Agriculture
Presenter: Frikkie Calitz
When: Monday, July 11, 2016 Time: 4:00 PM - 5:30 PM
Room: Salon A Carson Hall (Level 2)
Session Synopsis:
The incorrect application of experimental designs and the alternative application thereof
My experience as a consulting Biometrician has increasingly exposed me to statistical experiments that are applied incorrectly. As a result the data of such experiments is meaningless or cannot be analysed. The purpose of this presentation is to demonstrate why such so-called statistical experiments of randomised design (CRD) or randomised block designs (RBD) are not statistically defendable. Such experimental layouts are often statistically undefendable due to a misunderstanding of the definition of replication and randomisation. One of the most common errors in agricultural experiments is that experimental units are simply subdivided into sub-plots and then regarded as experimental replications (also called random replications). The following is a typical example of incorrect application: An experiment was done in Food Science to test the effect of two cake recipes on the texture of the cake. The different cakes were baked simultaneously. Each cake was then cut into four quarters (for the researchers four replications). The texture of each quarter was measured, a quarter of the cake was considered to be a random replication of the recipe and thereafter the data was analysed as a complete randomised design (CRD). When analysing data from the above mentioned so-called statistical experiment the statistical analysis programs will accept the wrong data entered. The calculation is then a sampling error and not a statistical experimental error. It is expected that the sample variation will be much smaller than the experimental variation. Using sampling error to compare treatment will result in significant differences amongst treatments that may not be true and therefore may lead to misinterpretation of results. Similar experiments with animals, crops, fruit and agricultural engineering and how alternative statistical designs can be applied will be discussed in this presentation. It is recommended that biometricians should be aware of misconceptions that exist amongst researchers and therefore it is very important to be familiar with the experimental layout (field plan) before analysing the data.
Agriculture
Presenter: Catherine Lloyd-West
When: Monday, July 11, 2016 Time: 4:00 PM - 5:30 PM
Room: Salon A Carson Hall (Level 2)
Session Synopsis:
An Exploration of Statistical Methods for Genomic Selection in Perennial Ryegrass: A New Zealand Study
Genomic selection (GS) is a breeding strategy that uses genome-wide DNA markers to predict the genetic value of selection candidates in breeding programs. The central process of GS is to first establish models for the training population that have both genotypic and phenotypic data and then use these models to calculate GEBVs (genomic estimated breeding values) for individuals (e.g. in breeding populations) that have genotypic data only. These GEBVs are then used to identify and select elite parental genotypes for random mating to generate the next generation in the breeding cycle. It is argued that the combination of affordable, high-throughput genotyping and GS prediction methods has resulted in marker-based prediction that show great promise for increasing genetic gains from selection when integrated with existing animal and plant breeding systems. In this study, we investigate the potential of Genomic Selection using a genotyping-by-sequencing (GBS) SNP marker system in conjunction with phenotypic data from samples of five advanced ryegrass breeding populations that constitute part of a larger, composite GS reference population currently in development. Our objectives were to assess a number of GS models (single as well as two-stage models, including GBLUP and Bayesian, shrinkage and machine learning regression methods) and compare the efficacy of these predictive models for a small number of agronomically important traits. Since the GBS marker system usually results in missing data, we also considered a number of imputation methods to estimate the missing SNP values. This presentation briefly highlights the insights and difficulties associated with building such GS predictive models for genomic selection, and reports on the findings of the current study.
Agriculture
Presenter: Janusz Go?aszewski
When: Monday, July 11, 2016 Time: 4:00 PM - 5:30 PM
Room: Salon A Carson Hall (Level 2)
Session Synopsis:
The progressive meta-method and meta-analysis in application to crop production engineering
The objective of the study was to compile a procedure for a valid testing a cost-and-time-efficient crop production technology on the basis of a progressive meta-method and meta-analytic approach. The consecutive three-step procedure consisted of: 1) a survey among farmers in order to select the key agro-technical factors; 2) generation of factorial design (FD) or fractional factorial design (FFD) for the key factors, conduction of field on-station experiments; 3) on-farm FD/FFD experiments, statistical analyses for the estimation of the main and interaction effects, followed by the estimation of a proportional share of each factor in the total yield variability, as well as the economic analysis of profitability. Pea (Pisum sativum L. sensu lato) was chosen as a test crop. The assumptions for the observational studies (1st step) were to gather information on the state-of-the-art of pea production technology for fodder (dry seeds) and consumption (green pea) purposes as well as the selection of key agrotechnical factors of pea production. The survey study was carried out 243 farms in the region of north-eastern part of Poland with an area of pea production higher than 1 ha. The experimental study (2nd step) was based on the results of a two-year two single-replicated fractional factorial completely randomized designs of type 35-1, and on spatial measurements of soil properties (each plot). In the third stage, a distributed system of on-farm experiments would be built on the basis of the FD and FFDs of type 23 and 23-p (Tab.1) located at different farms while taking into account the organizational customizability of a given farm. In the conclusion it was stated that the meta-analysis approach when testing new crop production technology may be adopted by processing companies which contract feedstock from farmers because for a given feedstock the crop production technology at the contracted farms is unified and any innovation changes in production factors may be quickly and efficiently verified.
