Agent-based modeling of migrant integration through community sport Testing program design before implementation
Abdullah Tadmuri · UC3M MSc in Computational Social Science · 2026
An agent-based model of a structured calisthenics programme in urban public parks where migrants and locals train together. It tests which programme-design choices best retain participants and build cross-group ties, evaluated across 14,110 BehaviorSpace runs before any real-world pilot.
Navigate: Overview · Requirements · How to reproduce · Repository structure · Scenarios · Key parameters · Open access · Citation · License
CIM v6.4 simulates a structured calisthenics programme in urban public parks where migrants and locals train together, calibrated to conditions observed in Istanbul (2013--2022), and provides a second-domain Berlin BAMF integration-course calibration for cross-domain portability testing. Over 52 weeks (1 tick = 1 week), the model tracks:
- Attendance and dropout (4 mechanisms: motivation collapse, distance, work conflict, winter)
- Motivation dynamics (peer influence + decay)
- Language acquisition (CEFR scale, diminishing returns)
- Social tie formation (cross-group friendships, tie strength)
- Programme cost per retained participant
22 scenario conditions + 1 contextual variable (23 scenarios total) test programme design choices across infrastructure, support services, targeting, group structure, seasonal conditions, and Phase 3 robustness extensions (dose-response, buddy timing, open-cohort churn). The full simulation budget: 7,500 primary policy runs + 810 sensitivity + 100 auxiliary = 8,410 main-pipeline runs, plus 5,700 Tier 3 framework-generality runs (γ-bracket + Berlin second domain) = 14,110 BehaviorSpace runs in total.
- NetLogo 7.0.3: ccl.northwestern.edu/netlogo
- Java 21 (required by NetLogo 7)
- R ≥ 4.3 with packages installed via
R/00_setup.R
Two model files.
CIM_v6_4.nlogois the canonical reproducibility artifact; all results, the 36 BehaviorSpace experiments, and the thesis reference this file.CIM_v6_4_showcase.nlogois a presentation build with an upgraded interface (thematic shapes, park-catchment territories, thesis-palette plots, legend) for browsing and demos; its model logic is bit-identical to the canonical file (verified across the full 810-run sensitivity sweep). Use the canonical file to reproduce results.
Interactive dashboard. An exploratory web dashboard (policy ranking, animated year replay, network growth, and a browser spatial replay) lives in
viz/and builds todocs/for GitHub Pages. It reads exported data only; seeviz/README.md.
Open CIM_v6_4.nlogo in NetLogo 7.0.3. The pre-v6.4 source CIM_v6_3.nlogo is preserved at the repository root for reference.
- Set
config-domainchooser tocalisthenics-istanbul(default),language-course-berlin, orcustom. Forcustom, editconfig/custom.csvfirst and the model takes every parameter from that file. Or just click the Load Config CSV button to import any.csvas the custom config in one step (seeconfig/schema.md→ "Adapting to a new domain") - Set
scenario-typechooser to desired scenario - Click Setup → Go
- Exports written automatically to
data/{scenario}/on run end
Two options:
(a) From NetLogo GUI: Tools → BehaviorSpace → select experiment → Run.
(b) From the command line (recommended for batch runs):
bash run_all_tier3.sh # all 4 Tier 3 experiments, ~20 min on Apple Silicon
bash run_all_experiments.sh # the original v6.4 experiment suiteBoth scripts invoke org.nlogo.headless.Main directly via Homebrew's openjdk@21 + netlogo-7.0.3.jar rather than the native NetLogo_Console binary, which can hang in sandboxed environments at JVM init.
BehaviorSpace experiments in CIM_v6_4.nlogo:
| Group | Experiments |
|---|---|
| Main pipeline, original 16 | Baseline_300runs, NoIndoor_300runs, MinimalSupport_300runs, LowParkDensity_300runs, LowPark_topup200runs, WeakPeer_300runs, WeakPeer_topup200runs, SuboptimalComp_300runs, HighSES_300runs, WomenOnly_300runs, WomenOnly_topup200runs, NoIndoorMinimal_300runs, Targeting50_300runs, Targeting70_300runs, Targeting90_300runs, BuddyProgram_300runs, RotatingGroups_300runs, Winter50_300runs, WomenChildcare_300runs |
| Main pipeline, Phase 3 (7) | Composition2_300runs, Composition3_300runs, Composition4_300runs, OpenPopulation_300runs, SuboptimalOpen_300runs, CentralityBuddy_300runs, RandomBuddy_300runs |
| Auxiliary | Sensitivity_3level (810 runs), Equifinality_ContactContagion_100runs (100 runs) |
| Tier 3 framework-generality | GammaBracket_Low_300runs (900 runs), GammaBracket_High_300runs (900 runs), Berlin_AllScenarios_300runs (3,900 runs) |
Run accounting: 23 main scenarios = 7,500 policy runs + 810 sensitivity + 100 auxiliary = 8,410 main-pipeline runs; + 5,700 Tier 3 = 14,110 runs total (OpenPopulation's 300 runs are counted once, in the 7,500).
