ChescoWx Studies · Raw vs NCEI, state by state · November 2026

Michigan's 83 Counties: Raw Readings vs NCEI, 1895–2025

What Michigan's observers actually recorded at every station in NCEI's own daily archive, averaged county by county into a statewide record, and compared with the statewide and county temperatures NCEI publishes.

The main finding

The NWS raw trained observer temperatures show significant warming in Michigan since 1895 (+0.07 °F per decade), and so does the anomaly method, which measures each station against its own 1951–1980 average (+0.15 °F per decade). NCEI's adjusted temperatures show significant warming (+0.26 °F per decade). NCEI's trend line is statistically significantly different from both the raw and anomaly lines. Averaging what Michigan's observers actually recorded, county by county and weighted by county area, gives +0.07 °F per decade from 1895 to 2025 (significant), about +1.0 °F in total. Measuring each station against its own 1951–1980 average first (the anomaly method, 107 stations) gives +0.15 °F per decade (significant), about +1.9 °F. NCEI's published statewide series gives +0.26 °F per decade (significant), about +3.4 °F in total. NCEI's series runs 1.2–2.5 °F below the actual readings in every decade from 1895 through the 1940s; 1.0–1.3 °F below them in the 1950s to 1970s; 0.3–0.6 °F below them in the 1980s and 1990s; within 0.4 °F of them from 2000 on.

NCEI's adjusted temperatures are lower than the actual readings in every decade from the 1890s through the 2000s.

Blue is the area-weighted average of each county's actual readings; green is the same readings with each station first measured against its own 1951–1980 average; red is NCEI's published Michigan series. Each is shown against its own 1991–2020 average. Over exactly the same years and counties that have raw data, NCEI's own county series average +0.21 °F/decade, so the gap is not caused by which counties had data.

The short version

1. NCEI's 83 counties add up to its statewide series. Weighting NCEI's county values by area reproduces its Michigan statewide series within 0.07 °F in every year. Building the state from counties is therefore a like-for-like comparison.
2. The trends since 1895. The actual readings give +0.07 °F/decade (95% range +0.00 to +0.14, significant). The anomaly method gives +0.15 (range +0.08 to +0.22, significant). NCEI's series gives +0.26 (range +0.19 to +0.33, significant).
3. NCEI's changes to the raw temperatures significantly change the trend: the two trend lines are significantly different. The gap between NCEI's statewide series and the actual readings grows at +0.19 °F per decade (95% range +0.17 to +0.21, p < 0.001), so NCEI's adjustments add a statistically significant amount of warming. Against the anomaly method the gap is +0.11 °F per decade (95% range +0.07 to +0.15, p < 0.001), also statistically significant. County by county, NCEI's adjustments add significant warming to the trend in 31 of 54 counties (57%) and significant cooling in 0.
4. NCEI cools the early record in most counties. From 1895 through the 1940s, NCEI's county series is cooler than the actual readings in 89–100% of counties in each decade, by a median of 1.1–1.6 °F. The share of counties where NCEI is cooler is 86–96% in the 1950s–70s, 60–73% in the 1980s and 1990s, and 17–29% after 2000.
5. Recent decades roughly match, with NCEI slightly warmer. From 2000 on, the median county difference is +0.1 °F. NCEI is cooler than the actual readings in 22% of county-decades and warmer in 52%.
6. County verdicts. 54 counties have long enough raw records for a trend. In the actual readings, 25 (46%) show significant warming, 29 (54%) no significant trend, and 0 (0%) significant cooling. In NCEI's county series, 54 (100%) show significant warming, 0 no significant trend and 0 significant cooling. Under the anomaly method, 27 counties have enough stations with a 1951–1980 baseline for a trend; 12 (44%) show significant warming, 15 no significant trend and 0 significant cooling. NCEI's trend is steeper in 53 of 54 counties; St. Clair is the only exception.
7. The 1930s lose their warmth. In the actual readings the 1930s average +0.3 °F against the 1991–2020 normal, the warmest decade before 2000. NCEI puts the 1930s at -1.1 °F.

Decade by decade

NCEI's adjusted temperatures are lower than the actual readings in every decade from the 1890s through the 2000s.

County comparisons match each county's 1991–2020 level first, because a county average and a set of station readings sit at different elevations. That matching pulls recent decades toward zero by design. A county counts in a decade when it has at least 5 complete raw years.

County map

Same-station check: Michigan's long-record stations

For the 24 Michigan stations in NCEI's long-record network (USHCN v2.5), NCEI publishes an adjusted version of each station alongside its raw values, so the adjustment can be read directly with no county averaging. Before 1950 the adjusted values are cooler than the raw values at 95–100% of these stations in each decade (median -1.8–-1.4 °F adjustment). From 2000 on, 25–33% are adjusted cooler and 53–75% warmer.

