Statistical Methods for Improving Sensitivity in

Dairy Cattle Feeding Experiments 1

P. J. C U N N I N G H A M and F. G. OWEN
Department of Animal Science, University of Nebraska
Lincoln 68S03

Abstract experimental performance as a eovariable

This study compares four methods of sta- is relatively sensitive, especially for milk
tistically analyzing performance data from yield.
a dairy cattle feeding experinmnt. The ba-
sic design (modified switch-back) involved
a pre-experimental (Period I), an experi- Introduction
mental (Period I I ) , and a postexperimental Researchm's are continually searching for
(Period I I I ) period. Periods I and I I I methods to increase sensitivity of their experi-
were standardization, or control periods, ments. Because of the high cost of maintaining
during which all cows were fed the same. a dairy animal under experimental conditions,
Thirty-sLx Holsteins were used to determine it is important to minimize animal numbers in
the effect of six treatments (2 X 3 fac- feeding studies. Cochran and Cox (5) list three
torial) on milk and solids-corrected milk methods for increasing the accuracy of experi-
yield, milk fat and solids-not-fat percent- ments. These are to A) increase the size of the
age, dry matter and estimated net energy experiment, B) refine the experimental tech-
intake, and estimated net energy/solids- nique, and C) utilize an experimental design
corrected milk. The four methods of analy- or method of analysis which minimizes the ef-
sis ntilized different data as follows: (A) fects of variability. The size of an experiment
average performance in Period II, (B) is a function of the available facilities, number
average performance in Period I I adjusted of animals, labor, and other factors. Refine-
by covariance for average performance in ment of the experimental technique is depen-
Period I, (C) average performance in Pe- dent upon the precision of the equipment nsed
riod I I adjusted by covariance for both and the experimenter. Of the three methods,
average performance in Period I and body more sensitive experimental designs or meth-
weight at the end of Period I and (D) ods of analysis seem to offer greatest potential
difference in average performance in Pe- for improvement. Many different designs and
riod I I relative to that in Periods I and I I I . methods of analysis have been proposed for
The difference method of analysis yielded dairy cattle feeding experiments. Bartlett (1)
the smallest error variances and coefficients and Broster and Curnow (3) have discussed
of variation for each of the traits studied. covariance as a means of error reduction.
Compared to Method A, Method D reduced Switch-back trials were originally proposed by
the error variance for milk yield by 94%, Brandt (2) and were later extended by Lucas
estimated net energy intake by 65%, and (7). This type of design is effective in remov-
fat percentage by 52%. Use of covariance ing between-cow sources of variation. Lucas
adjustment for Period I performance also (8) presented an excellent discussion on the
reduced error variances to almost the same design of dairy cattle feeding experiments. The
extent without requiring data from Period number of cows required per treatment to de-
I I I . The inclusion of body weight as a tect significant differences of various magni-
second eovariable did not result in any ap- tudes for several designs has been discussed by
preciable reduction in error variance com- Gill (6). Numerous other types of designs or
pared to pre-experimental performance as modifications appropriate to feeding experi-
a single covariable. Consequently, from the ments can be found in most statistical textbooks.
standpoint of sensitivity of dairy cattle
The purpose of this investigation was to com-
feeding experiments, the difference method
pare the efficiency of A) a completely random-
would be the preferred method among those
compared; however, analysis using pre- ized design with B) an analysis of eovariance,
C) multiple covariance, and D) a modification
of the switchback type of design in the measure-
:Receive(] for publication October 12, 1970. ment of differences in the principal criteria for
503
504 0UNNINGHAM AND OWEN

evaluating lactation performance. By utilizing Results and Discussion

data from the same experiment for all four Expectation of error mean squares. An ex-
methods of analysis, the differences in sensi- amination of the expectations of the error mean
tivity between methods of analysis can be eval- squares of Ylethods A, B, and D should provide
uated with minimal effects of sampling. some insight into the theoretical possibility for
improving sensitivity using these methods of
Materials and Methods
analysis. (Method C was not included since
Data for this investigation were from an ex- Method B represents the covariance type of
periment with 36 Holstein cows on six treat- analysis.)
ments. The experimental design (Table 1) was
a modification of the switch-back design de- Method A : E ( E r r o r M.S.)
scribed by Lueas (7). Each of the three ex- = ~2 with t ( b - - 1 ) (If
Y
perimental periods was 4 weeks long with Method B: E ( E r r o r M.S.)
Periods I and I I I being control periods. Per-
formance data were recorded during the last 2 t(b--1) °"2
xy

