Choosing an analysis from your estimand
Choose an analysis for your estimand, the exact quantity you want to learn, rather than choosing a familiar model first.
Fit
When it matters, and when it does not
It matters when
- You have defined the population, outcome, comparison and time horizon.
- Several models could fit the same outcome but answer different questions.
It matters less when
- The question is still vague. Define the estimand before choosing a model. Read: From research question to estimand
- The needed measurements or comparison are absent. Revisit the design before fitting a model. Read: Choosing a study design from the research question
Before you start
What to check in your study
- The goal and summary measuretargetsummaryMeasureSay whether you seek a cause, an association, a description or a prediction. Name the measure, such as a risk difference.
- How observations ariseassignmentclusteredrepeatedMeasurescompetingEventsState assignment, clustering, repeated measures and competing events. Each can change the analysis.
- The design and data can identify the chosen quantity.How to check: For a causal effect, explain the comparison and confounding assumptions. A fitted model alone cannot supply them.
- The model output matches the chosen scale and population.How to check: A conditional odds ratio is not a population risk difference. Explain any averaging of predicted risks.
Plain language
How it works
Start with the target, then check the outcome type and how observations are linked. The same binary outcome can support different questions. A risk difference, risk ratio and odds ratio are different summaries.
Mallard compares the recorded summary measure with its sizing rules. It also checks whether the analysis names a conflicting measure. These checks do not establish causal identification. Old plans without a structured estimand receive no estimand findings; silence means not checked.
Worked example
A worked example
- Setting
- A clinic compares two follow-up programs at a fixed horizon.
- Question
- How much does the risk of an urgent visit differ?
- Approach
- For simple arithmetic, suppose follow-up is complete and each group contains 100 people.
- Result
- There are 20 of 100 events with one program and 30 of 100 with the other. The risk difference is -10 per 100; the risk ratio is 0.67.
- What it means
- These are observed group contrasts. Calling them causal effects still needs a defensible assignment mechanism and assumptions about confounding.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
A model name does not define the target estimate. A compatible sizing formula does not prove that the design identifies a causal effect.
- Calling every logistic coefficient a causal effect.Instead: Choose adjustment for the specific causal contrast. Other coefficients can answer different questions.
- Treating a green compatibility check as proof of validity.Instead: Check the design and assumptions yourself. Automated checks compare recorded claims and can miss errors.
Inside Mallard
How Mallard uses this
Mallard checks the recorded estimand against sizing, analysis and figures. A blocking sizing contradiction withholds the headline sample size.
concept.choosing_analysis, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Using Big Data to Emulate a Target Trial When a Randomized Trial Is Not Available. (opens in a new tab) · Am J Epidemiol · doi 10.1093/aje/kwv254
- The table 2 fallacy: presenting and interpreting confounder and modifier coefficients. (opens in a new tab) · Am J Epidemiol · doi 10.1093/aje/kws412