Distinguishing Overall Patterns from Subgroup Patterns

Why an Overall Summary May Not Tell the Whole Story

An overall average or rate is a useful starting point, but it combines observations from potentially different groups. If those groups have different outcome levels, the overall summary depends not only on the outcomes within each group but also on how many observations come from each group. As a result, two populations can show one pattern overall and a different pattern within meaningful subgroups.

This matters in predictive analytics because an aggregate can conceal important variation in a portfolio, customer population, or operational process. A responsible exploratory summary therefore asks two questions: What is the overall pattern? and Does that pattern also appear within relevant subgroups? This tutorial is aligned with the SOA Exam PA (Predictive Analytics) syllabus.

A subgroup should have a substantive reason for being examined, such as a known difference in exposure, product design, or operating conditions. Dividing data into many arbitrary groups can produce unstable or distracting comparisons. The aim is not to report every possible slice; it is to check whether a meaningful grouping helps explain the aggregate.

How Subgroup Summaries Combine

For a numeric outcome, the overall mean is a weighted average of subgroup means. The weights are the subgroup observation counts divided by the total count. For a rate, the same principle applies when subgroup rates have denominators such as policies, exposures, or cases: weight each rate by its denominator, not by an unrelated count.

Here, \(G\) is the number of subgroups, \(n_g\) is the number of observations in subgroup \(g\), and \(\bar{x}_g\) is that subgroup's mean. The weights are nonnegative and sum to one. This formula assumes the groups partition the observations being summarized: each included observation belongs to exactly one subgroup, and the outcome and inclusion rules are consistent across groups.

For subgroup event counts \(c_g\) and corresponding denominators \(n_g\), the overall event rate is

A rate is a ratio of totals, so its natural weights are the subgroup denominators. An unweighted average of subgroup rates answers a different question: what is the average of the subgroup rates if each subgroup is given equal weight, regardless of its size? That may be useful in a specific analysis, but it is not generally the rate across all observations.

Why the Weighted-Average Formula Holds

The subgroup mean is \(\bar{x}_g=S_g/n_g\), where \(S_g\) is the sum of the outcomes in subgroup \(g\). Since the subgroups partition the observations, their outcome sums add to the overall sum, and their counts add to the overall count. Therefore,

Dividing each subgroup count by the total count gives the weight \(w_g\), producing the weighted-average expression. The same argument applies to a rate: add the subgroup event counts for the numerator and the subgroup denominators for the denominator. This derivation also shows why simply averaging subgroup means or rates is generally incorrect: it discards the sizes of the groups.

When the Overall Comparison Reverses a Subgroup Pattern

Suppose two groups have event rates reported separately for simple and complex cases. The overall rate for each group depends on its mix of case types. If one group contains many more complex cases, its overall rate may be lower even when its rate is higher within both case types. This is a form of Simpson's paradox: an association seen in separate subgroups can reverse when the groups are combined, because the subgroup proportions differ.

The reversal is an arithmetic fact about the observed data, not evidence that one summary is automatically the “right” one. The overall rate describes the actual combined populations. The subgroup rates describe outcomes within the named groups. Which summary is most relevant depends on the question being asked, so both the rates and their denominators should be visible.

Using a Common Composition to Compare Groups

When two groups have different subgroup proportions, one descriptive comparison is to calculate a standardized summary: apply the same set of subgroup weights to both groups' subgroup summaries. This answers a specific question: how would the weighted summaries compare if both groups were represented using the chosen common composition? It does not replace the observed overall summaries, and the choice of common weights should be stated.

For a mean or rate, let \(m_{Ag}\) and \(m_{Bg}\) be the summaries for groups A and B within subgroup \(g\), and let \(w_g^*\) be common weights that sum to one. The standardized summaries are

Common weights might be the proportions in a chosen reference population, or equal weights if the goal is to give each subgroup equal representation. These choices can lead to different standardized results; a comparison must identify which weights it uses.

A Practical Review of Aggregate and Subgroup Results

A useful exploratory summary preserves the link between each statistic and the observations that produced it. For an overall rate, report the event count and denominator, not only the percentage. For a subgroup comparison, report the subgroup counts as well as the subgroup values. Small groups can have unstable rates, and a large-looking percentage difference based on very few observations deserves careful qualification.

If records are missing the subgroup variable, the observed subgroups may not partition the full population. In that case, report how many observations are unclassified and whether they are included in the overall summary. Do not silently combine a full-population total with subgroup summaries that cover only part of it.

A subgroup comparison is descriptive: it identifies how summaries differ across the recorded groups. By itself, it does not establish why the differences occurred. For this topic, the central task is to accurately describe the aggregate, the subgroups, and how their composition affects the summaries.

Common Mistakes

• Taking an unweighted mean of subgroup means or rates. This gives each subgroup equal influence, even if the subgroup sizes differ. Use the appropriate observation counts or rate denominators for the actual overall summary.
• Reporting only the aggregate. A single summary can conceal subgroup variation or even an opposing within-group pattern. Inspect meaningful subgroups before describing the result.
• Reporting subgroup rates without denominators. The same percentage may be based on many observations or only a few. Include the counts so readers can assess the scale of the evidence.
• Confusing observed and standardized summaries. Observed summaries use each group's actual composition. Standardized summaries use a stated common set of weights. Label them distinctly.
• Over-interpreting a reversal. A reversal describes how the aggregate and subgroup comparisons differ. It does not, on its own, explain the reason for that pattern.