Why Starting Values Matter
A GLMM fit is obtained by numerically optimizing an objective associated with the model likelihood. Numerical optimization begins from an initial point, called a starting value. Depending on the model and fitting implementation, that point may include initial values for fixed effects and for parameters governing the random-effects distribution. The algorithm then takes a sequence of steps intended to reach a stationary solution.
Starting values are not additional data, prior information, or assumptions about the true parameter values. They are computational inputs. They can nevertheless affect a fit: the objective may be difficult to optimize, its surface may have flat or poorly curved regions, or an algorithm may stop at different solutions when initialized differently. A restart means fitting the same model again from a deliberately changed initial point.
For a GLMM with fixed-effect parameters \(\beta\), random effects \(b\), and random-effects parameters \(\theta\), the marginal likelihood has the form
Here \(f\) describes the conditional response model and \(\phi\) the random-effects distribution. In general, the integral must be evaluated numerically or approximated. A fitting procedure therefore optimizes a computed objective, such as the negative log likelihood or its approximation. Restart comparisons are meaningful only when runs evaluate the same objective for the same model.
A Controlled Restart Protocol
The purpose of a restart is not to search randomly until a preferred answer appears. It is to test whether the fitting result is robust to reasonable changes in initialization. Before fitting, specify which starting values will vary, how they will vary, and what quantities will be compared. A practical protocol is:
Keep the response, predictors, random-effects structure, link, likelihood approximation, optimizer, and control settings unchanged. This makes the starting point the intended source of variation.
Use the software's default or standard starting values as a reference, then specify a small set of distinct, finite starting values within the parameter domain. Avoid extreme values that cause numerical overflow or are not valid for the parameterization.
Save the initial values, convergence status, objective value, parameter estimates, and any diagnostic messages for every run. A warning or failed run is part of the evidence, not a result to omit.
Among runs that have reached an acceptable fitted solution, compare objective values and estimates on the same scale. Check whether substantive conclusions depend on which start was used.
The word acceptable matters: an optimizer's success flag alone does not establish that a fit is trustworthy. The earlier material on convergence messages, scaled gradients, Hessian checks, and ill-conditioned fits provides complementary checks. Restart agreement is an additional reproducibility check, not a replacement for those diagnostics.
What Agreement Across Starts Shows
Suppose a fitting routine minimizes the negative log likelihood \(Q(\beta,\theta)=-\ell(\beta,\theta)\). Smaller values of \(Q\) correspond to larger likelihood values. If two runs return \(Q_1\) and \(Q_2\), then
This identity gives a direct way to interpret an objective difference: if \(Q_1=Q_2\), the two runs have equal likelihood values to the reported precision; if \(Q_1<Q_2\), run 1 has the larger likelihood value. It does not establish that either run found the global optimum. In numerical work, also check that objective values are calculated using the same approximation and are reported with enough precision to distinguish meaningful differences.
Parameter estimates matter alongside the objective. Nearly identical objective values can occur with noticeably different estimates when the objective is relatively flat in some direction. Conversely, slightly different objective values can accompany estimates that are practically indistinguishable at the precision relevant to the analysis. Compare estimates on their natural scales and in context, and do not rely on an objective value rounded too aggressively.
Worked Example: Nearly Identical Solutions From Four Starts
Suppose a hypothetical data set records a binary outcome for 200 patients in 20 clinics, with 73 outcomes coded 1. The prespecified model is a logistic random-intercept GLMM with no additional predictors. Using the same optimizer and fitting settings, four starting values for the intercept and random-intercept standard deviation give the following results. The standard deviation is in log-odds units.
| Run | Initial intercept | Initial standard deviation | Negative log likelihood | Fitted intercept | Fitted standard deviation |
|---|---|---|---|---|---|
| 1 | -0.50 | 0.20 | 146.3821 | -0.57 | 0.31 |
| 2 | 0.00 | 0.50 | 146.3824 | -0.57 | 0.31 |
| 3 | -1.00 | 0.80 | 146.3822 | -0.57 | 0.31 |
| 4 | 0.50 | 1.00 | 146.3823 | -0.57 | 0.31 |
All four fits returned the same displayed estimates and nearly identical objective values. The largest objective difference is \(146.3824-146.3821=0.0003\). The likelihood ratio between the best and worst displayed objective values is
Thus, the displayed likelihood values are practically indistinguishable, and the reported parameter estimates are reproducible across these four starts. This is reassuring evidence about sensitivity to these initial values. It is not proof that no other solution exists or that all model assumptions are appropriate.
When Restarts Disagree
Disagreement needs investigation rather than selective reporting. A lower negative log likelihood is preferable for the same objective, but a run with a lower objective is not automatically trustworthy: it may have failed a diagnostic, evaluated a different approximation, or been affected by a numerical problem. Compare the full record of each run, including warnings and estimates.
Worked Example: Two Distinct Fitted Solutions
Consider hypothetical binary outcomes for 96 patients in 12 clinics, with 31 outcomes coded 1. Six clinics receive a treatment and six do not. The same prespecified logistic random-intercept model and fitting settings are used in both runs. One start is close to an intercept-only fit; another begins with a more negative treatment coefficient and a larger random-intercept standard deviation.
| Run | Initial treatment coefficient | Initial standard deviation | Negative log likelihood | Fitted treatment coefficient | Fitted standard deviation |
|---|---|---|---|---|---|
| A | 0.00 | 0.20 | 58.72 | -0.18 | 0.22 |
| B | -1.50 | 0.90 | 58.19 | -1.42 | 0.71 |
Run B has the smaller negative log likelihood, by \(58.72-58.19=0.53\). The likelihood ratio is
For this computed objective, run B has a likelihood about 1.70 times the likelihood from run A. The fitted treatment coefficients also differ substantially: \(-1.42\) versus \(-0.18\) log-odds units. This is initialization sensitivity that must be resolved before interpreting a single reported fit. The comparison alone does not establish which solution is the global optimum. First verify that both runs used the same model and objective, then review their convergence evidence and attempt additional planned starts. If the solutions remain different, report the instability rather than presenting one result as settled.
