Why Burn-In Requires a Decision
An FCS run begins from an initial set of values for the missing entries and then updates variables through successive iterations. Early iterations may reflect that starting state more than the chain’s later behavior. A burn-in period is the initial segment of iterations that is discarded before summarizing the retained imputations. The practical question is not simply how many iterations to delete, but whether the chain’s behavior gives a defensible reason to regard its early segment as transient.
Burn-in selection is a diagnostic judgment, not a universal numerical rule. Trace plots and iteration summaries can reveal drift or changing variability, but a stable-looking plot cannot prove that the chain has reached a particular target distribution. In FCS, this caution matters especially because conditional models need not define a compatible joint distribution. A burn-in choice cannot repair incompatible or misspecified conditional models; it only addresses evidence that the observed chain is still moving away from its early behavior.
The word “summary” is deliberately broad. A single statistic can look settled while another feature of the imputed data continues to change. Depending on the analysis, relevant summaries can include means, standard deviations, proportions in categories, and parameters from conditional models. Inspecting summaries connected to the variables and quantities that matter helps avoid choosing a cutoff from an uninformative trace.
Assumptions and a Practical Assessment
Interpreting a burn-in as removal of transient behavior presumes that the FCS updates are implemented as intended and that the chain can approach a stable regime under those updates. For an ordinary stationary Markov chain with an appropriate limiting distribution, convergence means that the distribution of the chain’s state approaches that limiting distribution as iterations increase. FCS diagnostics often use the same practical language of a chain “settling,” but apparent stability in a finite run is evidence, not proof, of convergence. If the FCS conditionals are incompatible, apparent stability does not establish a coherent joint distribution.
A useful assessment compares plots and summaries across iteration blocks. A plot can show whether values drift, whether their spread changes, or whether there are abrupt shifts. Block summaries make those patterns easier to describe and compare. If multiple independent chains are available, their traces can also be compared, using the chain-comparison ideas introduced earlier. The objective here is specifically to locate any early segment that behaves differently; it is not to declare convergence from a single visual impression.
Include features that could matter for the imputations or analyses, not just one convenient scalar. For example, inspect both the location and spread of a continuous variable.
Look for directional movement, changing spread, abrupt changes, or distinct early and later patterns. Compare independent chains when available.
Summarize successive blocks for the same quantities. Ask whether the apparent changes are meaningful in context, rather than relying on exact equality between block summaries.
Discard the earliest segment whose behavior is plausibly transient, while retaining enough iterations to assess the later pattern. If important summaries continue to move, the run does not support a confident cutoff yet.
Why a Burn-In Can Reduce Transient Bias
A short derivation clarifies what burn-in can accomplish. Let \(\mu\) be the limiting mean of a scalar summary \(S_t\), when that mean exists. The expected value of the retained average is the average of the retained expectations. Its difference from \(\mu\) is therefore
Thus, a burn-in reduces transient bias only insofar as the retained iterations have expectations closer to \(\mu\) than the discarded early iterations. For example, suppose the deviations satisfy \(\left|E(S_t)-\mu\right|\leq C\rho^t\), where \(C>0\) and \(0<\rho<1\). The triangle inequality and the finite geometric sum give
Under this illustrative decay condition, increasing \(B\) decreases the bound on transient bias. But discarding more iterations also leaves fewer values for estimating summaries. Further, the bound depends on a decay condition that cannot generally be verified from a finite trace. This is why the derivation explains the rationale for burn-in but does not supply a universal cutoff.
Worked Example: A Continuous Summary Settles After an Early Shift
Consider a hypothetical FCS run of 200 iterations for a dataset with missing systolic blood pressure values. For each iteration, the analyst records the mean of the imputed blood pressures, in mmHg. Successive 20-iteration block means are 126.4, 122.1, 120.5, 119.9, 120.2, 119.8, 120.1, 119.9, 120.0, and 119.8 mmHg. Suppose the trace plot shows a gradual decline through iteration 60, followed by variation around approximately 120 mmHg without a visible trend.
The first three block means, for iterations 1–60, average \( (126.4+122.1+120.5)/3=369.0/3=123.0 \) mmHg. The next three, for iterations 61–120, average \( (119.9+120.2+119.8)/3=359.9/3=119.97 \) mmHg. The last four, for iterations 121–200, average \( (120.1+119.9+120.0+119.8)/4=479.8/4=119.95 \) mmHg.
The large difference between the first and later blocks, alongside the trace’s declining pattern, supports treating the early segment as transient. A provisional choice is \(B=60\), retaining iterations 61–200. The later block averages are close, and the plot is described as having no visible trend after iteration 60. These observations support the choice; they do not prove convergence. The analyst should also check other relevant summaries, since stability in this mean alone does not establish stability in spread or other features.
Use More Than One Feature of the Imputed Data
A burn-in decision based only on a mean can miss changes in distribution. A variable can retain a similar mean while its spread changes, or while the proportion of imputations in an important category shifts. The next example shows how examining several summaries can change the assessment.
Worked Example: A Stable Mean Hides Changing Spread
In a hypothetical run, the analyst records the imputed mean and standard deviation of a continuous biomarker, along with the proportion of imputed values above a clinically chosen threshold. The summaries are:
| Iteration block | Mean (units) | Standard deviation (units) | Proportion above threshold |
|---|---|---|---|
| 1–20 | 50.2 | 7.0 | 0.18 |
| 21–40 | 50.1 | 7.2 | 0.20 |
| 41–60 | 49.9 | 8.4 | 0.23 |
| 61–80 | 50.0 | 10.1 | 0.31 |
| 81–100 | 50.1 | 10.0 | 0.30 |
| 101–120 | 49.9 | 9.9 | 0.32 |
The mean varies only from 49.9 to 50.2 units across all six blocks. Its stability alone might suggest no burn-in is needed. However, the standard deviation rises from 7.0 to around 10 units, and the threshold proportion rises from around 0.20 early to around 0.31 late. For the last three blocks, the average standard deviation is \( (10.1+10.0+9.9)/3=30.0/3=10.0 \) units, and the average threshold proportion is \( (0.31+0.30+0.32)/3=0.93/3=0.31 \).
