Linear Models¶

Summary¶

PyFLASH linear-model pipelines fit ordinary least squares models for numeric dependent variables. They are used to estimate group terms, covariate terms, interactions, adjusted means, and model summaries across one or more endpoints.

Where PyFLASH Uses It¶

linear_model is the main pipeline for covariate-adjusted ordinary least squares analysis.
run_linear_model_pipeline is the legacy wrapper for the same modelling path.
plot_multivariable_regression_matrix displays multivariable regression summaries.
iterative_model_sweep is separate classification/discovery modelling; see Classification.

Inputs¶

Inputs include one or more dependent variables, a modelling formula or term lists, optional group terms, predictors, interaction terms, categorical settings, reference levels, row filters, and sentinel values to exclude.

Categorical handling can be automatic. Object, category, and boolean columns are treated as categorical by default, and the configured group term is categorical. Numeric predictors stay numeric unless explicitly converted.

Rows with missing values in the dependent variable or required model terms are dropped for that model. Sentinel values are filtered before fitting and recorded in summary metadata.

Outputs¶

Linear-model outputs include:

coefficient tables with estimates, standard errors, t-values, p-values, confidence intervals, q-values, and rejection flags.
model-summary tables with sample counts, dropped rows, R-squared, adjusted R-squared, model degrees of freedom, residual degrees of freedom, F-statistic, AIC, and BIC.
model metadata describing formulas, terms, categorical handling, covariance settings, and filters.
optional adjusted-mean tables and adjusted-mean contrast tables.
adjusted-mean figures and coefficient summary figures.
report records when structured reporting is active.
manifest.json describing the run.

Interpretation¶

A coefficient estimates the expected change in the dependent variable for a one-unit change in a numeric predictor or relative to a reference level for a categorical predictor, conditional on the other model terms.

Adjusted means estimate group-level means after applying the configured reference-grid or observed-profile strategy. Their contrasts are often easier to read for group questions than raw model coefficients, especially when several covariates are present.

Robust covariance options change standard errors and p-values, not the fitted mean structure. Q-values in coefficient tables reflect the configured false discovery rate family.

Limitations¶

Linear models are associative unless the experimental design supports causal interpretation. Collinearity, small group sizes, sparse categorical levels, outliers, nonlinear relationships, and complete-case filtering can change estimates and uncertainty.

Adjusted means depend on the chosen profile and weighting mode. They should not be mixed with raw group means without noting how they were computed.