A Simple Statistical Procedure for Partitioning Soil Test Correlation Data Into Two Classes

TLDR

Soil test laboratories usually classify results into categories to simplify fertilizer recommendations, yet the boundaries are often subjective. The study proposes a statistically sound method to determine class limits for soil test data. The method partitions data into two groups by iteratively selecting critical levels that maximize overall R², using the group means as predictors, and is illustrated with typical soil test data. The Low–High split produced the highest R², outperforming continuous models, and the approach was validated across 200 data sets, suggesting broad applicability.

Abstract

Abstract Most soil test laboratories divide soil test results into two or more classes for the purpose of making fertilizer recommendations. This is usually done for the practical reason that it reduces the number of different recommendations necessary. However, the basis for defining the different classes (e.g., Very Low, Low, Medium, etc.) is often subjective or arbitrary. This paper explains a simple, yet statistically sound, method for setting the class limits. The procedure is to split the data into two groups, using successive tentative critical levels to ascertain that particular critical level which will maximize overall predictive ability ( R 2 ), with the means of the two groups (classes) as the predictor values. The paper presents an actual example, using data believed typical of this kind of problem. Several continuous correlation models were also fitted to the same data. None gave as high an R 2 as a single Low‐High split defined by the suggested procedure. Similar results have been obtained with a majority of 200 sets of soil test correlation data, indicating that the new procedure may be widely applicable.