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

[ChescoWx]

You will often here folks say well while our average daytime highs may not be warming significantly....we are certain that the nightime lows are warmingg....Well I went back and did a statistical analysis to see if between 1890 and 2025 we saw statistically warming of overnight lows and daytime highs along with the overall average temperatures. We used both a pure raw all station mean which is just the raw untouched data and an anomaly composite which while still raw data, treats each individual station referenced to its own baseline so the changing station roster and locations can't fabricate some false trend.

Raw all-station mean (the literal, unadjusted county average): none of the three are statistically significant.

• Average high: +0.055 °F/decade, p = 0.073 — warming, but just short of significance (borderline).

• Average low: +0.022 °F/decade, p = 0.49 — not significant.

• Overall average: +0.039 °F/decade, p = 0.15 — not significant.

So by the pure raw mean, there's no statistically significant warming or cooling in any of the three over 1890–2025 — a slight (non-significant) warming lean in all three.

 

Anomaly composite (still raw data, but each station referenced to its own baseline so the changing station roster can't fabricate a trend): two of the three are significant, and they point in opposite directions.

• Average high: +0.094 °F/decade, p = 0.003 — significant warming.

• Average low: −0.075 °F/decade, p = 0.008 — significant cooling.

• Overall average: +0.010 °F/decade, p = 0.72 — not significant (the high and low cancel out).

The one thing both methods agree on firmly: the overall average temperature shows no statistically significant trend either way across the 135 years across Chester County from 1890-2025. Below are some charts that highlight all of the above.

[Mike Cycle]

What smooth ramp do you refer to? I just glanced at the breakpoints at BEST for the stations in the table and none of them look like "smooth ramps".  

Coatesville's breakpoint are fully documented at this point.  I have shown you that there is hard documented real world events driving the breakpoints for this station.  If you think there is something off about the breakpoints for the other stations, and that the published work of scientists around the world is flawed,  then it is on you to do the deep dive on the histories through all available sources and make your case.  Don't just wave your hands saying "absolutely no metadata backing, that's an algorithm asserting station changes that are clearly not able to be documented"--without first scouring all source for that documentation.  

For West Chester, I just took a quick look at the breakpoints at BEST, the 530-1, and HOMR records.  About half the breaks are clear just from the 530-1 and HOMR.  There are breakpoints at ~1978, ~1982 and ~1988 that don't have explanations yet and may be in the paper forms, or other records. Its the Daily Local--there are going to be a lot of records about where they were located. Have AI write a tool to download all the paper forms from batch "AA92589" or just do it manually around those dates.  Anyway, it is on you to prove there is no documentation for these moves, you cant just say there aren't any without looking at the available documentation first.

 

[ChescoWx]

1 hour ago, Mike Cycle said:

What smooth ramp do you refer to? I just glanced at the breakpoints at BEST for the stations in the table and none of them look like "smooth ramps".  

Coatesville's breakpoint are fully documented at this point.  I have shown you that there is hard documented real world events driving the breakpoints for this station.  If you think there is something off about the breakpoints for the other stations, and that the published work of scientists around the world is flawed,  then it is on you to do the deep dive on the histories through all available sources and make your case.  Don't just wave your hands saying "absolutely no metadata backing, that's an algorithm asserting station changes that are clearly not able to be documented"--without first scouring all source for that documentation.  

For West Chester, I just took a quick look at the breakpoints at BEST, the 530-1, and HOMR records.  About half the breaks are clear just from the 530-1 and HOMR.  There are breakpoints at ~1978, ~1982 and ~1988 that don't have explanations yet and may be in the paper forms, or other records. Its the Daily Local--there are going to be a lot of records about where they were located. Have AI write a tool to download all the paper forms from batch "AA92589" or just do it manually around those dates.  Anyway, it is on you to prove there is no documentation for these moves, you cant just say there aren't any without looking at the available documentation first.

 

 

 

 

No hand waving Mike....just no clear documented reasons as only four of 51 (8%) trace to a documented real physical cause — three moves and only one TOB change. There are 2 zero break stations that are our controls - Coatesville 2W and Hopewell demonstrate that the PHA is indeed selective as it left these 2 stations untouched: NCEI applied no adjustment to either - which proves the 14 breakpoints at Coatesville 1SW were a deliberate finding rather than a default. The Coatesville 1SW dates, all code 4: 1896-12, 1902-02, 1917-08, 1920-09, 1922-08, 1930-10, 1936-05, 1941-06, 1946-04, 1959-11, 1968-05, 1971-04, 1973-07, 1976-02. Fourteen inferred discontinuities against six documented HOMR relocations, coinciding with none of them. 

The linear ramp clearly fits far better than any step. And when you allow both, the step collapses to −0.050 °C while the slope survives essentially intact at +0.182 °C/dec. A one-time convention change explains basically none of it. And more importantly the actual magnitude is out of range. The Coatesville adjustment swings +1.51 °C from first decade to last (West Chester, from NCEI's own file, swings +2.38 °C). Published TOB corrections for a full afternoon-to-morning switch run roughly 0.2–0.5 °C in annual mean. So TOB, even granting it entirely (which is not valid), accounts for perhaps a third of the Coatesville adjustment and a fifth of West Chester's.

