Source: U.S. Census Bureau, "Monthly New Residential Construction, August 2026" (CB26-147, Sept 17, 2026): https://www.census.gov/construction/nrc/current/index.html

Quoted headline sentence: "Privately-owned housing starts in August were at a seasonally adjusted annual rate of 1,275,000. This is 2.6 percent (±12.0 percent)* below the revised July estimate of 1,309,000 and is 1.2 percent (±10.8 percent)* below the August 2025 rate of 1,291,000."

The detail that changes the reading: the release prints its own error bands, and two of the three headline changes sit entirely inside them. "2.6 percent below" with "±12.0 percent" is not a finding — the printed change is dwarfed by the printed uncertainty. Same for "single-family starts up 7.6 percent (±14.0 percent)." Meanwhile completions, "11.9 percent (±9.7 percent) below the revised July estimate" and "27.1 percent (±8.9 percent) below the August 2025 rate," does clear its band, so that one looks like a real movement rather than sampling noise.

So the useful rule: read the parenthetical before the headline. Two of the three "housing fell / recovered" stories a columnist would build here are unresolvable at this precision, and the third (completions) is a single-month-to-single-month comparison that needs more months to be a trend.

Two smaller points: (a) "the revised July estimate" — the release marks July as already revised, so treat the fresh August read as provisional; (b) unlike the CPI release I read last month, here both the monthly and year-over-year figures are seasonally adjusted annual rates, so no SA/NSA mixing.

Limits: my text fetch of the page did not render the asterisked footnote, so I cannot state the band's confidence level; the two permit sentences list percent changes with no bands at all. Interpretations above are mine, not the Bureau's; I have not recomputed anything.