Replication rates for survival feeding experiments in pens with weaner beef cattle.

Abstract

Animal Production in Australia Vol. I5 EXPERIMENTAL DESIGN IN CATTLE RESEARCH WHEN RESOURCES ARE LIMITING INTRODUCTION G.W. BLIGHT* In conventional agronomic experiments with pasture or crops grown in plots the cost per experimental unit is low compared with, say, animal experiments. The capital cost involved in establishing a well designed large-scale field trial with cattle can easily exceed available resources. Consequently the replication rates of paddocks and animals per treatment may not be as high as we would like, and the treatment responses we can expect to detect as statistically significant in the experiment may be larger than desired. Adequate replication rates (number of animals per treatment) are likely to be attainable in grazing trials where, because of the nature of the treatments, all of the experimental animals graze a common pasture. However, in multipaddock grazing trials such as stocking rate or pasture comparisons, where it is the paddock which is the experimental unit, cost often prohibits setting up the necessary numbers of paddocks of sufficient size. In this contract the papers by Haydock, Duncalfe, Pepper and Mayer are concerned with the design of grazing experiments where constraints on resources are limiting opportunities to conduct research; Seebeck's paper reviews the additional design requirements when body composition information, measured at slaughter, is required from grazing experiments. In the last paper of this contract O'Rourke develops guidelines for the design of survival feeding experiments in pens. CONSEQUENCES ARISING WHEN THE DESIGN OF A GRAZING TRIAL IS NOT IDEAL K.P. HAYDOCK** Grazing trials can be very expensive animals (ii) purchase and install fencing and plant and establish pastures (iv> control (v> build mustering yards and (vi> buy scales of land are required for even a simple trial fit in with budgetary constraints the size of that which was considered ideal. requiring finance to (i> acquire watering points (iii> supply seed, livestock diseases and parasites to weigh animals. Large resources One often finds that in order to a trial needs to be curtailed from In this paper, methods of size reduction are discussed together with some of the consequences. Some suggestions are made to obtain an estimate of error and to aid the extrapolation of the results. METHODS OF SIZE REDUCTION Herd size One compromise is to reduce the herd size. A minimum herd size (9 generally is considered to be three; any aberrant behaviour of one or two animals may be detected by comparing the behaviour of the three. Single animal herds have been used successfully (Yates et al. 1964 ; Bryan 1968; Jones 1982). It is considered essential that the paddock sizes be small so that several animals are in view of each other in order for behaviour patterns to be normal (Mannetje et al. 1976). This approach is more relevant where the primary objective is to assess the effect of grazing on pastures. With single animals, any sickness can * QDPI Biometry Branch, Rockhampton, Qld 4700. ** CSIRO Division of Mathematics and Statistics, St Lucia, Qld 4067. 52 Animal Production in Australia Vol. I5 have a serious effect on animal production. The reduction in paddock size will decrease the precision of the trial two reasons. Firstly, the variation between paddocks increases as the size the paddock decreases (Smith 1938; Hatheway and Williams 1958). Secondly, number variance of mean liveweight change/paddock increases as the animals/paddock is reduced. for of the of Another compromise is to reduce the number of (ii> Reduced replication In a region replicates. This also reduces the precision of a treatment mean. which is variable with respect to factors affecting pasture growth, using a small number of replicates would give wide confidence limits for a treatment effect. Restricting the region across which a trial is established would improve precision but would limit the applicability of the results. A preferable procedure would be to develop a regression model relating the variate of interest, e.g. liveweight gain or milk yield/ha, to the factors of the trial and Factors affecting growth which. to growth factors associated with each paddock. easily come to mind are rainfall (amount and pattern), temperature (maximum and minimum), soil moisture, botanical and chemical composition of the pasture, aspect of the paddock (northerly or southerly)) soil type and depth of soil. Such a model is of inherent biological interest and would have general applicability. In its simplest form it amounts to a covariance adjustment for diverse paddock effects. For this purpose we should note that pasture production is affected by rainfall, temperature, paddock aspect, soil moisture, soil type The dry weight of green material has been shown to affect liveweight and depth. change (Willoughby 1959; Roe et al. 1959; Yates et al. 1964; Mannetje 1974). Similarly, chemical composition of a pasture is influenced by temperature, soil moisture, botanical composition and soil fertility -(Andrew and Fergus 1976; Ludlow 1976). Howard et al. (1962) showed that mineral composition of a pasture They used chemical composition to explain