To summarize how it all fits together Mike:

1. Raw agreement across stations. Coatesville 1SW −0.366, West Chester 2NW −0.409, Phoenixville 1E −0.339 over 1941–75, all p ≤ 0.0012, inside each other's error bars.
2. Both systems erase it. Berkeley and NCEI each render all three non-significant, with adjustment trends correlating at r = 0.9986.
3. The correction is a "smooth ramp" which you question above but certainly not a step — which no single documented event can explain.
4. Its size exceeds TOB by 3–5×.
5. The breakpoints lack documentation. Fourteen at Coatesville, all code 4, against six documented relocations in HOMR. At West Chester, April 1921 and November/December 1927 are documented observer-and-site changes classified as empirically detected, and the March 1936 relocation produced no breakpoint at all.
6. Phoenixville is the cleanest case — no detected convention change anywhere in its record, and it still cools at −0.339 (p=0.0012) - why is that? So many questions and still so few answers!!

The unanswered questions that follows from this is narrow and should be answerable: exactly what documentation supports the fourteen breakpoints at USC00361589 - Coatesville 1SW and what portion of the adjustment at Coatesville is attributed to time-of-observation versus other causes? That last question is the one they can't deflect — because the shape of their own adjustment says it isn't mostly TOB because if TOB was factored in it would actually be supporting a warming not a cooling adjustment at Coatesville! 

[chubbs]

13 hours ago, ChescoWx said:

 As you know homogenization remains nothing more then a hypothesis-driven statistical correction, it is of course not a measurement.

Nope.  Homogenization is accepted science. Thoroughly tested over decades. Proven to obtain the correct climate information from raw temperature when station changes have occurred. The raw temperature data in Chester County is contaminated by myriad station moves and changes. Assuming that Coatesville and other Chesco COOPs stayed in exactly the same spot with zero equipment and method changes over decades will give you the wrong climate answer every time.

The cooling of Coatesville between 1945 and 1949 is well documented in this thread. The regional raw temperature data is 100% conclusive on this point. Coatesville was much cooler in 1949 than it was in 1945 relative to all stable regional stations. Comparison of Coatesville to Phoenixville over the period 1941-75, is immaterial; because, as the bottom chart shows, Phoenixville had a station break in 1948. The post-war  Phoenixville and Coatesville station breaks occurred at different times proving that they are not weather-related.

Mike's analysis and the NCEI bias adjustments are completely independent, derived from 2 different sources of information. Mike uses the station records. Bias adjustments use raw temperature data. The fact that they agree so well shows how powerful and accurate homogenization is in Chester County.

Your AI slop is wasting our time if it doesn't focus on the station breaks themselves. Regression over 35 years is not the right tool to identify station breaks. You need much more focused analyses to look at station breaks. The 1945-49 period for the Coatesville post-war move/changes for instance. 

[ncjoaquin]

Sometimes I think my life is not always that exciting.  Then I come here to see that this weird ass discussion is still going on after more than 2 years.

[Mike Cycle]

Paul, you have made many claims in the last few comments.  These look like strawmen to me, and I have summarized them below.  To me, it appears you are making these comments to show that the accepted process of dealing with raw noisy data is flawed, and therefore that published climate trends are based on a flawed process.  

• "86% of breakpoints are undocumented" → implies the algorithm should mostly find documented breaks. It's designed to catch undocumented ones — a high rate is expected, not damning.
• "Chester County's raw data shows no significant county-wide trend, so warming is manufactured by adjustment" → implies mainstream science predicts a clean, significant trend in every small local raw sample. Nobody claims that — a handful of stations in one county is noisy by nature, which is the whole reason homogenization exists.
• "The correction is a smooth ramp, not a step, so no single event explains it" → implies a legitimate correction must be one discrete jump. Many small breaks stacked over decades naturally produce a ramp-shaped cumulative correction.
• "Some stations (Coatesville 2W, Hopewell) got zero adjustment, proving selectivity" → implies adjustments should be applied roughly evenly everywhere. A detector that only flags real breaks and leaves clean stations alone is working correctly, not cherry-picking.
• "Highs are warming while lows are cooling — that's a contradiction" → implies day/night trends must move together. Diurnal-range divergence is a documented, separately-studied climate phenomenon, not a paradox.

Yet, the Chester County raw data is in good agreement with published climate trends, more so when the urban makeup of the pre-1940 station group is taken into account.  

I suggest looking into the West Chester 2 NW records to document the rest of the breakpoint.

 

[ChescoWx]

Mike, this is in my view the most useful thing anyone has put in this thread, and it changes my position on several points. Thank you for doing the archival work rather than just asserting it.

Taking your points in order. You're right that "undocumented" in the algorithm's metadata linkage is not the same as "no documentation exists." I was treating the two as equivalent and that was a wee bit sloppy. Your figure shows some actual events behind most of the Coatesville breaks, and the burden argument is fair — I shouldn't claim an absence of records without going through the 530-1s, the deeds and the aerials first. I hadn't. You had - thanks!.

You're also right that a high rate of undocumented detections is expected from a detector built to catch undocumented changes, so the 86% figure doesn't carry the weight I was giving it. I'd already dropped that argument after NCEI sent me their changepoint list for this station — it shows 4 of 12 linked to documented events, not zero, so the Berkeley code counts were never a good proxy for what the PHA is doing.

On the ramp point: Mike - you're right - many small stacked breaks produce a ramp-shaped cumulative. I proved your point without meaning to — cumulating NCEI's twelve discrete changepoints gives a series that fits a straight line at r-squared 0.495. That's a stack of steps looking smooth, exactly as you say. Zero-adjustment stations and the diurnal range point are both fair points also. A detector leaving clean stations alone is working, and DTR divergence is a studied phenomenon rather than a contradiction. And I accept that a handful of stations in one county is a noisy sample by nature.

