Why old data is tempting, and why it is doubted
A regional stream-sediment survey from the 1970s may cover thousands of square kilometres that nobody will sample again. Old drill logs may describe core that no longer exists. Ignoring this is expensive. Using it without checking is risky, because the numbers were produced by methods that differ from today's in ways that are easy to overlook.
What differs between old and modern assays
- Digestion. Many older methods used a weak or partial acid attack that released only part of the metal in the sample. Modern multi-acid digestions or X-ray methods measure nearly all of it. The two are not the same quantity.
- Measurement. Colorimetric methods, read by eye against a standard, give values in coarse steps. A table full of 5, 10 and 15 ppm is a sign of this.
- Detection limits. Old limits were higher, and values below them were recorded in several ways: as zero, as the limit, as half the limit or as a text flag.
- Batches. A survey run over years, by several laboratories or operators, carries systematic differences between batches.
- Location. Positions were read from paper maps in a local datum. Errors of hundreds of metres are common, and a wrong datum shifts a whole sheet.
- Quality control. Blanks, standards and duplicates were often not used, or not recorded.
The test: paired samples
The question can be answered where some samples were measured twice, once by the old method and once by a modern one. Re-assayed archive pulps, resampled sites and twinned drillholes all provide such pairs.
The pairs show how the two measurements relate. Sometimes there is a steady offset or ratio, and the old values can be translated with a known margin of error. Sometimes the relation holds within each batch but differs between batches. Sometimes there is no relation at all, and the old values carry no usable information about that element.
Why one correction factor is not enough
The common shortcut is to compute a single factor for the whole dataset and rescale everything. This fails when batches differ, which they usually do. A factor that is right on average is wrong for most batches, and the rescaled values can be worse than no values. Corrections should be fitted at the level where the differences arise, and each translated value should carry its own uncertainty so that it counts for less than a modern measurement.
Checking that it helped
A correction that looks reasonable is not yet shown to be useful. The test is to hide some modern results, predict them with and without the old data, and compare the errors. If including the old values makes predictions better, they earn their place. If it makes no difference or makes them worse, they are left out for that element.
In one piece of our work on a regional survey, this test gave different answers for different metals: the old copper values improved the map, while the old lead and zinc values did not and were dropped. The detail is in the story Can fifty-year-old assays be trusted?
Practical rules
- Keep the original values untouched and store any translated values in separate columns, with the method recorded.
- Decide element by element. A survey can be usable for copper and useless for lead.
- Verify locations before chemistry. A good assay in the wrong place is a bad data point.
- State the margin of error with the result, and let it widen where only old data exists.
- Treat legacy data as a guide to where to sample, not as a substitute for sampling.
Where this fits
Reconciling old and new data is part of data integration. For public disclosure, the use of historical data is also a regulatory matter: reporting codes require its source, reliability and verification to be discussed by a qualified or competent person.