Under-the-Hood Disaggregation
Slice aggregate indicators into groups that reveal the real economic mechanism.
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 6
- Confidence
- 93%
Under-the-Hood Disaggregation replaces a single aggregate statistic with a mechanism-based view of the groups beneath it. First define the apparent puzzle, such as slow wage growth despite low unemployment or changes in the national savings rate. Identify the channels that should influence the outcome, then segment the data by meaningful exposure to those channels—for example income group, industry, geography, asset ownership, or dependence on labor income. Examine each segment separately and determine which group contributes most to the total. The analyst then reconnects those findings to the aggregate result and validates the explanation with another measure. This avoids treating households, workers, or companies as interchangeable and exposes cases where one group's strong gains are offset by another group's stagnation.
Origin
Extracted from Masters in Business during Ellen Zentner's explanation of Morgan Stanley's post-financial-crisis economic analysis.
Core principles
- 01Aggregates can conceal opposing experiences.
- 02Choose segments that correspond to distinct causal mechanisms.
- 03Connect each group to the income, wealth, or cost channel affecting it.
- 04Recombine segment findings before making a system-level conclusion.
How to run it
- 1
Define the aggregate puzzle
Specify the headline statistic and the behavior that appears inconsistent with it. State what a conventional aggregate interpretation would predict.
Pro tip Frame the puzzle as a concrete question, such as who is spending or where wages are accelerating.
Watch out Do not segment data without first defining the decision or puzzle the analysis must resolve.
- 2
Identify causal channels
List the mechanisms that could produce different outcomes across groups, such as labor income, financial wealth, housing equity, debt costs, or regional living costs.
Pro tip Use domain knowledge to choose mechanisms before inspecting every possible subgroup.
Watch out Arbitrary segmentation can produce correlations with no explanatory value.
- 3
Select meaningful segments
Divide the population according to exposure to the causal channels. Examples include income quintiles, industries, states, age cohorts, or asset ownership.
Pro tip Prefer a small set of interpretable groups over dozens of thin slices.
Watch out Check sample sizes and avoid groups too small to support reliable conclusions.
- 4
Measure each group's trajectory
Compare levels, changes, and turning points for every segment. Look for offsetting movements that disappear in the aggregate.
Pro tip Use the same time window and definitions across groups.
Watch out Do not compare nominal and real measures or incompatible datasets without adjustment.
- 5
Find the aggregate driver
Estimate each segment's weight and contribution to the headline result. Distinguish the largest group from the group with the largest marginal influence.
Pro tip Contribution analysis is more informative than comparing growth rates alone.
Watch out A fast-growing small segment may barely move the aggregate.
- 6
Validate the mechanism
Test the explanation with an independent measure or related behavior. Use the result to refine the system-level conclusion.
Pro tip Pair confidence data with actual spending, wages with job creation, or asset values with savings behavior.
Watch out Do not declare causation solely because two segmented series move together.
In the wild
An analyst sees only modest national wage growth despite falling unemployment. The workforce is segmented by industry, revealing faster wage pressure in service sectors where job creation is concentrated. The aggregate remains subdued because the higher increases occur among relatively low-paid workers whose total wage bill carries less weight.
→ The apparent contradiction is resolved without assuming the labor data are wrong.
Instead of treating the national savings rate as representative of every household, an analyst segments households by income and asset ownership. The analysis shows that wealthy households perform most saving and adjust spending when financial assets rise, while lower-income groups remain closer to paycheck-to-paycheck behavior.
→ Forecasts focus on the group that actually moves the aggregate savings rate.
Common mistakes
Choosing cosmetic segments
Demographic categories are useful only when they correspond to a plausible mechanism. Segmenting without a causal reason creates descriptive trivia rather than insight.
Ignoring contribution weights
A segment can exhibit a dramatic trend while having little effect on the total. Weight each group before attributing the aggregate movement to it.
Never recombining the findings
Segment analysis must ultimately explain the system-level result. Isolated subgroup charts do not answer the original question by themselves.
Is it for you?
Best for
Analysts studying heterogeneous populations, customers, workers, or market participants.
Not ideal for
Very small datasets where segmentation would create unstable or personally identifiable groups.
From the transcript
“you just can't look at anything in the aggregate anymore, and wages are a great example of that.”
“What we do Bury and I find that a lot of our work on the U S Economics team at Morgan Stanley is now we…”
“We we scrub three hundred different categories of consumer spending, and that's how we know who's spending, who's doing the spending.”
From the episode
Ellen Zentner's Shift From Public to Private Sector
Ellen Zentner