Now a couple things I will push back on a little, not on it's substance but on how firmly you state it. Your own figure marks 1920 unresolved. And I'd read the 1959-1976 group differently to the header on it. The box says five breaks, but NOAA only has two in that window — 1959 and 1971. The other three are BEST-only. And BEST's five nearly cancel: -0.50, +0.37, +0.02, +0.72, -0.71 comes to -0.10 F net across seventeen years, which isn't doing much work. NOAA's two come to +0.40 F, and that's the direction you'd correct for spurious cooling, not for a station warming under encroaching infrastructure.

Which brings me to the part I can't pull together. The station moves to progressively cooler sites implies the raw acquires spurious cooling and needs the later data warmed — which matches what NOAA applied. Infrastructure growing up around the site implies the opposite. Both are in your account of this window. Which one do you think is actually driving it?

For what it's worth, your 1959 and 1971 values match NCEI's changepoint list for this station exactly once converted — -0.54 and +0.94 F against -0.28 and +0.53 C. So we're working from the same PHA output, which is worth knowing! Thanks again Mike!!

That window also stops something I was about to post. I ran a segmented breakpoint analysis on the raw daily data and it attributed only one concordant discontinuity to Coatesville, January 1973. I then fitted the 1952-82 trend with that single step in the model. If there are more real breaks in that window, my model is under-specified and the number isn't worth quoting. I'm going to refit it against the full break set before I use it. If it doesn't survive, I'll say so here. Stay tuned for that later!

The one thing still open is narrower and comes out of NCEI's own files rather than the station history. Cumulating their twelve changepoints reproduces +0.123 C/decade of adjustment trend. The actual adjusted-minus-unadjusted trend for the same station and years is +0.184. About a third isn't in any listed changepoint. It isn't rounding, and it isn't run-to-run variation — I compared two pulls of the adjusted series two days apart and they were identical across all 119 years at West Chester. I've asked NCEI what accounts for it rather than guessing, and I'll post their answer either way. I sent the email this morning to my contact there...stay tuned for more of that later!

Now my ask - would you be willing to share your compiled 530-1 and deed work for Coatesville, and whatever you have for West Chester? I'd rather build on it than duplicate it badly, and I'll take your point about the 1978, 1982 and 1988 West Chester breaks — the Daily Local should have left a trail worth chasing.

Charlie, I think most of the above answers you as well. You were right that a 35-year regression is not a breakpoint test, and running it properly is what led here. On 1945-49 specifically, NCEI's list has a documented move at 1946-01 with a +0.79 C step, so something documented did happen there. I still can't test that window myself — only three Chester County stations cover 1930-1975 with continuity and attribution needs four — so at this point we can't claim it either way.

Paul

[Mike Cycle]

The main question of this thread is whether the adjustments to raw data are a) valid and b) essential.  The deep dive on Coatesville upthread answered these two questions for Coatesville with two resounding YESs for that station. 

Here is the breakpoint analysis for Phoenixville 1E.  It also demonstrates that the adjustments done to the raw data are a) valid and b) absolutely necessary.  Note that all of the aerials from 1927 forward show changes at the site--roughly every ten years between aerials--so there is going to be microsite changes. 

One has to wonder, for all of the time some folks have put into explaining why the adjustments are arbitrary, ad hoc, suspect, etc etc, why they did not just take two hours and look at the easily available documentation on line. 

 

[ChescoWx]

Not so fast Mike!! We are not close yet to what think is being answered! The deep dive above clearly has not yet answer these two questions for Coatesville! First the refit I promised with some very important findings!! and then where I think you and I are answering different questions. At the end I will address your Phoenixville box above.

I said I wouldn't quote that 1952-82 trend until I'd refit it against the full break set, because if there are several real discontinuities in the window then a single-step model is under-specified and the number is worthless. I've now done it, and the result is quite surprising!! The mechanism first, because it's the interesting part. A step term and a trend term compete for the same variance in a regression. If Coatesville's apparent cooling were really just the station stepping down at a few dates, fitting those steps should soak up the drop and leave the residual slope flattened toward zero. That is exactly the test. But what actually happened is the complete reverse. Every fitted step comes out positive - the model wants the later years lifted, not lowered - and lifting the later years forces the underlying line to fall more steeply to still pass through the data. Coatesville 1952-1982, annual, trend and steps estimated jointly in one model rather than subtracting steps first:

  no steps                              -0.594 F/decade   p = 0.0001
  1973 only (my attributed break)       -0.735 F/decade   p = 0.0048
  NOAA's two (1959, 1971)               -0.959 F/decade   p = 0.0272
  Berkeley's five (1959/68/71/73/76)    -1.657 F/decade   p = 0.0188

Significant in all four. Fitted steps +0.45, +0.46, +0.89, +0.91, +0.81 F, every one positive. So the more of your documented breaks I grant, the more raw cooling is left over needing some other explanation.

