- Published
- Updated
Measurements from the same person over time can be correlated. How a repeated-measures model represents that relationship affects uncertainty estimation and practical feasibility. Specify the model structure and assumptions rather than relying on the name of the statistical method alone.
Define the question and assumptions
Consider visit timing, data distribution and anticipated missingness patterns. Prespecify the covariance rationale and the sequence of alternatives if the model does not converge. A highly flexible structure can be unstable with limited information, while a simpler one may not represent every plausible pattern. Select and assess the approach with appropriate statistical expertise.
Hypothetical study example
In a hypothetical trial, many visits are scheduled but relatively few complete follow-up profiles exist. The planned model is unstable. The team applies its prespecified alternative sequence and diagnostics rather than choosing the structure producing the smallest p-value, documenting the connection between the decision and observed data characteristics.
Design review checks
- Specify time, group and covariance components explicitly in the model definition.
- Plan alternative-model rules without using comparative outcome knowledge to choose them.
- Assess convergence, uncertainty estimates and missingness patterns together.
Record the decision and its limits
Accepted code implements the reported model consistently. Do not ignore software warnings; show alternatives and reasons. Explain the clinical question addressed by different models and how uncertainty changes. Numerical completion of one run is not proof that every assumption holds or that the model provides an appropriate scientific interpretation for the available records.
Frequently asked questions
- Is the most flexible covariance always best?
- No. Information quantity and model stability also matter.
- Does convergence prove every assumption?
- No. Numerical completion differs from scientific suitability.