Those summaries are comparatively stable over iterations 61–120, whereas earlier blocks show a change in spread and threshold proportion. A provisional \(B=60\) is therefore better supported than a decision based on the mean alone. The threshold proportion is a proportion, not a measurement in the biomarker’s units. The chosen cutoff remains specific to the observed plots and summaries; a different analysis target could make other summaries important.
Why a Fixed Rule Can Be Inadequate
A rule such as “discard the first 10 iterations” is easy to apply consistently, but it has no universal statistical justification. One FCS run might show little evidence of a transient after a few iterations; another might continue to drift well beyond an arbitrary cutoff. A fixed rule can therefore discard useful iterations in one setting and retain visibly transient iterations in another. It may also give a false sense of security when the retained trace still shows changes.
The opposite mistake is to keep extending the burn-in until a trace looks perfectly flat. Random variation is expected, and exact equality among summaries is neither necessary nor realistic. The relevant question is whether there is meaningful residual structure, such as a systematic trend or sustained change in variability. Visual judgment should be accompanied by numerical summaries and interpreted in the context of the scale and purpose of the variable.
Worked Example: No Defensible Cutoff Within the Run
A hypothetical run records the mean of an imputed outcome, in points, in five successive 20-iteration blocks. The block means are 102.5, 104.0, 105.7, 107.4, and 109.2 points. The trace plot also shows a continuing upward trend through iteration 100.
The difference between the first and last block means is \(109.2-102.5=6.7\) points. The increase does not stop at an obvious boundary: the successive changes are \(104.0-102.5=1.5\), \(105.7-104.0=1.7\), \(107.4-105.7=1.7\), and \(109.2-107.4=1.8\) points. Choosing \(B=20\), for example, would retain blocks whose means continue to rise from 104.0 to 109.2 points.
These summaries do not support a defensible cutoff by iteration 100. The appropriate conclusion is not that a larger burn-in has solved the problem; it is that this run has not shown a stable later segment for this summary. The analyst should investigate the run and obtain additional iterations before making a supported burn-in decision. Discarding the entire run is not a solution if the next segment also continues the same pattern.
Interpretation, Limitations, and Common Mistakes
Burn-in selection is different from deciding how many total iterations to run. Once a cutoff is selected, the number of retained iterations is \(m=T-B\). Increasing \(B\) reduces the retained count at fixed \(T\), which can increase Monte Carlo uncertainty. The retained draws can also be autocorrelated, so their count is not generally the same as the number of independent draws. The earlier material on Monte Carlo error in FCS chain summaries explains why uncertainty depends on both the retained length and dependence across iterations.
Comparing plots and block summaries is a practical diagnostic, not a formal test that certifies convergence. A trace may conceal poor exploration, and a chain may appear stable in one summary while changing in another. Nor does burn-in remove bias from inappropriate conditional models or establish compatibility. These are distinct issues; discarding early iterations cannot repair them.
- Mistake: use the same burn-in for every run. A fixed number ignores the observed behavior of the chain. Correction: choose a cutoff in light of traces and summaries, and report the evidence for it.
- Mistake: judge only the mean. A stable mean does not imply stable spread or category proportions. Correction: inspect multiple summaries relevant to the imputed variables and analysis.
- Mistake: claim that a flat-looking trace proves convergence. A finite diagnostic plot cannot establish convergence or compatibility. Correction: describe the observed evidence and its limits.
- Mistake: treat a longer burn-in as a cure for continuing drift. If the retained segment still trends, the cutoff is not supported. Correction: acknowledge that the run has not demonstrated a stable segment and investigate further.
Check Your Understanding
Answer each question using the definitions and diagnostic logic above. Solutions follow the questions.
- A 120-iteration chain has a proposed burn-in of 30 iterations. Which iterations are retained, and how many are retained?
- Block means of an imputed variable are 8.2, 8.1, 8.2, 8.9, 9.0, and 8.9 across six equal blocks. What feature should prompt caution before choosing a burn-in?
- Why might a stable trace of an imputed mean be insufficient evidence for selecting a cutoff?
- A researcher discards 50 iterations because that number is used in every analysis, although their trace shows a continuing trend after iteration 50. Is this cutoff supported? Explain.
- Under the geometric-deviation condition in the derivation, what happens to the displayed upper bound on transient bias as \(B\) increases, with \(m\) held fixed?
Solutions. (1) Retain iterations \(31\) through \(120\), for \(120-30=90\) iterations. (2) The first three means are near 8.2, but the later blocks rise to about 8.9–9.0. This shift suggests a changing chain summary, so the proposed retained segment and any cutoff need to be assessed against the trace and additional summaries. (3) Spread, category proportions, or other relevant features may still be changing even when the mean appears stable. (4) No. A continuing trend after iteration 50 indicates that the retained segment still has systematic change for that summary; the fixed rule does not justify the cutoff. (5) The bound decreases because it contains the factor \(\rho^{B+1}\), with \(0<\rho<1\). This is a mathematical consequence of the assumed geometric decay, not proof that an observed FCS run satisfies that condition.