results.params
results.bse
results.robust_bse
results.summary()Summary (results object)
model.fit() returns a Summary object. In examples below it is assigned to results. Its everyday attributes are direct, named pandas.Series or scalar values; optimizer implementation details are kept under optim_result.
Everyday use
| Attribute | Description |
|---|---|
params |
Estimated coefficients indexed by variable name. |
bse |
Conventional standard errors indexed like params. |
robust_bse |
Robust sandwich standard errors indexed like params. |
bse_at(params, robust=False) |
Standard errors evaluated at supplied parameter values. False (default) omits robust adjustments; True uses the configured robust covariance. Does not optimize. |
bse_at_diagnostics(params) |
Both SE variants plus score, Hessian conditioning, weakest scaled-Hessian direction, and variance clipping diagnostics at supplied values. |
tvalues |
Test statistics based on robust standard errors. |
pvalues |
Two-sided p-values based on robust standard errors. |
conf_int(alpha=0.05) |
Two-sided confidence intervals based on robust standard errors. |
nobs |
Number of observations used in the fit. |
df_resid |
Residual degrees of freedom. |
llf |
Maximized log-likelihood. |
aic, bic |
Information criteria. |
converged |
Whether optimization converged. |
random_effects |
Object exposing .group and .time residual effects and standard deviations. |
resid |
Regression residuals. |
fittedvalues |
Fitted dependent-variable values. |
forecast(steps=1) |
Future dependent-variable values for each panel unit. |
summary() |
Formats a human-readable regression summary. |
bse_at() accepts a complete dict or pandas.Series keyed by the same parameter names as results.params. It evaluates the likelihood, gradient, and Hessian at those values without re-estimating the model. By default, robust=False returns standard errors without HAC or cluster adjustments. Set robust=True to use paneltime’s configured robust covariance, including its default Newey-West lag and panel-cluster adjustments. Built-in likelihoods use analytic derivatives; custom likelihoods may use the finite-difference fallback. This is useful in simulations where the true coefficients are known. The returned standard errors remain sample-specific; they are not theoretical or Monte Carlo standard errors.
For example, start with a complete fitted-parameter vector and replace values whose simulation truths are known:
true_params = results.params.copy()
true_params["X0"] = 1.0
true_params["X1"] = 1.0
se_at_truth = results.bse_at(true_params)The current initial variance and random-effect variance components are not free entries in params; they remain determined by the fitted model/data. conf_int(alpha=...) uses the requested significance level, where alpha must be between zero and one. Forecast requests must fit within the future prediction rows prepared by the model; unavailable horizons raise ValueError instead of returning a partial panel slice.
bse_at_diagnostics() accepts the same complete parameter mapping and returns both bse and robust_bse, the log-likelihood, total and per-observation scores, Hessian diagonal, raw Hessian condition number, scaled condition index, weakest direction of the diagonally scaled Hessian, and fraction of included conditional variances clipped at model bounds. The scaled condition index is the eigenvalue-ratio diagnostic used by paneltime’s multicollinearity report.
Optimizer internals (advanced)
Advanced diagnostics are available from results.optim_result, including hessian, gradient_matrix, gradient_vector, dx_norm, its, msg, t0, t1, and log_likelihood_object.
For compatibility, results.output and results.table expose engine objects. Prefer the direct result attributes above for routine analysis; the engine objects are implementation details.