| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Department of Population Medicine and Diagnostic Sciences, College of Veterinary Medicine, Cornell University, Ithaca, NY 14853
1 Corresponding author: rcb28{at}cornell.edu
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
5 mo before lameness was detected, respectively; 1, 2, 3, 4, and 5 = 1, 2, 3, 4, and 5 mo after diagnosis, respectively; and 0 = cows that were never lame. With the binary lameness classification analyzed by repeated measures ANOVA, there was no effect of lameness on milk yield. The model was biased because lame cows had higher milk yields before lameness compared with nonlame cows. When the LAMIX was used, milk production before lameness was greater than after lameness (3.1 ± 0.28 kg/d). Yet, point estimates generated for LAMIX were inaccurate because of the multicollinearity detected between LAMIX and week of lactation and because of the inability of adjusting the least squares means for the interaction of LAMIX and week of lactation. Therefore, the most appropriate models were the ANCOVA models (both for the matched and nonmatched retrospective-cohort designs). The estimated losses associated with lameness were 314 and 424 kg/cow per 305-d lactation, respectively, for the matched and nonmatched designs. Furthermore, high milk yield in the beginning of the lactation was a risk factor for lameness.
Key Words: dairy cow • lameness • analysis of covariance • milk loss
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
Most of the pertinent peer-reviewed literature that associated lameness with increased milk production used similar analytical methodology: a binary classification of lameness (lame at least once versus never lame) and ANOVA (Dohoo and Martin, 1984; Rowlands and Lucey, 1986; Hultgren et al., 2004). Dohoo and Martin (1984) reported that cows affected with foot and leg disorders had increased milk production of 1.6% when compared with control cows. Rowlands and Lucey (1986) reported a positive association of sole ulcer and white line disease with milk production. More recently, Hultgren et al. (2004), using a binary classification for the occurrence of sole ulcer and ANOVA, reported that sole ulcers were positively associated with milk production. In contrast, most studies that used a disease-indexing methodology for lameness (instead of the binary lameness variable), and repeated-measures ANOVA found lower milk production in lame cows after the lameness event compared with unaffected herdmates (Rajala-Schultz et al., 1999; Warnick et al., 2001; Green et al., 2002). Gröhn et al. (1999) reported that when using a binary classification variable for ketosis (yes or no), the effect of ketosis on milk production was masked due to the short-term impact of the disease on milk production, and sick cows were actually milking better before the incidence of ketosis; however, the effect that ketosis had on production became evident when a more detailed indexing methodology was used for ketosis. Lameness was generally not a short-term disorder, and often chronic and recurring. Therefore, lameness could have a long-term effect on milk production.
The typical cohort study with binary classification of lameness clearly exposed 2 kinds of bias: confounding (lame cows are higher producing cows than nonlame cows) and selective survival bias (Warnick et al., 2001). Confounding bias can be controlled by using appropriate statistical analysis (ANCOVA or lameness indexing). Selective survival bias notes that lame cows must be alive and milking at least up to the moment the disease is detected. On the other hand, the cohort of nonlame cows will include cows that have left the herd because of several other reasons (i.e., lower milk production, unresolved sickness of any kind, and reproductive failure). Hence, it is possible that lame cows were high-producing cows and consequently kept in the herd longer. Selective survival bias is problematic (if not impossible) to control during statistical analysis (Kleinbaum et al., 1982). Still, it is possible to prevent selective survival bias (also known as loss to follow-up bias) by designing a specific study that prevents selective survival. For instance, selective survival bias would not occur if diseased cows were matched with control cows that were alive at least up to the moment the disease was detected.
