aaag-research-design

Research Design (aaag-research-design)

The Annals spans four areas and accepts many methodologies, but is demanding about each. The design must credibly connect the geographic argument (aaag-theory-building) to the evidence, and must take space and scale seriously — spatial dependence, the MAUP, projection, and sampling are design issues, not afterthoughts. This skill is mode-aware: pick the section that matches your work.

When to trigger

• Specifying identification, sampling, case selection, or measurement
• A reviewer questioned spatial autocorrelation, scale/MAUP, edge effects, validation, or a confound
• Justifying why the design adjudicates the rival account from aaag-literature-positioning

Spatial / quantitative analysis & GIScience

• Take space seriously. Test and model spatial dependence (Moran's I, spatial lag/error, GWR/MGWR where heterogeneity is the point); state how the MAUP / scale could change conclusions.
• Geography of the data. Document projection/CRS, areal units, edge effects, and the support of measurements; spatial sampling and its biases.
• Inference. Cluster or use spatial SEs at the right level; for spatial autocorrelation, report diagnostics; for prediction, use spatially-aware cross-validation (blocked/spatial CV), not random folds.

Remote sensing / physical-environmental

• Measurement validity. Sensor/resolution choices, atmospheric/geometric correction, and ground truth; quantify accuracy (confusion matrix, kappa/F1, RMSE) with an independent validation sample.
• Process linkage. Tie observed pattern to an earth-surface process and its scale; state the uncertainty budget end to end.

Qualitative / human-geography

• Case selection by design logic (typical, extreme, paired, regional contrast) — say what the case is a case of. Convenience is not a rationale.
• Positionality, reflexivity, and rigor appropriate to the method (ethnography, interviews, archives, discourse/textual analysis); state how interpretations were checked.
• Source/field transparency: plan how fieldnotes, interviews, and archives are documented and cited (see aaag-transparency-and-data), including consent and geoprivacy.

Nature-society / mixed methods

• Integrate, don't staple. Specify how the biophysical and social strands inform one another (e.g., land-change observation + livelihood interviews), and how convergence/divergence is handled.

The adjudication test (Annals-specific)

For the single strongest rival explanation, write: "If the rival held rather than my argument, the [spatial pattern / measurements / accounts] would look like ___; instead they look like ___." If the design cannot distinguish them — including ruling out a scale or spatial-autocorrelation artifact — it does not yet identify the contribution.

Referee pushback → Annals-specific fix

Likely objection	Area	The fix
"Your OLS ignores spatial autocorrelation."	Methods/Human	Test residual Moran's I; move to a spatial model and report diagnostics.
"This is a unit-of-analysis artifact (MAUP)."	Methods/Nature-Society	Re-run across areal units/bandwidths; show stability or scope the claim by scale.
"Random CV overstates accuracy on spatial data."	Methods/RS	Use blocked/spatial CV; report the spatial structure of error.
"No independent validation of the classification."	RS/Physical	Add a held-out reference sample + area-adjusted accuracy.
"Convenience case; what is it a case of?"	Human/Nature-Society	State the case-selection logic and the population it represents.
"Whose voice / positionality?"	Human	Make reflexivity and interpretation-checking explicit.

Calibration anchors

• Space is a design issue, not a covariate. Dependence, scale, projection, and sampling are decided in the design, not patched in robustness.
• Each tradition on its own terms. A qualitative design is not weaker for lacking an estimand; it needs case logic, reflexivity, and disconfirmation criteria instead.
• Mixed means integrated. Two parallel analyses are not mixed methods; specify the linkage.

Anti-patterns

• Ignoring spatial autocorrelation, then reporting OLS SEs as if observations were independent
• No MAUP/scale sensitivity when the result could be a unit-of-analysis artifact
• Classification/prediction with no independent validation, or random CV on spatial data
• Convenience case selection dressed up as theory-driven; positionality omitted in interpretive work
• A nature-society design that never actually links the two strands

Output format

【Mode】spatial-quant / remote-sensing-physical / qualitative / mixed
【Estimand or claim】what is identified/shown
【Spatial integrity】dependence / MAUP-scale / projection / validation handled? [Y/N]
【Rival ruled out】the adjudication sentence (incl. scale/spatial-artifact)
【Robustness】planned checks
【Next】aaag-data-analysis

Supplementary resources

• ../../resources/external_tools.md — spatial-analysis, GIS, remote-sensing, and CAQDAS tooling by area
• ../../resources/official-source-map.md — areas, scope, and review model