Evaluating Adjustments in Global Temperature Datasets

Hello. The earlier modules developed the physics of warming and the carbon cycle. This module turns to the evidence: how temperatures and other climate quantities are measured, assembled, and tested. Here the question is not simply whether a thermometer reading is accurate. It is whether readings made under changing conditions can support a reliable estimate of a long-term trend.

You will examine three potential sources of distortion in land-temperature records: changes in the time thermometers are read, changes at the station itself, and urban development around it. The aim is to assess both the justification for correcting these effects and the ways a correction could fail. Your statistical background will be useful: an adjustment is an estimate of a measurement bias, not an instruction to make the resulting trend larger or smaller.

The quantity we want is not quite the quantity recorded

A weather station measures temperature at a particular place, with particular equipment and observing practices. A climate analysis typically asks a different question: how has temperature changed across a region, measured as consistently as possible over time?

A useful conceptual model for station $i$ at time $t$ is

Here ${\mu }_i$ represents the station’s usual local temperature, $C_{r(i),t}$ the changing climate signal in its region, $b_{it}$ a bias associated with how or where the measurement is made, and ${\epsilon }_{it}$ short-term variation and measurement noise. The decomposition is not directly observable; it frames the inference problem. If $b_{it}$ is constant, it affects the station’s level but not its trend. If it changes, an unadjusted trend can partly reflect the measurement system.

This distinction also explains why a large correction to one station does not, by itself, establish a large correction to a global average. The effect on an aggregate depends on when and where biases changed, their signs, and how observations are spatially combined.

Time of observation: a bias with a testable mechanism

Many historical stations recorded the highest and lowest temperatures since the previous reading, rather than measuring a midnight-to-midnight calendar day. Consider an unusually warm afternoon followed by a cooler day. If the observer resets the maximum thermometer in the late afternoon, warmth just after that reset can be counted in the following day’s maximum. A morning reading can similarly lead to repeated counting of a cold minimum. Monthly averages calculated from these daily maxima and minima can therefore depend on the reading hour.

The important issue for a trend is that observation schedules changed. In the U.S. volunteer network examined below, a widespread move from afternoon toward morning readings introduced an artificial cooling tendency into unadjusted records. Correcting that tendency tends to increase the estimated warming trend. That direction follows from the measurement mechanism; it is not a general rule that temperature adjustments increase warming.

The validation design matters as much as the reported result. With hourly observations, researchers can construct a reference daily maximum and minimum, then reconstruct what would have been recorded under alternative reading schedules. They can apply the adjustment and inspect what bias remains. In the study’s test, residual bias is smaller than the original time-of-observation bias, although performance varies by reading hour. This supports the correction without implying that every historical station-month has been recovered exactly.

There are limits to the inference. Hourly reference stations are not a random sample of every historical observing environment, and knowledge of past reading times can be incomplete. Comparing trends with other datasets is informative, but shared stations and changes in station coverage can weaken claims of independence. The strongest case combines a plausible mechanism, documented schedule changes, and tests against observations detailed enough to reveal the bias.

Station moves and instruments: look for breaks relative to neighbours

Now suppose a station moves a short distance, or its thermometer and shelter are replaced. The new site may be systematically warmer or cooler even though both sets of readings were correctly recorded. A step in the station series should not automatically be interpreted as a step in regional climate.

Nearby stations provide a way to investigate. If two stations share much of their weather and climate variation, subtracting their series suppresses the common signal:

An abrupt, persistent change in this difference is evidence that something changed at one station relative to the other. Comparing the target with multiple correlated neighbours helps identify which series contains the break. Where station histories exist, a recorded move or instrument change can corroborate the statistical finding; where histories are missing, pairwise comparisons can still detect some changes.

PHA does not label every temperature jump “artificial.” Its working assumption is narrower: a persistent shift unique to one station relative to correlated neighbours is likely to be non-climatic. The documentation describes estimating the correction from pairwise differences with neighbours that do not show a break at the same time, and not applying corrections for statistically insignificant changes.

That is a defensible approach, but it has identifiable failure modes. Sparse networks offer fewer useful comparisons. A change shared by most neighbouring stations may be difficult to distinguish from a genuine regional change. Conversely, a real local climatic change that is not shared by neighbours could be treated as a station artifact. A gradual bias may not appear as a sharp break at all. Correlation and station density affect how persuasively the data identify a change; neither makes the classification infallible.