Outputs written to data/{scenario}/ for per-scenario experiments, or data/second_domain/{scenario}/ for the Berlin second-domain run:
CIM_results_{scenario}_{run}.csv: per-run summary metricsCIM_timeseries_{scenario}_{run}.csv: weekly aggregate panelCIM_agents_{scenario}_{run}.csv: agent-level cross-sectionCIM_panel_{scenario}_{run}.csv: agent-week person-period panelCIM_edges_{scenario}_{run}.csv: friendship edge list
Run from the CIM_Model/ directory:
source("R/00_setup.R") # install packages (once)
source("R/01_load_data.R") # load CSVs → RDS cache
source("R/02_descriptive_stats.R")
source("R/03_hypothesis_tests.R") # Welch t-tests, Holm family split (4+11)
source("R/04_sensitivity_analysis.R") # PRCC
source("R/05_survival_analysis.R") # Kaplan-Meier + Cox (cluster(run_id) + cox.zph)
source("R/06_visualization.R")
source("R/07_thesis_tables.R")
source("R/08_network_analysis.R")
source("R/09_policy_outputs.R")
source("R/10_validation_table.R")
source("R/11_agent_hazard.R")
source("R/12_distributional.R")
source("R/13_surrogate.R")
source("R/14_internal_invariants.R")
source("R/15_equifinality.R") # equifinality check (vs contact-contagion)
source("R/16_tier3_blockj_analysis.R") # open-population + gamma-bracket robustness
source("R/17_tier3_blocki_analysis.R") # Berlin second domain + hero figure + ranking invariance
source("R/18_tier3_integrate.R") # fills Discussion template from robustness outputs
source("R/19_tier3_splice_and_ship.R") # end-to-end: analysis → splice → render → package
source("R/20_sna_graph_analysis.R") # structural network metrics (modularity, clustering, etc.)
source("R/21_r_naught.R") # R0 / SIS-threshold derivation
source("R/22_did_parallel_trends.R") # difference-in-differences corroboration
source("R/23_randomization_inference.R") # randomization-inference p-values
source("R/24_dose_response.R") # composition dose-response sweep (Kendall, AIC/LRT)
source("R/25_link_prediction.R") # static link-prediction validation (4 predictors)
source("R/26_open_population_analysis.R") # open-cohort robustness (6 outcomes)
source("R/27_centrality_buddy_analysis.R") # 4-way buddy comparison (timing vs criterion)
source("R/28_open_pop_sna.R") # open-population multi-metric SNA
source("R/29_suboptimal_open_test.R") # SuboptimalOpen ranking-preservation testOr run all at once:
source("R/00_run_all.R")rmarkdown::render("thesis_CIM_v6.Rmd",
output_format = c("pdf_document", "word_document"))Produces thesis_CIM_v6.pdf (85 pages, PDF/A-1b).
CIM_Model/
├── CIM_v6_4.nlogo ← NetLogo model (current version, v6.4)
├── CIM_v6_3.nlogo ← Archival v6.3 source (preserved for reference)
├── thesis_CIM_v6.Rmd ← Thesis R Markdown document
├── thesis_CIM_v6.pdf ← Rendered thesis output (PDF/A-1b)
├── run_all_experiments.sh ← Canonical headless batch runner (all experiments)
├── run_all_tier3.sh ← Tier 3 experiment runner (γ-bracket, open-pop, Berlin)
├── README.md
├── CITATION.cff
├── LICENSE
├── apa.csl / references.bib
├── renv.lock ← Locked R package versions
├── R/
│ ├── 00_setup.R ← Package installation
│ ├── 00_run_all.R ← Run full pipeline
│ ├── 01_load_data.R ← CSV → RDS (handles duplicate blocks)
│ ├── 02_descriptive_stats.R
│ ├── 03_hypothesis_tests.R ← Welch t-test, Cohen's d, Holm correction
│ ├── 04_sensitivity_analysis.R ← PRCC + tornado plot
│ ├── 05_survival_analysis.R ← Kaplan-Meier + Cox PH
│ ├── 06_visualization.R ← core publication figures
│ ├── 07_thesis_tables.R ← LaTeX + CSV tables
│ ├── 08_network_analysis.R ← Degree distribution, assortativity
│ ├── 09_policy_outputs.R ← Cost-effectiveness, equity frontier
│ ├── 10_validation_table.R ← Face validity + pattern validity
│ ├── 11_agent_hazard.R ← Discrete-time dropout hazard (GLM)
│ ├── 12_distributional.R ← Quantile ribbons, tail risk
│ ├── 13_surrogate.R ← Scenario comparison summary table (surrogate regression removed in v6.3)
│ ├── 14_internal_invariants.R ← 10 mechanism invariant checks (11 sub-checks)
