ergmx.ErgmFit

class ergmx.ErgmFit(model, estimate, mple, control, seed)

A fitted ERGM, as returned by ergmx.ergm().

Print summary() for the table of coefficients, like R’s summary(fit); check the fit with mcmc_diagnostics() and gof(); compare it with others with ergmx.compare().

save(path)

Save the fit to a file, to reload with ergmx.load_fit(): its model and network, estimates, MCMC sample and settings, with Python’s pickle (so only load files you trust).

property names

Names of the coefficients, as in ergm.

property params

The estimated coefficients.

property coef

The estimated coefficients by name.

property cov

Covariance matrix of the estimates, including MCMC error.

property stderr

Standard errors by name, including MCMC error (NaN for fixed coefficients).

property offset

Whether each coefficient is fixed by offset() rather than estimated.

property constraints

The model’s sample space constraints.

property mple

The maximum pseudo-likelihood estimate, the starting point of the MCMC MLE.

property method

"MLE" (exact, for dyad-independent models), "MCMLE", "MPLE" or "CD".

property iterations

Iterations of the Monte Carlo MLE (or of contrastive divergence).

property converged

Whether the estimation met its convergence criterion.

property loglik

exact for dyad-independent models, estimated by path sampling for the others (None if fitted with eval_loglik=False).

Type:

Log-likelihood

property loglik_se

Monte Carlo standard error of the log-likelihood (0 if exact).

property loglik_relative

Whether the log-likelihood is relative to the null model (the uniform distribution on the networks the constraints allow), as with dyad-dependent constraints in ergm. Relative log-likelihoods compare models with the same constraints.

property df

Number of estimated coefficients (offsets and constant statistics excluded).

property formula

The model’s terms.

property aic

Akaike’s information criterion, -2 loglik + 2 p.

property bic

Bayesian information criterion, -2 loglik + p log(number of free observed dyads), as in ergm.

property sample

Statistics sampled in the last iteration (chains x samples x statistics).

property observed

Statistics of the observed network (for curved terms, the counts their parameters weight).

simulate(nsim=1, *, seed=None, output='network', time_slices=None, nw_start='last', **options)

Simulate networks from the fitted model, starting from the observed one.

See ergmx.simulate(). For a model fitted to a series of networks (ergmx.tergm()), time_slices= simulates the process forward in time from the network nw_start instead, as tergm’s simulate(fit, nw.start=, time.slices=): "last" (the default) or "first" network of the series, its 1-based position, or a network. See ergmx.simulate_dynamic(), whose options apply.

mcmc_diagnostics()

Diagnostics of the MCMC sample of the last iteration, like R’s mcmc.diagnostics(). Print the result, or call its plot().

gof(nsim=100, **options)

Goodness of fit of the model. See ergmx.gof().

summary()

The table of coefficients, standard errors and tests, with the log-likelihood, AIC and BIC. Print it (it prints itself in notebooks).

to_frame()

The table of coefficients as a pandas DataFrame (needs pandas), with R’s columns: Estimate, Std. Error, MCMC %, z value and Pr(>|z|).

predict(conditional=True, type='response', nsim=100, *, seed=None, **options)

Tie probabilities of every dyad, as R’s predict(fit).

With conditional=True, each dyad’s probability of a tie given the rest of the observed network, computed exactly from its change statistics (type="link" for the log-odds); with conditional=False, the share of nsim networks simulated from the model in which the dyad is a tie. As in ergm, conditional probabilities ignore the sample space constraints, and dyads whose value is missing are predicted too; unconditional ones are simulated with the model’s constraints (ergm ignores them).

Returns:

TiePredictions – tail, head and p arrays, with .matrix(), .mean_by(attribute) and .to_frame().

marginal_effects()

Average marginal effects on the tie probability, as R’s ergMargins (ergm.AME()): for each estimated parameter, the change in a dyad’s conditional tie probability per unit of the term’s statistic, theta * p (1 - p), averaged over the dyads, with its delta-method standard error. (ergMargins holds the probabilities fixed in the delta method; ergmx also counts how they change with the parameters.) Curved terms’ decays have no marginal effect.

Returns:

NumericTable – Columns AME, Delta SE, Z and P; .to_frame() for pandas.

odds_ratios(level=0.95)

Odds ratios, exp(coefficient), with Wald confidence intervals: the factor by which a unit increase in a term’s statistic multiplies the conditional odds of a tie.

confint(level=0.95)

Wald confidence intervals of the estimated coefficients, as R’s confint(fit).