gcb-data-analysis

Data Analysis (gcb-data-analysis)

GCB reviewers are quantitatively sophisticated, and because **data and code are archived publicly with

a DOI** (see gcb-reporting-and-data-policy), the analysis must be reproducible by a third party.

Analyze as if both are true — because they are. This skill covers execution and reporting norms; design

decisions live in gcb-study-design.

When to trigger

• Running main and supporting analyses; building the results
• Choosing the right model for nested/repeated/spatial ecological data
• Synthesizing effect sizes for a meta-analysis or evaluating a process model
• Making the analysis reproducible before deposit

Analysis norms GCB expects

1. Respect the data structure. Use mixed / hierarchical models (lme4, glmmTMB, brms, INLA)

for nested, repeated-measures, and spatially/temporally autocorrelated data; do not ignore random

effects or autocorrelation.

1. Report uncertainty honestly. Effect sizes with confidence/credible intervals, not just p-values

or stars; state the magnitude and its ecological/biogeochemical meaning.

1. Quantify, propagate, and partition uncertainty. For models, separate **parameter, structural,

and scenario uncertainty; prefer ensembles**; show measurement error where it matters.

1. Meta-analysis discipline. Appropriate effect size (log response ratio, Hedges' g), random/mixed

effects, heterogeneity (I^2, tau^2), moderators pre-specified, and a publication-bias check.

1. Evaluate models against observations. Report skill metrics and where the model fails, not only

where it succeeds.

1. Right inference for the unit. Match the analysis to the experimental/sampling unit; avoid

pseudoreplication carrying through from design.

Reproducibility while you work (not at the end)

• One master script regenerates every table and figure from raw/constructed data.
• Set and report seeds for any stochastic step (bootstrap, MCMC, simulation, model ensembles).
• Pin software/package versions (renv.lock, conda/requirements.txt, model version + forcing).
• Keep manuscript table/figure numbers matched to script outputs — they will be archived together.

Matching the method to the global-change question

GCB referees expect the analysis to fit the data-generating process. Use this as a routing table from

question shape to the inferential machinery a quantitatively literate reviewer will look for.

| Question shape | Expected machinery | What a reviewer checks |

|----------------|--------------------|------------------------|

| Effect of a manipulated driver across randomized plots | Mixed model with plot/block random effects | Random structure matches the design; no pseudoreplication |

| Trend in a flux time series | Autocorrelation-aware regression / state-space | Residual autocorrelation modelled, not ignored |

| Spatial pattern across a gradient | Spatial random field (INLA/spaMM) | Spatial dependence handled; CRS and area stated |

| Synthesis across many studies | Random/mixed-effects meta-analysis | Effect-size choice, I^2/tau^2, bias check |

| Future projection from a process model | Multi-model ensemble | Structural + parameter + scenario spread shown |

Worked micro-example (illustrative)

A warming-experiment meta-analysis pools log response ratios (lnRR) of aboveground biomass from 64

studies. A defensible GCB workflow: fit a random-effects model, report the pooled lnRR back-transformed

to a percentage with its interval, and quantify heterogeneity. Illustrative output — pooled lnRR 0.12,

i.e. a +13% biomass response (95% CI 6–20%), I^2 = 71% with tau^2 = 0.04, and a moderator showing the

effect halves in water-limited sites. The funnel plot and trim-and-fill leave the sign unchanged. The

71% heterogeneity is the result, not noise: it motivates the moisture moderator. All numbers illustrative.

Referee pushback patterns and the GCB-appropriate fix

• "Pseudoreplication: chamber treated as replicate" → move the treatment effect to a random-effect or

split-plot structure at the true unit of inference.

• "Heterogeneity ignored in the synthesis" → report I^2/tau^2 and pre-specified moderators, not a single

pooled mean.

• "Projection has no uncertainty band" → run an ensemble and partition parameter, structural, and

scenario spread rather than reporting one trajectory.

• "Skill claimed but never tested out-of-sample" → report validation against held-out observations and

the conditions where the model fails.

Anti-patterns

• Treating nested/repeated/spatial data as independent observations
• Stars-only results with no effect sizes, intervals, or ecological magnitude
• A single model run reported as if it had no structural or scenario uncertainty
• A meta-analysis with no heterogeneity or publication-bias assessment
• Code that cannot reproduce the printed tables/figures ("works on my machine")

Output format

【Main estimate】effect size + interval + ecological/biogeochemical meaning
【Data structure】random effects / autocorrelation handled? [Y/N]
【Uncertainty】measurement + parameter + structural + scenario partitioned?
【Model evaluation / heterogeneity】skill metrics or I^2 reported?
【Reproducible】master script + seeds + pinned versions? [Y/N]
【Next】gcb-figures-and-tables

Supplementary resources

• [../../resources/external_tools.md](../../resources/external_tools.md) — mixed-model, meta-analysis, spatial, and modelling packages
• [../../resources/official-source-map.md](../../resources/official-source-map.md) — data/code archiving policy