The Grammar of Graphics

Abstract

Statistical graphics is, or should be, an transdisciplinary field informed by scientific, statistical, computing, aesthetic, psychological and sociological considerations. The sciences (natural, applied or social and indeed also various literary, linguistic and historical disciplines) generate data relevant to various problems that are amenable to graphical display. Statistics advises or instructs on what should be plotted. Software and hardware are needed to produce displays in practice: very few would now reach for their pens and graph paper as was still common even some twenty or thirty years ago. We clearly should ponder how far graphics are attractive or not, and especially how far graphics work or not, in terms of being interpreted appropriately by users. But sociological considerations also? I suggest that a full appreciation of the situation of graphics within statistical science depends on recognising and trying to explain a variety of attitudes and behaviours. 'Statistical science' is understood here to range from the science of statistics to any part of science done statistically. Graphics is prominent in almost every introductory course or text and central to a large part of statistical practice. Indeed, I regularly attend scientific presentations largely innocent of p values, t statistics, confidence intervals or even figures of merit such as R 2 or AIC, but which include curve or model fits kept honest, mostly, by intense graphical analysis, large datasets and attention to the underlying science.

Comparison of causes of stillbirth and child deaths as determined by verbal autopsy and minimally invasive tissue sampling

“…After the respective underlying causes of death that DeCoDe assigned were mapped and matched to its corresponding verbal autopsy standard, the agreement of both methods was evaluated using their concordance and accuracy of CSMF. We compared the CSMF of InterVA4 against DeCoDe using Lin’s Concordance Correlation Coefficient (LCC), [ 42 ] which was calculated using a user-defined command made for Stata–“Concord” [ 43 ].…”

“…The program (“Concord”) produces the LCC by multiplying the “Pearson correlation coefficient, r” with the bias-correction factor. Whereas the “Pearson correlation coefficient, r” is the measure of precision, the bias-correction factor is for accuracy [ 43 ].…”

Practical tools for exploring data and models

“…Wilkinson's grammar successfully describes a broad range of graphics, but is hampered by a lack of an available implementation: we can not use the grammar or test its claims. These issues are discussed by Cox (2007), which provides a comprehensive review of the book.…”