Starting in the mid-1980s, most US weather stations traded their liquid-in-glass thermometers for electronic sensors — and wrote down the date. Compare the 385 cleanly measurable switches against unchanged neighboring stations, and the size of the bias falls straight out of the data.
In the companion piece to this page (Two ways to erase a hot day) we simulated the effect of the 1980s instrument transition on hot-day counts. Here we measure the real thing. NOAA's station metadata archive records the equipment history of the volunteer network, including the date each station's Cotton Region Shelter and liquid-in-glass thermometers gave way to the electronic Maximum–Minimum Temperature System (MMTS). Among the 1,218 long-record stations of the US Historical Climatology Network (USHCN), around three in four (920) have a documented switch to MMTS — nearly all between 1982 and 2005, and most clustered in 1984–1986. (The remainder mostly kept liquid-in-glass longer and later jumped straight to newer electronic systems.) Of those 920 transitions, 385 are cleanly measurable: they also have at least one neighboring station within 75 km that kept the same equipment for years on either side of the switch, providing an undisturbed reference to difference against.
The measurement trick is the oldest one in climatology: neighbors share weather. Two stations 30 miles apart experience nearly the same monthly temperature swings, so if you subtract one station's monthly anomalies from the other's, the climate cancels out and what remains is the difference between the stations: their instruments, their siting, their local quirks. If that difference series is flat for years, then jumps exactly when the paperwork says a new thermometer arrived, the jump is the instrument's fingerprint. So for each documented transition we paired the station with every USHCN neighbor within 75 km that had no documented equipment change within ±4 years, and averaged the before-vs-after jump*.
Here is the average difference series across every pair, aligned so that month zero is each station's documented swap date. Daily highs (TMax) and daily lows (TMin) are shown separately.
Mean difference in monthly anomalies (transitioning station minus unchanged neighbors), aligned on each station's documented MMTS installation month · 703 pairs, 385 transitions · USHCN monthly data, TOB-adjusted
Two things are worth staring at. First, the series are flat for four years before the swap: these station pairs were measuring the same climate, stably. Second, the two panels move in opposite directions. The new electronic sensor reads daytime highs cooler (its shield does not soak up sunshine the way a wooden box does) and nighttime lows slightly warmer. An instrument artifact does that; a change in the actual climate does not. It is also the opposite of what you would expect if, as is sometimes alleged, adjustments for these changes existed to manufacture warming: correcting the highs warms recent decades, but correcting the lows cools them.
The average hides real spread: shield differences interact with local sun, wind, and siting, and sometimes the MMTS was installed closer to the house than the old shelter (the cable had to reach the indoor display). Here is the distribution of the measured step at all 385 stations.
One value per transitioning station (averaged over its reference pairs) · bin width 0.1°C
The typical station's highs dropped by around a third of a degree C, but individual stations range from near zero to beyond a full degree, in both variables. This is why modern homogenization (Menne and Williams 2009) doesn't apply one blanket correction: it compares every station against its neighbors, exactly as done here, and corrects each detected break by its own locally measured size.
Every documented transition, with its best-matched unchanged neighbor. The dashed lines are the average difference before and after the documented date; their gap is the measured step.
Monthly anomaly difference (transitioning station minus neighbor), ±48 months around the documented installation date
Spread across thousands of stations in under a decade, a −0.34°C step in daily highs is enough to erase roughly one in seven days at or above 95°F from the raw record (see the companion demonstration, where this bias, replayed on pristine modern data, flips a warming hot-day trend to an apparently cooling one). It also explains a chunk of why raw and homogenized US temperature records diverge starting in the 1980s: the raw record contains a nationwide, one-way, non-climatic step. The swap was worth making; the new sensors are better. But you cannot splice two different instruments together, call the result raw truth, and read a climate trend off the seam.
*Method. Station equipment histories from NOAA HOMR (TEMP element records; transition = first period whose equipment is MMTS, preceded by non-MMTS equipment, 1982–2005). Temperatures are USHCN v2.5 monthly TMax/TMin with time-of-observation adjustment only (no homogenization), so observation-time changes cannot masquerade as instrument steps; monthly anomalies are relative to each station's own 1961–1990 means. Reference stations: USHCN neighbors within 75 km with no documented TEMP equipment change within the ±48-month window; the two months around the documented date are excluded from the step fit, and each side must have at least 30 valid months. Steps are averaged per transitioning station before computing the mean and its 95% confidence interval. Sharp-eyed readers may notice the composite begins to diverge two or three months before month zero: documented dates record when the metadata change took effect on paper, which often lags the physical installation by a month or more (297 of the 385 dates are day-precise, so this is paperwork lag, not rounding). Beyond three months out, the pre-period is flat; widening the excluded window from ±2 to ±9 months deepens the mean step only ~5% (−0.36 to −0.38°C), so the values reported here are, if anything, slightly conservative. Using fully raw (non-TOB-adjusted) data instead gives −0.39°C / +0.20°C. Code and data: github.com/hausfath/us-hot-days.