Mix Effects in Blended Quality Targets

From WFM Labs
Two nodes each hold their own target; as the share shifts toward the node with the lower target, the blended line falls by arithmetic alone.

Mix effects in blended quality targets describes what happens to an organization's headline quality score when its delivery nodes — locations, delivery arrangements, or suppliers — are assigned different targets and the headline is a volume-weighted blend of them. As the share of work handled by a node with a lower target grows, the blended score falls even when every node sits exactly on its own target. The decline is not a performance signal; it is the arithmetic consequence of the target architecture. The effect is distinct from Simpson's Paradox in Contact Center Metrics, which concerns a trend that reverses when data are disaggregated because of a hidden confounder; the mix effect described here involves no reversal and no confounder, only a weighted average whose weights are moving, though composition effects of this kind are often filed loosely under the paradox's name. This page gives the arithmetic, explains why location-differentiated targets create the effect and tier-differentiated targets do not, and sets out the target architecture that prevents it. The general discipline of setting targets from measured baselines is at Baselines Before Targets.

The arithmetic

A blended score is a share-weighted average: the headline equals the sum over nodes of each node's share of volume multiplied by its score. If each node is held to its own target and achieves it, the headline equals the sum of each node's share multiplied by its target. Any movement of share from a node with a higher target to a node with a lower one lowers the headline, with no node having done anything different.

An illustration with invented figures: two nodes, one held to 93 and one to 90. At a ten-percent share for the lower-target node the blend is 92.7; at twenty percent it is 92.4; at forty percent it is 91.8. Every node has hit its target throughout. A headline that reads "down nine tenths of a point" describes nothing but the mix.

The effect has a well-known form in demography and epidemiology, which deal with differences in crude rates between two populations. That difference is decomposed into a component due to differences in composition and a component due to differences in the underlying rates;[1] the remedy long used there — reporting directly standardized rates that hold composition fixed — is the remedy here.

Why location targets create it and tier targets do not

A target that varies by service tier is a product decision: the customer bought a level of service, and the target says what was bought. A target that varies by location or delivery arrangement is a different kind of statement. It says that the organization expects less of one of its own nodes, and it has three consequences that a tier target does not.

  • It builds the decline into the growth plan. If the lower-target node is also the lower-cost node, every step of the cost strategy lowers the headline by construction. The organization has arranged to look worse as it succeeds.
  • It is an institutional admission. A location-differentiated target concedes that the node is expected to deliver less, and it makes the commercial objection — why would a client agree to be served from a node the provider itself rates lower — unanswerable. The comparability argument is developed at Comparing Delivery Arrangements.
  • It contradicts a location-agnostic service promise. An organization that tells its clients service quality does not depend on where the work is done, while holding its nodes to location-varying targets, has a goal architecture that contradicts its positioning, and the commercial side of the business notices the headline before it notices the reason.[2]

A tier target has none of these properties because the tier follows the customer, not the node: work of a given tier is held to the same standard wherever it is delivered.

What node-level data usually shows

The location-differentiated target is usually defended as a reflection of measured differences between nodes. In practice the measured differences are often smaller than the target differences, or reverse by work type — a lower-cost node matching the established one on the same accounts and channel, or a supplier outperforming a captive center on one work type and underperforming on another.[2] Where that is so, the location target is not tracking a performance gap; it is creating the appearance of one. The comparability conditions — same instrument, case-mix adjustment, sufficient observations — are at Sample Size and Detectable Difference in Quality Measurement and Comparing Delivery Arrangements.

The target architecture that prevents it

  1. Standardize the instrument across every node. One measurement method, one definition, one coverage rule; a blend of scores from different instruments is not a blend at all.
  2. Vary the threshold by service tier only, never by location. What the customer bought determines the standard; where the work is done determines nothing about it. The instrument-versus-threshold distinction is developed at Sourcing Design Axes: Node and Client Ownership; this page supplies the arithmetic cost of getting it wrong.
  3. Case-mix adjust before comparing nodes. Nodes holding harder work — disruption, high-stakes, complex — are compared on adjusted scores, so that the standard target does not penalize whoever takes the difficult work.
  4. Report the blend beside its composition. The general rule — that a blended number is read next to the mix that produced it — is at Simpson's Paradox in Contact Center Metrics. What this page adds is the stronger form: where the mix is a deliberate output of sourcing strategy, publish a composition-standardized headline that holds the mix fixed at a reference period, so the strategy cannot move the number.
  5. Set the headline target on the standardized figure. The organization's number should not be one that management's own sourcing strategy is designed to lower. The practice for deriving that figure from a measured distribution is at Performance Management.

Failure modes

  • Reading the blend as performance. The headline falls; a servicing problem is declared; remediation is aimed at nodes that are each on target.
  • Fixing the headline by raising the lower node's target. Without changing the instrument or the work, the lower node is now failing a target it was never measured against (see Baselines Before Targets).
  • Blending across instruments. Nodes measured by survey, by analytics platform and by end-of-chat prompt are averaged into one number.
  • Defending the location target with a case-mix argument. If the node genuinely holds harder work, the remedy is case-mix adjustment on a common target, not a lower target — the two produce very different incentives.[2]

Maturity Model Position

A location-differentiated target architecture is a Level 2 condition on the WFM Labs Maturity Model™ — targets exist and are tracked, but they are indexed on the organization rather than on the work or the customer. Correcting it is a precondition for the Level 3 discipline of reading blends beside their composition, and for the Level 4 practice of expressing targets as bands and hit-rates, which cannot be compared across nodes that are held to different points.

See Also

References

  1. ↑ Kitagawa, E. M. (1955). Components of a difference between two rates. Journal of the American Statistical Association, 50(272), 1168–1194.
  2. ↑ 2.0 2.1 2.2 Practitioner observation from multi-node service estates; a consistent pattern rather than a measured result.
Retrieved from "https://wiki.wfmlabs.org/w/index.php?title=Mix_Effects_in_Blended_Quality_Targets&oldid=4353"