Why an Objective-Function Difference Needs Context
A run comparison table may show that one model run has a lower objective function than another. That is a numerical observation, not yet an interpretation. Before treating it as an improvement, check what the objective measures, whether both runs evaluated the same likelihood on the same observations, and whether the reported difference is larger than any relevant numerical uncertainty.
In pharmacometric estimation, a commonly reported objective function is the negative two times the log likelihood. Other programs or workflows may report a log likelihood, a negative log likelihood, a penalized objective, or another criterion. Labels alone do not guarantee that two values have the same meaning or scale. The exact definition used by each run must be established from its recorded settings and output.
The subtraction order must be stated. Some tables instead report \(OFV_A - OFV_B\), which reverses the sign. Record the convention in the table or report rather than relying on the phrase “the OFV decreased.” Also preserve the full-precision values available from the run output: displayed values may be rounded.
When a Lower Value Means a Higher Likelihood
Suppose both runs report \(OFV=-2\log L\), where \(L\) is the likelihood evaluated for the same observed data under the same likelihood definition. Let \(L_A\) and \(L_B\) be the likelihood values for runs A and B. Then the difference has a direct relationship to the likelihood ratio.
Because the logarithm is increasing, \(\Delta OFV < 0\) implies \(\log(L_B/L_A) > 0\), and therefore \(L_B/L_A > 1\). Under this specific convention and these comparability conditions, run B has the higher likelihood at the evaluated estimates. The magnitude can also be translated: \(L_B/L_A=\exp(-\Delta OFV/2)\). This describes a ratio of likelihood values; it does not, by itself, establish statistical significance or prove that one model is scientifically preferable.
The comparison is meaningful when both runs use the same observations and response definitions, the same likelihood formulation and scale, and compatible treatment of missing records, weights, transformations, and constants. The fitted parameter values may differ—that is usually the point of comparing estimated runs—but the target likelihood must be the same. If the objective includes a penalty, prior contribution, or approximation beyond \(-2\log L\), the likelihood-ratio interpretation may not apply.
Worked Example: Interpreting a Difference Under the Same Likelihood
Two hypothetical runs use the same data and likelihood, and both report \(OFV=-2\log L\). Run A reports 1240.6 and run B reports 1234.2. Calculate and interpret the change.
Difference. Using the stated B-minus-A convention, \(\Delta OFV=1234.2-1240.6=-6.4\). The negative sign means B has the lower objective.
Likelihood ratio. Since the objective is negative twice the log likelihood, \(L_B/L_A=\exp(-\Delta OFV/2)=\exp(3.2)\approx24.53\). Thus, at the estimates represented by these outputs, B's likelihood is about 24.5 times A's likelihood, provided the runs truly share the stated likelihood definition and data.
This calculation does not supply a p-value, a confidence interval, or a general measure of model adequacy. It reports the likelihood ratio implied by these two objective values. Whether a formal inferential comparison is appropriate depends on additional conditions not established by the arithmetic alone.
Check the Scale and Sign Convention
Never infer the objective's scale just from the magnitude of a number. Confirm the definition in the run output, configuration, or analysis record. For example, if one output gives \(-2\log L\) and another gives \(-\log L\), the values differ by a factor of two even when they refer to the same likelihood. A smaller number may still correspond to a better likelihood under both conventions, but the numerical difference is not directly interchangeable.
Likewise, if an output reports \(\log L\), larger values correspond to higher likelihood; if it reports \(-\log L\) or \(-2\log L\), smaller values correspond to higher likelihood, assuming the same likelihood. A penalized objective can have a different interpretation because a change may reflect both the likelihood contribution and the penalty. Inspect the calculation rather than assuming every field named “objective” is a raw likelihood measure.
Worked Example: Converting Between Two Objective Scales
Two hypothetical reports concern the same data and likelihood. Run A reports \(-2\log L=1234.6\), while run B reports \(-\log L=614.1\). Put the values on the same scale before comparing them.
Convert B. Since \(OFV_B=-2\log L_B\), multiply its reported negative log likelihood by two: \(OFV_B=2(614.1)=1228.2\).
Compare. On the \(-2\log L\) scale, \(\Delta OFV=1228.2-1234.6=-6.4\). This is a lower objective for B and, subject to the shared likelihood and data conditions, corresponds to \(L_B/L_A=\exp(3.2)\approx24.53\).
Comparing 614.1 directly with 1234.6 would mix scales and produce a meaningless difference. If the reports do not define their scales clearly enough to perform a justified conversion, the correct conclusion is that the objective values are not directly comparable.
Verify That the Likelihood Has Not Changed
A lower objective does not establish improvement if the runs evaluate different likelihoods. Changes that can alter the likelihood include changing the included records, response variable, observation units, residual-error model, weights, censoring treatment, or data transformations. Some changes also alter data-dependent constants in a reported objective. Even if each run is internally valid, subtracting their values may not isolate a difference in fit.
A shared data file alone is not enough: the analysis may select different rows or use different likelihood contributions. Conversely, a change in an estimation setting does not automatically make the likelihood incomparable; the key question is whether the reported objective still evaluates the same likelihood on the same observations and scale. Document what changed, and verify the effective records and objective definition rather than relying on run labels.
