The statistical models described here use data from the NAHMS Dairy 2014 study. NAHMS is a non-regulatory program of the United States Department of Agriculture, Animal and Plant Health Inspection Service, Veterinary Services (USDA APHIS VS) that conducts national surveys to collect health and production information for U.S. livestock, poultry, and aquaculture.

In partnership with the USDA’s National Agricultural Statistics Service (NASS), NAHMS conducted the study in 17 states representing 80.3 percent of the U.S. milk cow inventory and 76.7 percent of operations with milk cows [22,23]. The study was conducted in two phases, though the analysis presented here focuses on data collected during Phase I of the study using the General Dairy Management Questionnaire (GDMQ). There were 1,261 complete responses from dairy farmers to this questionnaire.

See the NAHMS Dairy 2014 Report 1 [22] for additional details regarding the statistical design of the study and the NAHMS Dairy 2014 Report 4 [23] for baseline descriptive statistics pertaining to MMS.

Working alongside members of the NAHMS Dairy 2014 team, including veterinary epidemiologists and statisticians, the final working dataset used in the analysis was prepared using SAS [24] to include the original study data along with variable edits and additions specific to the MMS analysis (see Supporting information, S1 Appendix for specific details on variable edits and additions). The list of MMS-specific variables used in this analysis are given in Table 1, including cow manure handling, storage or treatment, and application methods. “Other” categories in each of the MMS components were omitted from this analysis and generally showed low representation in the population [23].

Each component has been categorized as primarily handling solid or primarily handling liquid or slurry manure. Solid manure usually contains more than 15 percent solids and is typically found in areas where dirt, pasture, or bedding absorb moisture from the manure. Slurry manure usually contains from 5 to 15 percent solids and is generated when there is limited or no material to absorb moisture from manure. Liquid manure contains less than 5 percent solids and is generated when wastewater or rainwater is mixed with manure. For the purposes of this report, liquid and slurry manure are combined [23].

Descriptive analyses were performed using SUDAAN [25], SAS-callable survey software that allows for the proper analysis of data from complex surveys by accounting for the study design. Survey-weighted relative frequencies, along with their estimates of standard errors were computed for respondents to the questions of interest (Table 1).

Estimates of the percentages of operations by manure handling, storage, and application method type were computed to assess the estimated relative frequency of each MMS. In addition, networks were estimated using relative frequencies using the R package igraph [26] and visualized using the R package ggraph [27] within the R statistical software [28,29]. Networks were created using estimates of percentages of operations at two levels. The first set of estimated relationships were among components of manure handling, manure storage, and manure application methods separately (Figs 1–3). The second set of estimated relationships were between components of manure handling and storage (Fig 4), and between components of manure storage and application (Fig 5).

These relationships were estimated using percentages of operations as the closeness metric to investigate the most common components of manure systems in the population. Nodes in the networks are represented by individual components and diameters of nodes are proportional to the estimated percentage of operations that had that component as a part of their MMS. Edges in the networks represent links between components and their widths are proportional to the estimated percentage of operations with the given pair of components. Component pairs representing 5 percent or less of operations for relationships among the strategies of each individual manure management component (Figs 1–3) and pairs with 10 percent of operations or less for the relationships between strategies of different components (Figs 4 and 5) were suppressed to depict the more common relationships more clearly.

To assess the relationship between manure handling, storage, and application strategies, a series of logistic regression models were fit, which fall into two sets. In the first set, a single manure storage strategy was regressed on the group of eight manure handling strategies; a total of nine models were fit (Fig 6). In the second set of models, a single manure application strategy was regressed on the group of nine manure storage strategies; four models were fit (Fig 7).

In this way, we assessed the set of relationships between the three key phases of MMS as they are actually practiced on dairy operations. The results are reported as odds ratios, where ratios above one indicate a “positive association” between the two MMS components. That is, the expected odds that a dairy producer used the two strategies together is greater than the odds that the strategies were not used together. Ratios below one indicate a “negative association” between the two MMS components. That is, the expected odds that two strategies were used together is lower than the odds that the two strategies were not used together. These interpretations assume that other independent variables are held constant.

The PROC RLOGIST procedure in SUDAAN was used to run all of the logistic regression models described in this paper. In contrast to a standard logistic regression model, PROC RLOGIST incorporates the survey design and weights of the data to produce accurate point estimates and estimated standard errors. The survey-weighted logistic regression models used in this paper were modeled in the form: logit(p)=Xβ pih=P(yih=1|Xih=x) where y_ih is thebinary dependent variable (1 = the method was used; 0 = the method was not used), X is the design matrix, β are the regression coefficients to be estimated, i indexes the operation, and h indexes the stratum. The design matrix contains a column of 1s for the intercept, followed by columns of observed survey response data for each independent variable used in the logistic regression model, with each column corresponding to one survey variable. The design matrix has the same number of rows as there are producers represented in the underlying dataset. The independent variables were either be the 8 handling methods (Fig 6) or the 9 storage methods (Fig 7) in the two sets of models. SUDAAN estimates β by solving the weighted score equations. Standard errors are estimated using the Taylor series linearization method, assuming a without replacement design (i.e. once an operation is selected for inclusion in the sample, that same operation cannot be selected again); unequal survey weights are accounted for, and finite population corrections are applied at the stratum level. Unequal survey weights were used, as the NAHMS Dairy 2014 study design used a stratified random sample of dairy operations across State and size category strata with operations having different probabilities of selection across strata.

