Back The method

Causally honest analytics

Eight auditable pipeline stages implementing the DCCA loop, indicators that pair outcomes with the behaviors meant to produce them, and one discipline underneath it all: no claim presented above its evidence.

S0 to S7

A decision flow, not a linear sequence. Stages 3 and 4 are blocking gates: bad data halts the cycle, records the failure with diagnostic detail, and never reaches a score.

S0
Data preparation
Format conversion, deduplication, structural alignment, per source
S1
Ingestion
Extract and load, with run records and freshness timestamps
S2
Correction
Deterministic fixes for known source quirks, logged
S3
Validation
Schema, completeness, consistency, business rules. Blocking gate
S4
Statistical validation
Simpson's paradox, distribution sanity, outliers. Blocking gate
S5
Signal modeling
Criterion scores, with SHAP and partial-dependence explainability
S6
Composite scoring
Effective weights applied per entity, attribution recorded
S7
DCCA finalization
Drift check, weight update, snapshot persisted append-only

A bad extract fails loudly. Silent degradation would corrupt the weight snapshots and propagate for cycles before detection, so the gates make failure the visible outcome.

The DCCA loop

Four controls keep the weights honest

Dynamic Correlation-Causation Adjustment answers one question: how should weights be maintained while the world underneath them drifts? Keep scrolling. It comes back around.

Control 1

Stability monitoring

Drift measured against a fixed baseline, with bands published before any observation: stable below 0.10, flagged to 0.25, forced recalibration above. Falsifiable, not post-hoc.

Control 2

Weight recalibration

A smoothed update every cycle, single-cycle shift capped, vector renormalized. Large moves take several cycles and leave a clear trail.

Control 3

Data-quality validation

Nothing enters a weight update without passing the pipeline's two gates. Failures are explicitly non-silent.

Control 4

Change logging

Every weight update persists an append-only snapshot with the observed signal that drove it. No row is ever modified, and the schema makes mutation detectable. Then the next cycle begins, back at control one.

Indicators

Outcomes, paired with the behaviors that produce them

Every strategic objective carries at least one of each, and the causal link between them is explicit: performing this activity, at this cadence, produces that outcome.

Lagging, outcome-based

Ethical performance indicators

Whether a result the gate identified as mattering in itself was achieved: emissions reduced, disputes resolved, externalities compensated, findings closed. The "ethical" mark means the organization can explain why this outcome was selected for measurement, not merely that it was.

Leading, activity-based

Behavioral event signals

Whether the upstream behaviors expected to produce the outcome are occurring at the expected cadence and quality: training completions, tickets filed within policy, reviews conducted on time.

The pairing prevents two failures. Signals without outcomes is checklist compliance: the box ticked, the result never verified. Outcomes without signals is accountability without diagnosis: results the organization cannot explain. And when signals are healthy but the outcome never moves, the hypothesis was wrong, and it is the activity that gets redesigned.

Composite scores

Weights you can challenge

A weighted sumRen, J. (2021) Multi-Criteria Decision Analysis for Risk Assessment and Management. Cham: Springer. is a policy decision. Anyone disagreeing with a conclusion can disagree with the weights, not the arithmetic.

Three layers

Base weights. Set by the governance council, published in policy, static between reviews.
Domain weights. Adapted from observed data by the DCCA loop, smoothed every cycle.
Effective weights. Derived from both. Ordering within a domain is preserved; emphasis between domains adapts.

Missing data is not bad news

A criterion with no data scores the neutral sentinel, the exact midpoint of the scale. Zero-filling would punish absence as if it were failure; dropping silently would rebalance the score with no audit evidence.
Anywhere a reader might mistake it for a score, the sentinel is labeled insufficient data. The scalar is an operational convenience; the label keeps it from being read as meaning.
From correlation to cause

Three methods, earned in order

The causal hypotheses already live in the business map and the metric registry. The methods formalize them and put them to statistical scrutiny, each becoming available as the organization matures.

From the metric catalog

Structural equations

The hypothesized links become a system of equations, constrained by the business map's topology. Fit tests either confirm the structure or send the governance body back to revise its map.

From the first controls

Difference-in-differences

Every control deployed to some entities and not others is a natural experiment. Treated against untreated, before against after, with matching when trends diverge. Answers the audit question: did this control produce its effect?

After thirty stable cycles

Bayesian networkPearl, J. (2013) Causality: models, reasoning, and inference. 2nd edn. Cambridge: Cambridge University Press.

The confirmed structure is promoted to a probabilistic model with full posterior distributions, able to answer intervention questions: if we move this, what happens to that, and with what confidence?

Correlation is not causation Simpson's paradoxPearl, J. (2013) Causality: models, reasoning, and inference. 2nd edn. Cambridge: Cambridge University Press. Multiple comparisons Survivorship bias False precision

Five pitfalls the method defends against by construction, not by promise.

Every claim is labeled correlational or causal. A claim that cannot be supported at the level it is presented is downgraded or retracted.