Agriculture
Presenter: Hans-Peter Piepho
When: Monday, July 11, 2016 Time: 4:00 PM - 5:30 PM
Room: Salon A Carson Hall (Level 2)
Session Synopsis:
One step at a time: Stage-wise analysis of agricultural experiments using mixed models
Many trials are replicated in multiple environments in order to broaden the inference space. For example, plant breeding and variety trials are typically performed in multiple locations and in several years. A joint analysis of such multi-environment trials (MET) can be done in a single stage by a linear mixed model for the plot data. Such an analysis is commonly considered to be the most efficient because all sources of variation can be accounted for simultaneously in a single model. An alternative method of analysis is to proceed in two stages, where in the first stage genotype means are computed per trial and in the second stage genotype means from all trials are subjected to a joint analysis. In both cases, individual trials are first analysed separately, paying due attention to all specifics of a trial, including outlier detection, the particular experimental design and randomization scheme used, and selection of a preferred analysis model among contending candidate models. In two-stage analysis, only the means and some measure of precision are saved and carried forward to the second stage. By contrast, in a single-stage analysis, the preferred analysis models identified for each individual trial are integrated into an overall model for analysis of the MET plot data, and hence the plot data are analysed a second time. Researchers wanting to analyse MET are frequently faced with the question whether to use a single-stage or stage-wise analysis. When both types of analysis are tried, it may turn out that results are not exactly the same, and sometimes the discrepancy of results appears unacceptably large. This then raises the question which analysis is preferable. In this paper it will be argued that, while single-stage analysis can justly be regarded as the gold standard, a stage-wise analysis, if done properly, is perfectly valid and typically very close to a single-stage analysis. Several papers have been written comparing single-stage and two-stage analysis. In this presentation, the key results, facts and arguments justifying a stage-wise analysis will be briefly reviewed and the important practical implications discussed. Links with network meta-analysis will be made. Several worked examples serve to illustrate the similarity between single-stage and two-stage analysis.
Agriculture
Presenter: Kenneth Russell
When: Monday, July 11, 2016 Time: 4:00 PM - 5:30 PM
Room: Salon A Carson Hall (Level 2)
Session Synopsis:
The truth is out there if you look: a barley near-infrared calibration analysis
Near infrared (NIR) calibration models are widely used to predict or describe chemical properties of a substance (response variable). Absorbance values (explanatory variable) from the NIR region of the electromagnetic spectrum are measured on samples of the substance. Peaks and troughs in the NIR spectrum are produced by absorbance of the electromagnetic energy in specific chemical bonds (usually O-H, N-H and C-H) (Hirschfeld & Stark, 1984). When the energy from a particular frequency is absorbed by a molecule, the atoms in the molecule vibrate. The nature of these vibrations results in energy absorbance at related wavelengths producing highly correlated absorbance information (Scotter, Worsfold) referred to as fundamentals, overtones and combinations. The calibration is obtained from a multivariate statistical method linking the scanned absorbance data (often pre-processed using one or more transformations (Osborne, Fearn & Hindle, 1993) with the chemical properties of a substance. Typically, in this process the absorbance values result from a single scan of each sample. A multiphase experiment was conducted to investigate the variance associated with the different stages involved in calibration from NIR spectra, to predict the amount of protein in barley grain. A linear mixed model was fitted to the data in ASReml-R to estimate the contribution to the variability by each of the design and treatment factors from the field and laboratory phases. The largest variance component (averaged across the wavelengths) was associated with the duplicate scan in the laboratory - accounting for on average in excess of 25% of the total variability. Modelling of variance in a linear mixed model framework provides a novel application in this area of chemometrics. The results from this experiment enable efficient resource allocation in the design of experiments to improve the accuracy and precision of NIR calibration models. References Hirschfeld, T. & Stark, E. W. (1984). Near-infrared reflectance analysis of foodstuffs. In G. Charalambous (Ed.), Analysis of foods and beverages: modern techniques (chap. 16). US. Osborne, B. G., Fearn, T. & Hindle, P. T. (1993). Practical NIR spectroscopy with applications in food and beverage analysis (2nd ed. ed.). Essex: Longman Scientific and Technical. Scotter, C. N. G., Worsfold, P., Townshend, A. & Poole, C. (2005). Near-infrared. Oxford: Elsevier.
Agriculture
Presenter: Raul Macchiavelli
When: Monday, July 11, 2016 Time: 4:00 PM - 5:30 PM
Room: Salon A Carson Hall (Level 2)
Session Synopsis:
Bayesian Semiparametric Mixed Beta Regression for Severity in Plant Disease
Severity in plant diseases is typically measured as the proportion of plant material damaged, and the severity index (SI) can be expressed in a 0-1 scale. In order to study how severity changes as the plant grows, and to compare different conditions, it is measured repeatedly during the season (for example, weekly). Existing approaches for the analysis use semiparametric models with normal response after a logit transformation to analyze this type of data. However, this approach could be inefficient and inappropriate due to the asymmetric distribution of the response even after transformation. We propose a Bayesian semiparametric mixed beta regression model to describe the severity disease progress. The proposed model allows the response variable to follow a beta distribution within the Bayesian paradigm. An advantage of the Bayesian approach is that it facilitates the comparison of curves across time. We applied the proposed models to study severity in Black Sigatoka, a fungal disease on banana crops, under different treatments.
