Start with what the overall number measures
Suppose two customer support teams have different rates of resolving requests on the first reply. It is tempting to rank them by their overall percentages. But one team may receive mostly straightforward questions, while the other handles more complicated problems. The overall rate describes what happened across each team’s actual workload; it does not, by itself, tell you which team would perform better with the same mix of requests.
This distinction matters whenever groups contain different proportions of cases that affect the outcome. Schools may serve students with different starting needs, shops may sell different product mixes, and survey respondents may differ in age or experience. A group average combines two things: results within each type of case and the share of each type in the group. Before drawing a conclusion, ask which of those things your comparison is meant to assess.
Look inside the groups
A useful first step is to break each group into meaningful subgroups, often called strata. For the support teams, you might separate routine requests from complex ones. Then compare the teams within each request type, while also showing how many requests of each type they handled. Counts matter: a percentage based on a handful of cases is less stable than one based on many cases, even if both are displayed with the same precision.
Consider a simplified example. Team A resolves 90 of 100 routine requests and 20 of 100 complex requests on the first reply, for 110 of 200 overall. Team B resolves 24 of 30 routine requests and 63 of 70 complex requests, for 87 of 100 overall. Team B has the higher overall rate, 87% versus 55%. Yet Team A does better on routine requests, 90% versus 80%, while Team B does much better on complex ones, 90% versus 20%. The overall ranking alone conceals both differences.
Decide whether to adjust the comparison
Sometimes the question is about the actual experience of all cases each group handled. In that case, the overall rate is relevant and should remain visible. Other times, the question is how the groups compare under a common case mix. You can estimate that by applying the same weights to each group’s subgroup rates. If routine and complex requests each receive half the weight in the example, Team A’s adjusted rate is 55% and Team B’s is 85%. Those figures answer a hypothetical question about an equal mix, not a question about either team’s real workload.
There is no universally correct set of weights. An equal mix is easy to understand, but a combined workload or another clearly defined reference group may be more useful for a particular decision. State the reference mix you chose and why. If changing reasonable weights changes the apparent winner, that sensitivity is worth reporting rather than hiding behind a single adjusted figure.
Choose subgroups with care
A comparison is only as useful as its categories. Subgroups should relate to the question and be defined consistently across groups. Labels such as “complex” can be subjective, so check how cases were classified. A category may also hide differences of its own: two teams’ complex requests might involve different issues. Breaking the data into ever smaller slices is not automatically better, however. Small slices can make rates noisy or leave no comparable cases at all.
Also consider when a characteristic was measured. A case’s difficulty at arrival might help describe the workload. A label assigned after the team has worked on the case could partly reflect the team’s actions, making it a questionable basis for adjustment. Comparing like with like requires thought about how the cases entered each group, not just a spreadsheet operation.
Report both the pattern and its limits
A clear summary can present the overall rates, the subgroup rates and counts, and any adjusted comparison side by side. Explain in ordinary language what each view answers. If one group has a higher overall rate mainly because it handles a different mix, say so. If results vary across subgroups, describe that variation instead of forcing a single verdict.
Adjustment cannot account for differences you have not measured, and an observed gap does not establish its cause. The practical goal is a comparison that is honest about its denominator, its case mix and its assumptions. When those are visible, readers can tell whether an apparently simple ranking is answering the question they actually care about.
