Survey Trend Analysis Across Waves

Last updated October 30, 2026·4 min read

A repeated cross-sectional survey measures the same population at successive times with fresh samples each wave: a state health survey every year, a tracking poll every quarter. The trend question sounds simple, has it changed, and is surrounded by ways to answer it badly. Comparing only the first and last waves discards the path. Eyeballing a chart of point estimates ignores that each point carries its own error. Pooling waves into one regression without the design understates every standard error in the building.

The estimates come first, one per wave

Each wave is estimated on its own design, with its own weights, and reported with its interval. A three-wave series on a behavior prevalence:

• Wave one: 200 respondents, 60 events, 30.0 percent, interval 24.1 to 36.7
• Wave two: 210 respondents, 71 events, 33.8 percent, interval 27.8 to 40.4
• Wave three: 205 respondents, 80 events, 39.0 percent, interval 32.6 to 45.8

The point estimates climb steadily. The intervals overlap generously, which at these sample sizes is expected and settles nothing by itself. Overlap of intervals is not a test; two intervals can overlap while their difference is significant, and the trend test below exists precisely because the eyeball comparison of intervals is the wrong instrument.

The trend test: contrasts on the actual times

The module regresses the outcome on orthogonal polynomial contrasts of the actual time values. Linear contrasts ask whether the series trends steadily; quadratic contrasts ask whether it bends. Using the actual times rather than wave numbers one, two, three matters whenever spacing is uneven: waves at years one, two, and five are not equally spaced evidence, and the contrast coefficients should know it. The estimation is design-based throughout, with strata redefined as wave crossed with stratum, because each wave is an independent draw and the variance machinery must treat them that way. This follows the approach the CDC documents for trend analysis in its youth risk behavior surveillance, the reference the module was verified against.

On the worked series, the nine-point climb across three waves reads as a genuine upward trend in direction, with strength that the per-wave intervals already warned would be modest at n near two hundred per wave. The deliverable of the test is a slope with an interval and a stated model, not a verdict stamped on the chart.

Two-wave change: the honest arithmetic

Where only two waves are compared, the change is a difference of two independent estimates, and its standard error combines the two in quadrature: the square root of the sum of the squared standard errors. Independence is what licenses the simple formula, and repeated cross-sections supply it by construction. The same formula misapplied to overlapping panel samples, where the same respondents appear in both waves and the estimates covary, understates the error; panels are a different design and a different analysis.

Pooling for small subgroups

A subgroup too small to estimate in one wave can sometimes be estimated across pooled waves: weights divided by the number of waves pooled, the estimate read as an average over the period rather than a value in any year. The label matters as much as the arithmetic. A pooled estimate reported as the current year figure is a quiet misstatement, and the module output names the pooled span so the report can too.

What breaks trends that are not real

• Instrument changes. A reworded question manufactures a jump that no test can separate from real change. Trend programs live or die on question stability, and a methods note belongs beside any series that spans a redesign.
• Mode changes. A move from telephone to web shifts responses on sensitive items. The shift lands in the trend unless a bridge study measures it.
• Weighting changes. A new calibration source moves estimates. Consistent reweighting of back waves, where the program supports it, is the clean cure.
• Composition drift presented as behavior change. Population change is real change in the population estimate, and may or may not be the story the report wants to tell. Age-standardized companions to headline trends settle the argument where it matters.

Run it

Module 12 of Analysis Studio stacks the waves, estimates each on its design, and runs the trend and change tests with the chart and intervals on one screen, and it is a free module:

Module 12: Trend AnalysisOpen in Analysis Studio →

Sources and method notes

Method follows the CDC trend-analysis approach for repeated cross-sectional surveys, as recorded in the module verification file. Worked figures are the banked three-wave example for this module, re-verified at build.