Accounting for measurement errors when harmonising incongruent soil data − a case study

When collating soil data from different sources, the data should be congruent. Ordinary linear regression (OLR) has often been used to harmonise incongruent data. To do so, one of the sources is nominated as the reference and so is assumed to provide data that are determined without error despite evidence to the contrary. Alternative approaches that can handle errors in both variables, such as constructing a maximum likelihood functional relationship (MLFR), are seldom used. Two scenarios compared these two approaches using soil organic carbon data determined by the Walkley and Black method or the Dumas method. An inter-laboratory proficiency program provided data to represent an ideal scenario of complete information on precision, i.e. a mean and standard error of multiple determinations for each method as applied to each soil sample. In this scenario, it was found that the recovery of carbon was not consistent between laboratories or methods, nor was the precision of determinations consistent. Importantly, the precision data showed how neither method had an advantage and so could serve as a reference. Unfortunately, soil researchers are more likely to be trying to harmonise data from single determinations and have no data on the precision of either method. This second scenario was explored using legacy data and new data from re-analysis of 116 archived soil samples, with precision data from different external sources. Here the OLR regression coefficients were found to be much less accurate than those from using the MLFR harmonisation model. We concluded from these scenarios, that MLFR should be used to harmonise incongruent data when data on measurement errors are available. MLFR gave different predicted values to OLR while accounting for measurement errors in both variables. Where sufficient information on precision is lacking, OLR yields similar results and so may be an easier but less rigorous option. However, more research is needed to establish when OLR can be used versus when MLFR should be used.