Description
Before modelling counts, plot simple functions of the frequency table to see whether a binomial, Poisson or negative binomial shape fits, then test variance against mean.
Count data such as errors per session, tickets per user or visits per week are often modelled as Poisson by default. Ord's plot checks this cheaply: for frequencies f(0), f(1), ... plot r f(r)/f(r - 1) against r. A roughly flat line suggests Poisson, a falling line binomial and a rising line negative binomial; Klugman and colleagues use the same ratio plot for their (a, b, 0) class. Hoaglin, Mosteller and Tukey's Poissonness plot graphs log(r! f(r)/N) against r; for Poisson data it is straight with slope giving the log mean, and it resists outlying cells better than ratio plots. Ord's diagram places a sample, from its first three moments, among regions for several families. The index of dispersion test then formalises the variance check: (N - 1) times the variance over the mean, against chi-squared on N - 1 degrees of freedom.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Latent
- Stage
- Evaluation, Post-launch
- Data
- Quantitative
- Effort
- Hours
Steps
- Tabulate how many units have each count value 0, 1, 2 and so on.
- Compute r f(r)/f(r - 1) for low r, smoothing adjacent values if noisy, and plot against r; skip ratios where a frequency is zero.
- Read the slope: about flat suggests Poisson, falling binomial, rising negative binomial.
- Optionally draw a Poissonness plot of log(r! f(r)/N) against r, levelled by the sample mean, to spot outlying cells.
- Test dispersion with the index (N - 1) s squared over the mean against chi-squared on N - 1 degrees of freedom.
- Fit the shortlisted distribution and verify with a Pearson chi-square or empirical distribution function test before using it in regression or simulation.
Sources
- Univariate Discrete Distributions (Johnson et al., 1992)
- Loss Models: From Data to Decisions (Klugman et al., 2012)
Also known as
- Ord plot
- Ord ratio plot
- (a, b, 0) ratio plot
- Poissonness plot
- Hoaglin-Mosteller-Tukey count metameter plot
- Ord's (S, I) diagram
- Binomial probability paper
- Index of dispersion test
- Variance-to-mean test
- Graphical and test-based choice of a count distribution: Ord ratio plot, Poissonness plot and index-of-dispersion test
Categories
- Analytics & Behaviour
- Statistics & Modelling