Hundreds of recent papers report the estimated annual percentage change (EAPC), a one-number summary of how fast a disease rate is rising or falling. It comes from fitting a straight line to the logarithm of an age-standardized rate, meaning a rate adjusted so that populations with different age structures can be compared, over twenty to thirty years. These papers then label each trend as "increasing," "decreasing," or "stable" depending on whether the 95% confidence interval excludes zero. Getting there involves choices the confidence interval never sees: which release of the Global Burden of Disease (GBD) estimates to download, which start and end years to use, which standard population to age-standardize against, ordinary least squares on log rates or a weighted or count-based fit, and how many breakpoints the trend-segmentation software may place. Machine learning and clinical informatics already know this pattern. Preprocessing choices routinely move results more than modeling choices do. Two neuroimaging pipelines that agree well at each single timepoint have been shown to disagree several-fold on the annual rate of change.
We propose to measure that gap for epidemiological trends. The claim is simple and falsifiable. For the smoothed, model-derived annual series that most EAPC papers use, the spread of EAPC values across reasonable analyst choices will often be larger than the reported confidence interval. When that holds, the increasing/stable/decreasing label is partly an artifact of the pipeline. We will build an open tool that runs the full set of choices, reports a choice-aware interval and a label-stability score, and check whether that interval predicts what a later data release will show.
Core Claim and Why It Is Not Already Done
The closest existing work is Longitudinal assessment of cortical thickness in healthy older individuals: a comparison between CAT12 and freesurfer. It followed 184 older adults across a scanner change from Siemens to GE. Its two software pipelines correlated at r ≥ 0.60 cross-sectionally, yet produced annualized percent changes of about -2% to -3% (CAT12) versus about -0.4% (FreeSurfer), with only weak-to-moderate agreement on change. That paper motivates our proposal but differs in three ways. It compares two fixed tools over two timepoints, it works with individual-level imaging, and it offers no way to fold pipeline disagreement into the reported uncertainty. We work with multi-decade, population-level, log-linear trend estimates. We treat the set of analyst choices as a quantity to be measured, and we validate the result against later data releases.
The clinical informatics audits play a different role: they show that definitional choices dominate in related settings. Subtle variation in sepsis-III definitions markedly influences predictive performance within and across methods found that onset definitions moved cohort size from 867 to 2,178 and AUROC by 0–6%, more than switching model families did. Reproducibility in critical care: a mortality prediction case study found that half of reproduced MIMIC cohorts differed in size by more than 25% from what was reported. Yet Another ICU Benchmark standardized ICU extraction for exactly this reason. All three address short-horizon prediction, not long-run rates of change. The GBD trend papers supply the application: tuberculosis in China (Long-Term Trends and Determinants of Tuberculosis Burden in China, 1990–2023), psoriasis (Global Psoriasis Burden 1990–2021), and autism (Global, Regional, and National Burden of Autism Spectrum Disorder). Each reports a single specification with a confidence interval computed only from the residuals of the fitted line.
There is a specific reason to expect those intervals to be too narrow. GBD annual values are outputs of smoothing models, not independent yearly observations. Residuals around a log-linear fit are therefore small and strongly correlated from one year to the next. That shrinks the standard error of the slope $\beta$, and with it the interval around $\text{EAPC} = 100(e^\beta - 1)$. Our central quantity is the ratio
$$R = \frac{\mathrm{SD}_{s \in S}\left(\text{EAPC}_s\right)}{\mathrm{SE}_{\text{OLS}}\left(\text{EAPC}_{s_0}\right)},$$
where $S$ is the set of defensible specifications, $\text{EAPC}_s$ is the estimate under specification $s$, $s_0$ is the conventional default (latest GBD release, full 1990-to-latest window, OLS on log rates), and the denominator is that default's reported standard error. When $R > 1$, the choice of specification contributes more uncertainty than the published interval admits.
Research Program
We first fix the space of specifications $S$, using only choices that published EAPC/AAPC papers in the file actually vary or could defensibly vary. The first choice is data vintage: GBD 2019, 2021, and 2023 releases restricted to their overlapping years. Each release re-estimates the entire back series, which makes the comparison a natural two-pipeline experiment like CAT12 versus FreeSurfer. The remaining choices are the start and end years within ±5 years of the conventional window, the age-standardization reference population, and the estimator: OLS on log rates, weighted least squares, or a Poisson-type count model. The last is the trend-segmentation setting. Joinpoint regression, which finds the years where the trend changes slope and summarizes the whole period as the average annual percentage change (AAPC), has its own options: maximum number of breakpoints, minimum segment length, and permutation versus information-criterion model selection. The set is fixed before any results are computed.
Second, we report two outputs for each disease-location-measure series. One is a choice-aware interval that combines the default sampling variance with the between-specification variance. The other is a label-stability score: the share of specifications that give the same increasing/stable/decreasing label as the default.
Third, we test whether the cross-sectional-versus-trend divergence seen in neuroimaging also appears in epidemiology. For pancreatic cancer in the countries studied by Global temporal patterns of pancreatic cancer and association with socioeconomic development, we compare registry-based rates from CI5 and GLOBOCAN with GBD modeled rates. We ask whether the two sources agree on rate levels across countries but disagree on their 1998–2007 AAPCs. That paper reported, for example, an AAPC of 10.4 for Brazilian men and 4.7 for French men.
