Companion Animals Symposium I
Statistics for Reproducible Research in Companion Animal Nutrition
Presented by Nora M. Bello, DVM, MS, PHD- Kansas State
Nora Bello of Kansas State presented a riveting discussion focused on “Statistics for Reproducible Research in Companion Animal Nutrition”. The major topics of the discussion explored current understanding of P-values and hypothesis testing with a lens of companion animal nutrition and the associated argument that statistical analyses must be fully understood by the user. Bello began by addressing the importance of understanding that a P-value is a measure of incompatibility between observed data and working assumptions but does not atone for how correct the research is. To interpret the probability value (p-value), one must acknowledge that the p-value falls on a continuum from 0-1 in which the closer the value to 0 data is, the more surprising the results are which is then interpreted into the probability leading to the rejection of a null hypothesis. The current standard in research, as it pertains to a p-value, establishes significance at 0.05 which correlates to a 1in 20 chance of occurring. Bello established that this numerical assignment of what is considered significant is arbitrary since small p-values can be obtained by increasing the sample size, which brings us the conjecture that statistically significant results may not be relevant. However the reverse is not necessarily true: that claiming practical significance or numerical differences in the absence of statistical significance is “bogus” as it prevents us from hearing about significant events that are not due to chance. Bello makes the point that absence of evidence does not equate to the evidence for absence with the example that “no evidence for effects is not the same as treatment has no influence”. With nuances in research and the use of data from one experiment to answer multiple questions, researchers slip into type 1 and type 2 error. As a scientific community it is our goal to control error to keep the error at a minimum. In the context of companion animal nutrition, typical experimental setups include 2x3 factorials with numerous response variables over a relatively long timespan leading to a minimum of ~270 pair wise comparisons which increases the probability of type 1 errors. Increasing the probability of type 1 error requires adjustments for multiple comparisons which account for that type 1 error. In summary, statistical analysis is a powerful tool that requires its users to learn when to use varying methods and what methods are best suited for their data set.
Analytical Strategies for Survey Data in Pet Nutrition and Management
Presented by Sandra L. Rodriguez-Zas, PhD University of Illinois
Sandra Rodriguez-Zas from the University of Illinois spoke on analytical strategies for survey data in pet nutrition and management as part of Thursday mornings companion animal symposia. Survey data collection focuses on experiments which complement survey studies due to the nature of survey-based research which requires you to travel to populations to collect qualitative data focused on individual views and opinions. These collection processes vary in nature and can range from in-field in-person surveys, online, or mail-in questionnaires. This form of sample collection can be challenging as it can be a random, opportunity, or selective sample. A benefit of this form of data collection is the opportunity for large scale samples which can represent the population to a larger extent. Unfortunately, you lose control of distribution conditions or response rates which makes sample representation of the population variable. With this variability, analysis of complex survey data requires various methods of analysis. Rodriguez-Zas discusses four methods of analysis which included sampling weights, strata, survey data concep4ts, and higher population correction. Survey weights are not always highly representative of the population imposing the use of weights on the data so the sample approaches population representation. Weight implementation which was referred to as probability weight is the invers of the sampling fractions (n/N). Strata, as a form of analysis, is best suited for non-overlapping groups because measurements are homogeneous within the strata and are independent. Estimations using this method are across the strata and tend to reduce variance of estimates. Survey data concepts is best suited for sampling units and clusters with the primary sampling unit at first level, but it is important to note that clustering in this method will impact the estimate variance. Finally, Rodriguez-Zas discussed finite population corrections. FPC is best suited for samples that are a substantial fraction of the population. When the FPC is higher (>5% of the population) a lower variance of the sample mean can be seen. There are numerous software packages suitable for complex survey designs such as R and SAS that utilize procedures to identify variance of estimates. As research moves forward, Rodriguez-Zas recommends careful considerations of questionnaires and content with regard to complex survey analysis to minimize misinterpretations, keeping in mind inclusion and exclusion criteria as it pertains to the population.