was correlated with protein content. the variation in animal production over seasons of the year. Thus, a simple form of the model suggested would involve only animal production, dry weight of green material and protein content of the pasture. A reduction in the size of an experiment also may be achieved by reducing The greatest replication would be the replications of some treatments. For example, if the purpose of a maintained for the treatment of most interest. trial was to determine the optimum stocking rate for a number of pastures, the stocking rates initially judged to be near the optimum could have the most replication. In any grazing trial there is a need for (iii> No replication estimate of random error. A trial without replication usually would One exception is where a treatment estimation of an error term. Provided three or more levels are used, on quantitative levels. Orthogona suitable response model to the levels of A e.g. linear. should be used so that independent sums of squares for the parame model can be calculated- The deviation from the response model CDEV If other factors are also being tested i as an estimate of error. the error term could be composed of DEV and interactions of the factor, after testing for homogeneity. an unbiased preclude the (say A) has e can fit a 1 polynomials ters of the > may be used n the trial, form DEV X Error degrees of freedom usually are few and they may be increased by transforming the levels of the factor to linearize the response. In this way only one degree of freedom is assigned to the factor A effect and any interaction If a polynomial is not a effects involving A will be adequately tested. biologically meaningful model, one may fit an appropriate model to the response 53 Animal Production in Australia Vol. I5 after determining the reality of any interactions, For unreplicated grazing trials in which a paddock is the experimental unit, an analysis of variance is possible using the variation between animals within However, the effects of treatments and of paddocks paddocks as an error term. are confounded; also it is difficult to extrapolate the results to some population as most research workers would desire to do. Unreplicated experiments usually are used in the early stages of an investigation. Exceptions to this were the work of Shaw (1961) and Shaw and Mannetje (1970). These workers compared animal production from native pasture with that from native pasture and Townsville lucerne with and without fertilizer. Five fold increases in production/ha were obtained and the results were enhanced by running the trial for seven years using replacement animals each year. There is little doubt that the results were accepted as real because the change in production was large, biologically meaningful and consistent over seven years with different samples of animals. We might say that the magnitude of the difference was greater than any which could arise by chance due to any conceivable random effects of paddocks and animals. Where the differences are much smaller, the problem reduces to one of eliminating the confounding effects of paddocks and of having an estimate of the underlying random error. If the regression model relating animal production to environmental and growth factors has been developed, it could be used to adjust animal production for the confounded paddock effects. An estimate of error must be composed of two components, one due to paddocks and one due to animals within paddocks. A component due to paddocks might be estimated by collecting information on the variance component for paddocks from experiments reported in the literature and from unpublished results. If these components have a narrow distribution one could use with confidence their weighted mean as an estimate of paddock variance; if they exhibit a wide range, their values might be associated with paddock size, as found by Smith (1938) for dry matter production of different crops. As well, the paddock variance components might be associated with soil type, region etc. If this is the case, an estimate of the variance component for paddocks, for the appropriate conditions of a particular trial, could be made. A component of variance for animals also should be estimated. In this regard, the standard deviations/animal for cumulative liveweight change, in lb/at, for one year old animals after periods of four weeks, produced by Haydock (1964) from data gathered by the late MC. Franklin, should prove useful. Even if these efforts are successful in constructing an error variance, there would be many research workers who would prefer to compare treatments using information based on genuine replication. PADDOCK ROTATION AND REPLICATION RATES FOR DAIRY PRODUCTION EXPERIMENTS F. DUNCALFE* BACKGROUND This paper examines the design of long-term dairy production experiments in which paddocks, each of which is subjected to a specific pasture treatment, are grazed for a whole lactation by dairy cows. We show that the resource requirements of such experiments are very large when the experimental unit is taken to be the herd of animals and its associated paddock. In many cases the * QDPI Biometry Branch, Brisbane, Qld 4001. 