Factoring out an effect
Objective
Where the evidence identifies an effect, factor it out so that the measurement moves closer to independent reference values, within its stated uncertainty.
Mechanism shown Current focus See the results How stages are defined
Taking the method from designed test data to real satellite data, at a scale large enough to check.
An adjustment that improves agreement with the data it was learned from proves little. The difficulty is showing that it improves agreement with evidence it never saw, and that a systematic offset is not simply absorbed into a wider uncertainty.
An effect left in the data shifts every conclusion drawn from it. Factoring it out is worth doing only when the result agrees better with independent evidence.
For each case we record the original value, the size of the effect factored out, the result and its remaining uncertainty, and check the result against reference observations that were not used to estimate the effect. A shift in the centre of the values is assessed separately from a change in their spread, and both are compared with existing processing.
On cases held back from estimating the effect, results agree better with independent reference measurements than the original values and existing processing do.
Each result is dated and published on its own page with its data, conditions and limits. Results that went against the objective are listed here with the rest.
-
A known effect can be recovered from designed test data
RefCal can recover a known effect from designed data without disturbing anything already live. This establishes the mechanism on data where the answer is known. The next step is a real-data run large enough to support a checkable result.
Full result, data and credits ->
Established work this area builds on. Results from the programme are published separately, with their data and limits.
- Learning about physical parameters: the importance of model discrepancy (2014) J. Brynjarsdottir and A. O'Hagan Why an adjustment must be checked on evidence it was not fitted to.
Questions about this work?
If you work on this problem, or hold reference data that bears on it, contact us.