Our objective was to test the hypothesis that lameness incidence in lactating dairy cows decreased milk production in comparison to nonlame cows. To test the proposed hypothesis we used 2 different study designs and 3 different statistical methods. It was our objective to evaluate the different study designs and statistical methods.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
Lameness was defined as the first time a lactating dairy cow received an orthopedic block as a form of treatment for a laminitic-type disease. Orthopedic blocks were used to treat claw diseases such as sole ulcers, white-line disease, white line abscesses, toe ulcers, heel ulcers, and other laminitic-type disorders. Most cows treated with an orthopedic blocks were visually detected as lame before treatment. But, if cows were diagnosed with any of the above-mentioned laminitic-type disorders during a routine hoof trimming, the cows were added to the cohort of lame cows. Cows visually detected as lame, but without a laminitic-type disorder detected during hoof trimming, were excluded from the cohort of lame cows. The disease-specific incidence was not available because all animals that were treated with an orthopedic block had their information entered under a Dairy Comp 305 (Valley Agricultural Software, Tulare, CA) event called BLOCK. Infectious diseases (such as digital dermatitis, interdigital dermatitis, and foot rot) did not receive orthopedic blocks, and treatment data from these diseases were stored under different Dairy Comp 305 event codes. Therefore, cows affected with these infectious diseases were not included in the cohort of lame cows unless the infectious disease was concurrent with a disease suitable for orthopedic-block treatment (these were rare concurrent events).
Farm and Management
Data were from a dairy farm located near Ithaca, NY, and from cows that calved between January 1, 2004, and January 1, 2007; only 1 lactation per cow was evaluated. This farm was selected because of its long history of a working relationship with the Ambulatory and Production Medicine Clinic at Cornell University. The farm milked 2,800 Holstein cows 3 times daily in a double 52 milking parlor. The cows were housed in free-stall barns with concrete stalls covered with mattresses and bedded with waste-paper pulp. Feed alleys were grooved-concrete flooring and cleaned by automatic scrapers. All walkways to and from the milking barn and the holding pen were covered with rubber. Footbaths were located in the exit lanes of the milking parlor. The footbath consisted of a 5% formalin solution and applied a minimum of 4 times weekly. Every cow was scheduled to receive routine hoof trimming twice yearly. Lame cows were identified and treated by trained farm employees.
The cows were fed a TMR consisting of about 55% forage (corn silage, haylage, alfalfa silage, alfalfa hay, and wheat straw) and 45% concentrate (corn meal, soybean meal, canola, cotton seed, and citrus pulp).The diet was formulated to meet or exceed the NRC nutrients requirements for lactating Holstein cows weighting 650 kg and producing 45 kg of 3.5% FCM.
Data Collection and Study Designs
Data were collected from the dairy-records database Dairy Comp 305. Two study designs were used: a retrospective-cohort design (RCD) and a retrospective matched cohort design (RMCD). For the RCD, cows were identified as being either having received an orthopedic block (Lame) or not having receive an orthopedic block (Not Lame). For the RMCD, all cows that receive an orthopedic block were matched with 1 herdmate that served as a control. Control cows were matched as closely as possible for calving date, and perfectly matched for parity number, and being alive and lactating at the time of treatment. If more than 1 suitable control cows was identified, the first cow on the list generated by Dairy Comp 305 was used. Date of calving was used to generate the list, and if no cows matched perfectly, the cow in the same lactation with the closest date of calving was chosen. If there were no control cows to fit these criteria, the lame cow was excluded from the analysis; 19% of the lame cows were excluded because of imperfect matching. A similar matching strategy was used elsewhere (Raizman and Santos, 2001; Bicalho et al., 2006).
Statistical Analysis
The outcome variable was daily milk yield (DMY; 1-wk average DMY); DMY was extracted from Dairy Comp 305. Data were structured so that the data for each cow had 45 lines of information, each line containing the DMY and a milk-test number, and cows had from 1 to 45 wk of lactation available for analysis. A single table for the entire data set was created in JMP (SAS Inst. Inc., Cary, NC). A separate table was generated containing only 1 line of data per cow; this spreadsheet was used to perform descriptive statistics.
The LAMIX was defined as –1, –2, –3, –4, and –5 = 1, 2, 3, 4, and 5 mo before lameness was detected, respectively; 1, 2, 3, 4, and 5 = 1, 2, 3, 4, and 5 mo after diagnosis, respectively; and 0 = cows that were never lame. To assess multicollinearity between LAMIX and week of lactation, a Spearman correlation was calculated using the FREQ procedure of SAS. Spearman correlation is a nonparametric measure of correlation that assesses how well an arbitrary monotonic function could describe the relationship between 2 variables, without making any assumptions about the frequency distribution of the variables.
