The THOR dataset is a high-value historical source, but like all digitised military archives, it carries systematic biases that must be understood before drawing operational or archaeological conclusions. THOR originates from paper mission reports digitised decades after the war, and the chain from combat log to spreadsheet row introduced errors, omissions, and classification inconsistencies that are visible in the data itself. A.D.I. has identified four categories of bias that any team using this data should account for before deploying findings in the field.Documentation Index
Fetch the complete documentation index at: https://mintlify.com/HelenDiMo/archaios-data-Intelligence/llms.txt
Use this file to discover all available pages before exploring further.
The Four Biases
1. Representativeness Bias
1. Representativeness Bias
Only documented missions are present in the THOR record. Some operations were never archived — either because record-keeping was incomplete in the field, because documents were lost, damaged, or destroyed during or after the war, or because certain Allied air forces contributed records at different levels of completeness.THOR covers USAAF and RAF missions most thoroughly, along with some contributions from Australian, New Zealand, and South African air forces. Operations conducted by other Allied forces, or missions for which the source documentation was never located during the digitisation programme, are absent.Impact for A.D.I. users: Zones showing low mission density in the THOR-derived heatmaps are not necessarily inactive — they may simply be under-documented. An archaeological team should not rule out a low-density area as a prospection candidate on THOR data alone. The absence of a record is not evidence of absence of activity.
2. Geographic Bias
2. Geographic Bias
The dataset records target points where bombing was documented. It does not represent the full spatial footprint of destruction — bomb patterns spread across areas substantially wider than the recorded target coordinate, and many
Target Latitude / Target Longitude fields contain centroid approximations rather than precise impact points.A significant number of records have NULL or missing coordinate values (Target Latitude: 169,565 non-null out of 178,281 total records in 2_operations.csv; Target Longitude: 169,543 non-null). These missing values mean that spatial analysis underestimates the true geographic distribution of activity.Additionally, coordinates in the original paper records were sometimes expressed as grid references or approximate area designators rather than precise GPS-equivalent positions. When digitised, these became centroid coordinates that place many missions at the same point on a map — creating artificial density clusters that do not reflect the real dispersion of impact.Impact for A.D.I. users: Heatmap “cold zones” — areas appearing empty in the Power BI spatial view — require independent verification before being ruled out as archaeological prospects. Cross-reference with aerial photography and local military maps before concluding that a zone was not bombed.3. Labelling Bias
3. Labelling Bias
Many records carry vague or absent target classification. In
2_operations.csv, 176,018 of 178,281 records have a Target City value, but a substantial fraction of those values are the string UNKNOWN. Similarly, the Target Type column (146,370 non-null) contains the value UNIDENTIFIED TARGET as the second most common classification across the France scope.A specific digitisation artefact: the “unidentified target” class appears under two different spellings in the raw data:UNIDENTIFIED TARGET(574 occurrences in the France/1943–1944 scope)UNIENTIFIED TARGET(1,181 occurrences — a typo from original paper digitisation)
Target Type frequency counts without this normalisation will misrepresent the true distribution of target classes. Any analyst running independent queries against the raw CSV files must apply the same normalisation to reproduce A.D.I.’s figures.4. Temporal Bias
4. Temporal Bias
Record quality is not uniform across time. The 1943 portion of the dataset was digitised from operational records that were less consistently maintained than those produced in the run-up to D-Day. As Allied planners prepared for Operation Overlord in 1944, documentation standards improved significantly — more complete mission logs, more detailed target information, and better after-action reporting.The consequence is that the apparent increase in recorded mission activity between 1943 and mid-1944 partly reflects improved record-keeping, not necessarily a proportional increase in actual bombing sorties. The June 1944 period is particularly well-documented because of the invasion’s logistical and historical significance.Impact for A.D.I. users: Time-series comparisons between 1943 and 1944 must account for this documentation improvement effect. A spike in recorded missions in May–June 1944 should not be interpreted purely as a spike in actual bombing intensity — some of that increase is an artefact of better records. Normalising by estimated total sorties flown (from secondary historical sources) is recommended for any trend analysis.
Recommended Supplementary Sources
Before making any site-selection, safety, or investment decisions based on A.D.I. outputs, the following additional source categories should be consulted:- U.S. National Archives — Military maps and operational planning documents from the European Theater of Operations (ETO)
- RAF Operational Record Books — Held at the National Archives, Kew (London); provide mission-level detail for RAF sorties not fully captured in THOR
- French local archives and municipal damage records — Regional and departmental archives in Normandy hold post-war damage surveys and administrative records documenting destruction at the commune level
- Imperial War Museum aerial photography collection — Pre- and post-raid aerial photographs from the IWM allow visual cross-validation of impact crater locations against THOR coordinate data
- Post-war Allied bomb damage assessment (BDA) reports — Produced by Allied air forces in 1944–1945, these systematic photographic surveys document actual damage and can confirm or contradict THOR target coordinates
The THOR dataset has proven its value for locating unexploded ordnance in Southeast Asia. For Normandy, supplementing with French EOD (Explosifs et Engins de Guerre) clearance records from the French Sécurité Civile is strongly recommended before planning any ground work. French EOD services maintain ongoing records of UXO finds and clearance operations that post-date any historical dataset.
Data Quality Handling in the A.D.I. Pipeline
The notebook and Power BI model incorporate the following cleaning and governance decisions, documented here for reproducibility: 1. Dataset exclusion —1_cleanbombww2.csv: This dataset is excluded from all analytical outputs. When filtered to France in 1944, it yields a single record with a coordinate-format city name (44 00 N 001 00 E) mapping to central France — not Normandy. The dataset is heavily biased toward the Mediterranean Theater of Operations (MTO) and contains no usable Normandy signal.
2. UNIDENTIFIED TARGET typo normalisation: The two variant spellings (UNIENTIFIED TARGET and UNIDENTIFIED TARGET) are merged into a single canonical UNIDENTIFIED TARGET class before any target-type aggregation is performed. This affects counts in the Power BI dashboard and in the adi_master_1943_1944_CLEAN.csv export.
3. 999-ton outlier flagging in 2_operations.csv: The maximum value of 999.0 in the Total Weight (Tons) column is identified as a legacy documentation placeholder — a sentinel value used when the exact bomb load was unknown at the time of digitisation. This value is flagged and excluded from tonnage-based spatial density calculations. The field’s natural distribution (mean ≈ 44t, median ≈ 30t, max clean ≈ 602t) is used for the UXO risk layer.
4. NULL ordnance field handling via weapon_type classification: Many records have NULL values in High Explosives Weight (Tons) and Fragmentation Devices Weight (Tons). Rather than dropping these records, A.D.I. fills NULLs with 0 and applies the weapon_type classification logic. Records with he_tons = 0 and frag_tons = 0 receive the label Sin Datos de Armamento (no ordnance data), preserving them in the mission count while clearly marking their ordnance data as unavailable for the UXO risk layer.