Choosing a dose-response analysis
Choose a dose-response approach from the exposure scale, clinical question and plausible curve; a trend p-value cannot establish shape or causation.
Fit
When it matters, and when it does not
It matters when
- Exposure has meaningful amounts or ordered levels, and you must decide how to describe its relationship with an outcome.
- You need to distinguish a scored linear trend from category contrasts or a flexible curve.
It matters less when
- The exposure has unordered kinds, such as different drugs. Compare each with a reference group instead.
- You need to show that the exposure causes the outcome. A gradient supports a causal reading; only design and adjustment address confounding. Read: Confounding
Before you start
What to check in your study
- The exposure, in its own unitsKeep the measured amount, such as milligrams or pack-years. Cutting it into categories throws information away.
- The reference levelEvery estimate is compared with it. Choose one that means something clinically and holds enough patients, such as never exposed.
- How the shape will be modeledCompare a prespecified straight-line target, category contrasts and a flexible curve. The clinical question should govern the choice.
- The outcome and its summary measureoutcomeTypesummaryMeasureSuch as a risk ratio for a yes-or-no outcome. It decides the regression model and the scale of the curve.
- For cumulative exposure, whether intensity and duration act alikeTen cigarettes a day for 20 years and 20 a day for 10 years are both 10 pack-years. Model them separately where you can.
- The fitted curve follows the true shape closely enough.How to check: Compare a spline with a straight line. A straight line can hide a threshold or a plateau.
- Factors that change with the amount of exposure are measured and adjusted for.How to check: Heavy exposure often travels with other risks, such as smoking with drinking. List them from a causal diagram.
- Exposure is measured well across its whole range.How to check: Recalled or estimated doses are often least accurate at the extremes. Error there can flatten or bend the curve.
- Each part of the range holds enough patients.How to check: Show the exposure distribution. A steep curve at the top may rest on a handful of people.
Plain language
How it works
A category comparison estimates contrasts with a reference group. A linear trend asks about a component along prespecified exposure scores. A flexible curve can show departures from a straight line. These choices answer different questions.
Keep the measured exposure when possible. Choose a linear trend for the stated score scale, or a flexible curve when shape matters. Category summaries can help readers, but may hide changes within groups. Consider intensity and duration separately when equal cumulative doses could act differently.
A gradient can support a causal argument alongside other evidence, but confounding and measurement error can also create one. A threshold or plateau can be clinically meaningful without a linear trend.
Worked example
A worked example
- Setting
- 800 factory workers, split into four equal groups by their cumulative solvent exposure and followed for liver disease.
- Question
- Does the risk of liver disease rise with exposure?
- Approach
- Compare each exposure quarter with the lowest, and plan a flexible curve using the original measured exposure.
- Result
- In each of the three lower quarters, 8 of 200 workers developed liver disease. In the top quarter, 24 of 200 did: a risk difference of 8 per 100 and a risk ratio of 3.
- What it means
- The lower three quarters have equal observed risk; the top one is higher. These summaries alone cannot distinguish a smooth rise from a threshold.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
A trend test assumes a shape, and the usual one assumes the effect is linear in whatever scale the categories were coded on, which is rarely defensible when the categories are quartiles of a skewed exposure. Assigning the category midpoint as a score builds in an assumption about spacing that the data may contradict. And a dose-response gradient strengthens a causal argument but does not establish it: confounders that themselves increase with exposure produce the same gradient.
- Cutting the exposure into quarters and reporting only the categories.Instead: Keep measured exposure when possible. Choose a linear term or flexible curve by the question; categories assume flat risk within cutpoints.
- Choosing cutpoints after looking at the data.Instead: Set them, or the spline's knots, in advance. Cutpoints chosen for the best P value overstate the effect.
- Reporting a top-category effect without saying how many people are in it.Instead: Give the count in each category, or the exposure distribution, so a large effect resting on few people is visible.
Inside Mallard
How Mallard uses this
Mallard lists Choosing a dose-response analysis under Core research concepts. Its related methods include a scored trend test and restricted cubic splines.
concept.choosing_dose_response, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Methods for trend estimation from summarized dose-response data, with applications to meta-analysis (opens in a new tab) · Am J Epidemiol · doi 10.1093/oxfordjournals.aje.a116237
- Cubic splines to model relationships between continuous variables and outcomes: a guide for clinicians (opens in a new tab) · Bone Marrow Transplant · doi 10.1038/s41409-019-0679-x
- Dose-response analyses using restricted cubic spline functions in public health research (opens in a new tab) · Stat Med · doi 10.1002/sim.3841