Chester County PA - Analytical Battle of Actual vs. Altered Climate Data

[ChescoWx]

6 hours ago, chubbs said:

Nope. The Chesco COOPs cool significantly relative to the SE PA Division data from the PA climatologist. Regression isn't a good tool because the station changes are sudden not gradual and you picked an interval that started after the initial Coatesville move and ended before the West Chester move took full effect. If you widen the interval to 1941-1975, the difference in slope is large.

Per table below, comparing the five years before the first move and the five years after the last move, the Chesco COOPs cool by 1.7F relative to the SE Pa Division. Note that the Pa SE Division includes the Chesco Coops so the two items compared are not independent. For that reason it is not surprising that the delta for the SE PA is slighly smaller than ABE. 

Funny seeing your persistent effort to dismiss what the the raw data from Chester County clearly shows.  The 3 stations each cooled significantly after the 1945-1970 moves: first Coatesville cooled relative to Phoenixville and West Chester; then Phoenixville cooled relative to Coatesville and West Chester; and finally West Chester cools relative to Coatesville and Phoenixville.  Clear as day in the raw COOP data.

 

 

All three stations cooled at statistically the same rate as the surrounding stations. Additionally, in looking closer at potential TOB adjustments one may wonder why the West Chester 2NW and Coatesville 1SW records which predominantly relied on morning (AM) observations...so one would think the Raw TOB Distortion of an AM reading captures a naturally lower minimum, meaning the raw data was technically already too cool correct? The NCEI Adjustment you would think logically would be to correct for morning bias at Coatesville and West Chester by warming the historical record. However, NCEI's automated, grid-wide homogenization algorithms actually did the opposite by applying additional sweeping chilling adjustments to both of these stations data from 1895 through 1982. 

[ChescoWx]

Of course @chubbsWidening the interval to 1941 - 1975 does not change the previous answer. Both series show a real, statistically significant cooling trend on their own — this isn't noise, and it fits the well-documented mid-century cooling that shows up across the eastern US from the 1940s into the 1970s.

Are the two trends statistically the same?

This is the more interesting question, and the answer depends on the test:

• Naive comparison (treating the two trends as independent): the difference in slopes (−0.66 vs. −0.46 °F/decade) gives t = −0.92, p ≈ 0.36 — not remotely significant. By this test they're indistinguishable.
• Paired comparison (the right test here, since both series describe the same regional weather in the same years — they're correlated at r = 0.83, so a paired test on the year-by-year difference is more powerful): the difference series trends at −0.021 °F/yr (−0.21 °F/decade), 95% CI [−0.042, +0.001], p = 0.057.

So the more rigorous test lands just on the wrong side of the conventional 0.05 cutoff — the station data cooled somewhat faster than the division average, and that gap is close to, but doesn't quite clear, statistical significance. The honest read: the two trends are not clearly, statistically different so no need to cool the raw data further as they are already a strong match.

[chubbs]

19 hours ago, ChescoWx said:

Of course @chubbsWidening the interval to 1941 - 1975 does not change the previous answer. Both series show a real, statistically significant cooling trend on their own — this isn't noise, and it fits the well-documented mid-century cooling that shows up across the eastern US from the 1940s into the 1970s.

Are the two trends statistically the same?

This is the more interesting question, and the answer depends on the test:

• Naive comparison (treating the two trends as independent): the difference in slopes (−0.66 vs. −0.46 °F/decade) gives t = −0.92, p ≈ 0.36 — not remotely significant. By this test they're indistinguishable.
• Paired comparison (the right test here, since both series describe the same regional weather in the same years — they're correlated at r = 0.83, so a paired test on the year-by-year difference is more powerful): the difference series trends at −0.021 °F/yr (−0.21 °F/decade), 95% CI [−0.042, +0.001], p = 0.057.

So the more rigorous test lands just on the wrong side of the conventional 0.05 cutoff — the station data cooled somewhat faster than the division average, and that gap is close to, but doesn't quite clear, statistical significance. The honest read: the two trends are not clearly, statistically different so no need to cool the raw data further as they are already a strong match.

Repeating what  I said above garbage in, garbage out.  Your analysis and AI's conclusion is faulty. You aren't telling AI that the Chesco COOPs are included in the SE PA climate division. You can't conclude that Chester County is the same as the non-Chesco stations if you include Chester County with the other stations. Second the regression slopes depend on the end-date. You need to include enough years before and after the moves for the station moves to be fully reflected in the regression. For the period 1941-1975 I get a much bigger difference in slopes -.69 for Chesco and only -.32 for the Division. Excluding Chester County from the Division would lower the Division slope further. Finally AI already has a Pvalue of 0.057, i.e, there is a 94% chance that Chesco is different using your faulty comparisons. Correct the errors and the P value will shrink considerably.

