9  Chapter 9 — Is It Really the Cause?

Applied Epidemiology in Aquaculture — Part III: The Investigator’s Toolkit

Where this sits — Anticipate (primordial/primary prevention). Anticipation only works if you act on real causes. A control measure aimed at a confounded, non-causal association wastes money and leaves the true driver untouched. This chapter separates genuine causes from the impostors.

9.1 Overview

Chapter 5 gave us associations; this chapter asks whether they are causal. Three things can manufacture an association that is not a cause: chance (handled by confidence intervals), bias (a flaw in how data were collected), and confounding (a third factor driving both exposure and outcome). We work a real, important example — the link between sea lice and SRS — show how confounding by season inflates it, adjust it away, and close with the criteria epidemiologists use to argue causation when a randomised experiment is impossible.

We use srs_confounding.csv: 1,000 cage-cycles with sea-lice burden, SRS status, and season.

9.2 9.1 Preparation

pacman::p_load(rio, here, tidyverse, janitor, epiR)
conf <- import(here("data", "srs_confounding.csv"))

9.3 9.2 Confounding — the impostor

A confounder is associated with the exposure and an independent cause of the outcome, without being on the causal path. Warm water is the classic one here: it raises sea-lice burdens and, separately, raises SRS risk (through gill stress, faster transmission). So cages with high lice also tend to be warm-season cages, which would have had more SRS anyway.

The crude association:

tab <- table(conf$high_lice, conf$srs)
epi.2by2(tab, method = "cross.sectional")
#> Odds ratio (crude)   2.84

A crude OR of 2.84 looks alarming. But stratify by season and adjust with the Mantel–Haenszel estimator — which base R does directly:

mantelhaen.test(table(conf$high_lice, conf$srs, conf$season))
#> common odds ratio (MH, adjusted for season) = 1.55

The adjusted OR falls to 1.55 — close to the real effect built into these data (OR ≈ 1.5). Within each season the OR is modest (cool 1.98, warm 1.45); only the crude figure was inflated, because season was doing most of the work. The lesson is exact and it comes straight from the field: lice do raise SRS risk, but naïvely comparing high- and low-lice cages without accounting for temperature roughly doubles the apparent effect (the “honest caveats” in your keynote — temperature and season drive both).

9.3.1 Controlling confounding

  • In design: randomisation (rare in field aquaculture), restriction, or matching.
  • In analysis: stratification (Mantel–Haenszel, above) or multivariable regression (Chapter 11’s logistic models), which adjust for several confounders at once.

A directed acyclic graph (DAG) is the tool for deciding what to adjust for — drawing the assumed causal arrows (season → lice, season → SRS, lice → SRS) makes the confounder, and the need to condition on it, visible. Adjust for confounders; never adjust for a variable on the causal pathway itself.

9.4 9.3 Bias — the crooked ruler

Where confounding is a real-but-mixed signal, bias is a systematic error baked in by how data were gathered. Two families matter most:

  • Selection bias — the sampled animals are not representative. A case-control study of SRS that draws controls only from healthy, well-managed sites will overstate the effect of any risk factor.
  • Information (misclassification) bias — exposure or outcome is measured with systematic error. If the qPCR cut-off differs between labs (Peña et al., 2025), “SRS-positive” is misclassified differently across sites, distorting any comparison. When misclassification is non-differential (equal in both groups) it usually biases toward the null; when differential, it can push either way.

Bias cannot be fixed in analysis — only avoided by good design (Chapter 7). This is why design discipline is an Anticipate activity, not a formality.

9.5 9.4 Arguing causation without an experiment

Field epidemiology rarely allows randomised trials, so causation is argued with a set of viewpoints (after Bradford Hill). Applied to lice → SRS:

  • Strength — a clear, adjusted association (OR ~1.5; RR ~1.4 in meta-analysis).
  • Consistency — seen across field cohorts and experimental challenges (Lhorente 2014; Arriagada 2019).
  • Temporality — the non-negotiable one: the rise in egg-bearing lice precedes the SRS mortality (Diethelm-Varela et al., 2025). Cause must come before effect.
  • Dose-response — higher lice counts, higher SRS mortality.
  • Plausibility — lice lesions open the skin, a known entry route; louse secretions dampen local immunity.
  • Experiment — co-infection challenges raise mortality when lice are added.

No single criterion proves cause, and context matters: the lice–SRS link is strong in Chile but absent in Ireland at record-low lice levels. Causation is a weight-of-evidence judgement, not a p-value.

9.6 9.5 Interpretation on the farm

For Anticipate, the pay-off is avoiding expensive mistakes. Treating the crude OR of 2.84 as the lice effect would over-credit delousing and ignore the shared driver, temperature — and delousing is itself a stressor that can worsen SRS (Price et al., 2020), so acting on a confounded estimate could backfire. The disciplined path is: estimate the association (Ch 5), rule out chance (CIs), design out bias (Ch 7), adjust for confounding (here), and weigh the causal criteria. Only then does “a risk factor” become “a control target.”

9.7 9.6 Exercises

Try it yourself (9.1). Reproduce the crude OR and the Mantel–Haenszel adjusted OR. By how much did adjusting for season change the estimate, and in which direction?

9.2 Compute the stratum-specific ORs for warm and cool season separately. They are similar to each other but both below the crude OR — explain why that is the signature of confounding rather than effect modification.

9.3 A case-control study recruits SRS cases from all farms but controls only from farms with no SRS history. Name the bias and explain which direction it pushes the odds ratio.

9.4 Of Bradford Hill’s viewpoints, which one is logically essential (an association failing it cannot be causal), and why can a cross-sectional study rarely establish it?

9.5 Draw a simple DAG for season, sea lice, and SRS. Which variable must you adjust for, and which must you not adjust for, and why?

9.8 Summary & what’s next

  • An association can arise from chance, bias, or confounding before it reflects a cause.
  • Confounding (a shared driver) is handled by stratification (Mantel–Haenszel) or regression; here it inflated the lice–SRS OR from a true ~1.5 to a crude 2.84.
  • Bias is a design flaw and cannot be fixed in analysis — only prevented.
  • Causation is argued by weight of evidence (strength, consistency, temporality, dose-response, plausibility, experiment), and is context-specific.

This closes Part III. In Chapter 10 — Vaccines, R₀ and Herd Immunity we move to Prevent: measuring vaccine effectiveness, estimating how fast disease spreads (R₀), the herd-immunity threshold, and whether vaccination pays.

9.9 References & further reading

  • Dohoo I, Martin W, Stryhn H. Veterinary Epidemiologic Research. 2nd ed. VER Inc.; 2009. — confounding, bias, causal inference.
  • Thrusfield M, Christley R. Veterinary Epidemiology. 4th ed. Wiley-Blackwell; 2018.
  • Diethelm-Varela B, et al., Mardones FO. Epidemiology of SRS: temporality of lice and SRS mortality. J Fish Dis. 2025;e70097.
  • Lhorente JP, et al. Co-infection challenge: Caligus and P. salmonis. PLoS ONE. 2014.
  • Arriagada G, et al. Sea-lice abundance and subsequent SRS mortality. Prev Vet Med. 2019;171:104771.
  • Price D, et al. Sea-lice baths and SRS onset. Aquaculture. 2020.

Files for this chapter: data/srs_confounding.csv and scripts/make_srs_confounding.R (new).