Mixed general linear models were fitted to the data using the MIXED procedure of SAS (SAS Inst. Inc., Cary, NC). The outcome variable was DMY (kg), which was modeled as a Gaussian (normally distributed data) variable. The assumption that the residuals were normally distributed was assessed by visually evaluating the distribution plot of the Studentized residuals. The data were longitudinally collected and had a series of repeated measures of DMY throughout lactation. This implied that data points were correlated within each research subject. To account appropriately for within-cow correlation of the DMY, the error term was modeled by imposing a first-order autoregressive covariance structure for all statistical models. The model described below was fitted to all models evaluated.
 |
where Y = 7-d average of total DMY; X = the matrix of all independent variables; β = the vector of all fixed effect parameters; Z
= random effect of matched pair. (This term was only used for the RMCD and it represents the random effect of matching group. The covariance structure assumed for this term was a compound symmetry); e = random residual. For all models, the within-cow correlation of the DMY was accounted for by imposing a first-order autoregressive covariance structure (assumed that the within cow correlation of the repeated measures of milk weights decreased as time between the test dates increased) to the error term. The correlations coefficients (
) depended on the time between repeated observations as described below:
 |
For the RCD, 3 different repeated-measures linear mixed models were used. The first model was an ANOVA and the following independent variables were used: Lame, parity (1, 2, and >2), and week of lactation (from 1 to 45). All possible 2-way interactions between the independent variables were added to the model. The second model was an ANOVA, but with LAMIX (from –5 to 5) as the independent variable of interest, parity (1, 2, and >2), and week of lactation (from 1 to 45) were offered to the model. Two-way interactions between independent variables were added to the model; however, the interaction of LAMIX and test-day number could not be tested because these 2 factors were not completely crossed (not all levels of the factor LAMIX were observed in all levels of the factor test-day number). The third analysis was an analysis of covariance (ANCOVA). The independent variables used in this model were Lame, parity (1, 2, and >2), and week of lactation (from 1 to 45). All possible 2-way interactions between the independent variables were added to the model. Additionally, the average milk production for the first 3 wk of lactation of each cow was used as a covariate. To assess the association of average milk production for the first 3 wk of lactation and mean DMY for the entire lactation a simple linear regression was fitted for these 2 variables using SAS.
To further investigate the potential occurrence of selective survival bias, the same cohort of lame cows used in the RCD were matched with herd-mates by parity, calving date, and being alive and lactating at the time when treatment was performed. The RMCD was analyzed by repeated-measures models. The first model was an ANOVA model fitted in SAS (SAS Inst.) using the MIXED procedure. The independent variables were; Lame and week of lactation (from 1 to 45); the interaction of lame and week of lactation on milk was offered to the model. Because cows were matched by parity this variable was not added to the model. The matching group was included in the model as a random effect with a compound-symmetry covariance structure; no interactions involving this factor were tested. The second analysis done for the RMCD was an ANCOVA. The independent variables used in this model were the same used in the first model described above. Nevertheless, the average milk production for the first 3 wk of lactation of each cow was used in the model as a covariate. For all ANCOVA models, the cows that were lame within the first 3 wk of lactation were excluded from the analysis. For all ANOVA and ANCOVA models, a sequence of contrasts was used to test hypotheses for milk-production differences in a series of combinations of levels for the categorical variables. Furthermore, variables and their respective interaction terms in all models were retained in the model when their P-values <0.05.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
3. The median DIM at lameness was 97, and the interquartile range was 58 to 165. For the RCMD study, 603 lame cows were matched with 603 nonlame control cows. For lame cows, the median DIM at the end of data collection was 250 and the interquartile range was 183 to 318. For controls the median DIM at the end of data collection was 245, and the interquartile range was 175 to 310. The median DIM at lameness was 101, and the interquartile range was 64 to 162.
Retrospective Cohort Design ANOVA with Binary Lameness
For all statistical models, the distribution of the Studentized residuals was approximately normally distributed; the homogeneity-of-variance assumption was satisfied in all statistical models (based on visual inspection of the scattered plot of the Studentized residuals against the fitted values). Moreover, all variables offered to the models were significant and were kept in the final models, including the interaction terms.
