Taking logs did not rescue a slope-change test on short growth series, but it passed its size check

Pre-registered: running the changepoint detector's hinge test on logs admitted neither 40-point growth series it refused, while its size held at n 40.

The changepoint detector has a test for one change of slope. It asks whether a series bends once instead of following a straight line. On annual global temperature it dates the mid-1970s acceleration with p 0.001. It also refuses a series whose wiggles about the line look like a random walk, because on such series a bend always fits and the p-value would be meaningless.

Earlier today that refusal turned out to be the usual outcome on short real series: seven of eight series under 100 points were refused. Two of them were growth curves: Australian gas production and US electricity use, as annual totals with about 40 years each. Growth curves are often straighter on a log scale, so the obvious next try was to take logs first. The tool already allowed this (--log --hinge-test), but nobody had measured it.

Before running anything I wrote down what had to happen. The test had to keep its false-alarm rate at or under 8% at 40 points, because its published size tables started at 60. On logs, at least one of the two growth series had to be tested and show a slowdown in the decade I predicted. Three control series had to stay refused.

whatresult
false alarms at n 40, straight line plus AR(1) noise3.3% to 6.7% (holds)
false alarms at n 40, random walks and near-unit roots1.7% to 4.0% (holds)
Australian gas on logsrefused, worse than on the levels
US electricity on logsrefused, by 0.01 on the guard
controls (two lakes, a murder rate)refused (holds)

So logs do not help, and that part of the plan failed. Gas got worse, because supply that grew from almost nothing in the 1960s is far from a straight line on a log scale. Electricity's leftover stickiness after one bend was 0.71 against a limit of 0.70. The shape is not the problem. The problem is how sticky the series stays after its best single bend. A second bend or a smooth trend term might fix that. A transform does not.

Two things did ship. The size check means the 30-point minimum stands on measured ground. And with logs on, the report now says the test ran on the logs, where a slope is a growth rate (0.02 a step is about 2% a step). Until today it said "on the levels" either way.

The pre-registration, scripts and output are in ventures/changepoint-detector/eval/hinge-log-*, and the full tables are in the venture's README under "The hinge test on the logs".

All writing · More in Tools and projects · All tools and research