Analysis
The same figures, with their uncertainty
The explorer pages report what CBS published. These pages add the statistics around them: intervals on every estimate, sample sizes stated, paired comparisons where two periods are compared, and effect sizes rather than bare p-values. The published figures are not changed; where an analysis needs a different basis (real dollars, ratios to the state rate) it says so.
Gambling Help Line 1800 858 858 or gamblinghelponline.org.au.
- STL, interrupted time series, paired monthsTrendsAt reopening in July 2020, real monthly NGR was $9.3m above its pre-closure path (95% CI +$5.8m to +$12.8m, n = 116 months) and its trend steepened by $3.4m a year.The first full year after reopening averaged $6.61m a month more than FY 2018/19 (paired over 12 calendar months, dz = 1.01).
- Funnel plots, ratios to the state rateCouncils39 of 48 areas sit outside the year-to-year 95% limits in FY 2024/25 (81%), where chance alone would put about 2.4: the differences between councils are lasting, not small-number noise.Of 55 areas with three or more years, 14 are consistently above the state rate and 34 below, with intervals widened for year-to-year dependence.
- Bootstrap intervals, change-point checkConcentrationManufacturer concentration stopped falling in Dec 2015 (95% CI Aug 2015 to Jun 2016). It fell by about 224 points a year until then and has risen by about 36 a year since.Moving-block bootstrap of the residuals in 20-month blocks (chosen from their autocorrelation), re-fitting the break in every resample.
Reproducible
Every resampled interval uses a fixed seed (20090701, the first month of the series), so the pages show the same numbers on every build. The statistical helpers are unit-tested against scipy, statsmodels and R, and the scripts that produce those reference values are in the repository.
Methods and decisions
How each analysis was set up, what it assumes and what I would change are on Methods, with decision records for the choices that matter (combined council groups, stock versus flow totals, the missing FY 2014/15 and the model specifications).