Causal Impact of Fishing Vessel Noise on Deep-Sea Migration

Course: IS 590 : Causal Inference Machine Learning

---

Causal Claim

COVID-19 disrupted commercial fishing vessel activity in the English Channel and North Sea, causing a measurable recovery in diel vertical migration (DVM) amplitude in those corridors relative to remote Mid-Atlantic control corridors where fishing effort was not comparably disrupted.

Diel Vertical Migration (DVM) is the largest daily animal migration on Earth by biomass. Mesopelagic organisms drive the biological carbon pump that sequesters CO2 in the deep ocean. Underwater radiated noise from vessel traffic suppresses DVM — but traditional ecological studies cannot isolate this from seasonal variation. This project uses COVID-19 as a natural experiment.

---

Dual Panel Design

Panel	Treatment variable	COVID signal	Purpose
Panel A	Real GFW fishing effort hours (174,026 records)	15.1% drop treated vs 5.8% control	Primary specification
Panel B	Simulated 60% COVID drop on real geographic grid	64.3% drop treated vs 0.5% control	Methodological validation

Panel B validates that the DML estimator correctly recovers a known true effect (TRUE_ATE = -0.35) on the real geographic structure. Both panels converge to essentially the same estimate — the central result of the project.

---

Method Stack

Layer	Method	Tool	Purpose
Identification	Difference-in-Differences	statsmodels	COVID fishing collapse as natural experiment
Estimation	Double/Debiased ML	EconML LinearDML	Residualises confounders nonlinearly
Validation	3 DoWhy refutation tests	DoWhy	Placebo, random cause, data subset
Extension	CausalForestDML	EconML	Heterogeneous effects by depth zone
Pipeline	DuckDB + dbt	dbt-duckdb	Staging and mart transformation layer
Frontend	Interactive dashboard	Streamlit + Plotly	Panel toggle, charts, results

---

Key Results

Method	Panel A (Real AIS)	Panel B (Simulated)
True ATE (DGP)	- 0.3500	- 0.3500
TWFE DiD	+ 0.037 (confounded)	NaN (multicollinearity)
LinearDML ATE	- 0.3489	-0.3460
Recovery error	0.32%	1.13%
Parallel trends	p = 0.913 (PASS)	p = 0.551 (PASS)
DoWhy refutations	3 / 3 PASS	3 / 3 PASS

Why TWFE DiD fails: SST and chlorophyll-a have nonlinear effects on DVM that linear regression cannot remove. DML's gradient boosting nuisance models residualise these flexibly, recovering the true ATE with under 1.2% error.

Why both panels converge: The DML estimator isolates the causal effect regardless of the treatment shock magnitude, provided confounders are adequately controlled.

---

Data

Source	Records	Coverage	Role
Global Fishing Watch API (public-global-fishing-effort)	174,026	2018-2020 Q2	Real treatment variable
Simulated DVM amplitude	118,221 cell-month obs	2018-2020 Q2	Outcome (known DGP, TRUE_ATE = - 0.35)
NOAA OISST (parameterised)	Per cell-month	2018-2020 Q2	SST confounder

Geographic Regions:

Region	Type	Bathymetry	Rationale
English Channel	Treated	~60m	High-traffic fishing corridor
North Sea	Treated	~95m	High-traffic fishing corridor
Mid-Atlantic zone 1 (35-40N, 40-30W)	Control	~4200m	Remote, low-traffic
Mid-Atlantic zone 2 (40-45N, 45-35W)	Control	~4500m	Remote, low-traffic

Note on treatment variable: The GFW free-tier API provides fishing vessel effort hours only — not all commercial shipping. Cargo ships, tankers, and container vessels (the primary URN sources and the vessels with the largest COVID collapse, ~60%) require licensed data access. The causal claim is therefore specific to fishing vessel acoustic disturbance. Panel B simulates the sharper cargo shock to validate the estimator under a known DGP.

---

Project Structure

midnight_zone/
├── data/
│   ├── fetch_ais.py          - Downloads real GFW fishing effort data
│   ├── build_panel.py        - Builds Panel A (real) and Panel B (simulated shock)
│   └── simulate.py           - Original simulation script (reference)
├── dbt_pipeline/
│   ├── models/
│   │   ├── staging/
│   │   │   └── stg_panel.sql       - Clean types, depth zones, derived fields
│   │   └── marts/
│   │       └── mart_analysis.sql   - Standardised features, group-week means
│   ├── dbt_project.yml
│   └── profiles.yml
├── analysis/
│   ├── 01_did_baseline.py    - TWFE DiD + parallel trends (visual + formal slope test)
│   ├── 02_dml.py             - LinearDML + DoWhy 3 refutation tests
│   └── 03_hte.py             - CausalForestDML by depth zone
├── app/
│   └── streamlit_app.py      - 6-page interactive dashboard with panel toggle
├── outputs/                  - Auto-generated plots and result CSVs
└── requirements.txt

---

Setup and Run

# 1. Clone
git clone git@github.com:PriyalManiar/midnight-zone.git
cd midnight-zone

# 2. Create virtual environment (Python 3.13 required)
python3.13 -m venv venv
source venv/bin/activate

# 3. Install dependencies
pip install -r requirements.txt

# 4. Download real AIS data from GFW
# Add your GFW API token to data/fetch_ais.py (GFW_TOKEN variable)
python data/fetch_ais.py

# 5. Build both panels
python data/build_panel.py

# 6. Run dbt pipeline
cd dbt_pipeline && dbt run --profiles-dir . && cd ..

# 7. Create sim mart table (run once after dbt)
python -c "
import duckdb
con = duckdb.connect('data/midnight_zone.duckdb')
con.execute('DROP TABLE IF EXISTS mart_analysis_sim')
con.execute('''CREATE TABLE mart_analysis_sim AS
    SELECT *, date::date as observation_date,
    avg(vessel_density) OVER (PARTITION BY week_index, is_treated) as group_week_mean_density,
    avg(dvm_amplitude) OVER (PARTITION BY week_index, is_treated) as group_week_mean_dvm,
    (sst-14.7)/4.3 as sst_std, (log_chla-0.48)/0.65 as log_chla_std,
    (bathymetry-1462)/1194 as bathymetry_std FROM raw_panel_sim''')
con.close()
"

# 8. Run analysis scripts in order
python analysis/01_did_baseline.py
python analysis/02_dml.py
python analysis/03_hte.py

# 9. Launch Streamlit dashboard
streamlit run app/streamlit_app.py

---

GFW API Token

Register at globalfishingwatch.org/our-apis for a free API token. Add it to data/fetch_ais.py:

GFW_TOKEN = "your_token_here"

The free tier gives access to public-global-fishing-effort : sufficient for this project.

---

Limitations

1. Parallel trends : another 2020 shock could confound; untestable in the post-period
2. Geographic dissimilarity : European shelf vs deep Mid-Atlantic; parallel trends rests on statistical evidence alone
3. Fishing vessels only : GFW free tier excludes cargo/tanker traffic, the primary URN sources
4. Simulated DVM outcome : recovery error validates on DGP, not real acoustic backscatter
5. SUTVA violation : acoustic spillover from treated to adjacent control cells
6. Spatial aggregation : 0.5-degree grid loses within-cell heterogeneity

---

Tech Stack

Python 3.13 · DuckDB · dbt-duckdb · EconML · DoWhy · statsmodels · scikit-learn · Streamlit · Plotly · GFW API v3 · pandas · numpy