weeks of each period. Six cows were randomly = [~ ---]

t(b--1)--i Y
a zx
allotted to each of the six treatments. The
nutritional findings from this study have been t(b--1)
reported previously by Owen and Appleman = (1--r~)~
t(b--1)--I "Y Y
(9) and will not be discussed here.
The data were analyzed by four different with t ( b - - 1 ) - - I df
methods to determine the relative sensitivity of Method D: E ( E r r o r M.S.)
each for measuring performance differences in
lactation trials. The four methods of analysis
= 4¢ + +
were : -- 40-xy - - 4 ~ y z -~ 20-x~
Method A, analysis of average performance with t ( b - - 1 ) df
for Period II. where, t ~ number of treatments
Method B, analysis of average performance b ~ number of cows per treatment
for Period I I with average performance
(r2 - - v a r i a n c e within treatment for
for Period I as a covariable.
Y Period I I
Method C, analysis o£ average performance
cr2 ---variance within treatment for
for Period I I with both average perfor-
x Period I
mance for Period I and body weight at the
end of Period I as covariables. a ~ - - v a r i a n c e within treatment for
z Period I I I
Method D, analysis of twice the average per-
formance for Period I I (Y) minus the %x "- covariance between Periods I
average performance for Periods I (X) and I I
and I I I (Z) (i.e. 2Y -- X -- Z). %.~ ~ covariance between Periods I I
and I I I
Seven performance characteristics were daily
milk yield, solids-corrected milk (SCM) yield, (rxz -- covarianee between Periods I
dry matter intake, estimated net energy (ENE) and III.
intake, ENE intake (Meal) (above mainte- If between cow variation is the same in all
nance) per kilogram of SCM ( E N E / S C M ) , three periods (0-2 -----0.2 : o-2z) and covariance
milk fat percentage, and solids-not-fat per- x Y

centage. between each pair of periods is equal ([Link] ~

¢ryz = 0-xz), the expectation for Method D sim-
plifies considerably.
TABLE 1. Experimental design.
E ( E r r o r M.S.) ---- 6(~ 2 -- ~ )
Treatment sequencea y xy
= 6(1-rxy) 0.~
Period 1 2 3 4 5 6 Y

Treatments Various questions can be raised concerning

the validity of these assumptions. The between-
I A A A A A A
cow variance within period, and treatment may
II B C D E F G
tend to increase with time, but the assumption
III H H H H H H that all are equaI to the variance in Period I I
a Six cows per sequence. may not be greatly in error. The assumption
JOURNAL O~ DAIRY SCIENCE ~rOh, 54, N o . 4
 E V A L U A T I O N OF L A C T A T I O N 505

TABLE 2. Mean and error variances for treatment differences for each trait.
Method of analysis
Trait Mean A B C D
Milk yield per day (kg) 21.50 31.00 2.12 2.06 1.95
Solids-corrected milk (kg) 20.23 24.87 3.02 2.88 2.10
Fat (%) 3.58 0.25 0.13 0.13 0.12
Solids-not-fat ( % ) 8.82 0.30 0.13 0.12 0.12
Dry matter intake (kg) 19.55 7.73 3.13 3.15 2.84
Estimated net energy
intake (Meal) 26.3 13.7 5.5 5.6 4.8
Estimated net energya/
SCM (Meal per kg) 0.88 0.034 0.017 0.017 0.015
Estimated net energy intake above maintenance.

