Design on What Was Actually Done

 Quick Analysis Plan Copied

 Long Analysis Plan

Evidence motivating the design

The Malagasy bee occurrence dataset (4,071 cleaned records, 218 species) shows 77.1% of records within 2.5 km of roads, 79.9% of 25×25 km grid cells entirely unsampled, and 68.4% of 231 KBAs with no bee records; sampled KBAs held a median of only 5 records and 3 species, and within-KBA penetration depth was not resolvable Roadside clustering and KBA sampling gaps. This makes a standardized-effort within-KBA/off-road comparison the decisive test of whether roadside bias is an artifact of collector access or a real ecological signal.

Experimental design

• Units and strata: 20 KBAs selected from the 73 historically sampled KBAs (prioritize those with ≥5 existing records) and stratified by region/vegetation type. Within each KBA, three matched plots: (a) boundary (<100 m from KBA edge, near-road), (b) interior (>500 m from boundary, off-road but within KBA), (c) >5 km off-road outside the KBA in matched habitat. Plot order randomized; observer-blind to stratum labels where feasible.
• Standardized effort: identical protocol per plot: 3 linear netting transects (500 m each, 60 min) plus 2 sets of 5 pan-traps (yellow/white/blue UV-bright) along 30 m line, deployed 6 h on 2 consecutive rain-free days, repeated in two seasons (wet/dry). Effort (trap-hours, net-hours) logged identically so effort is never a covariate confound.
• Primary hypothesis: with effort standardized, species richness and community composition differ between boundary/near-road and >5 km off-road plots; the roadside bias observed in occurrence databases is collector-access-driven, so bias magnitude (off-road minus boundary dissimilarity) shrinks toward zero relative to historical data.
• Falsification: if boundary and off-road plots show equivalent composition and richness (CI of paired effect includes zero), the roadside bias in the database reflects genuine ecological gradients or residual habitat confounds, not access bias.

Bioinformatics/analysis plan

• Clean occurrences with BeeBDC/bdc-style filters; georeference all new records; distance-to-road via GRIP lines and sf overlays (as in the source study) Pipeline: BeeBDC, sf, GRIP, vegan Chao1.
• Paired mixed models: richness/abundance ~ stratum (boundary/interior/off-road) with KBA as random blocking factor (lme4, brms); beta-diversity via PERMANOVA on strata pairs (adonis2, vegan) with KBA as strata.
• Chao1/iNEXT coverage-standardized richness per plot so sparse sampling (source KBAs median 5 records) does not distort comparisons Per-KBA sampling sparsity.
• COI metabarcoding or single-specimen barcoding of trap catches to guard against misidentification-driven bias, analogous to how COI barcoding exposed cryptic diversity in Neuroptera KBA assessments COI barcoding improves insect KBA reliability.
• Power: with 20 paired KBAs and 3 strata, detectable standardized effect (Cohen's d) ≈ 0.65 at α=0.05, power 0.8 for paired t-type tests; report per-KBA coverage to confirm adequacy.

Known unknowns and blindspots

Historical road networks cannot be reconstructed, so 'near-road' distances for legacy records are approximate against the contemporary network Contemporary-road limitation. Habitat type must be matched across strata or a vegetation covariate added, otherwise 'off-road' differences conflate access bias with habitat. Pan-trap color bias and seasonal timing are unresolved and should be balanced across strata within the same day-pairs. KBA Criterion E analyses show irreplaceability is sensitive to thresholds, metrics, and cost assumptions Criterion E sensitivity to analytical choices, so conclusions about KBA prioritization from this sample should remain bounded to the 20-KBA sample, not generalized to all 231 KBAs. Run Paired-Design Power Simulation