Fourth, a simulation study provides ground truth. We generate log-rate series with a known slope. We then apply GBD-like smoothing, and in some series insert artificial level shifts mimicking a change in disease coding (such as ICD-9 to ICD-10) or an expansion of screening. We measure how often each interval method covers the true slope.
Baselines
The naive reference point is the standard practice in every cited GBD trend paper: OLS EAPC on log age-standardized rates with its residual-based 95% interval and the zero-exclusion label. It is run on the default specification for every series. Beating it means better coverage in simulation and better prediction across releases, not a wider interval for its own sake.
The strongest existing method is Joinpoint AAPC with its segment-based interval, as used by Wong et al. and in the segmented tuberculosis analysis, which reported a +1.10% per year incidence rebound for 2021–2023. Joinpoint is the strongest competitor because it already accounts for non-linear trends, the most common objection to a single EAPC. We run it with its standard default settings on the same series and the same years. I found no reported result on Lacuna for the coverage or cross-release stability of Joinpoint AAPC intervals, or of EAPC intervals, so the baseline numbers will be produced here. Beating it means our interval reaches close to nominal coverage in simulation while Joinpoint's does not, and contains later-release AAPCs at a higher rate with no more than a modest increase in width.
The third baseline uses autocorrelation-robust (Newey–West) standard errors on the default OLS fit. This is the cheapest statistical repair for correlated residuals.
The fourth propagates GBD's own published uncertainty intervals by resampling each year's rate within its interval. The resampling is done under two assumptions about how errors relate across years: fully independent and fully correlated.
The key ablation isolates the data-vintage dimension. We recompute the choice-aware interval with vintage removed, leaving only window, standardization, and estimator choices. This tells us whether the effect comes from upstream re-estimation by the data producer or from downstream analyst choices. That distinction decides whether the remedy belongs with data producers or with paper authors. All methods are applied to the same series and the same year ranges. None involves tuning beyond the pre-registered set of specifications.
Why not just…?
Why not just use autocorrelation-robust standard errors? They correct the shrinkage caused by correlated residuals, which is part of what we claim. If that is the whole story, robust intervals will match ours in coverage, and the honest conclusion is "use robust standard errors." They cannot, however, see a level shift common to the whole series, or a revised back series in a new release. The CAT12/FreeSurfer result is a systematic pipeline difference, not residual noise, which suggests the same limitation applies. The simulations with inserted coding shifts and the cross-release test will settle this. This alternative is a real threat, and we report it as such.
Why not just propagate GBD's own uncertainty intervals? This is the most serious alternative, because those intervals are wide and come from the data producer. Two problems remain. The year-to-year correlation of GBD's errors is not published, so the result swings between the independent and fully correlated extremes. And GBD's intervals describe uncertainty within one release, not the revision between releases. The prediction test across releases checks directly whether GBD-interval propagation already covers later releases. If it does, our contribution narrows to the analyst-choice dimension, and the ablation will show how much that dimension adds.
Why not just report Joinpoint AAPC and move on? Joinpoint handles bends in the trend but adds its own free settings. The breakpoint it finds can itself be an artifact of a coding change rather than a real epidemiological shift. We include Joinpoint settings in the specification set and test, in simulation, how often inserted artifact shifts are reported as real breakpoints.
Proposed Experiments and Decisive Checks
The first decisive check is simulation coverage. With GBD-like smoothing and one coding shift, we predict the nominal 95% OLS interval will cover the true slope far less often than 95%, while the choice-aware interval will come close to nominal. If robust or GBD-interval-propagated methods match it, we say so.
The second is the distribution of $R$ and of label stability across the disease series in the cited GBD papers, which together cover tuberculosis, chronic kidney disease, psoriasis, autism, caries, skin diseases, and women's cancers. The hypothesis is falsified if $R < 1$ and labels stay stable for most series. That would be a useful result in its own right, licensing current practice.
The third is the prospective test across releases. We build intervals using GBD 2019 alone and check how often they contain the EAPC computed from GBD 2023 over the same years. This is the closest available stand-in for ground truth on real data.
The fourth is the registry-versus-model comparison for pancreatic cancer. Agreement on rate levels together with disagreement on AAPCs would reproduce the neuroimaging pattern in population surveillance.
References
- Longitudinal assessment of cortical thickness in healthy older individuals: a comparison between CAT12 and freesurfer
- Subtle variation in sepsis-III definitions markedly influences predictive performance within and across methods
- Reproducibility in critical care: a mortality prediction case study
- Yet Another ICU Benchmark: A Flexible Multi-Center Framework for Clinical ML
- Long-Term Trends and Determinants of Tuberculosis Burden in China, 1990–2023: Insights from the Global Burden of Disease Study 2023
- Global Psoriasis Burden 1990–2021: Evolving Patterns and Socio-Demographic Correlates in the Global Burden of Disease 2021 Update
- Global, Regional, and National Burden of Autism Spectrum Disorder: Trends and Decomposition Analysis From 1990 to 2021, and Projections for 2045
- Global temporal patterns of pancreatic cancer and association with socioeconomic development