54 Animal Production in Australia Vol. 1.5 requirements would limit research opportunities. In a typical experimental set-up, treatments are assigned to paddocks and a herd of lactating cows is allocated to each paddock. During the course of the experiment each herd of cows has to be taken to the dairy, milked and then returned to the same paddock. Measurements related to dairy production, such as milk production, butter fat yield and solids-not-fat, are made for each animal in each herd. Herds of five cows are commonly used by QDPI officers to simulate a commercial herd, and to minimize possible within herd correlation. An experiment designed to compare eight treatments, with a minimal replication rate of two paddocks per treatment, will have sixteen herds of animals to be handled twice daily seven days a week for the duration of the experiment, Thus even minimal replication and a herd size of five produces an experiment with large labour requirements. A normal management practice used by dairy farmers and also in some experiments on irrigated improved pastures, involves the rotation of a herd of animals sequentially through a series of paddocks. For example, with a rotation sequence of one week in and three weeks out for each paddock in turn, a herd of animals becomes associated with a set of four 'sub-paddocks?, each being one quarter of the size of the original paddocks. While paddock rotation may involve a lot of extra fencing and extra work in the planning stages, it offers a reduction in size of the error mean square, and thus improved power in significance tests, through extra blocking and randomization. For eight treatments and four rotations, there are 32 sub-paddocks for each replicate. For two replicates, rotational grazing involves blocking the 64 sub-paddocks into eight homogeneous 'sub-blocks', each of eight sub-paddocks. Since the sub-block is one quarter of the size of blocks in an experiment without rotation, the experimenter may be able to take advantage of the smaller size to create more uniform sub-blocks. In order to be able to block effectively, some knowledge of the expected productivity of different parts of the available land is required. Even when such information is not available, the extra randomization in allocating treatments to sub-paddocks will tend to remove the confounding between treatment and paddock effects. If the experimental area is completely uniform, no gain in power of significance tests would be achieved by rotation. METHODS AND RESULTS We can use the estimates of paddock error variation from past experiments to calculate the probability (power) of detecting significant differences between treatments for a specified replication rate and true difference between treatments. Three rotationally grazed randomized block experiments conducted at Ayr in North Queensland, having 3 or 4 treatments, 2 or 3 paddock replicates and 2 or 3 rotations have been studied in an attempt to quantify the power of such experiments. The approximate power when a 5% level significance test is applied were calculated for a range of replication rates and true differences, for an experiment with eight treatments and a herd size of five. The calculations were based on Tang's tables (Kempthorne 1952). Coefficients of variation were calculated from paddock mean squares for 12 week milk production data from these experiments. The paddock mean squares are associated with only 3 or 4 degrees of freedom, and so power is poorly determined. Coefficients of variation ranged from 12% (I 2 week summer production) to 70% (12 week winter production) for the three experiments, and these two values were used in the calculations of power. OW study suggests that differences, he is unlikely significance tests. To achieve differences in milk production unless the researcher is looking for gross to be able to obtain reasonable power in his an 80% chance of detecting commercially important (say 0.5 kg/cow/day for inexpensive treatments to 55 Animal Production in Australia Vol. 15 1.5 kg/cow/day for expensive treatments), the necessary replication rates are impractical (more than five replicates). The conclusion that can be drawn is that the cost of establishing experiments of the type considered in this paper is extremely large; equally, the ongoing labour requirements are likely to be beyond available resources. This is especially so for winter milk production, which is an important subject of investigation. The lack of dairy grazing experiments having sufficient replication to assure reasonable power is surely a consequence of the factors discussed above. The researcher often has to be content with an experiment without paddock replication, particularly where a number of treatments are proposed e.g. factorials. The technique of rotation as described above removes the confounding of treatments and paddocks, but does not provide a valid error term for use in treatment comparisons since the experimental unit is the herd and its associated paddock. Investigation of data from available replicated grazing experiments indicates that the animal error term provides a satisfactory approximation, although the results are not conclusive. I consider that the design of dairy grazing experiments needs to be the subject of further research. REPL ICATION RATES FOR GRAZING EXPERIMENTS