Diagnostic Plots for One-Dimensional Data
Günther Sawitzki · Contributions to statistics · 1994
How do we draw a distribution on the line ?We give a survey of some well known and some recent proposals to present such a distribution, based on sample data.We claim: a diagnostic plot is only as good as the hard statistical theory that is supporting it.We try to illustrate this point of view for some examples.Though the general contribution of diagnostic plots to statistics is accepted, sometimes diagnostic plots seem more of a fashion than a tool.There are uncountable possibilities to design diagnostic plots, not all being of equal use.Diagnostic plots can and should be judged the same way as any other statistical method.We have to ask: What is their power ?What is their reliability ?While we may have to stay with accidental notes or examples for some time now, in the end a diagnostic plot is only as good as the hard statistical theory that is supporting it.For many diagnostic plots, we are still far from having this theory.For some plots, we have to ask: What precisely are they trying to diagnose ?How do we judge their reliability or confidence ?For other plots, we know at least the statistical methods they are related to.In this paper, we consider plots as views on probability measures: We relate plots to functionals, operating on probability measures.If we have a functional defining a plot, we can proceed in three steps.We can ask which features are exhibited by the functional, and which are collapsed.As a second step, we can analyze what is retained by the empirical version and which stochastic fluctuation is to be expected.Third, we can optimize the functional and its empirical version to gain maximal power.Where possible, we try to indicate classical tests related to the plot.If these tests meet the core of the plot, the power of the plot may be identified with and judged by the power of these tests.The associated functional may even indicate a notion of distance, or a metric, associated with the plot.We can use this to find natural neighbourhoods of a given empirical plot, leading to confidence sets of compatible models.We restrict ourselves to a very modest case: Assuming a continuous distribution on the real line, we look at diagnostic plots based on a sample from this (unknown) distribution.We exclude some of the more difficult questions: we assume independent sample points with identical distribution.So we do not look at plots for the diagnosis of dependency, trend, heteroskedasticity or other inhomogeneities.We give a survey of some diagnostic plots, pointing to their related statistical methods.As is to be expected, the chance is taken to advocate some new plots: the silhouette, the densitogram, and the shorth plot. Diagnostic plots, what do we need them for ?We use diagnostic plots to investigate a data set by itself (a descriptive problem), or in comparison to a model distribution, or family (the one-sample-problem), or to compare two data sets (the two-sample-problem).In a classical framework, we may want to apply a certain method, like regression or analysis of variance.This method will depend on certain conditions, often on conditions which in principle cannot be verified.The role of a diagnostic plot is that of a detector: Since we are unable to verify the preconditions, we may use diagnostic plots to have at least a warning instrument.
Cited by 5 works. doi:10.1007/978-3-642-57991-2_14
Authors
References
- An Analysis of Variance Test for Normality (Complete Samples)
- Density Estimation for Statistics and Data Analysis
- Journal of the American Statistical Association
- X. On the criterion that a given system of deviations from the probable in the case of a correlated system of variables is such that it can be reasonably supposed to have arisen from random sampling
- The Future of Data Analysis
- Journal of the Royal Statistical Society.
- Variations of Box Plots
- The Visual Display of Quantitative Information.
- On the Choice of the Number of Class Intervals in the Application of the Chi Square Test
- Using Kernel Density Estimates to Investigate Multimodality