Two-lines test

(changes to the app)

How has the app changed?

v0.52 - 2018 11 23
  • Modified how fitted values are computed for plotting the scatterplot smoother. In particular, rather than fitting at mean values of covariates, as done until now, data are fitted for the "median" value of the covariate; in the case of a factor (i.e., categorical) variable, it is the value of the factor at the n/2 observation, after sorting. For example, if z=c("doctor", "nurse", "artist") the fitted value would be for z="doctor", the middle observation once the factor is sorted. This change follows the observation by Conor Wild from U of Waterloo that the app would give an error message when used with factor variables.

v0.51 - 2018 08 27
  • Fixed bug where missing values for one of the variables would cause the program to crash.

v0.5 - 2018 06 26
  • Modified template figure, it now shows results for two-lines as arrows with average slopes instead of as fitted regressions lines. This seeks to avoid confusion about what the two-lines test does. It does compute average slope, it does not seek to model the data.

v0.34 - 2018 04 20
  • Extended previous fix to models with covariates

v0.33 - 2018 04 02
  • Fixed bug that would lead the code to crash if the dependent variable was not named "y" (!).
    Thanks to Joseph Reiff (UCLA PhD student) for identifying the bug and the line that was generating it

v0.32 - 2018 03 20
  • Instead of relying always on HC1 for hetorskedasticity correction, goes back to 'HC3' by default and switches to HC1 only if a NA value is generated.

v0.31 - 2018 03 20
  • It was realized that using HC3 to estimate heteroskedasticity robust standard errors creates NA values under some circumstances leading the app to crash. For a few hours the app was switched to relying on HC1 instead HC3. But see App 0.32.
    Special thanks to Nathan Carter for sharing with me the data that were leading the app to crash, allowing me to identify the bug.
  • Fixed bug counting number of unique x values (used to prevent the mgcv:gam() function from crashing when x has few possible values)

v0.3 - 2018 01 24
  • Rewrote most of code to simplify syntax. Intead of u(x) to indicate the variable that is being tested for u-shapedness, now always the first variable is.
  • Moved from ShinyApp to standalone PHP server, running R in the background (this app is no longer a ShinyApp)
  • The interrupted regression now computes heteroskedasticity robust standard errors, addressing concerns shared via Andrew Gelman by Yair Heller. For more information read post-scriptum to the blogpost DataColada[62], or go straight to the R Code discussing the problem and solution.

v0.1 - 2017 ?? ??
  • Details lost to history