Skip to content

Dependence-Measure Selection Matrix

Decision tool — instantiates Correlation Structure Characterization

Maps the data's measurement scales and expected form to the dependence measure that is actually valid for them, so the coefficient fits the variables instead of the habit.

The default reflex is Pearson's r, and for a great many datasets it is the wrong tool applied without a thought. Dependence-Measure Selection Matrix is a lookup table whose rows are data conditions — the variables' measurement scale (nominal, ordinal, interval/ratio), the form the relationship is expected to take (linear, monotonic, arbitrary), sample size, and outlier-proneness — and whose cells name the appropriate measure together with the assumptions it demands. Its defining move is that it fixes the measure before any coefficient is computed, and records why: it is upstream policy, not a result. That is what separates it from the nonlinear screen, which detects unexpected structure after the fact — this matrix pre-selects the right instrument from what is already known about the data.

Example

An HR team wants to relate employee engagement (a five-point Likert item, ordinal) to tenure (months, ratio-scaled but heavily right-skewed). Habit says Pearson. The matrix routes the pair differently: an ordinal variable crossed with a skewed continuous one lands in the cell that recommends Spearman's rho or Kendall's tau, with a note that Pearson would be invalid because it presumes an interval scale and a linear form neither variable satisfies. The cell also records the reasoning, so the choice is auditable rather than reflexive — and if a later diagnostic shows a non-monotone shape, the matrix's fallback column points onward to a rank-free measure.

How it works

The matrix is a decision table keyed to two axes the analyst can judge before seeing results: the measurement scale of each variable and the assumed form of their relationship. Each cell carries three things — the recommended measure, the validity conditions it requires, and a fallback if those conditions fail. Because the recommendation is read off the data's properties rather than the coefficient's size, the measure cannot be reverse-engineered to maximize the number. The measurement context that drives the choice — scale type, units, coding — is captured as part of the selection, so the "why this measure" is recorded alongside the "which."

Tuning parameters

  • Scale granularity — how finely the rows split measurement types (e.g. treating ordinal-with-many-categories separately from few-category ordinal). Finer rows fit better but make the matrix heavier to maintain and apply.
  • Default and fallback measure — what the matrix recommends when conditions are ambiguous. A conservative default (rank-based, assumption-light) is robust but can understate a genuinely linear relationship.
  • Assumption strictness — how hard the validity conditions are enforced before a cell's measure is allowed. Strict enforcement blocks misuse but slows analysis.
  • Tie and small-n handling — whether ties and tiny samples reroute (e.g. Kendall over Spearman under many ties). Getting this wrong quietly biases the coefficient.
  • Robust variants offered — whether each cell exposes an outlier-resistant option. More options add power against messy data at the cost of choice overload.

When it helps, and when it misleads

Its strength is that it breaks the default-Pearson reflex and makes the measure choice explicit and auditable, matching the instrument to the variables' actual scale instead of assuming interval, linear, well-behaved data that rarely exist. It turns "which correlation should I run?" from a habit into a recorded decision.

Its central weakness is that a matrix is only as good as the expected form fed into it: judge the form wrong and it confidently pre-selects the wrong measure, offering false comfort ("I used the right test") while the real relationship is mis-specified. The classic misuse is reading the matrix backwards — trying cells until one yields the biggest or most significant coefficient. The discipline that guards against this is to fix the measure from the data's properties before seeing any result, and to let a diagnostic panel and a nonlinear screen validate the form assumption rather than trusting the matrix to have guessed it. The whole apparatus rests on getting the variables' level of measurement right in the first place.[1]

How it implements the components

Dependence-Measure Selection Matrix realizes the measurement-and-policy side of the archetype — deciding how co-variation will be quantified before it is:

  • dependence_measure_policy — the matrix is the policy for which measure applies to which data condition, with its assumptions attached.
  • measurement_context_record — its rows read from, and record, the scale, units, and coding of each variable that drive the choice.

It does NOT compute or display the coefficient (correlation_or_dependence_profile, relation_strength_indicator — Correlation Heatmap and Nonlinear Dependence Screen) nor vet raw-data integrity (paired_observation_alignment, missingness_and_measurement_error_record — Joint-Distribution Diagnostic Panel); it selects the measure, it does not run it.

Notes

The matrix's recommendation is provisional on the assumed form. Treat a surprising diagnostic panel or a hit from the nonlinear screen as grounds to re-select the measure, not as noise to be smoothed over — the selection is the start of characterization, not the end of it.

References

[1] Stevens' levels of measurement — nominal, ordinal, interval, ratio — are the standard taxonomy for what arithmetic a variable's values support. A coefficient that presumes more structure than the scale provides (Pearson on ordinal codes, say) reports a number that isn't meaningful; the matrix exists to keep the measure inside the scale's limits. ↩