The study-system term (sys_id) is fitted only in this sensitivity analysis. The primary model captures study-level dependence through ref_id and shared ancestry through the phylogeny, but does not include a separate study-system term. In PROCEED, sys_id, ref_id, and species identity are strongly associated. Most systems map to one species and often one source.
NoteWhat changed—and what did not
Changed: a study-system random intercept, (1 | sys_id), was added.
Unchanged: the full dataset, moderator definitions, phylogenetic effect, ref_id effect, priors, and MCMC settings match the primary models.
The comparison uses the fixed-effect tables. The three variance components are not reported separately.
D.2 Model specification
Models were fit on a remote server, not while rendering this book: 4 chains × 4,000 iterations (2,000 warmup), adapt_delta = 0.97, max_treedepth = 15, and the cmdstanr backend.
Six moderator models are reported (m01–m05, m07). The environmental-change model (m08) is excluded because its primary fit did not converge. Convergence criteria: max \(\hat{R} \leq 1.01\), bulk and tail ESS \(\geq 400\), and no divergent transitions.
Code
sens_diagnostics_table(variant)
Convergence diagnostics (4 chains, 2000 post-warmup draws each).
Moderator
Max R̂
Min bulk ESS
Min tail ESS
Divergences
Max-treedepth hits
Elapsed time — log₁₀ generations (m04)
1.0222
248
60
191
0
Disturbance context (m01)
1.0054
646
1311
21
0
Phenotypic vs genetic study (m07)
1.0063
742
1867
21
0
Trait type (m05)
1.0116
580
678
16
0
Comparison design (m02)
1.0194
571
1321
6
0
Elapsed time — log₁₀ years (m03)
1.0058
946
1518
0
0
Warning
Models requiring convergence caution: Comparison design (m02): max Rhat = 1.0194, min bulk/tail ESS = 571/1321, divergences = 6; Disturbance context (m01): max Rhat = 1.0054, min bulk/tail ESS = 646/1311, divergences = 21; Phenotypic vs genetic study (m07): max Rhat = 1.0063, min bulk/tail ESS = 742/1867, divergences = 21; Elapsed time in log10 generations (m04): max Rhat = 1.0222, min bulk/tail ESS = 248/60, divergences = 191; Trait type (m05): max Rhat = 1.0116, min bulk/tail ESS = 580/678, divergences = 16. See the table above for ESS and treedepth diagnostics.
D.4 Overall comparison
The comparison label uses a fixed rule: DIFFERS means that the posterior median changed sign or that the 95% credible interval changed between including and excluding zero. Consistent means neither occurred; it does not mean the two estimates are identical. Intercepts are omitted from the overall count.
Overall: 5 of 34 moderator coefficients change their credible-interval conclusion relative to the primary fit.
D.5 Paired estimates by moderator
Each table reports raw treatment-coded coefficients for the location (mean ln M) and scale (log-SD) submodels. See the sensitivity overview for column definitions.
---title: "Sensitivity C — Study-system (sys_id) random effect"---```{r setup, include=FALSE}source(here::here("Scripts", "00_packages.R"))source(here::here("Scripts", "19_sensitivity_tables.R"))variant <-"sens3_sysid"```## Recorded changeThe study-system term (`sys_id`) is fitted only in this sensitivity analysis. The primary modelcaptures study-level dependence through `ref_id` and shared ancestry through the phylogeny, butdoes not include a separate study-system term. In PROCEED, `sys_id`, `ref_id`, and speciesidentity are strongly associated. Most systems map to one species and often one source.::: {.callout-note title="What changed—and what did not"}**Changed:** a study-system random intercept, `(1 | sys_id)`, was added.**Unchanged:** the full dataset, moderator definitions, phylogenetic effect,`ref_id` effect, priors, and MCMC settings match the primary models.The comparison uses the fixed-effect tables. The three variance components arenot reported separately.:::## Model specificationModels were fit on a remote server, not while rendering this book: 4 chains × 4,000iterations (2,000 warmup), `adapt_delta = 0.97`, `max_treedepth = 15`, and the`cmdstanr` backend.```{r how-fit, eval=FALSE}phylo <-prepare_phylo_and_data(dat_mod, A_full)formula <-build_ls_formula(moderator, has_phylogeny = phylo$has_phylo,has_sys_id =TRUE)fit <-fit_ls_model( phylo$dat_model, formula,build_ls_priors(formula, phylo$dat_model, phylo$V, A = phylo$A_mod),V = phylo$V, A = phylo$A_mod, mcmc_args = default_mcmc_args)```## Convergence diagnosticsSix moderator models are reported (m01–m05, m07). The environmental-changemodel (m08) is excluded because its primary fit did not converge. Convergencecriteria: max $\hat{R} \leq 1.01$, bulk and tail ESS $\geq 400$, and nodivergent transitions.```{r diagnostics}sens_diagnostics_table(variant)```::: {.callout-warning}`r sens_diagnostic_flags(variant)`:::## Overall comparisonThe comparison label uses a fixed rule: **DIFFERS** means that the posteriormedian changed sign or that the 95% credible interval changed between includingand excluding zero. *Consistent* means neither occurred; it does not mean thetwo estimates are identical. Intercepts are omitted from the overall count.```{r stability-line, results='asis'}cat("**Overall:** ", sens_stability_summary(variant), "\n")```## Paired estimates by moderatorEach table reports raw treatment-coded coefficients for the location(mean `ln M`) and scale (log-SD) submodels. See the[sensitivity overview](../07_sensitivity_analyses.qmd#output-columns)for column definitions.```{r results, results='asis'}sens_all_moderator_tables(variant)```