Why Perturb Initial Estimates Systematically?
A fitted estimate is produced by applying an estimation procedure to a specified model, data set, and set of fitting controls. The initial estimates are part of that setup. If different plausible initial values lead to meaningfully different final estimates, the result may depend on where the procedure began. A structured perturbation experiment makes that dependence assessable rather than relying on a single alternative start chosen informally.
The purpose is narrow: vary selected initial values while holding the model, observations, estimation target, and other controls fixed. This is a sensitivity exercise about the starting point. It is not, by itself, a test that the model is correct, a diagnosis of why a fit behaves as it does, or proof that one resulting estimate is the unique solution.
Write the parameter vector as \(\psi\), with baseline initial vector \(\psi^{(0)}\). A perturbation scenario \(r\) supplies an initial vector \(\psi^{(r)}_0\). The fitting procedure returns an outcome such as a final estimate \(\hat{\psi}^{(r)}\), an objective value \(J^{(r)}\), and a status record. The superscript identifies the scenario, not a new model. The comparison is interpretable only when the fits target the same objective on the same data under the same settings.
Designing Perturbations That Respect Parameter Scales
A useful perturbation is large enough to test a meaningful range of starts, but not so arbitrary that it mainly probes invalid or implausible values. For a positive parameter such as clearance, a variance, or a residual standard deviation, multiplicative changes are often more coherent than adding a fixed amount. If \(q_0>0\) is the baseline initial value, define a perturbed value as \(q_0\exp(\delta)\). Choosing \(\delta=-\log 2\) and \(\delta=\log 2\) gives one-half and twice the baseline, respectively. This treats equal fold-changes symmetrically on the log scale.
For an unconstrained fixed effect, additive offsets may be suitable if their units and scale make the offsets interpretable. A dimensionless effect or a log-scale effect may instead be varied multiplicatively or by a prespecified offset on its modeled scale. The rule must reflect the parameter's definition, not merely the formatting of the software input.
Constraints require particular care. Variances must remain nonnegative, and many implementations require strictly positive initial variances. Correlations must lie between \(-1\) and \(1\), and the full random-effect covariance matrix \(\Omega\) must be positive semidefinite. Perturbing one raw correlation can make a multivariate matrix invalid even when the changed correlation is itself between \(-1\) and \(1\). Use the model's valid parameterization, or construct a valid covariance matrix for each scenario. For bounded fixed effects, perturb on a suitable unconstrained scale and transform back, or otherwise verify every generated value is admissible.
Two complementary designs are useful. A one-at-a-time design changes one selected component at a time while the other initial values remain at baseline; it helps attribute a difference to a specific perturbation. A joint design changes several selected components together; it can expose sensitivity to combinations that single changes do not probe. A small factorial set, for example low and high values for two parameters, can cover all four combinations. Neither design substitutes for the other, and neither should silently alter unselected initial values.
What to Compare Across Fits
Before fitting, specify which results will be compared and what counts as practically meaningful variation. Useful records include the final estimates, the objective value, the reported convergence status, boundary estimates, and any warnings. Compare each estimate in context and on a scale that is meaningful for that parameter. For a positive parameter, fold-change or log-ratio can be more interpretable than a raw difference; for a parameter in physical units, retain those units in reporting.
For two positive estimates \(\hat q^{(a)}\) and \(\hat q^{(b)}\), a symmetric comparison is the absolute log-ratio:
This quantity is zero for identical estimates and is unchanged if the two fits are swapped. It does not set a universal threshold for importance; define an acceptable difference before inspecting results, based on the parameter's scale and intended use. For objective values, compare values only when they represent the same objective with the same data and conventions. A lower value among two returned fits is not, alone, proof that it is the global optimum.
A short justification explains why the comparison is informative but limited. Let \(A\) denote the fixed data, model, target, and settings, and let \(F_A(\psi_0)\) denote the result returned when the procedure starts at \(\psi_0\). A deterministic procedure with identical controls will return the same result for an identical start. Changing only \(\psi_0\) and observing a changed result therefore establishes dependence of the returned outcome on that starting condition for this setup. Conversely, agreement across tested starts establishes only that the outcomes agree over the tested perturbation set; untested starts or other fitting behavior remain outside that evidence. This is why perturbation results are sensitivity evidence, not a certificate of global optimality or a diagnosis of model adequacy.
Worked Example: Multiplicative Perturbations of Clearance and Volume
Suppose a hypothetical one-compartment model has baseline initial typical clearance \(CL_0=5.0\) L/h and volume \(V_0=50\) L. The analyst prespecifies separate one-at-a-time factors of \(0.5\) and \(2\) for each parameter. All other initial estimates, the data, model, objective, and fitting controls are held fixed.
For clearance, the low and high starts are \(5.0(0.5)=2.5\) L/h and \(5.0(2)=10.0\) L/h. For volume, they are \(50(0.5)=25\) L and \(50(2)=100\) L. Together with the baseline, this gives five scenarios: baseline, low and high clearance, and low and high volume. It does not include a scenario changing both parameters at once.
Assume the hypothetical fits return typical clearance estimates of 4.8, 4.8, 4.9, 4.8, and 4.8 L/h, respectively, and typical volume estimates of 48, 48, 48, 49, and 48 L, in that scenario order. The largest clearance ratio is \(4.9/4.8=1.021\), or about a 2.1% difference. The largest volume ratio is \(49/48=1.021\), also about 2.1%. If the prespecified practical tolerance were 5% for each estimate, these outputs would be within that tolerance for the tested starts.
The conclusion is limited: under this hypothetical setup, the returned typical clearance and volume estimates are similar across the specified one-at-a-time perturbations. It does not establish that all parameter estimates are insensitive, that joint perturbations would behave the same way, or that the model is adequate.
