Dr Clair Alston-Knox
Assoc Prof Sama Low-Choy
The session. This pair of seminars address the issue of "sustainable statistics", and in particular the "small data" situation researchers are in before they can justify the outlay and resources to conduct a large study. We allocate 90 minutes to the whole session: each seminar will last 20-30 minutes, each followed by a 10-15 minute discussion time.
Perils. For the first seminar, we welcome back Adjunct Senior Lecturer Clair Alston-Knox, now Senior Statistician in the Predictive Analytics Group. She will be addressing the perils of low power, i.e. relying on very small samples (e.g. n=6), particularly when statistical inference revolves around p-values. "In recent years, we have seen many published and publicized papers that were essentially dead on arrival because they were attempting to identify small effects in the presence of noisy and often biased data. In such research projects, even findings that are statistically significant are likely to be in the wrong direction and can massively exaggerate effect sizes (Button et al, 2013, Nature Reviews in Neuroscience; Gelman & Carlin, 2014, Perspectives in Psychological Science)" -- Gelman (2017, CHANCE). It is an age old question in statistical analysis: How many units do I need to sample to yield a reliable statistical analysis? It is surprisingly difficult to answer, and is maybe best considered in the context of the consequences of sampling too few study units. In this talk we will examine common approaches to sample size estimation and talk about methods, such as power analysis, to help researchers conduct studies with an acceptance of the level of uncertainty their sample size provides. We will introduce some tools and resources that could greatly assist researchers examine and justify their pre-experimental sample size decisions.
Possibilities. The second seminar, is presented by A/Prof Samantha Low-Choy. Her talk will focus on the possibilities for statistics with small data. In particular, small data can be crucial when considered within a wider context of a series of studies. Small data may reveal the stories underlying the complex relationships amongst multiple variables, for instance in structural equation models, Bayesian networks or Bayesian regressions (e.g. Low-Choy, Riley & Alston-Knox, 2017, Educ Media Internat). Small data may also complement big data by ground-truthing predictions (from models or expert assessments), which is fundamental to triangulation and measurement error analysis (e.g. Caspi & Friebur, 2016, Internat J Behav Nut Phys Act). Small intensive data collection can be designed to distil expert knowledge, as a basis for input into risk assessment or as priors in Bayesian analyses (see reviews Low Choy et al, 2009, Ecology; van der Schoot et al, 2014, Child Development). The Bayesian statistical paradigm has a different logical and mathematical foundation, which makes it possible to deal with small data, albeit with care (e.g. Western & Jackman, 1994, Amer Political Sci Review).