The weaknesses, because I'd rather flag them myself. Not one of those steps is individually significant - best p is 0.17, standard errors 0.6 to 0.7 F against steps of 0.45 to 0.91. So this isn't proof the steps are positive. It's that the data, asked where steps would sit if they exist, puts them in the direction that deepens the cooling. Suggestive, not conclusive, and it neither confirms nor refutes your documentation, because at annual resolution this series can't resolve steps that size. AIC prefers the no-steps model over every step set, though with those SEs that's weak too. And the five-break fit is badly under-powered: seven parameters on 31 values at 0.75 F residual SD detects a genuinely present trend only about 14% of the time in simulation. It found one anyway, which is why I'm reporting it rather than quietly setting it aside - it's the most favourable number to me and the one I trust least.

So what I'll stand behind is a range: -0.59 to -1.66 F/decade over 1952-1982 depending on the break set assumed, significant in all four specifications.

Now, on validity and necessity. I think we're answering different questions.

Your work establishes that events occurred at those dates, and I've conceded that without reservation. Validity is a further step: it requires the magnitude applied at each date to be right. The PHA sizes each step from neighbour comparison, not from the site. A 0.3 mile move is a documented fact. The +0.86 C attached to it is an estimate. Nothing in the 530-1s, the deeds or the aerials tests that estimate, and that's the question I keep coming back to - it isn't answered by more documentation, however good the documentation is.

Necessity is different again. If the raw cooling at Coatesville were an artifact of the station changes, fitting those changes should dissolve it. It doesn't. It deepens it!!!!. Whatever else that shows, it isn't what "the adjustment is essential to get the right answer" predicts.

And about a third of Coatesville's adjustment trend isn't attached to any listed changepoint at all. NCEI's twelve cumulate to +0.123 C/decade against an actual adjusted-minus-raw trend of +0.184. Station history can't validate a component that no documented event is claimed to explain. I've asked NCEI what accounts for it rather than guessing, and I'll post their answer either way.

On Phoenixville I'd want to look at your figure properly before saying much, and I'll say up front that my own analysis can't support a claim about that station in either direction. When I tested its 1952-82 trend against a step placed at every possible date, 14 of 21 rendered it non-significant. That series is too short and too noisy to separate a step from a trend, so I'm not going to argue it either way. I'd also note that if the microsite is changing continuously - your reading of the aerials is roughly every ten years from 1927 forward - then the disturbance is gradual, and a step function is an odd model for gradual drift.

On the two hours: fair enough, and I've said as much. Though the changepoint list I've been working from came from NCEI directly on request, and isn't online anywhere.

The offer stands on your 530-1 and deed compilation. Your dates are better than the ones I read off your figure, and the magnitude question needs them.                                        

Now regarding the last Phoenixville block - that block does not say what your post said. Your comment claims the adjustments are "valid and absolutely necessary," but the header states ten breaks now carry a documented physical cause and seven smaller breaks remain unexplained, plus one marked UNRESOLVED at 1966 and a merged box labelled "UNDOCUMENTED — 7 BREAKS." That's 10 of 18, not a resounding yes!!!! Thanks for posting it — it's a more careful figure than the summary suggests.

What strikes me most is that the two sources diverge here in a way they don't at Coatesville. At 1940 — the largest documented move in the record, 2.1 miles — BEST flags +0.75 F and NOAA flags nothing at all. At 1942 it's the reverse. At 1985 you note the NOAA break is a single-station calibration artifact. If two implementations of neighbour-based detection disagree about whether the biggest documented move in the record produced a break, that seems worth more attention than either of us has given it.

The 1920 entry is the strongest piece of documentary work in either of your figures, incidentally — coords changing then reverting a year later while the offset doesn't is a real finding and I'd not have got there.

My own analysis cannot support the claim that Phoenixville cools — which matters because Phoenixville has been used as a comparator throughout this thread. Charlie dismissed my Coatesville–Phoenixville comparison over a 1948 break, and I cited Phoenixville's raw trend earlier in the exchange. Testing its 1952-82 trend against a step at every possible date, 14 of 21 rendered it non-significant. Too short and too noisy to separate a step from a trend, so to date we cannot find total support for these NCEI adjustments to Coatesville 1SW 

Paul

[ChescoWx]

What remains genuinely open regarding Coatesville 1SW remains  the magnitudes have never been independently checked. A documented 0.3-mile move does not actually tell us the offset should be +0.86 °C. That number comes from neighbor comparison, not from the site. Nobody in this thread has tested whether the applied steps match what the physical changes would produce. Also of note about a third of the adjustment trend attaches to no listed changepoint at all. +0.123 reconstructed against +0.184 actual. No station history can validate a component that no documented event is claimed to explain, and NCEI hasn't yet answered that from my earlier email. Lastly, and what can't be overlooked is that the raw cooling doesn't behave the way the artifact explanation predicts. Fitting the breaks actually deepens it rather than dissolving it!!

[Mike Cycle]

The Coatesville graphic I supplied says that in the period of 1952 to 1982 there was extensive construction going on at this station location.  its an electrical substation with megawatts of power going through it.  As components move, there will be step wise changes--up or down.  The station itself could be moving also as office space changes.  Here are the images showing that station and ongoing construction.   

Once again, we have 100% documented evidence of microsite impacts to a station record, and a) adjustments are valid and b) absolutely necessary.  Both a and b are within the context of using weather station data to develop climate information from multiple stations.

 

[ChescoWx]

3 hours ago, Mike Cycle said:

The Coatesville graphic I supplied says that in the period of 1952 to 1982 there was extensive construction going on at this station location.  its an electrical substation with megawatts of power going through it.  As components move, there will be step wise changes--up or down.  The station itself could be moving also as office space changes.  Here are the images showing that station and ongoing construction.   