concerning the covarianees is more likely to be Error variances and error coe~icients of vari-
in error. I t seems reasonable to assume O-xy ation. The error variances on a per cow per
O-yz, but ~xz will probably be smaller. I f this is Period I I basis for each of the seven traits are
true, the simplified expected mean square is in Table 2 and the error coefficients of variation
biased upward. in Table 3.
The relative efficiency of two designs can be The difference method (Method D) was the
estimated by comparing the inverse ratio of most efficient of the four methods of analysis
the error variances per unit (Cochran and Cox followed closely by the eovariance methods (B
1). To make comparisons on a per cow per and C). The error variances were smaller than
Period I I basis, the error mean square for for any of the other methods of analysis for
Method D nmst be divided by four. Dividing all seven traits. Compared to Method A, Method
the error mean square for Method D by four D improved sensitivity most for traits involving
would be analogous to analyzing the mean per- milk yield (milk yield per day and SCM) re-
formance in Period I I minus half the mean for ducing the error variances per unit from 31.0
Periods I and I I I ((2y -- x - - z ) / 2 ) . This to 1.95 (93.7%) for milk yield and from 24.9
makes the unit of comparison equivalent for to 2.1 (84%) for SCM, but only 65% for E N E
all four methods. intake and 52% for milk fat percentage. The
Method B t (b--l) error variance in the analysis of the data from
Period I I only (Method A) is influenced by
Method A = t ( b - - 1 ) - - l ( 1 - - r 2 y ) between-cow variation and random error. The
Method D difference and covariance methods (D, B, C)
- - 1.5 (1--rxy) are effective in removing a large p a r t of the
Method A between-cow variation for traits such as milk
Method B t(b--1) (I+r~)
TABLE 3. E r r o r coefficients of variation.
Method D t (b--1)--1 1.5
Methods o f analysis
These ratios indicate that Method D would
be more efficient than Method B if r~y > 0.5, Trait A B C D
but less efficient than Method B when rxy < 0.5 Milk yield per day (kg) 26.4 6.7 6.6 6.5
(ignoring differences in degrees of freedom). Solids-corrected milk
Method A is distinctly the least efficient if the (kg) 24.6 8.6 8.4 7.1
correlation between the response in Period I Fat(%) 14.0 10.1 10.1 9.7
and I I is high and positive and relatively large Solids-not-fat ( % ) 6.2 4.1 3.9 3.9
degrees of freedom are available for error. F o r Dry matter intake (kg) 14.2 9.0 9.1 8.6
most traits studied in feeding trials, the corre- Estimated net energy
lation is high and positive for successive period intake (Meal) 14.0 8.9 9.0 8.3
responses. Broster and Curnow (3) obtained
Estimated net energya/
correlations of 0.91, 0.73 and 0.62 for milk SCM(Mcal/kg) 21.1 14.8 14.8 13.8
yield, fat and solids-not-fat percentage, be-
tween production in Period 1 (Weeks 5 to 12 a Estimated net energy intake above mainte-
of lactation) and Period I I (Weeks 15 to 22). nance.
JOURI~AL OF DAIRY SCIEI~CE VOL, 54, NO. 4
506 C U N N I N G H A ~ AND OWEN

