Hello all. I have spent the last couple of months digging into one question and would welcome feedback and pushback. The question is not whether raw or adjusted data is more accurate. It is simpler than that... based on the data any of us can access, can we prove that NCEI's adjustments to our local temperature record are necessary, and sized correctly?
What the adjustments do here
Using every Chester County station, 1895 to 2025, the raw record warms about +0.05°F per decade, which is not statistically significant. NCEI's adjusted county series warms +0.25°F per decade. The gap between the two grows at about +0.2°F per decade and that gap is significant. Most of it comes from cooling the early decades, which NCEI runs 1.4 to 2.0°F below what the thermometers recorded.
I also compared the three nearest long-record airports, all outside the county, against their own NCEI adjusted series:
PHL (1941 to 2025): raw +0.44, adjusted +0.40. NCEI barely touches it.
ABE (1948 to 2025): raw +0.31, adjusted +0.60. NCEI cools the early record by up to about 2.5°F, which roughly doubles Allentown's warming trend.
RDG: NCEI only has an adjusted series from 2000 on, and it cuts Reading's trend from +1.36 to +0.77.
Over the same 1948 to 2025 window, NCEI's adjustments add roughly +0.2 to +0.3°F per decade of warming both at Allentown (+0.29) and in Chester County (+0.17 against all county stations, +0.33 against our two longest continuous stations, West Chester 2NW and Phoenixville 1E). So Allentown and Chester County get about the same boost. PHL is the one that gets almost none.
Against our most rural long-record stations, PHL warms roughly 0.2 to 0.3°F per decade faster since 1948. That looks like urban and airport growth that the adjusted PHL series keeps in, although the uncertainty range on it runs from about zero to +0.5.
What we can prove
Phoenixville shows a step of almost 2°F in 1987. Raw data has real artifacts in it.
What we cannot prove
We cannot show the adjustments are the right size, or that every one is needed. Coatesville 1SW is the clearest example. It had a documented move in January 1946 and NCEI lists a 1.4°F step for it, then lowers everything before 1946 to match. I compared Coatesville against every neighbor I could find with data on both sides of 1946... Phoenixville, Ephrata, Holtwood and Lancaster, all between 21 and 28 miles away. Each of those tracks Coatesville closely enough on its own to catch a step of about 0.4 to 0.7°F, roughly a third to a half of the size NCEI applies. Against those stations averaged together the step is only -0.25 to +0.07°F and not significant. So these neighbors were sensitive enough to see a step the size NCEI applies, and none of the ones without their own 1948 changes consistently show it. Phoenixville and Ephrata cooled right along with Coatesville. The only neighbors that show a bigger step are West Chester and Lancaster, and both of those had their own location changes in 1948.
Coatesville also carries 12 NCEI breakpoints and only 4 are documented. West Chester carries 15 with 4 documented. NOAA's own station history for observation times conflicts with the reading-time schedules commonly assumed for these stations.
To be fair, we also cannot show NCEI got it totally wrong. As NCEI compares against a much wider network of neighbors than we can see.
So my conclusion: adjustments are necessary in principle, but from public data we can neither verify nor disprove their size here.
What would close the gap
NCEI's pairwise log for each station, showing which neighbors, which years and what weights were used for each break. I reached out to NCEI who told me this is not something they publish.
The time-of-observation step on its own for each station, separate from the other adjustments.
The original observer forms (B-91s) for the years where NOAA's records conflict or are blank.
Dated site move and instrument change details, including the MMTS changes in the 1980s and ASOS at the airports in the 1990s.
Overlapping readings at the old and new sites when a station moved. The Climate Reference Network only began in the 2000s, so it is too short to help with the early record.
A re-run of NCEI's published code on the same inputs, to see whether we get the same breaks.
Reading's adjusted series before 2000.
Let me know if any questions or if you see anything that does not hold up!