Once again, we have 100% documented evidence of microsite impacts to a station record, and a) adjustments are valid and b) absolutely necessary.  Both a and b are within the context of using weather station data to develop climate information from multiple stations.

 

 

Mike, those aerials are genuinely useful and I'm glad to have them. They establish something I'd have had a hard time establishing myself: that the site was disturbed, repeatedly, across that whole period. I'm not disputing that. But again they just don't get you to either of the two claims you're making.

Documenting that something happened at a site is not the same as documenting how much it changed the temperature, or in which direction. The aerials show construction. They don't measure its thermal effect. Every step magnitude in the NCEI record was sized by neighbor comparison, not by anything visible in a photograph, and that is the number that has to be right for the correction to be right.

And there's a direction problem I raised earlier that I have not heard you answer or reconcile. An electrical substation with megawatts running through it, accumulating buildings and hard surfacing from 1946 through 1982, that is clearly a warming influence on the raw record. Correcting a progressive warming bias means pushing the later data down or cooler relative to the earlier. NCEI's correction goes the other way - the two NOAA changepoints in that window are +0.45 F at 1959 and +0.95 F at 1971, net +0.40 F, raising the later data, and the overall adjustment lifts the modern era relative to the early one.

You do offer a second mechanism - that the station itself may have moved as office space changed. If it moved away from the heat sources, that would produce cooling steps needing a warming correction, and the sign would work. But that's the opposite mechanism from the one the aerials show, and it's offered as a possibility rather than as something documented. Both can't be driving the same window. Which one do you think it is, and in which years?

The other thing that doesn't sit right: if construction were producing stepwise changes throughout 1952-1982, I'd expect the detectors to be finding them. NOAA places two changepoints in that window. Berkeley places five, and by your own figure they come to about -0.10 F net across seventeen years. That's a lot of documented disturbance producing very little net correction.

So where I see us in reality standing on this analysis is on validity: showing that an event occurred does of course not validate the size of the correction applied to it. That would take an independent measure of the effect, and I don't think you have so far produced one yet - yours truly also!  Of clear necessity: demonstrating that the adjustment is essential requires showing the uncorrected data gives the wrong climate answer, which requires knowing the right answer independently. Neither of those has been establish here.

I'm not claiming the adjustments are wrong. I'm saying "documented" and "validated" are different words, and the work so far has established the first, not the second. If you have something that speaks to magnitude rather than occurrence, I'd genuinely like to see it. But again we are far away from any proof you claim to see! 

 

[Mike Cycle]

Paul, you need to find a case where you can show PHA failed, and more generally, show that PHA is a bad approach.  The burden is on you to show that the station was incorrectly adjusted, not on me to walk you through the variables a Newlinville.  Here is a quick review of where we are at on your 2 year old thread:

Q1 — Is it necessary to adjust raw data before use for climate trends/model inputs?
Yes. Raw records carry known, non-climatic biases (station moves, instrument changes, time-of-observation changes, site changes) that don't cancel out and will systematically distort trends if left uncorrected.

Q2 — Is it perfect?
No. Breakpoint detection is demonstrably imperfect and even unstable run-to-run (shown in this thread's own evidence), and the method is explicitly statistical inference, not direct measurement.

Q3 — Is it pretty good, and has it undergone rigorous review/testing for over a decade?
Yes to both. Multiple independent blind benchmarks (Venema 2012, Williams 2012, Killick 2022) show it reduces error relative to known synthetic ground truth, and it's been continuously re-tested and refined for over 15 years, including by critics actively trying to find its weak points

Q4 — Given all that, does a cherry-picked local anomaly (minor adjustments at Newlinville era) amount to a valid attack on the process?
No. A single, non-representative case can't outweigh a decade-plus of adversarial, blind-tested evidence — however carefully that one case is analyzed.

[ChescoWx]

1 hour ago, Mike Cycle said:

 

Mike, I'll take your four questions in order

Q1 - yes. Raw records could potentially carry non-climatic biases that don't cancel but not always!

Q3 - yes. Williams 2012, Venema 2012 and Killick 2022 are real, adversarial, and they show homogenisation reduces error if a station has proven data issues, I accept that and I'm not going to pretend the literature says otherwise.

Q2 - one correction. You say breakpoint detection is "even unstable run-to-run (shown in this thread's own evidence)." It wasn't shown. I went looking for it: two pulls of the West Chester QCF taken a few years apart were identical across all 119 complete years, unadjusted series identical too. NCEI told me the output can change daily; I tested it and found none over that interval. That concession isn't mine to accept.

Q4 is where we really differ, and it's about scope. My claim is narrower: the specific magnitudes applied to specific Chester County stations haven't been independently verified, and showing an event occurred doesn't verify the size of the correction attached to it. The benchmarks measure whether homogenisation moves a NETWORK closer to truth on average - the right quantity for a national series, and a different question from whether one station in one decade got the right number. Venema's own write-up notes it modelled small European networks at high station density without network-wide trend biases; Williams found improvement in the presence of widespread systematic errors. Both are aggregate statements under stated conditions.

The conditions matter here. Before 1951 this county has three continuous stations. Menne and Williams say it themselves in the 2009 paper: false alarms at neighbouring stations reduce the number of homogeneous neighbours and thereby prevent the estimation of adjustments. That's the sparse-network failure mode described by the algorithm's own authors, in the era carrying the largest adjustments.

