Review Convergence Before Trusting a Fitted Model

Read more than the coefficient table

Many models require iterative optimization. A returned result object may coexist with warnings, boundary estimates or an inadequate stopping condition. Preserve the convergence status, optimizer, iteration count and relevant diagnostic information before interpreting coefficients.

The statsmodels pitfalls guide discusses incomplete convergence and cases where criteria can be misleading. A model’s numerical behavior needs inspection even when the software did not stop with an exception.

Identify what failed

Check whether the optimizer reached its iteration limit, whether parameters continue to change, and whether the objective or gradient supports convergence. Inspect nonfinite standard errors, unstable covariance estimates and parameters at boundaries. Separation, rank deficiency and poor scaling can create different symptoms that require different responses.

Do not respond to every warning by increasing the iteration limit. More iterations can help a slowly converging valid problem, but they cannot make an unidentified parameter identifiable or create information absent from the data.

Compare hypothetical restarts

Suppose a likelihood model returns log likelihood −120.0 from one starting point and −118.5 from another under the same model and data. The second objective is higher, but neither result should be accepted solely because it is higher. Check convergence diagnostics, parameter plausibility and whether repeated starts reach the same solution.

If the two solutions have materially different predictions or coefficients, retain both in the investigation log. An apparently stable objective can coexist with a flat direction in which parameters remain poorly determined. That matters when the scientific interpretation depends on those individual parameters.

Make the resolution reproducible

Record any changed starting values, tolerances, scaling or model constraints and their rationale. Test a small synthetic dataset with known behavior where possible. If the model remains unstable, use an appropriately justified alternative or report that the intended analysis could not be estimated reliably.

For simulations or resampling, count convergence failures rather than dropping them silently. A method that succeeds only on easier datasets can appear better when its failures are excluded. The final report should distinguish a scientifically appropriate model from a numerically successful implementation, and show the evidence supporting both before presenting an inferential conclusion.

Source verification: 6 November 2026. Examples are hypothetical.