Rote analysis vs. snooping

There’s a whole website about this

The best you can

• Identify scientific questions

• Distinguish between exploratory and confirmatory analysis

• Pre-register studies when possible

• Keep an exploration and analysis journal

• Explore predictors and responses separately at first

• Look at location and shape

• Maybe with different sets of grouping variables

• Contrasts
  • Parametric vs. non-parametric
  • Exploratory vs. diagnostic
  • Data vs. inference

• Built in to R

• Code

• Rental data

• Weather codes

• Code
  • Merge
  • Plot

• In general:
  • If your logged data span \(<3\) decades, use human-readable numbers (e.g., 10-5000 kilotons per hectare)
  • If not, just embrace ``logs’’ (log10 particles per ul is from 3–8)
    • But remember these are not physical values
• I love natural logs, but not as axis values
  • Except to represent proportional difference!

• See Rauser video

• This can work surprisingly well

• Bigger is not always clearer

• Since the point is to make patterns visually clear, trial-and-error can be OK here

Code (with built-in data)

• Data on lung capacity of schoolchildren

• Import data

• Show the confounding

• Scatter plots

• Density plots

• Depending on how many data points you have, scatter plots may indicate relationships clearly

• They can often be improved with trend interpolations
  • Interpolations may be particularly good for discrete responses (count or true-false)

• Local smoother (locally flat, linear or quadratic)

• Neighborhood size given by alpha
  • Points in neighborhood are weighted by distance

• Check help function for loess

• Loess is local, but not robust
  • Uses least squares, can respond strongly to outliers
• R is has a very flexible function called rlm to do robust fitting
  • Not local
  • But can be combined with splines

• Contours
  • use _density_2d() to fit a two-dimensional kernel to the density
• hexes
  • use geom_hex to plot densities using hexes
  • this can also be done using rectangles for data with more discrete values

• Use clear gradients

• If zero has a physical meaning (like density), go in just one direction
  • e.g., white to blue, white to red
    • or red to white with red borders (heat map)
  • If the map contrasts with a background, zero should match the background

• If there’s a natural middle, you can use blue to white to red, or something similar

• Three dimensional data is a lot like two-d with densities: contour plots are good

• Pairs plots: pairs, ggpairs

• iris code with built-in data

• Use boxplots and violin plots

• Make use of facet_wrap and facet_grid

• Use different combinations (e.g., try plots with the same info, but different factors on the axes vs.~in the colors or the facets)