Worked Example: A Lower Value After Changing Included Records
Two hypothetical runs report objectives of 1200.4 and 1189.7 on a nominal \(-2\log L\) scale. In run B, however, three observations were excluded after a data-processing change. What can be concluded?
The arithmetic difference is \(1189.7-1200.4=-10.7\). It is true that B's reported number is 10.7 lower. But the runs did not evaluate the same set of observations, so the difference does not establish that B has a higher likelihood for the same analysis. The changed record set changes the target likelihood.
A defensible comparison would first restore a common observation set and consistent likelihood settings, then evaluate both runs on that common basis. If exclusion is scientifically justified, it should be reported as a change in analysis, not presented as an objective improvement directly comparable to the original run. The number \(-10.7\) remains a descriptive subtraction, but it does not support the same-likelihood interpretation.
Numerical Precision, Optimization, and Diagnostics
Objective values are computed numerically. The last digits may vary because of printed rounding, numerical integration, approximation, or the optimizer stopping at different points. A small difference should therefore be evaluated against the precision of the stored output and evidence that both estimates are adequately optimized. A lower value from a run that stopped early or failed a convergence check is not automatically a reliable improvement.
Check the full-precision objective when available, the run status and termination message, and whether the reported value is the final value for the intended estimation. Investigate warnings, failed evaluations, or unstable objective values. If repeated runs of the same model and data produce differences of similar size, the observed change may be within run-to-run numerical variability. Do not create a threshold for “meaningful” change without a justified method-specific basis.
Worked Example: A Difference Near the Printed Precision
Two hypothetical outputs display objectives to three decimal places: run A is 981.247 and run B is 981.246. The displayed difference is \(981.246-981.247=-0.001\). How strong is this numerical conclusion?
The displayed values show a decrease of 0.001, but each value has been rounded to the nearest 0.001. The underlying value for A may lie approximately from 981.2465 to 981.2475, and the value for B from 981.2455 to 981.2465. Therefore, the underlying B-minus-A difference may range approximately from \(-0.002\) to \(0\), with endpoint details depending on the rounding rule.
The displayed values alone do not establish a strictly negative difference at full precision. Retrieve the unrounded values and check convergence and numerical stability before describing the change as a genuine objective decrease. Even if the full-precision difference is negative, its size should be reported as a numerical result, not automatically as evidence of an important or statistically significant improvement.
A practical interpretation sequence is to establish the objective definition and subtraction direction, verify that data and likelihood are comparable, calculate the signed difference from unrounded values, and assess numerical reliability. If any check fails, state exactly what is known and what cannot be inferred rather than forcing a model ranking.
Record whether each value is \(-2\log L\), \(-\log L\), \(\log L\), or a penalized or otherwise modified objective.
Check the effective observations, response definitions, likelihood formulation, weighting, transformations, and any additive terms that may differ.
For this tutorial's convention, compute \(OFV_B-OFV_A\) using the most precise available values.
Review rounding, termination, warnings, and numerical stability; report a likelihood interpretation only when its assumptions hold.
Common Pitfall and Appropriate Alternatives
Wrong reasoning: “Run B has a smaller objective, so it is a better model.” This skips both the definition of the objective and whether the runs evaluated the same likelihood. It can also confuse a numerical optimizer result with a scientific conclusion.
Correction: Say first what the quantity is and whether the runs are comparable. If both values are \(-2\log L\) for the same data and likelihood, a lower value means a higher likelihood at the evaluated estimates. If those conditions do not hold, report the values as non-comparable or describe the specific change that prevents interpretation.
When objective labels or scales differ, a valid alternative is to recover the underlying likelihoods or recompute a common, clearly defined objective at both sets of estimates, if the model and data permit it. When the full-precision values are close, use the stored precision and convergence diagnostics rather than the rounded table alone. These checks clarify the numerical comparison; they do not replace a separate method for formal inference or model selection.
Check Your Understanding
Answer each question by stating the relevant condition, calculation, or limitation.
- Run A has \(OFV=850.0\) and run B has \(OFV=846.0\), with both values defined as \(-2\log L\) for the same data and likelihood. What is \(OFV_B-OFV_A\), and what does its sign mean?
- One run reports \(-2\log L=700\), while another reports \(-\log L=345\). What conversion is needed before comparing the objectives?
- Two runs use the same objective convention, but one excludes observations that the other includes. Can the difference be interpreted as a likelihood improvement on the same analysis?
- Why is a displayed difference of \(-0.001\) potentially inconclusive when both objective values are rounded to three decimal places?
- If a lower objective includes a penalty term that differs between runs, what must be checked before interpreting the change as a higher likelihood?
Solutions. (1) \(846.0-850.0=-4.0\); run B has the lower objective and, under the stated conditions, the higher likelihood at the evaluated estimates. (2) Multiply the \(-\log L\) value by two, giving 690 on the \(-2\log L\) scale, then compare only if data and likelihood definitions also match. (3) No; the runs evaluate likelihoods on different observation sets, so the subtraction does not isolate a change for the same analysis. (4) Rounding uncertainty can be comparable to the difference, so the unrounded values may not establish a strictly negative change. (5) Determine whether the objective difference arises from the likelihood contribution, the penalty, or both; a penalized-objective decrease alone does not establish a higher likelihood.