Exploratory Data Analysis

Historical Development

The seminal works on exploratory data analysis come from John W. Tukey (1977) and Frederick Mosteller and John W. Tukey (1977). Tukey (1977, p. vii) noted that over the 20th century, confirmatory data analysis, which assesses how precisely sample statistics can be used to make inferences about population parameters, had come to dominate statistical research. He believed that although exploratory and confirmatory data analyses emphasize different principles, both should be used by practicing researchers. In aiming to complement the dominant methods of statistical analysis, he argued that researchers cannot get along without confirmatory data analysis but need not start with it. His central principle follows from this point (Tukey, 1977): “It is important to understand what you CAN DO before you learn to measure how WELL you seem to have DONE it” (p. v).

Consider the differences between confirmatory and exploratory data analysis (NIST/SEMATECH, 2003):

• Confirmatory approaches begin with a prespecified model and use the parameters obtained from the data analysis to evaluate the model, whereas exploratory approaches begin with analysis of the data to infer the model and specify its assumptions.
• Confirmatory approaches are formal and rigorous in efforts to evaluate a model, whereas exploratory approaches are more informal and active in the effort to discover meanings contained in the data.
• Confirmatory approaches impose a model on the data, whereas exploratory approaches let the data suggest appropriate models.
• Confirmatory approaches filter the data in reducing the information to a small number of parameters, whereas exploratory approachesrelyon techniques that reflect all parts of the data.
• Confirmatory approaches require strong assumptions, whereas exploratory approaches make few assumptions or attempts to validate assumptions before model testing.
• Confirmatory approaches have much power to precisely test a hypothesis and estimate parameters under narrow circumstances, whereas exploratory approaches have greater generality and less sensitivity.

[Page 360]

The two approaches represent ideal types that, in good research, both receive attention, but Tukey and others have had considerable influence in countering a belief that mere description and exploration of data were inferior to formal hypothesis testing and model confirmation. Today, most empirical social scientific research relies on confirmatory models, such as in the use of regression and analysis of variance to test theories, but exploratory techniques have become essential parts of the research process leading to the confirmatory models.

...