An Improvement to Coefficient Plots
I recently posted about coefficient plots, discussing my approach and providing some example R code to create the graphs. I had the good fortune of hearing Amanda Driscoll give a talk recently, and she made a small, but really nice improvement to her coefficient plots. I refer to the improved version of the plot as a "Driscoll plot."
The key idea is to highlight the area in which the researcher predicts the parameter lies. I've created an example an based on the simple logistic regression model. When creating this stylistic example, I have in mind a researcher testing the hypothesis that partisan strength increases the probability of turning out to vote. However, (for better or worse) researchers often present hypotheses about many of the explanatory variables in their regression models.
While studying this plot, the reader can quickly note five things.
- The size of the effect as each explanatory variable moves from its minimum to its maximum, holding all other explanatory variables constant.
- The relative size of the effects.
- The uncertainty around the effects.
- Whether the effect is statistically significant.
- Whether the researcher has support for the hypotheses.
For example, this plot indicates that the researcher predicted that education, union membership, and partisan strength would have positive effects. On the other hand, the researcher predicted that being African-American and female would have negative effects. From the plot, we can see that the researcher has strong evidence for her claims about education, union membership, and partisan strength, and little evidence for her claims about race and gender.
See the R code to create this plot here (P.P.S. Michael Bishop created a github page for the code here). P.S. As Luca and Stephano point out in the comments, the code requires the arm library, but doesn't load it. Be sure to add library(arm) at the top of the script.