A useful justification for using more than one start is that optimization follows a path determined partly by its initial point. If some starting values lead into different regions of the objective surface, repeated fitting can reveal that dependence. But the proportion of successful runs is not an estimate of the probability that the best solution is global: the starts are chosen rather than sampled from a known distribution over all possible basins, and some regions may be difficult to reach.
Planning Starts and Recording Results
A restart plan should be modest, deliberate, and interpretable. It is often useful to include the default start, plausible alternatives for fixed effects, and several valid initial values for random-effects parameters. Vary a manageable number of parameters at a time when the goal is to understand which initialization may be affecting the result. If a model-fitting function does not expose some internal starting values, do not imply that you varied them; document what the software allowed you to control.
Worked Example: A Compact, Documented Restart Set
A hypothetical study has 150 binary observations from 10 wards, with 44 outcomes coded 1. The planned model is a logistic random-intercept GLMM with one continuous predictor that has already been scaled. The analyst holds the model, optimizer, approximation, and controls fixed and plans six starts. The initial values are:
- Reference: intercept \(-0.25\), predictor coefficient \(0\), random-intercept standard deviation \(0.30\).
- Alternative 1: intercept \(-0.80\), predictor coefficient \(0.50\), standard deviation \(0.50\).
- Alternative 2: intercept \(0.20\), predictor coefficient \(-0.50\), standard deviation \(0.80\).
- Alternatives 3–5: use the reference fixed-effect values and standard deviations \(0.10\), \(0.60\), and \(1.00\), respectively.
Suppose all six fits return a negative log likelihood of \(91.307\) to three decimal places, and the fitted predictor coefficients range from \(0.412\) to \(0.415\). The displayed objective spread is \(91.307-91.307=0.000\), and the coefficient spread is \(0.415-0.412=0.003\) log-odds units per one-unit increase in the scaled predictor. This pattern supports reproducibility across the stated starts, subject to the precision reported and the other fit diagnostics. The result should be recorded with the six initial-value sets and the full-precision output, not only the rounded summary.
A useful run log includes the model formula, data version, software and fitting settings, the exact starting values, convergence status, objective value at full available precision, estimates, and diagnostic messages. This record allows another analyst to distinguish a genuine controlled restart from a run in which other choices changed as well.
Diagnostics, Alternatives, and Limitations
Restarts are most informative when interpreted alongside other evidence. Warning signs include materially different estimates across starts; a persistently higher objective from one or more apparently converged runs; repeated nonconvergence from a particular region; or objective values that change when only the start changes, without a clear explanation. Check whether estimates and objective values refer to the same parameterization and likelihood approximation. The earlier material on scaling predictors and centering predictors and coding factors is relevant when poor numerical behavior may be connected to the predictor representation.
A single default start is fast and may be adequate for an easy, well-behaved fit, but provides little evidence about initialization sensitivity. Controlled restarts offer a direct check while preserving the fitting method. By contrast, changing the optimizer changes the algorithm as well as potentially its path; that is a different diagnostic question. Changing the model or likelihood approximation also prevents a clean attribution of differences to starting values. Restarting is therefore complementary to, not a substitute for, those alternatives.
Several limitations should remain explicit:
- Agreement across a finite set of starts does not prove global optimality; all starts may reach the same suboptimal region.
- Similar objective values do not guarantee similar estimates if the objective is flat in a parameter direction.
- A lower objective does not excuse failed diagnostics or establish that the model is scientifically appropriate.
- Results depend on the numerical objective actually evaluated. Approximation error can remain even when optimization is reproducible.
- Initial values that are invalid, extreme, or poorly chosen can produce failures that reveal little about the model's attainable solutions.
A common mistake is to choose whichever restart gives the preferred coefficient or the smallest objective and then call it the answer. The correction is to predefine a sensible restart strategy, compare all runs on the same objective, investigate discrepancies, and report unresolved sensitivity. Restarts are a diagnostic tool for reproducibility, not a method for selecting a desired scientific conclusion.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In a controlled restart, which fitting choices should remain fixed, and which choice is intentionally changed?
- Two fits have negative log likelihoods \(42.10\) and \(42.35\). What is the likelihood ratio for the fit with the lower negative log likelihood relative to the other?
- Three starts produce nearly identical objective values but treatment coefficients of \(-0.2\), \(-0.8\), and \(-1.4\). What does this pattern suggest?
- Why does convergence to the same solution from eight starts not prove global optimality?
- An analyst changes the optimizer and starting values between two runs. Why can the difference not be attributed specifically to starting values?
Solutions. (1) Keep the data, model, likelihood approximation, optimizer, parameterization, and controls fixed; intentionally change the specified initial values. (2) The ratio is \(\exp(42.35-42.10)=\exp(0.25)\approx1.284\), so the fit with objective \(42.10\) has about 1.284 times the likelihood of the other fit under the same objective. (3) The similar objective values but widely different coefficients suggest that the estimates may be poorly determined along a relatively flat direction, or that the runs have reached distinct solutions. Review diagnostics and investigate rather than selecting one coefficient. (4) All eight starts could reach the same suboptimal region; a finite collection of starts cannot rule out other attainable solutions. (5) Both the algorithm and the initialization changed, so the observed difference has more than one possible computational cause.