Diagnostic Plots for Fitting Distributions

Thomas Roh

December 17, 2017

The fitur package includes several tools for visually inspecting how good of a fit a distribution is. To start, fictional empirical data is generated below. Typically this would come from a real-world dataset such as the time it takes to serve a customer at a bank, the length of stay in an emergency department, or customer arrivals to a queue.

set.seed(438)
x <- rweibull(10000, shape = 5, scale = 1)

Histogram

Below is a histogram showing the shape of the distribution and the y-axis has been set to show the probability density.

dt <- data.frame(x)
nbins <- 30
g <- ggplot(dt, aes(x)) +
  geom_histogram(aes(y = ..density..), 
                bins = nbins, fill = NA, color = "black") +
  theme_bw() +
  theme(panel.grid = element_blank())
g

Histogram vs Density Plot

Three distributions have been chosen below to test against the dataset. Using the fit_univariate function, each of the distributions are fit to a fitted object. The first item in each of the fits is the probabilty density function. Each fit is overplotted onto the histogram to see which distribution fits best.

dists <- c('gamma', 'lnorm', 'weibull')
multipleFits <- lapply(dists, fit_univariate, x = x)
plot_density(x, multipleFits, 30) + theme_bw() +
  theme(panel.grid = element_blank())

Q-Q Plot

The next plot used is the quantile-quantile plot. The plot_qq function takes a numeric vector x of the empirical data and sorts them. A range of probabilities are computed and then used to compute comparable quantiles using the q distribution function from the fitted objects. A good fit would closely align with the abline y = 0 + 1*x. Note: the q-q plot tends to be more sensitive around the “tails” of the distributions.

plot_qq(x, multipleFits) +
  theme_bw() +
  theme(panel.grid = element_blank())

P-P Plot

The Percentile-Percentile plot rescales the input data to the interval (0, 1] and then calculates the theoretical percentiles to compare. The plot_pp function takes the same inputs as the Q-Q Plot but it performs on rescaling of x and then computes the percentiles using the p distribution of the fitted object. A good fit matches the abline y = 0 + 1*x. Note: The P-P plot tends to be more sensitive in the middle of the distribution.

plot_pp(x, multipleFits) +
  theme_bw() +
  theme(panel.grid = element_blank())