On cherry-picking - you introduced Coatesville as the exemplar and called it two resounding yeses. If it's representative enough to validate, it's representative enough to examine.

You asked for a case. Here's the beginning of one, and it's a magnitude test rather than an occurrence test. So I built a regional reference for Coatesville from the raw, unadjusted Chester County records only - leaving Coatesville 1SW out so that it never contributes to its own comparator, and no homogenised product anywhere in the calculation. Measured drift of Coatesville against those raw neighbours over 1895-1982 is -0.190 F/decade using every available station, or -0.113 F/decade using a fixed panel that can't change membership. The applied NCEI adjustment runs +0.303 F/decade, which implies a drift of -0.303. That value falls outside the 95% confidence interval of the measured drift under both reference constructions.

The adjustment also tracks the drift year to year - r = -0.74 - so it's aimed at something real. It's the size I'm questioning, not the existence.

What this is so far in all of my multiple years of review which has been my analytic goal of this process of reviewing the adjustment of raw data at one targeted station Coatesville 1SW  is that the correction appears larger than the divergence it corrects, measured without using any homogenised data. If that's wrong I'd like to know how??

Paul

 

[Cobalt]

All this chat for 1 out of 3,144 counties btw

[chubbs]

The only way to estimate the magnitude of station breaks is from raw temperature data, by comparing stable and changed stations.  As documented in this thread, my tests using raw temperature data in Chester County agree very well with the NCEI bias adjustments. Not surprising because the bias adjustments are calculated from the raw data. My experience in Chester County has convinced me that the NCEI method is bullet proof. 

Below is raw temperature info on the Phoenixville 1948 move. NCDC has monthly means for Phoenixville for Janaury through July 1948, partial data from August through November but no monthly mean until December 1948. Phoenixville cooled relative to Coatesville and West Chester when the monthly mean started again in December 1948.  Phoenixville cooled 1.6F relative to Coatesville, and 2F relative to West Chester, due to the move. The stats using the monthly data are highly significant for both Coatesville and West Chester. 

The post-war moves at Coatesville (1946), West Chester (1970) and Phoenixville (1948) are all highly significant statistically. A coincidence that all of the major county COOP stations moved in the same direction and roughly the same magnitude after the war. Clearly not weather because the cooling occurred at different times at each station.

[Mike Cycle]

As an example of why breakpoint detection is absolutely essential, and where there are no alternatives besides something like a pictorial history of the station, Coatesville is a winner. From 1946 on to 1982, Coatesville station was at the Newlinville electric substation.  The electric substation is constructed about 1927 to supply the steel works in town, and takes high voltage transmission line power and steps it down.  BIG transformers are involved.  

The Newlinville electrical substation was frequently redesigned to meet the needs of the steel works.  Transformer technology improves over time, with less waste heat.  A cooling pool is installed at one point to take transformer heat during high loads, and is removed after 2005.

So, we have a station with massive heat sources, that are moving as the station expands, and we need to detect changes in the station environment.  This obviates all the musings about what caused a precise shift at time X and magnitude Y--this microsite is literally a thermal zoo.  Automated breakpoint detection is the only way to salvage this data. 

Attached is a figure showing estimated power increases over time, developed from historical uses.  These increases would have forced numerous changes over time at the substation.  And consider, if a new transformer/switchgear is being installed, it would occur at a new pad on the site, while the old transformer/switchgear remains in operation.  On change over, the newer transformer/heat source has moved abruptly, and the old transformer idled and eventually removed.  Breakpoint.  And the new heat source is very likely southwest from the earlier heat source, and very likely further away from the weather station. Attached also is a random thermal image of a substation. 

[Mike Cycle]

Chubbs, I did try to expand your manual breakpoint detection process by asking AI to look at surrounding stations and identify "islands" where multiple stations are moving together.  I explored various settings, such as the minimum amount of synchronized temperatures before declaring an island, minimum stations to make an island, rules for stations leaving an island and new stations joining, and new islands forming.  The geographical area grew too large, pulling in Allentown etc. It will work, but its not a clear demonstration of using semi-local stations to identify breaks.  

However, Reading and West Chester do synchronize from 1929 to 1960 and make what looks like a solid reference against which breaks can be detected.  I have not looked at the individual breaks, but these breaks should be those that would be found by your approach.

Edit:  this does not rule out the case where Reading and West Chester have overlapping breaks.  

 

[ChescoWx]

12 hours ago, Mike Cycle said:

As an example of why breakpoint detection is absolutely essential, and where there are no alternatives besides something like a pictorial history of the station, Coatesville is a winner. From 1946 on to 1982, Coatesville station was at the Newlinville electric substation.  The electric substation is constructed about 1927 to supply the steel works in town, and takes high voltage transmission line power and steps it down.  BIG transformers are involved.  

The Newlinville electrical substation was frequently redesigned to meet the needs of the steel works.  Transformer technology improves over time, with less waste heat.  A cooling pool is installed at one point to take transformer heat during high loads, and is removed after 2005.

So, we have a station with massive heat sources, that are moving as the station expands, and we need to detect changes in the station environment.  This obviates all the musings about what caused a precise shift at time X and magnitude Y--this microsite is literally a thermal zoo.  Automated breakpoint detection is the only way to salvage this data. 