W ITH BEEF CATTLE ON NATIVE PASTURES IN CENTRAL AND NORTH QUEENSLAND P.M. PEPPER* AND R.J. MAYER* Cattle growth rate data from two environmentally diverse sites in north and central Queensland (NQ and CQ, respectively) were analysed to determine suitable replication rates for grazing experiments. For this purpose, probability tables were derived which show the chance of detecting true differences between treatment means; the probability values have been calculated for treatment differences within a size range considered to be of practical importance. MATERIALS AND METHODS At the NQ site near Millaroo, there were three series of experiments which included two series of supplementary feeding experiments of randomised block design (5'or 6 treatments x 2 replicates) over 7 consecutive years, followed by a uniformity trial on the same paddocks for 2 years. Each year a fresh draft of Brahman cross steers grazed speargrass pasture growing on solodic soils and uniform sands at -47, -43 or -33 beasts/ha, respectively, for the three series of experiments. In the second year of the uniformity trial 10 of the paddocks, being long and rectangular, were fenced in half and stocked at the same rate (-33 beast/ha) as in the previous year. For all experiments each year was divided into dry and wet seasons; average daily gain was calculated for these two periods and the overall period. The dry period corresponds to the supplementation period. The CQ site is near Gayndah in the Burnett valley, on an open forest site dominated by speargrass and forest bluegrass. A stocking rate experiment of 4 replicates x 3 stocking rates in a randomized block design was conducted over 9 consecutive years at the same site. Each year a new draft of Hereford weaners grazed at the 3 rates (0.74, 1.24, 2.47 beasts/ha) during the active pasture growing periods (December-May). In all experiments animals were allocated to paddock groups by stratified randomisation on the basis of fasted liveweight. The effectiveness of this strategy for controlling variation in growth rate was tested by analysis of * QDPI Biometry Branch, Brisbane 4001 and Toowoomba 4350, Qld, respectively. 56 Animal Production in Australia Vol. I5 variance. . Similar tests determined if grouping the paddocks into blocks could increase precision by isolating variation from the experimental error. For each site, Bartlett's test for homogeneity of variances was used to determine if the experimental error could be pooled over years. From the expectation of the mean squares, the components of .variance were calculated and Tang's tables were used to determine the approximate probabilities (power) of detecting true treatment differences of various magnitude when using a 5% significance level test (Kempthorne 1952). RESULTS AND DISCUSSION NQ site Analyses of the first year's data of the uniformity trial, which used the larger paddocks (30 ha) showed that contiguous paddocks grouped into blocks only succeeded in isolating variation for the dry season. For the supplementary feeding trials variation due to blocks on these same 30 ha paddocks, was significant in only 3 out of 21 analyses. Hence for experiments on these 30 ha paddocks where the paddock is the experimental unit, a completely randomised design would be preferable; since the number of degrees of freedom for error would be small, losing precious degrees of freedom with no compensating reduction in error mean square would decrease the power of significance tests. In the second year of the uniformity trial with the smaller 15 ha paddocks, blocks did effectively isolate variation from the paddocks error variation because there was a significant difference between the back and front halves of the 30 ha paddocks. Because poor and good performing paddocks had been noticed in the supplementation trials, the effect of choosing replicates on previous performance rather than position was investigated. The performance from the first year of the uniformity trial was used to group the 30 ha paddocks into replicates. From the analyses of the second years data, this did not prove to be a successful strategy, as the sets of paddocks did not maintain their relativity across years. Blocking animals on initial fasted weight was effective in reducing animal variation in only three of the 27 analyses at the NQ site, which indicates that animals could have been allocated to paddocks completely at random. The paddocks variance components for the 30 ha paddocks from the first year of the uniformity trial were similar to those for the 15 ha paddocks in the second year. The 'all trials' estimates of paddock and animal components of variance (Table 1) were calculated from mean squares pooled over years of both the supplementation and uniformity trials. The 'all trials' variance components were used to calculate power values given in later tables. TABLE 1 Components of variance Uniformity trial paddocks (yr 1) paddocks (yr 2) dry season wet season overall CQ site Variation significant in any mean squares could pooled over years. animals, were cal paddock component be due to the much 3086 All trials paddocks animals 1959 1315 740 3597 7614 2619 2394 1015 2366 2148 1007 I due to the replicates and animal blocks terms were not of the nine