PUBH5010 · Epidemiology Methods And Uses
Systematic Reviews, Meta-Analysis and Causation
The final block turns a pile of individual studies into one defensible conclusion. A systematic review answers a focused question with an explicit, reproducible protocol — a defined search, pre-set inclusion criteria, and a critical appraisal of each included study — which is what separates it from a casual narrative review. When the studies are similar enough, a meta-analysis pools their results into a single weighted estimate, larger and more precise studies counting for more, and displays them on a forest plot: each study a box (size = its weight) and a horizontal line (its confidence interval), with the pooled estimate as the diamond at the bottom; the vertical line of no effect tells you at a glance which studies and which pooled result are significant. The catch is heterogeneity — if the true effects genuinely differ across studies, pooling them into one number can mislead, so you inspect the scatter (and statistics like I²) before trusting the diamond, and watch for publication bias, the tendency for positive results to be published more readily. The chapter closes on causation: a statistical association is necessary but not sufficient, and the Bradford Hill considerations (strength, consistency, temporality, dose–response, plausibility, and the rest) are a framework — not a checklist to tick — for arguing whether an association is likely causal, with temporality (cause precedes effect) the one genuinely required condition.
What this chapter covers
- 01Systematic review vs narrative review: the explicit protocol
- 02Meta-analysis: pooling studies into one weighted estimate
- 03Reading a forest plot: boxes, lines, the diamond and the line of no effect
- 04Heterogeneity and when pooling is (in)appropriate
- 05Publication bias
- 06Association is necessary but not sufficient for causation
- 07The Bradford Hill considerations, with temporality required
Worked example: reading a forest plot and judging causation
- +2(a) The diamond. The pooled diamond lies entirely to the right of the line of no effect (RR = 1), so the pooled association is statistically significant — the exposure is associated with higher risk. A diamond crossing the line would be non-significant.
- +2(b) Two Hill considerations supported. Consistency (the studies mostly agree, intervals overlapping above 1) and strength (RR = 1.6 is a moderate, non-trivial association).
- +1(c) The required one. Temporality — the exposure must precede the outcome. It is the single Bradford Hill consideration that must hold for causation; the cohort design supports it here.
Key terms
- Systematic review
- A review that answers a focused question using an explicit, reproducible protocol — a defined search strategy, pre-specified inclusion/exclusion criteria, and a structured critical appraisal of each study. The transparent method is what distinguishes it from a narrative review and lets others reproduce or update it.
- Meta-analysis
- A statistical pooling of the results of similar studies into a single weighted summary estimate, giving more weight to larger, more precise studies. It increases power and precision but is only valid when the studies are sufficiently alike (low heterogeneity).
- Forest plot
- The standard display of a meta-analysis: each study is a box (area proportional to its weight) with a horizontal line for its confidence interval, and the pooled estimate is a diamond at the foot. A vertical line of no effect lets you read significance at a glance — intervals or the diamond crossing it are non-significant.
- Heterogeneity
- Genuine variation in the true effect across studies, beyond chance. High heterogeneity (e.g. a large I²) warns that a single pooled estimate may be misleading, prompting subgroup analysis, a random-effects model, or a decision not to pool at all.
- Bradford Hill considerations
- A set of viewpoints — strength, consistency, specificity, temporality, dose–response (biological gradient), plausibility, coherence, experiment, analogy — used to weigh whether an association is likely causal. They are a framework for judgement, not a checklist to tick; temporality (cause before effect) is the one that must hold.
Systematic Reviews, Meta-Analysis and Causation FAQ
What makes a review 'systematic'?
An explicit, reproducible method: a pre-specified question, a documented and comprehensive search strategy, clear inclusion and exclusion criteria applied consistently, and a structured critical appraisal of every included study. This transparency — anyone could repeat the process and get the same set of studies — is what separates a systematic review from a narrative review, where selection and emphasis are at the author's discretion and prone to bias.
How do I read a forest plot?
Each study is a box (its size shows the weight it carries in the pool) with a horizontal line for its confidence interval; the pooled result is the diamond at the bottom. A vertical line marks no effect (RR/OR = 1). Any study whose interval crosses that line is individually non-significant; if the pooled diamond clears the line, the pooled estimate is significant. The spread of the study estimates also gives a visual read on heterogeneity.
When is it wrong to pool studies in a meta-analysis?
When the studies are too heterogeneous — their true effects genuinely differ because of differing populations, exposures, outcomes or designs — a single pooled number can be meaningless or misleading, averaging apples and oranges. Statistical heterogeneity (e.g. a high I²) and clinical judgement both inform this; the responses are subgroup analysis, a random-effects model, or simply not pooling. Publication bias (positive studies published more) is a further threat to any pooled estimate.
Does the Bradford Hill framework prove causation?
No — it is a structured aid to judgement, not a proof or a checklist to tick off. You weigh considerations like strength, consistency, dose–response, plausibility and coherence to build a case that an association is causal rather than due to chance, bias or confounding. Only temporality — the cause must precede the effect — is a genuine requirement; the others strengthen or weaken the argument but none is individually necessary or sufficient.
Exam move
Separate the three jobs: a systematic review is defined by its explicit, reproducible protocol; a meta-analysis pools comparable studies into one weighted estimate; a forest plot displays them — practise reading significance off whether a study line or the pooled diamond crosses the line of no effect, and use the scatter to judge heterogeneity. For causation, remember an association is necessary but not sufficient, treat the Bradford Hill considerations as a weight-of-evidence framework rather than a checklist, and know that temporality is the one condition that must hold.