My biggest complaint though is that your statistical model is inadequate. Temperature observations are highly correlated across a region because all stations experience the same weather. The Chesco and SEPA stations should be very close. Instead the chart clearly shows that Chesco COOPS cool significantly vs the SEPA Climate Div. The different temperature response by the Chesco Coops is closely associated with the timing of station moves. End of story.

[ChescoWx]

AI knew the Chesco COOPs are included in the SE PA climate division but did not feel it material enough to change the statistical fact that the raw data at Chesco was cooling the same as the surrounding stations so again no need for further cooling. Plus with all those AM TOB readings at Coatesville and West Chester why didn't NCEI apply some warming adjustments? Again only non-stop chilling adjustments. 

[ChescoWx]

A member asked for an analysis of 100 degree days by decade since the 1890's. This is normalized as days per station per year (total 100°F+ days across all stations in a decade, divided by the number of station-years of data in that decade) rather than a raw sum — otherwise a decade with 22 stations would look "hotter" than a decade with 4 stations just from having more thermometers out there, which would be misleading.

The pattern holds up clearly at the county level, same as it did for individual stations: the 1930s stand out sharply at about 1.1 days per station per year — nearly double any other decade — consistent with the well-documented Dust Bowl-era heat. From there the rate declines fairly steadily, bottoming out near zero in the 1970s and again in the 2020s (0.01 and 0.02 days/station/year respectively), with the 2000s-2010s sitting in a modest middle range around 0.11–0.15.

Worth flagging directly: the two lowest decades in the entire 130-year record are the 1970s and the current 2020s — and the current decade of the 2020s (partial decade).

[chubbs]

18 hours ago, ChescoWx said:

AI knew the Chesco COOPs are included in the SE PA climate division but did not feel it material enough to change the statistical fact that the raw data at Chesco was cooling the same as the surrounding stations so again no need for further cooling. Plus with all those AM TOB readings at Coatesville and West Chester why didn't NCEI apply some warming adjustments? Again only non-stop chilling adjustments. 

Clear that you do not understand the science behind bias adjustments. Bias adjustments are not made in an ad hoc manner or to raise or lower the temperature of a station. Bias adjustments are only made to correct station changes. If the station observation time doesn't change, then no adjustment is needed. Adjustments are based solely on raw data. A station has to change with time in a manner that is different from all other stations that experience the same weather. The goal is to include weather and exclude station changes when calculating long-term climate trends.

The only way to evaluate bias adjustments is to inspect raw station data at the time of the station change. Here's an example using the West Chester 1970 move. The old West Chester COOP shut down in December of 1969 and re-opened in May of 1970 at a new location. Subtracting Coatesville from West Chester isolates the relative temperature changes at the two stations at the time of the move.  The raw data shows that the new West Chester site cooled by significantly relative to Coatesville after the move. That's all the proof needed to identify a major station change at either Coatesville or West Chester.  We know from raw data from Phoenixville that West Chester is the guilty party. The Chesco raw data clearly shows that West Chester cooled significantly after its 1970 move.

The NOAA bias adjustment is in complete agreement with the raw data from Coatesville and West Chester.  That shouldn't be surprising. NCEI is using all the raw data that experiences the same weather as West Chester, hundreds of stations potentially. I've looked at multiple large station changes at the Chesco COOPs and NCEI nails every one.

Note that NCEI is asking a similar question as your AI analysis. Does West Chester respond to local May 1970 weather in the same as the other regional stations? However the NCEI procedure is much more targeted than AI, is using more data, and it is using the right conceptual model. The temperatures at West Chester do not vary independently of the temperatures at Coatesville or other stations. This gives the NCEI analysis much more power. It can correctly flag the West Chester move as a major station change. While your AI analysis can not. 

The raw data is the final arbiter and in this case NCEI is clearly superior to AI.

 

[chubbs]

17 hours ago, ChescoWx said:

A member asked for an analysis of 100 degree days by decade since the 1890's. This is normalized as days per station per year (total 100°F+ days across all stations in a decade, divided by the number of station-years of data in that decade) rather than a raw sum — otherwise a decade with 22 stations would look "hotter" than a decade with 4 stations just from having more thermometers out there, which would be misleading.