Worked Example: Perturbing a Between-Subject Variance
Consider a hypothetical initial variance for the random effect on clearance, \(\omega^2_{CL,0}=0.16\), on the log-clearance scale. Use half and double the baseline variance as the prespecified alternatives. The starts are \(0.16(0.5)=0.080\), \(0.16\), and \(0.16(2)=0.320\). All are positive, so they satisfy a strict positivity requirement.
Suppose the fitted variance estimates are 0.12, 0.13, and 0.14, and the corresponding objective values to be minimized are 1024.6, 1024.4, and 1024.5. The estimated variance range is \(0.14-0.12=0.020\). Relative to the middle estimate, this range is \(0.020/0.13=0.154\), or about 15.4%. The best objective among these three is 1024.4, which is 0.2 lower than 1024.6 and 0.1 lower than 1024.5.
If a 10% relative range had been specified in advance as the tolerance for this variance, the observed 15.4% range would exceed it. The fits therefore show some sensitivity in this parameter over the tested starts, even though the objective values are numerically close. This result does not identify the cause of the sensitivity, nor does it establish that the lowest reported objective is a global minimum. It signals that the estimate merits careful reporting and further investigation within the modeling workflow.
Joint Perturbations and Reproducible Records
A joint design can be useful when fixed effects and variance parameters may interact in the fitting process. The point is not to change everything indiscriminately. Select a small set of parameters with a clear rationale, define admissible levels for each, and enumerate the combinations before fitting. A complete record makes it possible to tell which parameter values were actually tested and to repeat the exercise.
Worked Example: A Two-Parameter Factorial Set
Suppose the baseline starting typical clearance is 5 L/h and the baseline random-effect variance on log clearance is 0.16. A hypothetical \(2 \times 2\) joint design uses clearance starts of 2.5 and 10 L/h, and variance starts of 0.08 and 0.32. These are all positive, and each value is a half or double of its baseline.
The four scenarios are \((2.5,0.08)\), \((2.5,0.32)\), \((10,0.08)\), and \((10,0.32)\), where the entries are clearance in L/h and variance on the log scale. Suppose the four fitted clearance estimates are 4.9, 4.8, 4.9, and 4.8 L/h. Their range is \(4.9-4.8=0.1\) L/h, or \(0.1/4.85=2.06\%\) of their midpoint. Suppose the fitted variance estimates are 0.12, 0.14, 0.13, and 0.15; their range is \(0.15-0.12=0.03\), and their midpoint is 0.135, giving \(0.03/0.135=22.2\%\).
In this hypothetical set, clearance estimates vary little by the stated summary, while the variance estimates span 22.2% of their midpoint. The design reveals the combined variation across the four tested starts, but this range alone does not separate a clearance-start effect from a variance-start effect or their interaction. The individual scenario results should therefore be retained and reviewed, not replaced by a single range summary.
A reproducible perturbation record should include the model and data version, the baseline initial vector, the parameters selected for perturbation, the transformation and levels used, and the full initial vector for every scenario. Also record the estimation target and relevant fitting controls, software version, run identifiers, and outcome fields. If the procedure includes stochastic elements, record the random-number seed and any settings governing that stochasticity. Reproducibility requires recording realized values, not just a verbal description such as “high start.”
Diagnostics, Alternatives, and Limits
Review outcomes for patterns rather than treating a single status field as the whole result. Differences in estimates, objective values, or boundary behavior may matter even when every run reports a successful termination. Conversely, similar estimates do not show that all parameters are well determined. Check that each scenario used the intended initial vector and the same target and controls; an accidental change to data, parameter bounds, model code, or settings defeats the comparison.
The main design alternatives answer slightly different questions. One-at-a-time perturbations help isolate sensitivity to selected starts but can miss dependence created by jointly changing parameters. A factorial design samples combinations and can reveal joint variation, but the number of runs grows with the number of parameters and levels. A broader multi-start design samples more starting points, but may be less interpretable unless the sampling rule and admissibility constraints are specified. Each is an experiment on starting values, not a replacement for reviewing the model, data, or estimation assumptions.
Common errors include changing the model or fitting controls between scenarios, using additive shifts that create invalid variances, perturbing a correlation without checking the full covariance matrix, and interpreting a small objective difference as proof that estimates are equivalent. Another mistake is declaring the model faulty because starts give different results. The corrected interpretation is narrower: the tested fitting setup returned different outcomes from those starts, and the differences should be quantified and investigated without assigning a cause from this experiment alone.
Check Your Understanding
Answer each question, then compare with the concise solutions below.
- A positive variance starts at 0.25. What are its half and double perturbations, and why is a multiplicative rule useful?
- In a one-at-a-time design, clearance is changed but the initial variance is not. What can this design help assess, and what can it miss?
- Two fits use the same model and data but different estimation targets. Is their objective-value comparison a controlled initial-value comparison? Explain.
- Two positive parameter estimates are 8 and 10 in the same units. Calculate their absolute log-ratio to three significant figures.
- Several tested starts return nearly identical estimates. What conclusion is justified, and what stronger conclusion is not justified?
Solutions. (1) The perturbations are \(0.25(0.5)=0.125\) and \(0.25(2)=0.50\). Multiplicative changes preserve positivity and express a symmetric fold-change. (2) It helps assess whether the returned outcome changes when the clearance start alone changes; it can miss sensitivity to joint changes in clearance and variance. (3) No. Because the target changed, objective values are not from the same controlled comparison. (4) \( |\log(8/10)|=|\log(0.8)|=0.223 \). (5) The tested starts returned similar estimates under the recorded setup. This does not prove insensitivity to every possible start or establish a unique global solution.