Trace Plots for Joint Model Chains

Why Trace Plots Matter

A joint-model posterior can involve association parameters, baseline hazard terms, and random-effects variance components, all sampled by an MCMC algorithm. A trace plot puts sampled values in iteration order: it can reveal whether a chain appears to be drifting, repeatedly staying at the same value, or visiting only a narrow part of the parameter’s plausible range. These are practical warning signs that a chain may not yet be exploring its target distribution well.

A trace plot is a diagnostic display, not a test that certifies convergence. A plot that looks steady cannot establish that a chain has explored every important region of the posterior. Its value is that it makes some problems visible early, so a posterior summary is not accepted uncritically.

In a shared-random-effects joint model, \(\alpha\) can denote an association coefficient linking a longitudinal feature to the event hazard. Write \(\lambda_j\) for a baseline hazard rate in interval \(j\) of a piecewise-exponential model, and \(\sigma_b^2\) for a random-effects variance component. These parameters have different meanings and scales, so inspect their traces separately. A trace should be labelled with the parameter, chain, scale, and—where relevant—units.

Plot the retained draws in their original order, usually with a different colour or panel for each chain. Do not sort values before plotting: sorting destroys the sequence information that makes a trace useful. If the parameter is constrained, such as \(\lambda_j>0\) or \(\sigma_b^2\geq 0\), its trace must respect that constraint. It can be useful to view a positive parameter on its natural scale and on a transformed scale, but state clearly which is shown. For instance, \(\log(\lambda_j)\) has no rate units, while \(\lambda_j\) is a rate.

What a Trace Should Show

When a chain is sampling a stationary posterior distribution, its values fluctuate around a stable range rather than showing a persistent systematic movement across iterations. A stationary chain does not have to look flat: posterior uncertainty creates variation, and successive draws can be dependent. Nor does every short upward or downward run imply a problem. The relevant concern is a sustained shift, a long plateau, or an apparent failure to visit portions of the range that other chains visit.

A short justification clarifies why drift is concerning. Let \(\pi\) be the posterior distribution targeted by the sampler and \(P\) its transition rule. If the chain has reached stationarity, then \(\pi\) is invariant under \(P\): a draw with distribution \(\pi\), followed by one transition, still has distribution \(\pi\). Thus, for a parameter with finite posterior mean,

The equality does not say that each realized value, or even every short segment average, must be constant. It says there is no systematic change in the marginal distribution from one stationary iteration to the next. A persistent change in the trace is therefore a reason to investigate initialization, warmup, sampler behaviour, or difficult posterior geometry. Conversely, a finite trace can look steady by chance even if the chain has not reached the relevant parts of the posterior. Visual inspection can flag concerns, but it cannot prove stationarity.

Reading Hazard and Variance Traces

Baseline hazard parameters can be difficult to inspect because several interval-specific rates may have different posterior ranges and can be correlated. Plot each \(\lambda_j\) with its interval identified. If a log-rate is sampled, a trace of \(\log(\lambda_j)\) is often easier to display when the rate spans orders of magnitude; exponentiating a plotted value restores the rate on its original scale. A stable trace for one interval does not guarantee that other hazard terms are well explored.

Variance components need particular care because they are nonnegative and their posterior can concentrate near zero or have a long right tail. A chain that spends time near zero is not automatically stuck: that may be a real feature of the posterior. Concern arises when the trace stays in one narrow region for a long time and does not visit other regions that are plausible or visited by other chains. Inspecting both \(\sigma_b^2\) and \(\sigma_b=\sqrt{\sigma_b^2}\) can aid interpretation, but the transformation changes the scale and shape of the displayed distribution.

Comparing Chains and Other Displays

Overlaying traces from multiple chains makes it easier to see whether they visit broadly similar ranges and whether one chain behaves differently. Use distinct colours, clear labels, and a common scale for the same parameter. If chains were initialized at different values, early separation may reflect initialization; focus on the retained portion after warmup, while still watching for later separation or persistent chain-specific patterns. A single-chain trace cannot reveal disagreement between chains.

Trace plots answer a different question from marginal posterior plots. A histogram or density plot shows how often values occur across a chain but discards their order: it can look smooth even when the chain spent a long time stuck and then moved to a different region. A running mean retains some order information, but averaging can hide local sticking and abrupt transitions. Use these displays as complements, not substitutes. Numerical convergence and sampling diagnostics provide additional information beyond visual inspection; a trace alone does not quantify sampling precision or establish agreement.

For a joint model, a practical review should include the association parameter or parameters, each relevant baseline hazard term, and variance components that affect the shared random effects. A sampler can behave differently across parameter blocks. Also inspect important derived quantities if they are central to interpretation, clearly defining their calculation. Keep axes and transformations consistent when comparing chains for a given quantity, and do not infer a parameter’s meaning from the shape of its trace alone.

A Practical Inspection Sequence

Common Pitfalls and Limitations

• Calling a trace “converged” because it looks like a fuzzy band. A chain can remain in one region and produce a visually steady trace while missing another important region. Visual stability is reassuring only in a limited sense and is not proof of convergence.
• Treating every repeated value as an error. Repeated values can result from rejected proposals in a valid sampler. Long plateaus or repeated sticking are warning signs to investigate, not proof of a particular implementation fault.
• Ignoring the parameter scale. A hazard rate, log hazard, variance, and standard deviation have different scales and interpretations. Label the scale and units, especially when comparing plots.
• Reading a marginal plot as if it preserved iteration order. A histogram can conceal when values were sampled. Use the trace to inspect sequence patterns and the marginal display to inspect the distribution’s shape.
• Assuming one well-behaved parameter represents the whole model. Association, hazard, and variance terms may mix differently. Inspect parameters relevant to the scientific interpretation and model structure.

Trace plots are most useful as an early, interpretable screen for problems that can undermine posterior summaries. They are limited by the finite run, subjective visual judgment, and their inability to measure sampling precision. A good-looking trace supports continued evaluation; it does not replace further diagnostics or careful review of the model and computation.