The target quantity matters here. If a station’s immediate surroundings genuinely become warmer after nearby land is developed, that warming is real at the station. It may nevertheless be unrepresentative of the broader regional climate that a gridded temperature estimate is intended to measure. Calling it a “bias” is a statement about the estimate’s purpose, not a claim that the thermometer malfunctioned.

Urban heat islands: separate a warm city from a warming bias

Built surfaces, reduced vegetation and evaporation, building geometry, altered airflow, and waste heat can make a city warmer than surrounding countryside. This is the urban heat island effect. Its existence is not, on its own, evidence that urban stations exaggerate a climate trend. A city could be consistently warmer throughout a record and contribute a higher temperature level without adding a spurious warming trend. The concern is change in the local urban influence over time—for example, development spreading around a formerly less-developed site.

Urban effects pose a particular challenge for relative homogenization. A sharp change in a station’s surroundings might appear as a detectable break. Slowly expanding development may instead produce gradual drift; if nearby comparison stations develop similarly, subtracting them will not expose a shared bias. Equally, one cannot assume every urban station has such drift or assign the same correction to every city.

An assessment should therefore distinguish three questions:

1. Is there an urban–rural temperature difference? Often yes, but its magnitude depends on place, weather, and time of day.
2. Has that difference changed during the station record? Only a changing difference can create the particular trend bias at issue.
3. Does that changing difference materially affect the spatial average? This depends on the stations used, their weighting and coverage, and checks against less-affected observations.

The video addresses the first question’s physical basis. It does not answer the second or third. Those require evidence from station histories, comparisons over time, and analysis of the resulting aggregate.

Reading adjustments without mistaking them for the climate signal

The following historical U.S. graph plots differences between successive processing stages, in degrees Fahrenheit—not the temperature record itself. Its labels distinguish a time-of-observation correction, an instrument-related stage marked MMTS, and further processing stages. It is an example of why “the adjustment” is too vague a description: separate procedures can have different signs and time patterns.

Read the legend literally. “TOBS minus RAW” compares time-of-observation-adjusted data with raw data. “MMTS minus TOBS” shows an additional stage, not an alternative estimate of the same correction. For a given year, successive differences can be added to compare the final and raw series, because the intermediate stages cancel algebraically. But a curve’s height at one year is not its contribution to a long-term trend. For that, examine how the curve changes across the period of interest.

The graph warrants careful boundaries on interpretation:

• It concerns a U.S. land network over the period shown. It cannot establish the adjustment to a global land-and-ocean mean.
• It displays processing differences, not whether each change is correct. Accuracy needs tests such as metadata checks, comparisons with detailed observations, and sensitivity analyses.
• Several stages partly offset one another. Neither “all corrections warm the past” nor “all corrections increase the trend” follows from this figure.

A global mean surface-temperature product introduces further decisions after station correction: constructing anomalies, handling uneven geographical coverage, and combining land observations with sea-surface temperatures measured through different systems. This lesson’s sources principally examine land stations, so their results should not be silently extended to ocean measurements.

A sound critical assessment consequently asks two separate questions: Does each adjustment estimate a plausible, independently testable measurement effect? And how much does applying it change the regional or global result? A large correction to an individual station can be justified, while an aggregate can be sensitive to many individually modest corrections. Both propositions must be checked rather than assumed.

Conclusion

Temperature adjustments address a real identification problem: observing schedules, equipment, sites, and surroundings can change while the climate is also changing. Time-of-observation bias has a specific mechanism and can be tested using hourly records. Pairwise homogenization uses neighbouring stations to detect otherwise undocumented breaks, but is less secure where comparisons are sparse or biases are shared or gradual. Urban warmth becomes a trend issue only when its influence changes over time and affects the quantity being estimated.

The standard is neither “raw data are always preferable” nor “adjusted data are automatically correct.” It is whether the bias mechanism, correction method, validation evidence, and effect on the final estimate withstand scrutiny. In the next lesson, you will apply that same discipline directly to a global temperature time series: estimating its trend while accounting for autocorrelation and natural variability.