Why Look at a Trace Plot?
An MCMC output is a sequence, not an unordered collection of estimates. For an NLME model, the sequence may include population parameters such as typical clearance, random-effect variance components, residual-error parameters, and derived quantities. A trace plot places each sampled value against its iteration number, preserving the order in which the sampler produced it. That order can reveal problems that a single average or histogram conceals.
A trace plot is a diagnostic view, not a test that certifies convergence. A chain that appears stable may still explore the posterior poorly, while an unusual-looking trace may reflect genuine posterior structure rather than a computational defect. Interpretation therefore begins with the model parameter, the scale on which it is plotted, the chain’s initialization and warmup history, and comparison across independently initialized chains.
What Patterns to Look For
A useful trace often looks like a dense, irregular band without a persistent direction, long periods of immobility, or chain-specific regions. This description is a visual target, not a formal definition of convergence. MCMC draws are dependent, and no finite trace can demonstrate that a chain has explored every relevant part of its target distribution.
- Drift: the trace moves systematically upward or downward over an extended portion of the retained iterations. It can indicate that the chain has not settled into a stable region, but a trend may also occur by chance in a finite dependent sequence.
- Sticking: the trace remains at, or very near, the same value for a long run of iterations and then changes abruptly or rarely. This can indicate poor exploration. It can also arise when a parameter is genuinely concentrated near a boundary, so the pattern needs context.
- Multimodality: the trace visits separated regions with relatively few transitions between them, or different chains remain in distinct regions. The posterior may have multiple modes, or the chains may not be traversing the target adequately. A trace alone cannot distinguish these explanations.
- Chain disagreement: chains occupy visibly different ranges or have different patterns. This is a warning that they may not be sampling the same posterior distribution adequately, even if each individual trace looks stable.
NLME parameters require attention to scale and interpretation. A population parameter and a variance component can have very different numerical ranges. Variance components are nonnegative, and a model may parameterize them as strictly positive; a zero variance is nevertheless a meaningful boundary case in which the corresponding random effect has no between-individual variation. On the natural scale, a variance trace can be visually compressed near zero. Plotting a transformed scale can clarify variation, but transformations change the visual impression and cannot be applied blindly at a boundary.
A Simple Check Using Blocks
When a trace seems to drift, dividing the retained sequence into consecutive blocks and comparing their means can make the visual impression more concrete. This is a descriptive check, not an independent-sample test: sequential MCMC draws are generally dependent, and arbitrary block means do not provide a convergence guarantee.
Worked Example: Checking an Apparent Upward Drift
Suppose one post-warmup chain for the typical clearance parameter, measured in L/h, has the following eight consecutive draws, shown only to illustrate the calculation:
The first four-draw block has mean \((1.8+1.9+2.0+2.1)/4=7.8/4=1.95\) L/h. The second block has mean \((2.4+2.5+2.6+2.7)/4=10.2/4=2.55\) L/h. The second block mean is \(2.55-1.95=0.60\) L/h higher.
The sequence and block means both suggest an upward trend over these iterations. They do not establish that the posterior mean is changing, nor do eight draws establish a general pattern. The appropriate response is to inspect the full retained trace and other chains, confirm that warmup was excluded correctly, and investigate possible sampling or model difficulties if the trend persists.
Worked NLME Trace Interpretations
Worked Example: A Variance Component That Sticks Near Zero
Consider a hypothetical trace for the interindividual variance of log clearance, denoted \(\omega^2_{\mathrm{CL}}\), on its natural scale. In 12 retained iterations, suppose the values are \(0,0,0,0.001,0,0,0,0,0.002,0,0,0\). The units are squared log-clearance units, since the random effect is on the log scale.
Ten of these 12 values are exactly zero, and the other two are \(0.001\) and \(0.002\). The trace therefore spends long stretches at zero, with only small departures. This is a sticking-like pattern and deserves investigation, but it is not enough to diagnose a sampler failure. The posterior may place substantial mass at or very near the zero-variance boundary, or the chain may be moving inadequately.
First check whether zero is permitted by the model representation and whether reported zeros are exact values or rounding. Then compare the pattern across independently initialized chains and inspect other relevant output. If a strictly positive parameterization was used, exact zeros in a displayed trace may instead reflect rounding or a transformation/reporting issue. A log-scale plot cannot display an exact zero, so it must not be used to hide or silently discard boundary draws.