Coverage

A station-year counts only when all 12 months are complete (3 or fewer missing days for both highs and lows). The fewest counties with a complete year is 10 (in 1895). Counties without data in a year are left out and the remaining counties are re-weighted by area. Antrim, Clare have no station with 10 or more complete years.

All 83 counties

Trends in °F per decade over the years each county has raw data. “Actual (raw)” is the average of the county's stations' actual readings. “Adjustment” is the trend of NCEI's county series minus the actual readings; * marks a statistically significant adjustment. County names link to NCEI's Climate at a Glance series for that county.

Stations used and sources

242 stations with at least 10 complete years. Each ID links to that station's daily file in NCEI's GHCN-Daily archive. Elevation is in feet.

• Raw daily highs and lows: NCEI Global Historical Climatology Network – Daily, version 3.35 (update of 24 Oct 2026), ncei.noaa.gov/pub/data/ghcn/daily/. Values with any NCEI quality flag were dropped; nothing else was changed.
• NCEI county and statewide series: nClimDiv, files climdiv-tmpccy-v1.0.0-20261006 and climdiv-tmpcst-v1.0.0-20261006, ncei.noaa.gov/pub/data/cirs/climdiv/; statewide series also at Climate at a Glance.
• Same-station check: NCEI USHCN v2.5 raw and FLs.52j monthly mean temperature, ncei.noaa.gov/pub/data/ushcn/v2.5/.
• County boundaries and areas: U.S. Census Bureau county boundaries, area in an Albers equal-area projection.

How this was done

• Stations: all 370 Michigan GHCN-Daily stations with daily highs and lows, placed in counties by their coordinates. A month counts with 3 or fewer missing days for both highs and lows; a year needs all 12 months; a station needs 10 or more complete years. That leaves 242 stations and 9,932 station-years.
• Combining: each county's value for a year is the average of the actual readings of its stations that have a complete year. The statewide value is the area-weighted average of the counties reporting that year. The same weighting applied to NCEI's county series reproduces NCEI's statewide series.
• Anomaly method: each station's annual values are first expressed as the difference from that station's own 1951–1980 average, using only stations with at least 15 complete years in 1951–1980 (107 of 242). County and state values are then averaged the same way. This removes differences in level between stations, so a change in which stations are reporting does not shift the average. 1951–1980 sits mid-record, catches the cooperative network near its largest, comes before most instrument and reading-hour changes, and is a standard baseline (NASA GISS). Using other 30-year baselines instead (1921–1950, 1961–1990, 1981–2010 or 1991–2020) changes which stations qualify; across all five baselines the Michigan anomaly-method trend ranges from +0.09 to +0.15 °F per decade, with significant warming under every one.
• Significance: linear trend with its uncertainty widened for year-to-year persistence, tested at 95%. Whether the two trend lines differ is tested on the year-by-year difference between them. County trends require 50 or more complete years, with at least 10 before 1931 and 10 since 1996 (54 of 83 counties qualify).
• Sensitivity: allowing up to 9 missing days a month gives +0.09 °F/decade. Using only stations with 25 or more complete years gives +0.11 °F/decade. NCEI statewide stays at +0.26.

Three ways to read the record, each with its own potential biases

• Raw: readings averaged as recorded. Uses every qualifying station. The average depends on which stations are reporting: if the network shifts toward higher, cooler sites or toward warmer valley and town sites over time, the average moves even if no individual site changed. It also carries any uncorrected station moves, reading-hour changes and instrument changes.
• Anomaly: the same raw readings, each station measured against its own 1951–1980 average. Removes the station-mix effect, but only stations with a long record in 1951–1980 qualify (107 of 242 here), so later and shorter records are left out. A station that moved or changed instruments carries that step into its anomalies. Choosing a different baseline period, for example one in a warmer or cooler phase of the natural climate cycles, changes which stations qualify and can shift the result.
• NCEI adjusted. Corrects reading-hour changes and uses neighbor comparisons to estimate and remove breaks in each station's record, most of them without documentation. The corrections are statistical estimates rather than measurements, and an error at one station can spread to its neighbors.
• All three involve choices made by people. Completeness rules, baselines, weighting and adjustment settings are all judgment calls, and any analyst, NCEI or ChescoWx included, can favor choices that confirm what they expect. That is why all three are shown side by side, with the data and method published.

Caveats

• Raw readings at the station level have some potential to carry occasional biases that NCEI believes may require the estimated adjustments that NCEI has chosen to apply to the actual raw data recorded by National Weather Service trained observers. This study quantifies how much the adjustments change the warming or cooling answer, not whether either the raw or adjusted data is correct.
• Coverage is thinner before 1900 and uneven by county. Stations whose coordinates fall outside every county were left out: Clinton.