The invisible variable that ruins your analysis
You run a Cox model with age, APACHE, SOFA, and treatment. The HR for treatment is 0.60 (95% CI 0.45–0.80). You declare the intervention beneficial. But the paper never mentions that patients with SOFA ≥ 8 were excluded from the analysis. Those patients were the sickest — and they were also the ones most likely to die. If you had included them, the HR might have been 0.92 (no effect). The missing variable is confounding — a third factor that distorts the apparent effect.
Confounding occurs when a third variable is associated with both the exposure and the outcome, and it is not adjusted for. The classic example: older patients (confounder) tend to receive more aggressive vasopressor therapy (exposure) and also have worse outcomes (outcome). If you don’t adjust for age, you attribute the age effect to the treatment.
How to spot confounding
- Check the baseline characteristics — compare exposed vs unexposed on all baseline variables. If they differ systematically, something is driving the difference.
- Look for known confounders — in ICU, common confounders include: APACHE score, SOFA score, lactate, vasopressor use, sepsis source, ICU admission route, and comorbidities.
- Use multivariate adjustment — include suspected confounders in the model. If the HR for the exposure remains stable after adjustment, the effect is likely genuine.
- Sensitivity analysis — if you cannot collect all confounders, quantify how much the estimate would shift if the confounder were stronger than you think.
Interaction (effect modification)
Sometimes the treatment effect differs by subgroup. For example, early goal‑directed therapy may be beneficial in moderate‑severity sepsis but harmful in severe shock.
Detecting interactions
- Statistical interaction: Include an interaction term (e.g.,
treatment * SOFA). If the coefficient is significant, the effect varies by SOFA level. - Clinical intuition: If the treatment seems to work in one subgroup but not another, an interaction is likely.
Interpreting interaction results
If the interaction is significant, you must report the effect in each subgroup (e.g., “In patients with SOFA ≤ 4, HR = 0.55; in SOFA ≥ 8, HR = 1.10”). A single pooled HR masks the reality that the intervention helps some patients and hurts others.
Model building in the ICU
Step‑by‑step workflow
- Start simple — fit a model with the exposure and a few key confounders (age, APACHE, SOFA, lactate).
- Check assumptions — linearity, homoscedasticity, independence, multicollinearity.
- Add covariates iteratively — use stepwise or LASSO-like pruning, but do not add covariates that are not theoretically justified.
- Assess discrimination — calculate the C‑statistic (area under the ROC curve). A C‑statistic < 0.70 suggests the model is not capturing the signal.
- Validate — use cross‑validation or bootstrapping to ensure stability.
Common pitfalls in ICU regression
- Over‑adjustment — adjusting for too many variables can absorb the treatment effect if those variables are themselves affected by the treatment (collider bias).
- Time‑dependent confounding — in ICU, treatment decisions often change over time (e.g., moving from vasopressors to steroids). Adjust for time‑varying confounders using g‑computation or marginal structural models.
- Multicollinearity — APACHE and SOFA are highly correlated; including both inflates standard errors and can flip signs of coefficients.
Your appraisal checklist for regression models
# | Question | Why it matters |
1 | Are confounders identified and adjusted? | Unadjusted confounding → biased effect estimate. |
2 | Is the model parsimonious? (too many covariates) | Overfitting; standard errors underestimated. |
3 | Is the functional form correct? (linear vs non‑linear) | Misspecification → biased estimates. |
4 | Are interactions examined? | Heterogeneous treatment effects may be masked by a single HR. |
5 | Is the model validated? (cross‑validation, bootstrap) | Ensures stability and generalizability. |
6 | Are assumptions checked? (homoscedasticity, independence, normality of residuals) | Violations invalidate inference. |
Go deeper
- StatPearls – “Confounding” (NBK540915): free, covers confounding, adjustment, and sensitivity analysis. https://www.ncbi.nlm.nih.gov/books/NBK540915/
- StatPearls – “Multivariable regression” (NBK560123): free, covers multicollinearity, interaction, and model building. https://www.ncbi.nlm.nih.gov/books/NBK560123/
- OpenIntro Statistics – Chapter 10 “Multivariable regression” (free PDF): practical guide to adjusting for confounders, checking assumptions, and interpreting multivariable models. https://www.openintro.org/stat/textbook.php
- PMC6620654 – “Confounding in clinical research” (free): detailed discussion of confounding mechanisms and sensitivity analysis. https://doi.org/10.1186/s12916-022-03391-5
Next: Chapter 14 — Practical tips & conclusion