Statistically significant relationships between manure handling and storage methods and between storage and application methods were assessed using adjusted Wald F-test p-values [30,31] from Type III ANOVA tests [32] of significance for effects of the dependent variables. The p-value threshold to determine statistical significance were Sidak-adjusted p-value thresholds [33,34], which adjust for multiple comparisons. Per Sidak’s method, multiple comparisons were adjusted for at the model level where the threshold was calculated using the equation: 1−(1−α)1m where α is the desired family-wise significance level, here chosen to be 0.05, and m is the number of independent variables in the regression model for which comparisons are desired. In this case, m = 8 and the p-value threshold equals 0.0064 for the models regressing storage methods on handling methods and m = 9 and the p-value threshold equals 0.0057 for the models regressing application methods on storage methods.

Statistical significance is a tool used here to focus the analysis and discussion towards common or likely relationships in the population in a complex, multivariate system of inter-related MMS components. Relationships that do not meet the significance levels are not implied to be absent or unimportant; they just did not meet the requirements to be statistically significantly associated given the methods described above and the observed data. Point estimates (and in some cases, standard error estimates) are given in order for the reader to be able to further investigate relationships beyond the binary decisions made regarding statistical significance.

The percentages of operations by manure management component strategies are presented in Table 2, along with the estimated standard errors. The estimates presented for the manure application strategies are for the 98.3 percent (SE: 0.4) of operations that applied manure to land owned or rented by the operation [23].

The most frequently practiced handling strategy was leaving manure on pasture (74.3 percent), followed by an open or dry lot that was scraped (59.0 percent), using a gutter cleaner (47.9 percent), and using an alley scraper (43.9 percent). The use of a bedded pack was used on less than a third of operations (30.8 percent).

The most frequently practiced manure storage strategies included spreading manure on a daily or nearly daily basis (61.3 percent), using a manure pack inside of a barn (57.6 percent), any of the solid storage strategies (47.3 percent), and using a slurry tank or basin (40.5 percent). By far the most frequently reported manure application strategy was use of a broadcast or solid spreader (87.2 percent) while surface application by tank wagon or tank truck was performed on 41.6 percent of operations.

More than one manure handling method and more than one storage method were used by 86.1 percent (SE: 1.3) and 81.4 percent (SE: 1.4) of operations, respectively. Also, of the 98.3 percent of operations that applied manure to land owned or rented by the operation, 42.2 percent (SE: 1.6) of operations used more than one manure application method.

The relationships among strategies within each manure management component are presented in Figs 1–3. The most common combinations of manure handling strategies (Fig 1): were

The most common combinations of manure storage methods (Fig 2) included:

The most common pairs of manure application methods (Fig 3) for the 98.3 percent of operations that applied manure to land owned or rented by the operation was a broadcast or solid spreader and surface application using a tank wagon or tank truck (32.5 percent, SE: 1.6).

The most common pairs of manure handling-storage strategies (Fig 4) reported included:

The most common pairs of manure storage-application strategies (Fig 5) reported included: broadcast or solid spreader combined with each of manure spreader (58.2 percent, SE: 1.7), manure pack (52.9 percent, SE: 1.7), other solid storage (43.3 percent, SE: 1.7), and a slurry tank or basin (31.8 percent, SE: 1.6).

Utilizing the significant odds ratios from handling, storage and application, we find clear patterns of related strategies, largely determined by the handling of manure as a liquid, slurry, or solid (Fig 8). While many of the associations in our data confirm the theoretical linkages between phases of MMS, we also find confounding results in some cases. These results suggest that farmers may often be using multiple overlapping systems, as can be observed in the NAHMS Dairy 2014 Report 4 (USDA 2018) and in Section 3.1 above. Note that the Phase I survey used in the NAHMS Dairy 2014 study did not explicitly ask detailed information on separate manure management systems within an operation, but instead focused on manure management system components used across the entire operation. Thus, sub-system analysis was not feasible given the data, and the analysis instead focused on the manure management systems used on operations, regardless if they were components in overlapping or independent systems within that operation.

As an example, operations frequently have cows in multiple housing systems, such as freestalls and dry lots where the MMS is unique to each housing system. Manure in the freestall might be removed using a flush system with a solids separator while manure in the dry lot is scraped. Thus, our results illuminate the more nuanced and complex interactions between overlapping systems that characterize MMS decision-making as it is practiced by U.S. dairy producers.