 

That is a great answer Mike! that transformer efficiency improving over time means declining waste heat, which puts spurious cooling into the raw record, which needs the later data warmed. That's the sign NCEI applied. It's a coherent mechanism and it's the first one offered in this thread that resolves the direction problem. Thank you for it.

However....I think it may cost you a bit more than you think this buys you! A few posts ago your position was "100% documented evidence of microsite impacts" and "Coatesville's breakpoints are fully documented at this point." Now the site is "literally a thermal zoo" and the complexity "obviates all the musings about what caused a precise shift at time X and magnitude Y." Those are different arguments. The first says the causes are documented and therefore the corrections are validated. The second says the causes can't be pinned down and therefore we have to trust the detector. I don't think both can be true of the same station, and the second is a much weaker claim than the one you started with.

It also runs into my actual objection rather than around it. If the microsite is too complex to attribute a given shift to a given cause, then it is also too complex to verify that the correction applied was the right size. You're offering unverifiability as a reason for confidence. Unverifiability is the thing I've been objecting to.

And your mechanism argues against the correction's shape. Transformer technology improving across decades is gradual. A cooling pool installed at one point and removed after 2005 is two events. Steadily declining waste heat produces a ramp, not a staircase. The PHA fits steps. By your own account the physical signal is the wrong shape for the correction being applied to it - and that's your description of the site, not mine.

There's a specific case I'd like your read on. The 1946-01 changepoint is +0.79 C, warming everything from 1946 forward. That's the move TO Newlinville. If the station relocated onto a substation with massive transformers and high loads, the raw record should have jumped warm at that moment and needed a cooling correction. NCEI warmed it instead. That only works if Newlinville was cooler than the water plant site it came from - which is your earlier "hot location to progressively cooler ones" account, and I don't see how it sits alongside "massive heat sources" at the same site in the same year.

Last thing, If Coatesville 1SW spent 36 years in a thermal zoo with industrial heat sources that moved, changed technology, and gained and lost a cooling pool, there's a conclusion available that neither of us has put on the table: that this record can't be salvaged for trend estimation and shouldn't be used for it. "Automated detection is the only way to salvage this data" assumes it can be salvaged. What's the basis for that, given everything you've just described? And is it a kind of a funny coincidence that the station carrying the county's largest adjustments is the one you're describing this way?

Paul

[ChescoWx]

So let's address Charlie @chubbs post above 

Charlie, of course no dispute that Phoenixville stepped down at the 1948 move. But you've written the very objection into your own post: "Not surprising because the bias adjustments are calculated from the raw data." Charlie, your test is pairwise comparison of a target against raw neighbors across a break date. But so is the PHA!  You've actually hand-implemented the method NCEI automates. Agreement demonstrates reproducibility, but of clearly not correctness as you are simpling showing us the same method, same data. It's the same issue as Berkeley and NCEI agreeing being offered as independent confirmation - they are of course not!

On your "bullet proof" claim - the four available estimates of this one documented move:

  Berkeley -1.23 F / yours vs Coatesville  -1.49 F
  NOAA  -1.88 F / yours vs West Chester   -2.03 F

Four pairwise estimates of the same event spanning 0.80 F; the largest is 65% bigger than the smallest. Same sign and rough scale but the real question here as always has been the magnitude is what's in dispute here, and that spread isn't what bullet proof looks like!!. Also, your 2 references disagree by 0.54 F. If both were stable through 1947-50 they should agree on that Phoenixville's step. They don't, so they moved relative to each other in your window - and you date West Chester to 1970 and Coatesville to 1946, before your "before" period even opens.

So you also say "all of the major county COOP stations moved in the same direction and roughly the same magnitude after the war". If so, doesn't that actually represent a real problem for the method, not support for it??. Pairwise comparison sees only RELATIVE change - a shift common to the whole network differences out and is invisible by construction. The more synchronised those moves were, the less your test and the PHA can detect, in exactly the era carrying the county's largest adjustments. So Charlie while none of this says your step estimate wrong. It means the method you are using is simply being validated against itself, the spread across implementations exceeds the confidence claimed, and the scenario you offer as clinching is the one the method handles worst.!!

Paul

[chubbs]

9 hours ago, ChescoWx said:

So let's address Charlie @chubbs post above 

Charlie, of course no dispute that Phoenixville stepped down at the 1948 move. But you've written the very objection into your own post: "Not surprising because the bias adjustments are calculated from the raw data." Charlie, your test is pairwise comparison of a target against raw neighbors across a break date. But so is the PHA!  You've actually hand-implemented the method NCEI automates. Agreement demonstrates reproducibility, but of clearly not correctness as you are simpling showing us the same method, same data. It's the same issue as Berkeley and NCEI agreeing being offered as independent confirmation - they are of course not!

On your "bullet proof" claim - the four available estimates of this one documented move:

  Berkeley -1.23 F / yours vs Coatesville  -1.49 F
  NOAA  -1.88 F / yours vs West Chester   -2.03 F

Four pairwise estimates of the same event spanning 0.80 F; the largest is 65% bigger than the smallest. Same sign and rough scale but the real question here as always has been the magnitude is what's in dispute here, and that spread isn't what bullet proof looks like!!. Also, your 2 references disagree by 0.54 F. If both were stable through 1947-50 they should agree on that Phoenixville's step. They don't, so they moved relative to each other in your window - and you date West Chester to 1970 and Coatesville to 1946, before your "before" period even opens.

