Comparing Imputation and Clustering Pipelines on Incomplete Suicide Data from Bangladesh

An exploratory study on 1,617 newspaper-reported suicide cases from Bangladesh — a country with no central suicide registry. We test 32 combinations of imputation, clustering, and evaluation to ask which pipeline recovers real structure from severely incomplete data, and which only recovers how the data was collected.

Why this study

Bangladesh has no centralised suicide registry. For a public-health problem this serious, that absence is the whole problem: the broadest record that exists is whatever newspapers happen to report. And that record is full of holes — ages missing, reasons unstated, locations vague, dates approximate. To study patterns in it at all, you first have to make two decisions: how to fill the gaps (imputation) and how to group what’s left (clustering).

Those choices are not neutral. Pick them carelessly and a clustering algorithm will happily “discover” structure that turns out to be an artifact of how the data was gathered rather than anything about the phenomenon itself. So this is an honest methods question wearing a public-health coat: when the data is this incomplete, which pipeline finds something real — and how would you even know?

The data

We assembled 1,617 newspaper-derived cases — to our knowledge the largest publicly compiled record of its kind for Bangladesh. Because raw reports are thin, we enriched them: missing weather and location fields were filled from external APIs (and the responses cached, so the work is reproducible without hitting those services again). Even after enrichment the missingness varies a lot feature to feature, which is exactly why the imputation choice matters so much.

What we tested

Instead of picking one pipeline and hoping, we built a factorial benchmark — every combination of three decisions:

• Imputation (3): MICE, KNN, MissForest
• Clustering (4): K-Means, Agglomerative, DBSCAN, Spectral
• Predictor strategy (2): context-aware vs. no-context (whether imputation is allowed to use related fields as predictors, or treats them independently)

That’s 32 full pipelines, plus two encoded baselines — each one run end-to-end and scored the same way.

How we scored them

A single clustering metric is easy to game — most of them will reward a degenerate “one giant cluster plus a few outliers” solution as nice and “tight.” So we scored each pipeline with a six-part composite: five standard validity indices (Silhouette, Calinski–Harabasz, Davies–Bouldin, Dunn, Xie–Beni) plus a custom balance penalty we added specifically to punish those near-degenerate partitions. The composite is what keeps the comparison honest across 32 very different pipelines.

What we found — and the catch

The best configuration was Agglomerative clustering on MissForest-imputed, no-context data, scoring 27.30 out of 32, and it produced a clean two-cluster split.

Here’s the part that matters most, and the reason this is framed as exploratory. That two-cluster boundary lines up almost exactly with a seam in our own dataset — the 2020 Kaggle batch versus the manually collected 2017–2019 and 2021–2025 cases. In other words, the structure is partly confounded with collection source: whatever the clusters capture is tangled up with reporting era and completeness, not just the underlying phenomenon. We report it as exploratory subgroup structure, not as discovered epidemiological categories — and we say so plainly, because the alternative is a confident result that isn’t true. Naming that confound is, honestly, the most useful finding here: it’s a trap that a single-method study would have walked straight into.

Reproducibility and ethics

The full pipeline, the assembled dataset, and the cached geocoding/weather lookups are open, so any stage can be re-run independently of the original sources — code under the MIT License, data under CC-BY 4.0. The dataset carries ethical safeguards throughout: no personally identifying information is published, and victim names and identifying details are stripped out. This is a sensitive subject, and the goal is to make the method reusable without putting individuals on display.

Status

This is ongoing work — defended as an undergraduate thesis, with a formal paper to follow. The repository is linked above as Source; the full writeup will go up once the work is formally published.