Data Analysis (aaag-data-analysis)
The Annals expects analyses that are spatially honest and reported with uncertainty, whatever the
area. The standard is that a competent reader in the area could follow the logic from data to claim and
see that the geography of the data was respected, not flattened.
When to trigger
- Estimating models, running spatial statistics, classifying imagery, or coding qualitative material
- A reviewer questioned uncertainty, robustness, spatial autocorrelation, accuracy, or interpretation
- Preparing the results section and deciding what to report
Spatial / quantitative
-
Diagnose space first. Report spatial autocorrelation in residuals; if present, move to a spatial
model (lag/error, GWR/MGWR, spatial regimes) and say why.
-
Uncertainty everywhere. CIs/SEs (spatially robust where needed), not stars alone; for prediction,
out-of-sample error from spatial/blocked CV.
-
Robustness. Re-estimate across plausible areal units and bandwidths (MAUP/scale sensitivity);
show the result is not a unit artifact. Report effect sizes in interpretable units.
Remote sensing / physical
-
Accuracy with an independent sample. Confusion matrix, overall/producer/user accuracy, kappa or
F1; for continuous outputs, RMSE/MAE and bias; map the spatial pattern of error, not just a scalar.
-
Propagate uncertainty from inputs through to the reported quantity; state the validation design.
Qualitative / interpretive
-
Transparent analytic trail. Coding scheme, how themes were derived, and how interpretations were
checked (negative cases, member checks, triangulation) — credibility over counting.
-
Evidence-to-claim mapping. Each interpretive claim is tied to identifiable (anonymized) evidence;
avoid quote-mining that over-generalizes from one informant.
Mixed methods
-
Show the integration. State where the strands converge and where they conflict, and how the
conflict was adjudicated — do not report two parallel analyses and call it mixed methods.
Cross-cutting reporting bar
- Match every claim in the text to an exhibit or statistic; no orphan assertions.
- Report negative / null / scale-dependent results honestly; geography rewards scope conditions.
- Keep analysis reproducible: master script, seeds, pinned versions (see
aaag-transparency-and-data).
Referee pushback → Annals-specific fix
-
"Are these effects just spatial autocorrelation?" → Show residual Moran's I before/after a spatial
model; report the spatial-error structure, not only a global coefficient.
-
"Would the result change at a different scale/unit?" → Provide a MAUP/bandwidth sensitivity panel
and state the scale at which the claim holds.
-
"How accurate is the map?" → Area-adjusted accuracy from an independent sample + a map of where
error concentrates, not a single kappa.
-
"How do I know the qualitative reading isn't cherry-picked?" → Coding scheme, negative cases, and an
excerpt-to-claim table.
Calibration anchors
-
Uncertainty is mandatory, not optional. A coefficient or accuracy number without an interval is
not yet a finding at this venue.
-
Scale dependence is a result, not a nuisance. If the answer changes with the unit, say so — that
is geographic knowledge.
-
The spatial pattern of error is itself a finding for remote-sensing and prediction work.
Checklist
Anti-patterns
- Reporting OLS on spatial data with no autocorrelation check
- Stars-only tables with no effect sizes or CIs
- A single global accuracy number with no spatial error map
- Cherry-picked quotes standing in for an analytic trail
- "Mixed methods" that never integrate the strands
Output format
【Mode】spatial-quant / remote-sensing / qualitative / mixed
【Headline result】effect/accuracy/theme + its uncertainty
【Spatial honesty】autocorrelation / MAUP / spatial-CV / spatial error map handled? [Y/N]
【Robustness】checks run and what held
【Reproducibility】master script + seeds + versions? [Y/N]
【Next】aaag-tables-figures
Supplementary resources
