A housing policy team needed to answer a deceptively simple question: where is housing least affordable, and for whom? Existing reports used different definitions, different geographies and different income measures, so two teams could reach opposite conclusions from the same national survey and both could defend them. The problem was not a lack of data. It was a lack of one agreed measure.
The decision
The team had to decide which places and household types to prioritise, and how to brief a programme whose eligibility rules were already under review. That required a measure that was comparable across regions, stable across survey waves, and honest about its own precision. It also required an answer to a question a city-level ranking cannot give: within a metro area that looks affordable on average, who is actually cost-burdened?
What the survey provided
Multiple waves of a national household survey, roughly one hundred variables per record:
- Household economics: household income, income relative to area median, and assisted-housing status.
- Housing cost: rent, utilities, other housing costs and several cost-burden indicators.
- Market context: fair market rent, area median income and income limits at 30, 50 and 80 percent of median.
- Dwelling and household: tenure, bedrooms, building age, structure type and household size.
- Geography and design: metro status, region and survey weights.
The file included imputed values, negative and implausible incomes, and a complex survey design. Ignoring the weights would over-represent some household types and distort every regional comparison.
Why survey data punishes shortcuts
- Row counts are not people. Without the survey weights, a large household in a sparse region carries the same analytical weight as a single-person household in a dense one. The reported shares change, and they change unevenly by region.
- Zero and negative incomes break ratios. Cost burden is a ratio, and a zero denominator is not a small inconvenience. Those households need a documented treatment, not silent deletion.
- Imputation is a finding. The share of imputed values varies by field and by geography. Hiding it makes every downstream estimate look more precise than it is.
- Definitions move. Metro boundaries and income limits change between waves. Smoothing over a definitional break turns an administrative change into a false trend.
How the measure was built
- Agree the definition before plotting. Cost burden was defined as housing cost above 30 percent of income, with a severe-burden threshold at 50 percent. Renters and owners with mortgages were analysed separately, because combining them hides two different cost structures.
- Respect the survey design. All estimates used the provided weights. Household counts, not row counts, drove the reported shares and the confidence intervals.
- Clean with a written rule set. Negative income, zero-income households, missing cost fields and imputed values were handled explicitly and flagged, so the team could see how much the exclusions mattered.
- Build comparable geographies. Metro areas were grouped consistently across waves. Where definitions changed, the break was documented and the trend was split rather than smoothed.
- Report by household type and income band. Affordability was reported for renters, owners, seniors, families with children and assisted households. A metro average conceals very different experiences.
- Quantify the gap, not just the ratio. Housing cost was compared with the income limits used by housing programmes, producing an affordability gap that a programme officer could use.
- Put the method beside the chart. The dashboard opens on the national trend, drills down to region, metro, tenure and household type, and carries a methodology panel with every view.
What this changes
- One definition replaced several. Teams that had been comparing incompatible figures could now argue about policy rather than about whose number was right.
- The unit of analysis moved inside the city. The most useful output was not a ranking of metros but a map of who was burdened within each metro, which changed how eligibility was discussed.
- Precision became explicit. Confidence intervals and imputation flags made it clear where the survey could support a conclusion and where it could only suggest one.
- Trend breaks became visible. Definitional changes were annotated instead of being absorbed into the line chart as apparent rises or falls.
What we delivered
- A weighted affordability index with 30 percent and 50 percent burden thresholds.
- Metro and regional profiles by tenure and household type.
- A multi-year trend view with documented definitional breaks.
- A dashboard with drill-down and a methodology panel.
- A limitations note covering self-reporting, imputation and geographic aggregation.
Where this applies
This fits policy teams, housing agencies, municipalities and researchers working with survey microdata. The minimum inputs are the survey weights, a stable geography identifier, income and housing-cost fields, and the programme's income limits. The first deliverable is usually the definition and weighting review — because with survey data, most disagreements about results are really disagreements about design.
This is an illustrative case study using anonymized data with no household-level information; it does not report results from a specific client engagement. It reflects our standard method for survey-weighted public-sector analysis.