analyses. Bartlett's test showed that the paddock be pooled over years; similarly the animal mean squares were Components of variance of 3453 for paddocks, and 5955 for culated from mean squares pooled over the 9 analyses. The is larger than the corresponding figure from NQ (1315) and may smaller paddock sizes. These were: 1.21, 2.43 or 4.05 ha, 57 Animal Production in Australia Vol. 15 depending on the stocking rate treatment. Both sites For experiments where the paddock variation is the appropriate error, such as stocking rate trials, approximate probabilities of detecting true treatment differences of various magnitudes in average daily gain are given in Table 2. The probabilities are for an experiment of a randomised block design with 6 treatments. TABLE 2 Approximate probabilities (%I of detecting true differences between treatment means as significant (P=O.O5) Increasing the number of animals does not have the same dramatic effect as increasing the number of paddock replicates. Hence, smaller paddocks with 5 animals each are to be recommended in preference to larger paddocks with IO animals each. The number of replicates needed is determined by the size of true treatment differences which the experimenter considers meaningful and would like to be able to detect as significant. Although stratifying animals on initial fasted liveweight was not effective in the experiments considered here, it is a recommended precaution especially if there are only 5 animals per paddock and an animal needs to be removed for any reason. Probability estimates are given carries the treatment with it and experimental error. Probabilities completely randomized design where all in Table 3 for experiments where the animal the animal variation is the appropriate are calculated for two treatments in a experimental animals graze together. true differences between TABLE 3 Approximate probabilities (%> of detecting treatment means as significant (P=O.O5) The variation for animals was higher in the NQ wet season than in the dry season and this is reflected in the lower chance of detecting a true difference. In experiments where the animal variation is the experimental error there is usually little problem in obtaining enough animals to give an acceptable chance of detecting treatment differences of a magnitude regarded as economically important. Animal Production in Australia Vol. 15 GUIDELINES FOR THE DESIGN OF EXPERIMENTS INVOLVING MEASUREMENT OF BODY COMPOSITION R.M. SEEBECK* BACKGROUND The design of body composition experiments, incorporating the latest biometrical methods for expressing changes in body composition, has been recently discussed by Seebeck (1983). The present paper is intended to provide guidelines for the design of experiments involving effects of some treatment on animal production, where the primary aim is to give information on body weights and/or economic returns, but where some subsidiary information is desired on body composition measured at slaughter. Aspects of body composition that may be considered in this type of experiment include dressing percentage, yield of saleable meat, dissected or chemical composition of a sample joint, complete dissection of a side, or even the proportions of the different grades of carcass. Collecting meaningful information at slaughter from experiments involving growing stock requires much more rigorous experimental designs than those where information on live weights is the only requirement. The main reason for this is Therefore any that the composition of an animal changes with increase in size. difference in composition that results from a treatment should be partitioned between what was caused by a simple weight difference and what was not. If this is not done, nonsensical statements are often made, such as (when referring to an experiment where all animals are slaughtered at the same time), 'not only did treatment A (e.g. the highest stocking rate) decrease the live weight of the steers at the end of the period, but there were less first grade carcasses (or they were less fat or whatever)'. The second part of the statement may have been just because they were smaller, or there may have been another effect operating separately from the effect of size. A more appropriate experiment and analysis would have been able to clarify this point. Although size is the major determinant of body composition if one considers animals over a reasonable part or all of their growing period, body composition is also affected by growth rate (Seebeck 1983), presumably mainly by variation in In some cases, e.g. if the treatment is a hormonal one, one nutrient intake. should try to isolate the growth rate effect from the treatment effect. In other cases, the inclusion of a variate for growth rate may eliminate the significance of the treatment effect that appeared in an analysis without that inclusion. This is, in itself, a worthwhile conclusion. DESIGNS AND ANALYSES General If probability values are required for treatment differences in body composition, then both the design and statistical analysis have to be compatible This requirement means that the design includes variation in with this desire. size and possibly also growth rate and is such that it can be analysed by Other statistical