Test whether a conclusion survives a change of index recipe

Runs the same samples through two or more index recipes and reports whether the ranking of samples survives. This is the practical question behind most methodological disagreements in the literature: it usually does not matter which scoring or aggregation route is "better" in the abstract, it matters whether your conclusion changes when you switch.

Usage

sqi_stability(..., labels = NULL)

Arguments

...

Two or more sqi_result objects or numeric vectors, all of the same length. Name them (e.g. linear = , sigmoid = ) and the names are used in the report.

labels

Optional character vector of names, overriding those taken from ....

Value

An object of class sqi_stability, a list with:

n

Number of samples

indices

Named list of the index vectors compared

pairs

Data frame with one row per pair: the two labels, the Spearman rho, and whether the top and bottom samples are preserved

stable

TRUE when every pair preserves both extremes

Details

For every pair of indices the Spearman rank correlation is reported, along with a flag for whether the best-ranked and worst-ranked sample stay the same. A high rank correlation with a changed top sample is still a changed conclusion if the top sample is what you act on.

Yuan (2026) used this kind of stability as a selection criterion in its own right: the EMDS route achieved fidelity R-squared of 0.74-0.77 with no combination of scoring and aggregation producing p > 0.05 – that is, it was stable, not merely accurate.

On the name. The CRAN package SQIpro exports an unrelated sqi_compare(), which tabulates the index value each of its methods produces. This function asks a different question and therefore carries a different name: not what value does each recipe give, but does my conclusion hold when the recipe changes. The unit of comparison here is the ranking, not the values.

See also

sqi_validate to assess a single index

Examples

props <- c("pH", "OM", "N", "P", "K", "CEC", "BD")

linear <- compute_sqi_properties(
  soil_data, properties = props, id_column = "SampleID",
  scoring_rules = standard_scoring_rules(props, scoring = "linear")
)
sigmoid <- compute_sqi_properties(
  soil_data, properties = props, id_column = "SampleID",
  scoring_rules = standard_scoring_rules(props, scoring = "sigmoid")
)

sqi_stability(linear = linear, sigmoid = sigmoid)
#> Soil Quality Index recipe stability
#>   Samples: 50 
#>   Indices: linear, sigmoid 
#> 
#> Pairwise rank agreement
#>   linear       vs sigmoid       rho =  0.884   top kept      bottom CHANGED
#> 
#>   At least one pair disagrees on the best or worst sample. A high
#>   rank correlation does not rescue this if the extreme is what you
#>   act on.