Seal Health Modeling — Human Impact on Alaska's Seals

The question

Does human activity move the needle on seal health?

The original research question behind this project, kept intact — everything downstream either supports, complicates, or extends this hypothesis with real numbers.

Research question

What is the relationship between human intervention — tourism, mineral extraction, fishing, construction — and seal population in Alaska?

Hypothesis

Seal population declines with fishing activity (bycatch, food competition), and is not meaningfully affected by drilling or tourism, which impact seals more indirectly. Tested below with correlation, ANOVA, multivariate regression, and cross-validated ML.

The data

Four public datasets, merged on year

Three power the original 2024 analysis. The fourth — 5,465 aerial survey rows across six real Alaska regions — sat unused until this extension.

Exploratory findings

Three trends worth worrying about

Estimated pup births, average seal mass, and the ocean economy driving human presence in seal habitat — all recomputed directly from the raw CSVs.

Estimated pups born, 2006–2018

Pribilof Islands fur seal pup production. Loading…

Average seal mass by year

Alaska Phocid Morphometrics, yearly mean ± field sample size.

Average standard length by year

Same morphometrics survey, body length in cm.

Ocean economy RealGDP by sector, 2005–2021 (indexed to first year = 100)

Six ocean-economy sectors, each indexed to its own starting value so growth is comparable — Offshore Mineral Extraction alone runs ~10x the raw dollar scale of the rest.

Statistical inference

Which relationships actually hold up?

Pearson correlation with significance testing, ANOVA on seal mass across years, and multivariate OLS isolating which ocean sectors move population, deaths, and mass.

Correlation matrix

Sector RealGDP vs. seal population, deaths, and mass. Blue = positive, red = negative.

Headline Pearson correlations

Filled dot = statistically significant at α = 0.10.

Predictive modeling — extended

An honest test of whether this actually predicts anything

The original notebook's bootstrapped Ridge model reported R² = 0.91 — but its held-out test set was 3 rows. Here, five models are compared on the same 13-year merged dataset using leave-one-out cross-validation, the right tool at this sample size.

Leave-one-out CV: predicting seal mass from sector RealGDP

Loading…

Permutation feature importance

From the best LOOCV model, fit on the full dataset — which sectors move its predictions most.

What this means

Read this before the bar chart above. A negative LOOCV R² means a model does worse than just predicting the training-fold average. That's not a bug — with 13 yearly observations, it's the honest result. It shows the original project's 0.91 R² was measuring how well Ridge regression memorized its own training data, not genuine predictive power on new years.

New for this site

The dataset nobody looked at

5,465 aerial survey counts of harbor seals across six Alaska regions, 1983–2006 — collected for the original project's dataset shortlist, then never touched again. Mined here for the first time.

Statewide non-pup count, all regions combined

Loading…

Trend by region

Linear slope of yearly counts. Blue = growing, red = declining.

Raw counts by region

Small multiples, each on its own scale.

Limitations & ethics

Where to be skeptical of this project

Statistical limitations

The merged modeling dataset has 13 rows. Every regression and correlation here is likely underpowered, and correlation between sector RealGDP and seal metrics does not establish causation — both could be driven by a third factor (e.g. general Alaska economic growth). The multivariate OLS models use a relaxed α = 0.15, which is looser than typical practice and would flag some false positives.

Coverage & sampling

Findings are specific to the surveyed Pribilof Islands and coastal Alaska regions and may not generalize elsewhere. Aerial survey coverage varies sharply by year (as few as 90, as many as 512 flights/year) and by region, so region-to-region trend comparisons should be read as directional, not precise.

Data provenance

All four datasets are public (NOAA Fisheries, ADF&G, and NOAA's Ocean Economy program) and used here for research purposes only, consistent with the original project's stated ethics review.

Model honesty

The leave-one-out comparison in the Modeling section exists specifically to counter over-claiming from a tiny sample — it's a deliberately more conservative number than the original notebook's in-sample bootstrap result, and it's the one we'd stand behind.