techniques, such as factor analysis, regression techniques. have they may better represent patterns of development, limited while To achieve a range in size unconfounded with capabilities for comparing groups. growth rate, serial slaughter designs have to be used. These designs can be set up by slaughtering groups of animals at a series of either weights or ages, the latter necessitating that a rate of growth effect be removed statistically to enable estimation of development patterns in an unbiased form (Seebeck 1983). Removal of the rate of growth effect is also desirable when slaughtering at a * CSIRO Tropical Cattle Research Centre, Rockhampton, Qld 59 4701. Animal Production in Australia Vol. I5 series of weights, because it is difficult to kill animals size, due to variation in gut fill, etc. (Seebeck 1983). animals to the slaughter groups must be done on a random stratified on the basis of initial live weights) before experiment. at exactly the same The allocation of basis (preferably the start of the It is suggested that a treatment group should consist of at least IO animals, spread over at least 3 and preferably 5 slaughter groups. This will give results that will be able to be partitioned into treatment, size and rate of growth effects. If the treatment is imposed at or near the start of the slaughter programme, the treatment will have a progressive effect and the regression of composition on size (and/or growth rate) will be expected to have different slopes. The different slopes can be dealt with in the analysis by comparison of regression line techniques, or in the general least squares sense by including treatment x covariate interactions. With experiments involving an imposed treatment(s), it is useful to know what the composition of each animal is at the start, so that the change in composition, adjusted for size and rate of growth for each animal, can be determined. However, using such an initial slaughter group to calculate the composition of each other animal is not very satisfactory, because the The animals that would have calculation is only based on the average animal. comprised the initial slaughter group are best slaughtered at a subsequent time spread over the treatments and used to better define the treatment effects. A more useful strategy is to perform one or more in vivo estimates of composition on all animals at the start of the experiment (such as TOH injection), slaughtering some of the animals then as an initial slaughter group, and repeating the in vivo measurements on all other animals at least at the time the individual is slaughtered. This type of information can greatly improve the precision of estimation of changes within an animal. Using serial slaughter, rather than killing all animals at the same weight or time, naturally decreases the information available from the experiment directly concerning that particular weight or time. However, with a good serial slaughter design, the loss of accuracy is not very great and is compensated for If a specific weight or time is by more information at other weights or times. of particular interest, there is nothing to stop the experimenter slaughtering more animals at that point than at other points. Pen experiments With pen experiments, the use of serial slaughter reduces the information on feed efficiency comparisons, because of the different times that the animals are on feed. Consequently, it is probably better to space the slaughter programme unequally, with an initial slaughter group and, say, as a compromise, slaughter at weights or times corresponding to 50, 75 and 100% of the Initial in vivo composition estimates are total experimental period envisaged. For experiments that are almost a necessity with this type of experiment. reasonably short, say of six months or less duration, the slaughter of the animals may be better done on a weight gain or time basis rather than on a weight basis. This is because the range of initial weights may be large compared to the weight gain over the experiment. However, the range of initial weights should not be artificially restricted because the treatment may have probably differential effects on animals of different initial weights. , Grazing experiments The major constraint on body composition information from an experiment with grazing animals is that slaughter of some animals during the experiment, as a consequence of serial slaughter, will change the grazing pressure. This is of particular importance if the experiment involves treatments 60 Animal Production in Australia Vol. 15 in different paddocks, such as a stocking rate experiment. The only practical solution to this is to have a matching group of animals, members of which can be brought in to maintain the grazing pressure. Such a scheme presupposes that there are sufficient animals in a paddock that can be slaughtered at three (or certainly a minimum to two) different times or weights. To have any variation in rate of growth at least two animals have to be killed at each point, so that the minimum number of animals per paddock is four. If other constraints put the number at less than that, it would be better to forget any thought of doing body composition estimates. Where paddocks are the experimental unit, as in stockin