Pitfalls in Experimental Design and Data Analysis
Keith Baggerly · Clinical Cancer Research · 2006
PL-26 High-throughput molecular assays have the potential to provide us with dozens of new disease markers, which in turn can be used to improve therapy. These markers are often found by using advanced tools to analyze the data, as these tools are very good at finding structure. However, tools that are sensitive enough to pick up subtle shifts in biology are also sensitive enough to pick up shifts due to other causes.In this talk, we will explore the continued need for zeroth-order data analysis, ensuring that potentially large but unwanted effects are excluded before subtle effects are sought. To do this, we will take a pictorial tour through the raw data from case studies, looking for interesting structure both in a single experiment and over mulitple experiments. Unfortunately, in these cases what the data most clearly shows is not biological structure, but rather the need for careful expeimental design, data cleaning, and data preprocessing to ensure that the structure found is not due to systematic bias.These issues are not new, but they remain important as we struggle to validate and regulate new markers.