yield for which there is appreciable between- other traits. Both the difference and covariance
cow variation. methods (D and B) remove from error between=
Except for E N E / S C M , the difference method cow variation in performance. However, when
was effective in reducing coefficients of varia- r~y is below 0.5, Method D gives more weight
tion to less than 10% (Table 3). The values to the periods before and after the treatment
with Method D are low compared to the varia- period than is justified by the association be-
tion generally found with biological data. W i t h tween periods of the same cow. That the differ-
Method D the coefficient of variation for E N E / ence method corrects for slope of the lactation
SCM is slightly larger (13.8%) than the values curve is of little consequence for comparisons
for other criteria, but it is considerably smaller between treatments within the same period of
than the 21.1% with Method A. In this study, the lactation. This would be the case in these
coefficients of variation for milk yield (6.5%) data.
and fat percentage (9.7%) were comparable The covarianee method (Method B) adjusts
for the difference method to the 7.0 and 9.3% for only prior period performance while the
obtained by Burr (4) from a Latin square de- difference method (Method D) includes both
sign. However, the coefficient of variation for prior and post period performance. Other ap-
solids-not-fat was larger than the value he re- propriate comparisons on the basis of time re-
ported (3.9 versus 1.5%). Lucas (8) reported quired to complete an experiment would be:
the normal between-cow variation in production 1) Method B compared with an analysis of the
rate to be around 25% but that this could be difference between the performance in Period I I
reduced to values which range from 1 to 7% and the performance in Period I (Y -- X ;
with switch-back designs. Method E ) , and 2) Method D compared with
The methods (B and C) involving eovariables an analysis of covariance with performance in
were also more efficient than Method A. Com- Periods I and I I I as eovariables (Method F ) .
pared to Method A, the reductions in the error The theoretical expectations of error mean
variances per unit ranged from a high of 93.4% squares can be compared assuming equality of
(31.00 to 2.06) for milk yield with Method C variances and covariances between periods.
to a low of 4S% (0.25 to 0.13) for fat percent-
Method E : E ( E r r o r M.S.)
age with either Method B or C. Inclusion of
body weight as a second covariable (Method C) = 2 (l--r~y)~ 2
Y
was not effective in further reducing error vari-
ances compared to the single eovariable of Method F : E ( E r r o r M.S.)
average performance in Period I (Method B). t(b--1)
The partial regression on body weight was not = (1-n~.~z)~ • t(b-1) 22
significant (P > .05) for any of the seven
= ( 1 - - r ~ ) ( l + 2 r ~ y ) ~2 . t(b--1)
characteristics. Body weight might be an im-
portant single covariable, but as a second co- l-~-rxy Y t ( b - - 1 ) --2
variable, in combination with average perfor- The relative efficiency of the covariance analy-
mance in Period I, it does not result in any sis to the corresponding difference method of
appreciable further reduction of error vari- analysis can be determined by comparing the
ance in these data.
ratio of expected mean squares.
The coefficient of variation of 6.7% for milk
yield using Method B was similar to the 8% re- Method E t(b--1)--I 2
ported by Bartlett (1) and Broster and Curnow Method B t(b--1) 1 + rxy
(3) with a similar covarianee method. Also, the
coefficients of variation for fat percentage Method D t(b--1)--2 1.5 ( l + r x y )
(10.1%) and solids-not-fat percentage (4.1%) Method F t(b--1) 1 -b 2rxy
were somewhat larger than the 8.0% and 2.2%
by Broster and Curnow (3). Ignoring differences in degrees of freedom, the
The error variances for treatment differences eovarianee method of analysis (Method B and
(Table 2) tended to be larger for the covariance F ) would be more efficient than the correspond-
methods (B and C) than for the difference ing difference method (Methods E and D) if
method (D). This was an expected result from rxy < l . The difference methods of analysis
theoretical expectation and the assumption of a are an approximation of the covarianee method
high positive correlation between performance of analysis, the accuracy of the approximation
in different periods. The advantage for the depending on the correlation between perfor-
difference method was greatest for SCM, sug- mance in different periods. Computationally,
gesting the r~y was higher for S C ~ than for the difference methods are simpler which prob-
JOURI~AL 01~ DAIRy SCIEI~CE VOI,. 54, NO. 4
 E V A L U A T I O N O F LACTATION 507