So you also say "all of the major county COOP stations moved in the same direction and roughly the same magnitude after the war". If so, doesn't that actually represent a real problem for the method, not support for it??. Pairwise comparison sees only RELATIVE change - a shift common to the whole network differences out and is invisible by construction. The more synchronised those moves were, the less your test and the PHA can detect, in exactly the era carrying the county's largest adjustments. So Charlie while none of this says your step estimate wrong. It means the method you are using is simply being validated against itself, the spread across implementations exceeds the confidence claimed, and the scenario you offer as clinching is the one the method handles worst.!!

Paul

Lets look at the big picture. All four estimates you've listed agree that Phoenixville experienced spurious cooling in 1948. Better to recognize that Phoenixville had a cooling move vs ignoring the move and using the raw data.

However things aren't as uncertain as the four numbers make it seem. First the BEST estimate is 1.58F on Mike's chart not 1.88F, closer to NCEI. Second  If you look at Mike's charts there are many small differences between NCEI and Berkeley that cancel out over a longer time period. I don't see a big difference in the NCEI and Berkeley local climate history. Third, I wouldn't expect my estimates to be that accurate. I haven't taken the time to optimize my approach and more importantly there are many other stations besides West Chester and Coatesville that can be used to estimate the impact of Phoenixville's move.  

I say the NCEI method is bullet-proof because it has a strong scientific foundation.  Science says that there is a strong correlation in year-to-year temperature changes over a wide area and that's exactly what I have found. This chart is an example. These stations are well correlated at the time of the Phoenixville move indicating that they all can be used to estimate the timing and impact of the Phoenixville move.  Moreover, I have only scratched the surface, there are just the ones that had the time to check. 

 

[chubbs]

17 hours ago, Mike Cycle said:

Chubbs, I did try to expand your manual breakpoint detection process by asking AI to look at surrounding stations and identify "islands" where multiple stations are moving together.  I explored various settings, such as the minimum amount of synchronized temperatures before declaring an island, minimum stations to make an island, rules for stations leaving an island and new stations joining, and new islands forming.  The geographical area grew too large, pulling in Allentown etc. It will work, but its not a clear demonstration of using semi-local stations to identify breaks.  

However, Reading and West Chester do synchronize from 1929 to 1960 and make what looks like a solid reference against which breaks can be detected.  I have not looked at the individual breaks, but these breaks should be those that would be found by your approach.

Edit:  this does not rule out the case where Reading and West Chester have overlapping breaks.  

 

Yes, Reading and West Chester are stable in the 1929 to 1960 period and I am going to add Reading to my analyses. A couple of comments. As I said above looking at shorter time periods significantly expands the list of reference stations for individual station breaks as all stations are stable between the station breaks. Second, there is no need to restrict to short distances from Chester County. Weather station data is well correlated over hundreds of miles. An interesting first step would be to check how far good correlation with Chester County extends.

[Mike Cycle]

49 minutes ago, chubbs said:

An interesting first step would be to check how far good correlation with Chester County extends.

AI can do that easily.  I will put it to work.  Should we look at daily, monthly, or annual?  Tavg, or Tmin and Tmax?

[Mike Cycle]

10 hours ago, ChescoWx said:

That is a great answer Mike! that transformer efficiency improving over time means declining waste heat, which puts spurious cooling into the raw record, which needs the later data warmed. That's the sign NCEI applied. It's a coherent mechanism and it's the first one offered in this thread that resolves the direction problem. Thank you for it.

However....I think it may cost you a bit more than you think this buys you! A few posts ago your position was "100% documented evidence of microsite impacts" and "Coatesville's breakpoints are fully documented at this point." Now the site is "literally a thermal zoo" and the complexity "obviates all the musings about what caused a precise shift at time X and magnitude Y." Those are different arguments. The first says the causes are documented and therefore the corrections are validated. The second says the causes can't be pinned down and therefore we have to trust the detector. I don't think both can be true of the same station, and the second is a much weaker claim than the one you started with.

More strawmen!  

1) The 1946-direction of temperature change. "If the station relocated onto a substation with massive transformers and high loads, the raw record should have jumped warm". Built on a mechanism I never asserted, but you attributed to me and then rebutted.

2) "Unverifiability is being offered as a reason for confidence". Same false premise as #1.

3) "Your mechanism implies a ramp, PHA fits steps — wrong shape".  same false premise as #1.

4) "you claimed to know the direction/magnitude". Not something I ever said.  I have offered a satellite-based 30 m pixel surface temperature map that showed progressively cooler station siting conditions for the day the satellite data was collected, and that "Coatesville 1 SW generally moves progressively from hotter to colder microclimates".  The satellite map is pretty neat I think, so I put it here again.

5) The "documented" vs. "thermal zoo" contradiction charge.  Not something I have claimed.  I have said repeatedly, and will repeat now, that the I have fully documented the breaks as valid and necessary.  I have never claimed to know or documented the precise direction or magnitude of breakpoints--that is your reading. To document the timing, direction and magnitude of each break I would need to have precise and accurate microsite details, precise and accurate knowledge of the instruments, and a precise and accurate records of every interaction of the observer with the instrument--probably much more--throughout the station's record.  Beyond ridiculous.  

[chubbs]

49 minutes ago, Mike Cycle said:

AI can do that easily.  I will put it to work.  Should we look at daily, monthly, or annual?  Tavg, or Tmin and Tmax?

I would look at Tavg, annual and monthly, to start.