What can scatterplots teach us about doing data science better?

Abstract

A scatterplot is often the graph of choice for displaying the relationship between two variables. Scatterplots are useful for exploratory analysis, but can do much more than just identifying correlations. As data sets get larger and more complex, relying solely on “eye power” alone may cause us to miss interesting associations, or worse, make wrong interpretations. We show that by combining scatterplots with statistical and logical reasoning (the sliding window and two-axis median bisection), we may identify interesting associations in a case study of Graduate Record Examination admission versus graduation outcomes, and whether low detectability of proteins in a biological sample are truly associated with low abundance. Due to subjective visual interpretability, we recommend graphing the data using a multitude of visual variables and graph types before concluding the absence of an association. Finally, even if associations are demonstrable, developing causal models that could explain the observed fuzziness and lack of apparent correlations in the scatterplot are helpful for better decision-making and interpretation.