Forward Addition With Conditional Covariate Effects

Why Forward Addition Is Conditional

A covariate can appear important when tested alone, yet contribute little after another covariate has entered the model. Conversely, a covariate whose initial test is modest can become informative once the current model accounts for a correlated covariate. Forward addition handles this by testing each candidate against the current model, not repeatedly against the original base model.

The result is a sequence of nested models and decisions. At every step, the test asks whether adding one candidate improves fit enough, conditional on the effects already included. The search path is therefore part of the result: it shows which covariates entered, in what order, and which remaining candidates failed the stopping rule at each stage.

Test Each Addition Against the Current Model

For a candidate covariate \(z\), let \(M\) be the current model and \(M+z\) the model that adds its effect. These models must be nested: \(M\) is obtained from \(M+z\) by setting the added effect parameter or parameters to their null values. Define the objective function as \(\mathrm{OFV}=-2\ell\), where \(\ell\) is the maximized log likelihood. The likelihood-ratio statistic is the reduction in objective function:

Because the larger model can reproduce the smaller one under the null, its maximized likelihood cannot be lower in an exact optimization. Thus \(D_z\) should be nonnegative, apart from numerical discrepancies. Under the null hypothesis that the added effect is zero, and under the regularity conditions for the likelihood-ratio test, \(D_z\) has an approximate chi-square distribution with \(df\) equal to the number of independently added parameters. The cutoff must be chosen in advance for the search rule; the earlier material on likelihood-ratio cutoffs covers its selection.

For the usual interpretation of this test, both fits should use the same observations, likelihood, structural and random-effects specification, and estimation setup. The added model must be identifiable and both models must be adequately optimized. Ordinary chi-square calibration may not apply when a tested parameter is on a boundary or other regularity conditions fail; in that situation, do not treat a routine cutoff as automatically valid.

In a population PK model, a covariate effect is conditional on the other effects in the model. For example, a weight effect on typical clearance might be represented as:

Here \(\theta_{WT}\) describes the weight relationship conditional on the other terms in the fitted model. If another covariate is added, the model is re-estimated; the estimate of \(\theta_{WT}\), as well as other model parameters, may change. A forward-search decision concerns the added effect's contribution given the current model. It does not establish that the effect is causal or that the selected model is uniquely correct.

A Stepwise Procedure and Its Justification

At a given step, the candidate tests are comparisons against the same current model. If the prespecified rule uses a common significance threshold, select the qualifying candidate with the smallest p-value (equivalently, the strongest evidence relative to its null distribution), or use another selection rule defined in advance. After that choice, the current model changes: all parameters are re-estimated, and the remaining candidates must be tested against this updated model. Reusing their old objective-function differences would answer the wrong conditional questions.

The test statistic follows directly from the nested likelihoods. Under the null, the restricted model is the appropriate model; under the alternative, the added parameters are estimated freely. The likelihood-ratio statistic compares their maximized likelihoods:

The chi-square reference distribution is an asymptotic result under the null and regularity conditions, not a guarantee for every nonlinear mixed-effects fit. In practice, the comparison also depends on successful estimation: a failed or materially different local optimum can make the computed difference misleading.

Worked Search Paths

Interpreting, Recording, and Diagnosing the Path

A useful search record makes each decision auditable. For every step, retain the identity of the current model, each candidate tested, its added parameter count, both objective functions, the difference, the test rule, and the decision. After an addition, record updated parameter estimates as well: a covariate coefficient describes its conditional relationship in that fitted model and can shift when another effect enters.

Several checks matter before interpreting a result:

• Check the comparison. The candidate model should contain the current model as a special case. The same observations and likelihood must underlie both objective functions; a change in records can make their difference unsuitable for the stated nested test.
• Check the fits. Review convergence, parameter estimates, standard errors, and relevant model diagnostics. An implausible estimate, large uncertainty, or unstable objective function can indicate weak information or estimation trouble, even when a numerical threshold is crossed.
• Check the statistic. A materially negative difference for a nominally nested addition is a warning to investigate optimization, model coding, or comparability. Do not silently replace it by zero or interpret it as evidence that the covariate worsens the maximized likelihood.
• Check the conditional meaning. A candidate that no longer qualifies after another addition has not necessarily become biologically irrelevant. Its test asks whether it adds detectable fit beyond the effects already in the model, in this data set and under this model specification.

Forward addition is useful when a scientifically motivated base model is available and the candidate set is manageable. It differs from fitting one large model and testing all effects jointly: forward addition constructs a sequence of one-at-a-time conditional comparisons. Backward elimination takes the opposite starting point, beginning with a fuller model, while a prespecified model avoids a data-driven search. Each approach answers a different modeling question. Forward addition can miss a useful effect that does not qualify early, and changing the order can change which conditional effects are available for later tests. A selected path is not proof that excluded covariates have no effect.

The cutoff used at each step should be stated as a search rule, not mistaken for a complete account of uncertainty after searching multiple candidates and steps. The nominal test result is conditional on the particular comparison. Selection can make ordinary post-selection interpretations too optimistic; the details of multiplicity adjustment and alternative selection criteria require separate treatment. Likewise, a statistically selected effect should be evaluated for clinical and biological plausibility and for adequate representation of the data, not judged by the test statistic alone.