For this analysis, lame and control cows did not differ in milk yields for complete lactation (P = 0.083; Table 1
). Nevertheless, the interaction of lameness cohort and week of lactation on milk production was highly significant; lame cows produced more milk than controls cows from the start of the lactation up to wk 15 of lactation (P < 0.001; Figure 1).
|
|
). To assess multicolinearity between LAMIX and week of lactation a Spearman correlation was calculated. When all levels of LAMIX were considered rsp = 0.60; when the control cows (LAMIX = 0) were excluded the rsp = 0.99.
Retrospective Cohort Design ANCOVA
Lame cows had lower milk yields than control cows (Table 1). Milk yields were similar for lame and control cows from the beginning of the lactation up to wk 12 of lactation; thereafter, milk yields for the cohort of lame cows dropped and remained below the yields for the control cows until the end of the lactation (Figure 2). The average milk production for the first 3 wk of lactation explained 51% (R2 = 0.51) of the variability found in the DMY (Figure 3).
|
|
). Furthermore, there was no significant difference in overall milk production for lame cows versus the control cows (Table 1).
|
).
|
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
3, 20.8% were diagnosed with digital dermatitis or foot rot (Bicalho et al., 2007). It is possible that the total incidence of lameness in the herd was underestimated by up to 20%. In the RCD ANOVA design and analysis lame cows produced more milk than controls in the beginning of the lactation. It is important to emphasize that median DIM at lameness occurrence was 97 d, while the drop in milk yields occurred at 105 DIM. It appears that no study that used binary classification of lameness and ANOVA concluded that lameness causes loss of milk.
The second methodology used to analyze data from the RCD was LAMIX instead of the lameness cohorts. A similar indexing approach was used by Rajala-Schultz et al. (1999), Warnick et al. (2001), and Green et al. (2002). Interpretation of general linear mixed model variables is difficult when the independent variables are highly correlated. LAMIX is a combination of the binary variable lame (yes or no) and the variable which represents time (in this case week of lactation from 1 to 45). Therefore, the variables LAMIX and week of lactation were clearly not independent (and the correlations of 0.55 and 0.99 confirmed that); therefore, their use in the same statistical model might disregard the model assumptions. Furthermore, least squares means generated do not take into account that milk production might depend on the interaction of LAMIX and week of lactation, but testing their interaction was impossible (because the factors were incompletely crossed). Therefore, the milk least squares means by LAMIX generated by this model lack potential adjustments from 495 parameters 45 wk of lactation multiplied by 11 levels of the variable LAMIX causing us to question the validity of the model outcome.
Although we have diagnosed inadequacies in the use of LAMIX in linear mixed models, all peer-reviewed studies we found that have made use of this disease-indexing methodology successfully identified a detrimental effect of lameness on milk yields. Using similar indexing methodology Rajala-Schultz et al. (1999) reported lameness had a negative effect on milk yields, Warnick et al. (2001) reported a milk loss between 104 and 295 kg per lactation for cows that became lame at 100 DIM. Green et al. (2002) concluded that lameness caused a total milk loss of 360 kg per cow over the entire lactation. Nonetheless, we are cautious about the validity of our point estimates for LAMIX because of the multicolinearity detected between LAMIX and week of lactation and because of the inability to test (and adjust the least squares means) for the interaction of LAMIX and week of lactation.
The third methodology used to analyze the RCD was the use of analysis of covariance. The ANCOVA was used to neutralize the effect of a continuous independent variable in the experiment (Lyman Ott and Longnecker, 2001). It is reasonable to assume that if the cohort of lame cows had not actually become lame, their milk yields would have remained above the yields of the controls cows throughout the lactation; the average milk production for the first 3 wk of lactation explained 51% of the variability found in the DMY for the entire lactation; hence, by adjusting for the milk average for the first 3 wk of lactation we were able to set an equal start for both cohorts (lame and nonlame) and neutralize most of the confounding effect of early lactation milk yields. With this methodology, we estimated that lame cows lost on average 1.0 kg/d of milk when compared with the controls. To the best of our knowledge this is the first time that a repeated-measures ANCOVA controlling for early-lactation milk yield was used to analyze the effect of lameness on milk yields.