The Use of Meta-analysis in Companion Animal Research
Presented by Emma N. Bermingham of AgResearch
Emma Bermingham, a representative of AgResearch, provided a detailed analysis of the use of meta-analysis in companion animal research. She began by addressing what all animal scientists are constantly questioned on: the social license for animal use which follows the 3 R’s of research. In a constantly evolving world, science is also at a pivotal point of evolution. In recent years we have seen an increase in data-heavy research such as sequencing technology research. Unfortunately, new reductions and restrictions in funding in addition to the unprecedented COVID pandemic have imposed unplanned limitations in larger animal studies which leaves us with the question of how can we get more out of existing research? Bermingham addresses two methods to answer the question. First a literature review could enhance knowledge of varying areas, adding to the breadth of knowledge available; but literature reviews are broad with no specific question. Additionally, literature reviews are qualitative assessments which can be heavily biased towards the author of the review compared to a systematic review, which is the second method discussed. Systematic reviews are more detail based, relying on a 2+ person screening process prior to data extraction. Meta-analysis furthers the work of systematic reviews by adding statistical analysis to the data set extracted. Implementing the preferred reporting items for systematic reviews and meta-analysis (PRISMA) guidelines assist researchers by quantifying and recording resulting in a clearly defined list of publications that can be used for statistical analysis. Bermingham challenges the audience by providing the question of how do you classify your data when conducting a statistical analysis? Specifically, what are the parameters that are likely to influence an observed result? It was concluded that statistical analysis undertaken depends on the question that is being asked, which should always be assessed for normality. Using a meta-analysis is a useful tool for presenting the gaps in the literature. But the gaps found using meta-analysis require a researcher to make assumptions of the data depending on what questions are being asked. Further implications include how data is published which can adversely affect whether the data can be used in future systemic reviews or meta-analyses’.
Nutrition Modeling: What Can the Pet Field Learn (or Steal) from Recent Directions in Other Species?
Presented by Jennifer L. Ellis PhD from the University of Guelph
Jennifer Ellis representing the University of Guelph finished the companion animal symposia with a glimpse into future research specifically on nutritional modelling and what the pet field can learn from recent directions in other species. Animal science, as an industry, has a long and established history of model application which has been used in the field as decision support and for opportunity analysis. Model application has its purposes in numerous fields especially academia as knowledge synthesis, hypothesis generation and testing, and knowledge transfer between research, academia, and government. The lines between model types have often been blurred, since research is seldomly a one answer question, but can be broken down into three overlapping categories: dynamic (measuring change over time) or static (measuring a stationary time), mechanistic (based on biology) or empirical (based on relationships in data), and stochastic (based on model randomness or variance) or deterministic (based on single variable). The major directions of modeling in animal nutrition implement statistical modelling (meta-analysis), mechanistic models used to describe biological processes and provide decision support which represents the animal scientists’ movement away from traditional requirement models. This shift towards statistical modelling allows for the incorporation of response variables, artificial intelligence and machine learning, and application of a hybridization to the above areas of statistical modelling described. There are advantages to empirical approaches including smaller numbers of inputs, which are ideal for identification of major drivers per given response and easy practical application. Unfortunately, empirical approaches have a limited scope and are database dependent. Moving past statistical modelling, mechanistic modelling is proving effective with problem solving by providing advantages in the improvement of underlying biological mechanisms, interplay with experimental research to advance the field, and forecast outcomes in scenarios not yet seen/isolated in practice. But mechanistic modeling has its drawbacks because they are slower to develop, generally require a larger number of inputs, and require training which increases complexity of the tool employed. The modern world requires modern solutions. With innovations in machine learning and artificial intelligences the future of research may rest with artificial intelligence and “big data”. AI and big data represent a whole knowledge field focused on the development of computer intelligence and knowledge acquisition set to make predictions that are data driven. With AI in the near future, Ellis suggests the addition of two categories to the current list of model types. These categories would be specifically incorporating AI and machine learning into research modelling and include supervised (output variable known) or unsupervised learning (no labelled outputs with the goal to infer the natural structures present within the data) and continuous data (individual numerical data points) or discrete data (categorial data). A niche has formed for machine learning and AI which fall under the areas of animal science focused on animal and environment monitoring, disease detection, health event detection, performance prediction and forecasting. While promising, limitations to AI and machine learning must be addressed. Ellis addresses that a lack of transparency in data processing and training is required for the present group of individuals (who currently lack a broad sense of skills in machine learning and algorithm readings) limit the potential of these systems. Ellis closes with the thought that now that we have big data, how do we generate value with it? The future may be a form of hybridization of AI and mechanistic modelling to create “intelligent” precision nutrition. Regardless as the world moves into a more technologically savvy era, expanding your digital skill set and exploring how modelling could contribute to and strengthen your literature review will benefit your scientific endeavors.