E. coli, S. dysgalactiae, or S. aureus.

When using the alternative approach for computing period at risk, we ended up excluding 8 RMS farms and 11 straw-bedded farms that sent <2 samples during the study period (Table 4). Using this alternative approach, the general estimated CM incidence was different between the two groups (Supplementary Table 2) with a LSM estimate of 26.5 cases/100 cow-year (95% CI: 19.2, 36.6) for RMS farms and a LSM of 46.2 cases/100 cow-year (95% CI: 30.2, 70.8) for straw-bedded farms. Furthermore, the total incidence of severe CM was not different between the two farms groups with a LSM estimate of 10.2 cases/100 cow-year (95% CI: 6.9, 15.0) in RMS farms and a LSM of 13.6 cases/100 cow-year (95% CI: 7.1, 26.1) in straw-bedded farms.

As in our first approach, the estimated incidence of CM due to K. pneumoniae was significantly higher in RMS farms with a LSM of 3.4 cases/100 cow-year (95% CI: 1.6, 7.1) compared with 0.6 cases/100 cow-year (95% CI: 0.2, 1.6) in straw farms. This was equivalent to a 5.9 times (95% CI: 1.6, 21.2) higher incidence in RMS farms. There was still no difference between the two groups concerning the incidence of CM due to S. uberis, E. coli, S. dysgalactiae, or S. aureus.