Simulate Events to Spot Bad Fits in Relational Event Models

🔍 Why standard fit tests miss the mark

Relational event models (REMs) blend survival analysis with network model terms, a combination that undermines conventional goodness-of-fit diagnostics. Standard methods therefore cannot reliably determine whether a REM is specified well enough to avoid biased parameter estimates.

🧭 A straightforward simulation-based checking procedure

A simple, model-based simulation procedure is presented to evaluate REM fit. The approach uses the model's estimated survival probabilities to generate predicted event timings and actor choices, producing simulated event sequences under the fitted model.

📈 How predictions are produced and used

• Compute survival probabilities from the fitted REM for each potential event.
• Draw events and timings from those probabilities to simulate new event sequences.
• Repeat simulations to produce a distribution of plausible sequences under the model.

📊 What to compare: simulated versus observed sequences

Comparisons between simulated and observed event sequences enable detailed model assessment, including:

• Whether parameter specifications reproduce observed dynamics
• Whether alternative model specifications improve or worsen replication
• Whether the model can replicate key network characteristics that matter for inference

⚖️ Key implication: diagnosing bias risk

If simulated sequences systematically diverge from the observed sequence on relevant network features, the fitted REM is unlikely to provide unbiased estimates. This simulation-based check therefore functions as a practical diagnostic to assess whether model specification suffices for valid inference.