│ ├── 15_equifinality.R ← Equifinality check vs. contact-contagion mechanism
│ ├── 16-19_tier3_*.R ← Tier 3: open-pop, γ-bracket, Berlin, integration, packaging
│ ├── 20-23_*.R ← SNA metrics, R0 derivation, DiD, randomization inference
│ ├── 24-29_*.R ← Phase 3: dose-response, link-prediction, open-cohort, buddy, SNA, SuboptimalOpen
│ ├── constants.R ← Single source of truth (targets, scenarios, palette)
│ └── extra_agent_fate.R ← Agent survival-landscape figure
├── data/ ← Simulation CSVs + RDS cache (gitignored)
├── figures/ ← Generated figures (committed)
└── tables/ ← Generated tables (committed)
| Experiment | Scenario | Runs | Purpose |
|---|---|---|---|
| Baseline_300runs | Baseline | 300 | Control condition |
| NoIndoor_300runs | No Indoor Continuity | 300 | Winter barrier (H3) |
| MinimalSupport_300runs | Minimal Support | 300 | Support cuts (H4) |
| LowParkDensity_300runs + LowPark_topup200runs (500 total) | Low Park Density | 500 | Spatial access |
| WeakPeer_300runs + WeakPeer_topup200runs (500 total) | Weak Peer Influence | 500 | Peer mechanism (H1) |
| SuboptimalComp_300runs | Suboptimal Composition | 300 | Group composition (H2) |
| HighSES_300runs | High SES Heterogeneity | 300 | SES inequality |
| WomenOnly_300runs + WomenOnly_topup200runs (500 total) | Women-Only Groups | 500 | Gender equity |
| NoIndoorMinimal_300runs | NoIndoor Minimal | 300 | Combined barrier |
| Targeting50_300runs / Targeting70_300runs / Targeting90_300runs | Targeting50/70/90 | 300 each | SES targeting accuracy |
| BuddyProgram_300runs | BuddyProgram | 300 | Local buddy assignment |
| RotatingGroups_300runs | RotatingGroups | 300 | Group reshuffling |
| Winter50_300runs | Winter50 | 300 | Partial indoor coverage |
| WomenChildcare_300runs | WomenChildcare | 300 | Women + childcare |
| Composition2/3/4_300runs | Composition2/3/4 | 300 each | Group-composition dose-response (H2) |
| OpenPopulation_300runs | OpenPopulation | 300 | Open-cohort churn robustness |
| SuboptimalOpen_300runs | SuboptimalOpen | 300 | Suboptimal under churn (ranking-preservation test) |
| CentralityBuddy_300runs | CentralityBuddy | 300 | Week-8 buddy by highest local degree |
| RandomBuddy_300runs | RandomBuddy | 300 | Week-8 buddy by random selection (timing control) |
Total: 7,500 runs across 23 scenarios (16 original + 7 Phase 3 robustness extensions)
| Parameter | Default | Range | Basis |
|---|---|---|---|
| Peer influence coefficient | 0.08 | 0.01–0.20 | Centola (2010) |
| Motivation decay rate | 0.018 | 0.01–0.04 | Gjestvang (2020) |
| Language gain rate | 0.019 | 0.01–0.025 | CEFR literature |
| Tie formation probability | 0.05 | 0.02–0.15 | Kossinets & Watts (2006) |
| Dropout threshold | 0.20 | 0.10–0.40 | SDT literature |
All outputs of this thesis are openly available and cross-linked:
- Code and model (this repository): https://github.com/ATadmuri-hub/CIM-Model (MIT)
- Simulation dataset (Zenodo, DOI): https://doi.org/10.5281/zenodo.20668921 (CC BY 4.0)
- Interactive dashboard (GitHub Pages): https://atadmuri-hub.github.io/CIM-Model/
- NetLogo model (Modeling Commons): https://modelingcommons.org/browse/one_model/7916
The complete raw simulation output (43,713 CSV files from 14,110 BehaviorSpace runs) is archived on Zenodo at the DOI above. This repository ships the analysis-ready figures, tables, and precomputed outputs; the raw per-run CSVs are regenerable from run_all_experiments.sh and run_all_tier3.sh.
@software{tadmuri2026cim,
author = {Tadmuri, Abdullah},
title = {Calisthenics Integration Model ({CIM} v6.4)},
year = {2026},
version = {6.4},
url = {https://github.com/ATadmuri-hub/CIM-Model}
}Dual-licensed, see LICENSE:
- Software (NetLogo model,
R/pipeline,config/, shell runners): MIT. - Thesis manuscript (text + figures;
thesis_CIM_v6.*): CC BY-NC-ND 4.0, as declared on the thesis cover and in its PDF/A metadata.