Trend or noise?

The three questions to ask any trend

1. How steep — and in what units? Our gage trends express each year as a percentage of that gage’s own seasonal median, so the slope is in points of that percentage per decade — comparable across a mountain creek and a plains river where raw cfs would not be.

2. Could noise alone produce it? That is the p-value, from a Mann–Kendall test. We rank findings by p-value, never by slope size — short, gappy records produce spectacular slopes by accident, so ranking by steepness is an artifact detector. The loudest-looking trend is usually the least trustworthy; the boring 90-year record with p < 0.01 is the finding.

3. Trend, or step? Rivers often change abruptly — a reservoir fills, an exchange starts, a diversion moves. A Pettitt changepoint test asks whether the record is better read as before/after than as a slope. A “decline” that is actually a 1957 step means something different for every argument built on it.

See the whole state ranked this way

The statewide Trends page is this lesson applied 2,300 times: click any row and the chart behind its numbers opens — Sen’s slope, the changepoint, the segment means. If a chart and a statistic ever disagreed, one of them would be wrong; they are drawn from the same series so they cannot.

What this does not let you conclude

Statistical significance is not attribution: a real declining trend does not say whether climate, pumping, administration or an upstream project caused it, and a non-significant result is absence of evidence, not evidence of stability. Advisory reading aid on public data — never an administrative or legal conclusion, and never a substitute for DWR administration or an engineer’s analysis. Our methodology has not yet had outside expert review: how this is built.