Worked Example: Chains Occupying Different Regions
Suppose two chains for a population absorption-rate parameter \(k_a\), measured in h\(^{-1}\), have these summaries of their trace ranges after warmup. Chain 1 fluctuates between \(0.42\) and \(0.58\) h\(^{-1}\), while chain 2 fluctuates between \(1.35\) and \(1.62\) h\(^{-1}\). Each path appears roughly level within its own range, with no visible transitions to the other range.
The ranges do not overlap: \(0.58 < 1.35\) h\(^{-1}\). Thus, the chains visibly disagree about the region being explored. Each chain’s local stability is not sufficient reassurance; different chains may be trapped in separate posterior regions or may not yet have reached the same target behavior.
This observation is a warning, not proof of two posterior modes. Inspect longer traces, confirm that the same parameter definition and units were plotted for both chains, and consider whether the model has weakly identified or multiple plausible regions. A pooled average would be particularly misleading here: it could fall between the two observed ranges, where neither chain spent time.
Worked Example: Distinguishing a Jump from a Sustained Trend
For a hypothetical population log-clearance parameter, suppose a trace fluctuates around \(0.7\) for many iterations, makes one short excursion to \(1.1\), and then returns to fluctuate around \(0.7\). The parameter is dimensionless because it is on the log scale.
A single excursion is not, by itself, evidence of sustained drift or multimodality. A relevant question is whether the trace repeatedly visits separated regions and whether chains show compatible behavior. If the value \(1.1\) occurs only briefly, it could be an ordinary tail draw; if distinct regions recur with few transitions, the possibility of multimodality or poor exploration warrants closer examination. The trace supplies this time-order information, whereas a histogram alone would not show when the excursion occurred.
Practical Reading Sequence and Common Pitfalls
Use a consistent sequence of checks rather than judging a trace from one striking feature. Plot chains separately or use clearly distinguishable colors, label parameters and units, and include enough of the post-warmup iterations to see both local and long-range behavior. For a model with many parameters, inspect important population parameters and variance components as well as quantities central to the scientific interpretation. A good-looking trace for one parameter does not guarantee good exploration of all parameters.
Check the parameter definition, units, chain identity, iteration range, and whether the plotted values are post-warmup draws.
Look for persistent trends, long flat stretches, separated regions, or abrupt changes, while remembering that finite sequences can look uneven by chance.
Ask whether independently initialized chains explore compatible ranges and show similar patterns, rather than relying on a pooled trace that obscures chain identity.
Check parameterization, boundary behavior, model geometry, and additional numerical diagnostics. Do not declare convergence or failure from the plot alone.
Two mistakes are especially common. First, a flat trace is not automatically a good trace: it may show a chain stuck at one value. Second, a trace that crosses a broad range is not automatically adequate: movement can coexist with trends, chain disagreement, or failure to explore another important region. Conversely, occasional spikes should not be removed just because they look unusual; they may be valid posterior draws.
Trace plots complement rather than replace numerical summaries and other diagnostics. Histograms summarize the values but erase their order; traces preserve the order but can be difficult to judge quantitatively, especially for subtle chain differences. Later numerical diagnostics address aspects of chain agreement and sampling efficiency more formally, but they also require interpretation and cannot repair an inadequately explored posterior. When visual warning patterns persist, investigate the computation and model rather than treating a summary statistic as a substitute for understanding the trace.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A post-warmup trace for typical clearance rises steadily across several long blocks. What does this pattern suggest, and what can it not establish by itself?
- Why can a variance component trace at zero not automatically be called an error?
- Two chains each look stable, but their ranges do not overlap. Why is pooling their draws into one mean potentially misleading?
- What assumption is needed to claim that a stationary chain’s block average has expectation \(\mathbb{E}_{\pi}[\theta]\)?
- What information does a trace plot retain that a histogram discards?
Solutions. (1) The sustained increase suggests possible drift or inadequate settling and should prompt inspection of the full trace and other chains. It does not prove nonstationarity, convergence failure, or a specific cause. (2) Variance components are nonnegative, and zero is a meaningful boundary corresponding to no variation for that random effect; the model’s parameterization and possible rounding should be checked. (3) The pooled mean may lie in a region neither chain explored, masking disagreement and possibly summarizing neither chain’s sampled region. (4) Require a stationary chain with \(\mathbb{E}_{\pi}[|\theta|]<\infty\); then each draw has the target mean and linearity of expectation gives the block-average result. (5) A trace retains the order of draws across iterations, making trends, long stays, transitions, and chain-specific patterns visible.