It is possible that high milk production in the beginning of the lactation is a risk factor for lameness. Green et al. (2002) reported that high milk production before lameness was a risk factor for lameness; lame cows produced 1.12 kg/d more milk before than after the lameness event. Still, the cohort study design using binary or indexing classification of lameness might be exposed to a selective-survival bias (Warnick et al., 2001). To test the hypothesis that high milk production before the lameness event is an artifact of selective survival bias and does not have a causal relationship with lameness, we designed an analysis based on cohorts matched to have comparable survival (RMCD). The group of survival-matched lame cows still produced more milk in the beginning of the lactation compared with nonlame cows. Although our matching method should have controlled selective survival bias in the current lactation, it was still possible that selective survival bias from previous lactations was present. Nevertheless, the outcome of our matched-cohort study analysis supports the hypothesis that high milk production in early lactation is a risk factor for lameness.
Although the epidemiologic studies have shown conflicting results of the effect of lameness on milk production (Green et al., 2002) the biological plausibility of such effect is clearly understood. Hence, lame cows are exposed to pain and impaired movement, and as a result the expression of natural behavior (eating, drinking, and interaction) is restricted. As a consequence, DMI should decrease and eventually milk yields should decrease. On the contrary, it is not as intuitive to hypothesize why increased milk production can expose high-producing cows to incremented risk of lameness. It is believed that claw horn lesions were a result of laminitis, which in theory can be caused by subclinical ruminal acidosis (Thoefner et al., 2004). Yet, a repeatable model of laminitis induction that reflects the reality of the commercial dairy farms has not been reported. Nevertheless, milk production is linearly associated with DMI, and high DMI was associated with lower ruminal pH (Oetzel, 2007). Therefore, higher producing cows could be at higher risk of laminitic-type disorders because of subacute ruminal acidosis triggered by higher DMI. It is important to acknowledge that the mechanism by which lameness can be triggered by high milk yield is not understood and further work needs to be done to elucidate the pathophysiology.
In summary, when a binary lameness-classification variable was used in repeated-measures ANOVA, there was no effect of lameness incidence on milk yields. Although, when a LAMIX or a repeated measures ANCOVA (controlling for mean milk production for the first 3 wk) was used, there was a significant detrimental effect of lameness on milk yields. Point estimates generated for LAMIX might be inaccurate because of the multicolinearity detected between LAMIX and week of lactation and because of the inability of testing and potentially adjusting the least squares means for the interaction of LAMIX and week of lactation. The milk losses estimated by the ANCOVA model were 424 and 314 kg per lame cow over the 305-d lactation for the RMCD and RCD, respectively. Furthermore, the results of the survival matched designs showed that high milk yield in the beginning of the lactation was a risk factor for lameness; the average milk production for the first 3 wk of the lactation for lame cows was 3.2 kg/d higher than the matched controls. After evaluating 2 study designs and 3 statistical methods, we concluded that both RCD and RMCD designs when analyzed by ANCOVA and controlling for early lactation milk yield generated the most credible estimates for milk losses.
Received for publication October 1, 2007. Accepted for publication March 11, 2008.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  K. K. M. O'Driscoll, M. M. Schutz, A. C. Lossie, and S. D. Eicher The effect of floor surface on dairy cow immune function and locomotion score J Dairy Sci, September 1, 2009; 92(9): 4249 - 4261. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
J. L. Kotschwar, J. F. Coetzee, D. E. Anderson, R. Gehring, B. KuKanich, and M. D. Apley Analgesic efficacy of sodium salicylate in an amphotericin B-induced bovine synovitis-arthritis model J Dairy Sci, August 1, 2009; 92(8): 3731 - 3743. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
R. C. Bicalho, V. S. Machado, and L. S. Caixeta Lameness in dairy cattle: A debilitating disease or a disease of debilitated cattle? A cross-sectional study of lameness prevalence and thickness of the digital cushion. J Dairy Sci, July 1, 2009; 92(7): 3175 - 3184. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
J. Liu, N. K. Neerchal, U. Tasch, R. M. Dyer, and P. G. Rajkondawar Enhancing the prediction accuracy of bovine lameness models through transformations of limb movement variables J Dairy Sci, June 1, 2009; 92(6): 2539 - 2550. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
T. C. Linden, R. C. Bicalho, and D. V. Nydam Calf birth weight and its association with calf and cow survivability, disease incidence, reproductive performance, and milk production J Dairy Sci, June 1, 2009; 92(6): 2580 - 2588. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