TABL~ 4. One-half confidence intervals, a then Method A. Both Methods B and D yield
intervals which are generally considerably
Method of analysis smaller than those for Method A. The loss of
A B D the degree of freedom from error for the co-
variance analysis (29 versus 30) did not appear
Milk yield per day (kg) 6.58 1.72 1.63 to have any noticeable effect on these values;
Solids-corrected milk (kg) 5.90 2.09 1.72 however, this factor becomes more important
Fat (%) 0.59 0.43 0.41 with smaller degrees of freedom for error.
Solids-not-fa~ ( % ) 0.65 0.43 0.41 None of the methods of analysis permitted
Dry matter intake (kg) 3.26 2.09 2.00 detection of extremely small differences. I n
Estimated net energy designing an experiment, a knowledge of the
intake (Mcal) 4.3 2.8 2.6 number of replications required to detect dif-
Estimated net cnergyb/ ferences of a given size is critical. Snedecor
SCM (Meal/kg) 0.22 0.16 0.13 (10) presents a method for determining the
number of replications necessary to have a
a p = 0.05. probability of 0.50 of detecting a difference at
b Estimated net energy intake above mainte-
5% significance. The number of replications
nance.
required for four magnitudes of difference in
milk yield for error variances in this study are
ably accounts for their past popularity. How- in Table 5. The numbers of cows required per
ever, with the advent of computers, this is no treatment are slightly larger for the com-
longer a valid criteria. F r o m expectations, co- pletely randomized design and smaller for the
variance methods should receive consideration covariance analyses compared to the data pre-
as a method of analysis in preference to differ- sented by Gill (6). The discrepancy is due to
ence methods. F u r t h e r investigation into the the magnitude of the error variances. Method
validity of the assumptions of the expectations A would be prohibitively expensive for the de-
would seem appropriate. tection of small differences. Methods B and D
require considerably fewer animals than Method
The possibility of carryover effects needs to
A. Even with these methods of analysis, con-
be considered in the difference method of an-
siderab]e cost would be incurred if extremely
alysis. I f the response in Period I I I is affected
small differences were to be detected.
by rations in Period I I , interpretation of the
results can be in error. However, in this par- These findings emphasize the necessity for
ticular type design only a differential carryover careful planning of dairy cattle feeding ex-
from different treatments would bias results. periments. The experimenter nmst assure that
an adequate number of animals is available to
Carryover effects of equal magnitude for each
detect differences which his objective dictates.
treatment would not bias comparison of treat-
I f relatively long term trials are required, this
ment differences. Allowing two weeks for transi-
study indicates that covariance adjustment for
tion at the beginning of each period is an at-
performance in a preceding standardization
tempt to minimize carryover effects. I n this
period would greatly reduce the number of ani-
experiment adjusted means for all four methods
mals needed to detect a given difference. I f
of analysis for all seven traits were essentially
the same. This suggests that carryover effects carryover effects are absent, minimal, or sub-
ject to rapid depletion, then a post-experimental
were absent or very small.
Detecting differendes. All researchers are
interested in detecting as small a difference as TABLE 5. Number of replications required to de-
possible. The size of the difference which can tect a difference between two treatments, a
be detected is dependent upon the error variance Difference in
and the degrees of freedom for error. The daily milk yield (kg)
larger the error, the greater the number of
replications required. Table 4 shows the mag- 0.91 1.36 1.81 2.27
nitude of the one-half confidence intervals (P ---- ( Cows per treatment)
.05) for the difference between two treatment
Method A 289 130 74 48
means for Methods A, B, and D. The average
Method B 22 11 7 5
standard error of a difference was used for
Method D 19 10 6 5
the eovariance analysis (Steel and Torrie (11)).
As expected, Method D permits detection of a A 1 : 1 chance of detecting a difference at 5%
the smallest difference, followed by Method B, level of significance.
JOUR/~AL OF DAIRY SCIENCE VOL, 54, NO. 4
508 CUNNINGIIA~¢I AND OWEN

s t a n d a r d i z a t i o n p e r i o d can f u r t h e r i m p r o v e ex- perimental Design. J o h n Wiley and Sons,

p e r i m e n t a l efficiency. Inc., New York.
(6) Gill, J. L. 1969. Sample size for experiments
on milk yield. J. Dairy Sci., 52 : 984.
References (7) Lueas, H. L. 1956. Switch-back trials for
(1) Bartlett, M. S. 1935. An examination of the more than two treatments. J. Dairy Sci.,
value of covarianee in dairy cow nutrition 39 : 146.
experiments. J. Agr. Sci., 25: 238. (8) Lucas, I-L L. 1960. Critical features of good
dairy feeding experiments. J, Dairy Sci.,
(2) Brandt, A. E. 1938. Tests of significance
in reversal or switchback trials. Iowa Agr. 43 : 193.
(9) Owen, F. G., and R. D. App]eman. 1970.
Exp. Sta. Res. Bull., 234.
Value of a gelatinized milo-urea mixture
(3) Broster, W. H., and R. X. Curnow. 1964. (extruded product) and DEITY-100 (pel-
A note on the use of the analysis of co- leted preparation) as carriers of urea in
variance to reduce standard errors in experi- lactation rations. Prec. X V I I I Int. Dairy
ments on milk production. J. Agr. Sci., Congr. 1: 540.
63 : 91. (19) Snedecor, G. W. 1956. Statistical Methods
(4) Burt, A. W. A. 1958. The effect of varia- (Jth ed.). Iowa State College Press, Ames.
tions in n u t r i e n t intake upon the yield and (11) Steel, R. G. D., and g. H. Torrie. 1960.
composition of milk. J. Dairy Res., 25: 1. Principles and Procedures of Statistics.
(5) Cochran, W. G., and G. M. Cox. 1957. Ex- McGraw-Hill Book Co., New York.

.~OUR-N-AL OF ])AIRY SCIENCE VOL. 54, No. 4