University project — Completed
Causal Inference — Smoking and Sleep Quality
Smoking and sleep study using 297,584 analyzed observations and Double Machine Learning.
01
The starting point
Study the relationship between smoking and sleep quality by distinguishing the raw association from an adjusted estimate. The central question is how confounding affects interpretation.
From an initial 413,768 observations, analysis of 297,584 records after excluding former smokers. Comparison of the raw difference, OLS regression and Double Machine Learning. Causal interpretation remains conditional on the causal graph assumptions and confounder adjustment.
413 768Initial observations
297 584Analyzed observations
5Cross-fitting folds
02
How it works
- 1413,768 observations
- 2Filtering & preparation
- 3DAG & adjustment
- 4Raw difference · OLS
- 5DML · cross-fitting
Architecture & implementation
- Cleaning and filtering: 413,768 initial observations, 297,584 analyzed.
- Causal DAG and adjustment variable selection.
- Comparison of raw difference and OLS regression.
- Double Machine Learning with Random Forest and five-fold cross-fitting.
Technical decisions
- Random Forest for nonlinear nuisance models.
- Five-fold cross-fitting to separate nuisance fitting from estimation.
Lessons learned
- Similar OLS and DML estimates do not by themselves prove causal validity.
- An explicit causal graph makes assumptions open to scrutiny.
Possible next steps
- Test additional causal methods (DiD, instrumental variables).
- Enrich the analysis with additional confounders (stress, physical activity).
- Study causal effect heterogeneity across subgroups.
03
Results & practical value
- A reproducible workflow: define the population, formalize assumptions and compare three approaches.
- Double Machine Learning implemented with Random Forest models and five-fold cross-fitting.
- Professional relevance: explain why correlation alone cannot support a causal decision.
Delivered so far
- Report and code produced.
- Three approaches compared: raw difference, OLS and DML.
- Estimates require interpretation in light of identification assumptions and unobserved factors.
04
Scope & limitations
- Results depend on the DAG, observed variables and identification assumptions. Unmeasured factors may change the interpretation.
- The reviewed deliverable does not report a DML confidence interval. This is an academic study, not medical guidance.
Source: causal inference project report and notebook. Counts after filtering; no clinical or commercial outcome estimate.