The pattern holds up clearly at the county level, same as it did for individual stations: the 1930s stand out sharply at about 1.1 days per station per year — nearly double any other decade — consistent with the well-documented Dust Bowl-era heat. From there the rate declines fairly steadily, bottoming out near zero in the 1970s and again in the 2020s (0.01 and 0.02 days/station/year respectively), with the 2000s-2010s sitting in a modest middle range around 0.11–0.15.

Worth flagging directly: the two lowest decades in the entire 130-year record are the 1970s and the current 2020s — and the current decade of the 2020s (partial decade).

Nope no dust bowl in Chester County, just Phoenixville with a bad shelter location. Philadelphia and West Chester don't agree at all with your plot. Not surprising since once again you are confounding station changes and weather. East Nantmeal and West Grove are a little different than the City of Coatesville or West Chester don't you think? 

[ChescoWx]

I also carved out Phoenixville in the 2nd chart based on concerns from @chubbs it's more urban location may skew the data. With or without Phoenixville the trend is clearly toward less 100 degree days in Chester County PA.  

 

[donsutherland1]

For an illustration on how Phoenixville has changed over time (Source: Berkeley Earth):

These changes mean that simply using raw data across time creates an "apples-to-oranges" comparison when it comes to data integrity.

Notice the larger difference in warming trend between the raw data and region and much closer figure between the breakpoint-adjusted data and the region.

That's very close to the observed warming for Chester County and Pennsylvania for the 1895-2025 period:

The raw data (0.11°C/century) is the clear outlier. That divergence from the regional and statewide data illustrates the flaws in treating a station's record identically despite breakpoints, station moves, etc.

 

[ChescoWx]

5 hours ago, donsutherland1 said:

For an illustration on how Phoenixville has changed over time (Source: Berkeley Earth):

These changes mean that simply using raw data across time creates an "apples-to-oranges" comparison when it comes to data integrity.

Notice the larger difference in warming trend between the raw data and region and much closer figure between the breakpoint-adjusted data and the region.

That's very close to the observed warming for Chester County and Pennsylvania for the 1895-2025 period:

The raw data (0.11°C/century) is the clear outlier. That divergence from the regional and statewide data illustrates the flaws in treating a station's record identically despite breakpoints, station moves, etc.

 

I removed Phoenixville and no statistical change to the raw data at all!!

[ChescoWx]

So let's look at a single station with long term data 1894-1982 Coatesville 1SW. I performed an analysis of that station to see if TOB bias was part of the adjustment for 77 of 78 years of cooling data that we have documented many times before from the raw actual data. I evaluated all of the changes in observation times for the entire period of record. Of note from 1946-1982 the station observed a consistent 8am reading whhich sits close to the −0.15°C (cool bias) end of the curve, meaning the TOBS-correct move would be to add temperature back (a small positive adjustment) — yet the actual 1946-82 adjustment stayed negative on average (−0.83°F). That's a clear red flag in the NCEI adjustments as TOBS alone would predict the sign should have flipped positive after 1946, and it didn't. So without TOBS alone there remains a clear and significant leftover unexplained cooling of roughly -1.1 to -1.8°F in every single period — evening, dual, and morning alike. That's the thing that needs a separate explanation, since TOBS alone doesn't account for it. So we have seen @chubbsoften say that station movements tied to  Coatesville's industrial history and station moves impacted the data and made it critical to chill the old data. The problem that NCEI and Charlie has is that in reality Coatesville was home to Lukens Steel — a continuously operating steel mill since 1810, one of the largest specialty steel producers in the country, still headquartered there through 1982 and beyond this is a station within a mile of significant industrial activity! This of course is a textbook candidate for a persistent local warm bias — waste heat, paved/industrial surfaces, altered local airflow. But here's the problem for NCEI and Charlie to address -  this doesn't match the data adjustments made by NCEI. The facts are that Lukens was arguably at its industrial peak in the postwar decades (1946–82) — yet that's exactly the period where the adjustment (and the residual) is smallest, not largest. If growing industrial heat were the main driver, the bias should have gotten worse over the 20th century, not better. So it is not TOB adjustments and clearly not site moves behind these adjustments so what say you Charlie?

[donsutherland1]

53 minutes ago, ChescoWx said:

I removed Phoenixville and no statistical change to the raw data at all!!

Coatesville 1SW has a strong cooling trend in the raw data, as you observed. However, there are multiple breakpoints. Adjusting for those breakpoints brings Coatesville's trend in line with regional trends.

From Berkeley Earth: