Why a Variance Near Zero Needs Careful Interpretation
In an NLME model, a random-effect variance describes the spread of individual-level deviations around the population model. When its posterior distribution concentrates near zero, the data and prior together support little variation on that particular model scale. But “little” is not the same as “none”: the posterior may still assign meaningful probability to larger values, and a continuous prior generally assigns probability zero to the exact boundary itself.
The interpretation is also scale-dependent. A variance can look numerically tiny because the modeled quantity has small units, while the corresponding standard deviation may be easier to interpret. The relevant question is not only whether the variance is close to zero, but how much between-subject variation remains plausible and whether that variation matters for the scientific use of the model.
For a log-normal parameter model, for example, \(\eta_{ik}\) may be a deviation on the log scale. Then \(\omega_k\) is a standard deviation in log units, and individual parameters vary multiplicatively on the original scale. The variance \(\omega_k^2\) has squared log units. Reporting both can prevent a small-looking variance from obscuring the magnitude of the implied individual variation.
What a Near-Zero Posterior Does—and Does Not—Say
A posterior distribution summarizes uncertainty about the variance component after combining the likelihood with the prior. A posterior median close to zero is a summary of location, not a complete description of uncertainty. A credible interval, posterior probabilities below scientifically relevant thresholds, and the implied standard-deviation scale help show whether the posterior is tightly concentrated near zero or has a long upper tail.
With a continuous prior supported on positive variances, an individual point has probability zero. In particular, an interval such as \([0, u]\) may have substantial posterior probability, but that is not the same as positive posterior probability at exactly zero. The distinction follows directly from the posterior density:
Here \(y\) denotes the observed data, and the final equality holds when the posterior has a density with respect to ordinary length on the nonnegative variance scale. The integral over a single point is zero. A model with a separate prior point mass at zero is different: it can assign positive posterior probability to the hypothesis of an exactly absent random effect.
Near-zero values can arise for several reasons. The population variation may genuinely be small; the study may contain limited information to distinguish individual variation from residual error; the design may offer little leverage on the relevant random effect; or the prior may pull estimates toward smaller values. These explanations are not mutually exclusive. Interpretation should therefore connect posterior uncertainty to the model scale, the study design, and reasonable prior choices.
Worked Example: Translating a Small Variance to the Model Scale
Consider a hypothetical model in which \(\eta_{i,\mathrm{CL}}\) is a random effect on log clearance, so that \(\mathrm{CL}_i=\mathrm{CL}_{\mathrm{pop}}\exp(\eta_{i,\mathrm{CL}})\). Suppose posterior draws give a median variance of \(0.0004\) log-clearance units squared and a 95% credible interval of \([0.00002,\,0.0049]\). In these same draws, 8,300 of 10,000 values are below \(0.0025\).
The posterior median standard deviation is the square root of the median variance:
The corresponding standard-deviation interval is obtained by taking the square root of each endpoint, since the square-root function is increasing:
For a log-scale standard deviation of \(0.0200\), a one-standard-deviation multiplicative factor is \(\exp(0.0200)\approx1.020\), or about 2.02% above the population value. At the upper endpoint \(0.0700\), that factor is \(\exp(0.0700)\approx1.073\), or about 7.25% above it. These are scale translations, not guarantees that a particular individual lies exactly one standard deviation from the population value.
If \(0.05\) log units is chosen in advance as a threshold for practically small between-subject standard deviation, the corresponding variance threshold is \(0.05^2=0.0025\). The posterior fraction below it is \(8{,}300/10{,}000=0.830\). Thus, under this model and prior, the posterior probability that the standard deviation is below \(0.05\) log units is approximately 83.0%. The remaining posterior probability is about 17.0%, so the result is not evidence that variation is certainly negligible. The threshold should be justified by the application, not selected merely because it makes the posterior look reassuring.
Assessing Prior Influence and Data Information
A near-zero posterior is informed by both the likelihood and the prior. The prior is especially consequential when the data provide weak information about the variance: many small variance values may fit nearly equally well, or the design may not separate random-effect variation from residual error. More observations per individual can improve estimation of within-individual patterns, but the number of individuals and the design’s ability to reveal between-individual differences also matter. Repeated measurements alone do not guarantee precise estimation of a population variance.
A useful assessment is to compare posterior summaries under defensible prior choices while keeping the likelihood and model structure fixed. Compare the same quantities on the same scale: for example, the posterior median of \(\omega_k\), an interval, and \(P(\omega_k<c\mid y)\) for a scientifically meaningful threshold \(c\). A noticeable shift is a sign that the conclusion depends on prior assumptions. Similar conclusions across plausible priors are reassuring, but do not prove that the model or design is adequate.
Worked Example: Near-Zero Results Under Two Priors
Suppose a hypothetical sparse study includes 18 individuals with two observations per individual. For a random effect with standard deviation \(\omega\), analysts fit the same likelihood twice, using half-normal priors with scales 0.50 and 0.10, respectively. A half-normal prior is supported on nonnegative standard deviations and has density proportional to the corresponding zero-mean normal density on that half-line. The posterior summaries below are hypothetical outputs from the two fits.
| Prior for \(\omega\) | Posterior median | 95% credible interval | \(P(\omega<0.05\mid y)\) |
|---|---|---|---|
| Half-normal scale 0.50 | 0.070 | [0.006, 0.240] | 0.39 |
| Half-normal scale 0.10 | 0.035 | [0.004, 0.110] | 0.67 |
The posterior probability below the chosen threshold changes by \(0.67-0.39=0.28\), or 28 percentage points. Both fits place substantial posterior probability on small standard deviations, but they do not support the same strength of conclusion. The difference is consistent with the data being insufficiently informative to overwhelm the prior. It would be misleading to report only the smaller posterior median and call the variability negligible.
Now suppose a different hypothetical design has 120 individuals, each with eight informative observations. Under the same two priors, assume the posterior probabilities below \(0.05\) are 0.08 and 0.09, and the posterior medians are 0.081 and 0.079. The probability difference is \(0.09-0.08=0.01\), or one percentage point, and the medians differ by \(0.081-0.079=0.002\). These results are less sensitive to the two specified priors than those from the sparse study. This comparison suggests greater data influence for these summaries, but it does not establish that every relevant feature of the model is well identified.
A Short Derivation: Why a Continuous Prior Cannot Prove Exact Zero
Let \(v=\omega_k^2\), and suppose the prior has a density \(\pi(v)\) for \(v>0\), with no point mass at zero. For observed data \(y\), Bayes’ rule gives a posterior density proportional to the likelihood times the prior:
Assume the denominator is finite and positive, as required for this posterior to be defined. The posterior probability of exact zero is the integral of this density over the singleton \(\{0\}\). A singleton has length zero, so that integral is zero. The posterior can nevertheless put high probability in an interval starting at zero, because an interval such as \([0,c]\) has positive length when \(c>0\).
To make exact zero a posterior hypothesis, the model must include it as a distinct possibility. For example, let \(H_0\) denote \(\omega_k^2=0\) and \(H_1\) denote \(\omega_k^2>0\), with prior probabilities \(P(H_0)=0.20\) and \(P(H_1)=0.80\). If the Bayes factor favoring \(H_1\) over \(H_0\) is \(BF_{10}=0.50\), then the prior odds for \(H_0\) versus \(H_1\) are \(0.20/0.80=0.25\). The posterior odds for \(H_0\) versus \(H_1\) are \(0.25/0.50=0.50\), giving:
In this hypothetical point-mass model, the posterior probability of exact zero is 33.3%. That conclusion depends on the model comparison and the prior probability assigned to \(H_0\); it is not obtained by reading an interval from a continuous-prior fit.
Worked Example: Choosing a Practical Threshold
A hypothetical oral-clearance model has a random effect on log clearance. Investigators define a standard deviation of \(0.10\) log units as the largest variation they would regard as practically small for their intended prediction task. Posterior draws show that 9,200 of 10,000 values of \(\omega\) are below \(0.10\).
The corresponding variance threshold is \(0.10^2=0.0100\) log units squared. The estimated posterior probability of being below the standard-deviation threshold is \(9{,}200/10{,}000=0.920\), or 92.0%. The multiplicative factor for a one-standard-deviation log difference at the threshold is \(\exp(0.10)\approx1.105\), approximately a 10.5% increase.
A defensible conclusion is that, under the specified model and prior, the posterior probability that the between-subject standard deviation is below the chosen practical threshold is about 92.0%. This does not mean there is a 92% probability that the variance is exactly zero, nor does it establish that the effect is irrelevant for every prediction or decision. If a different application makes 10.5% variation consequential, the same threshold-based conclusion may not answer the relevant scientific question.
Diagnostics, Alternatives, and Limitations
Before interpreting a near-zero posterior, check that the posterior summaries are trustworthy. Use the chain diagnostics covered in the earlier material on trace plots, rank-normalized Rhat, effective sample size, and Monte Carlo standard errors. In particular, a low effective sample size for the variance or standard deviation can make tail probabilities imprecise. A small Monte Carlo standard error addresses simulation precision, not whether the likelihood or prior provides strong scientific information.
Also inspect the model scale and design. Ask whether the random effect is defined on the scale used for the practical threshold, whether the data contain enough independent individuals, and whether residual error or other model components could account for similar patterns. Sampling pathologies such as divergent transitions can make posterior exploration unreliable, especially near constrained boundaries; a boundary-looking summary should not be treated as a cure for unresolved sampling problems.
Different approaches answer different questions:
| Approach | What it can support | Main caution |
|---|---|---|
| Continuous prior on a variance or standard deviation | Uncertainty over positive values, including probability below a practical threshold | Does not assign positive posterior probability to exact zero |
| Prior with a point mass at zero and a positive-variance alternative | Posterior probability comparing an exactly absent component with a nonzero component | The point-mass prior and model comparison affect that probability |
| Single boundary estimate, such as an estimate reported as zero | A fitted value at the boundary under a particular estimation procedure | Does not summarize the full uncertainty or, by itself, establish no variability |
A posterior near zero is conditional on the model being considered. It does not rule out misspecification, unmodeled sources of heterogeneity, or a different random-effects structure. Nor does it automatically justify removing a random effect: doing so changes the model and may change predictions or uncertainty for other parameters. The most informative report states the posterior on an interpretable scale, gives interval or threshold probabilities, describes relevant prior sensitivity, and explains what the data design can resolve.
Common Pitfall
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A continuous-prior posterior gives \(P(\omega<0.03\mid y)=0.96\). Does this mean the posterior probability that \(\omega=0\) is 96%?
- A posterior median for a variance is \(0.0009\) log-units squared. What is the corresponding standard deviation, and what is its unit?
- Two plausible priors give posterior probabilities 0.42 and 0.71 that \(\omega<0.05\) in a small study. What should this difference prompt the analyst to consider?
- What extra modeling feature is needed to assign positive posterior probability to an exactly zero variance component?
Solutions. (1) No. The 0.96 is probability below \(0.03\), not probability at zero. With a continuous prior and posterior, \(P(\omega=0\mid y)=0\). (2) The standard deviation is \(\sqrt{0.0009}=0.030\) log units; taking the square root converts squared log units back to log units. (3) The 29-percentage-point difference, \(0.71-0.42=0.29\), suggests the conclusion may be prior-sensitive because the data are not strongly determining that threshold probability. The analyst should compare defensible priors and consider what the design can identify. (4) The model must assign prior point mass to zero, typically alongside a separate positive-variance component; a continuous prior alone cannot do so.