Backward Elimination from a Covariate-Rich Model

Why Begin with a Covariate-Rich Model?

Backward elimination is a covariate-search strategy that starts with a prespecified model containing the candidate effects and then considers removing them. It can be useful when the candidate set is scientifically motivated and the full model is estimable. The key operation is not simply deleting a term from a formula: each proposed deletion is tested against the current model, and the reduced model is refitted before the next decision.

The method uses likelihood comparisons to ask whether the data support retaining an effect conditional on the other effects currently in the model. This makes the sequence important. Removing one covariate can change the estimates and the evidence for another. The search is therefore a model-building procedure, not a collection of independent tests of the original candidate effects.

Conditions for a Valid Elimination Step

A typical NLME model specifies the distribution of observations \(Y_{ij}\) for subject \(i\) at occasion \(j\), with subject-level random effects such as \(\eta_i\). Covariates may enter parameters such as clearance. For example, the following expression defines a covariate model for individual clearance:

Here, \(CL_{\mathrm{typ}}\) is typical clearance, the \(\beta\) terms describe covariate effects, and \(\eta_{CL,i}\) is the individual random effect on log clearance. The example centers age at 50 years, weight at 70 kg, and albumin at 4 g/dL. A deletion test for age, for instance, compares the model with \(\beta_{\mathrm{age}}\) estimated to the nested model with \(\beta_{\mathrm{age}}=0\), while leaving the other terms in place.

For the likelihood-ratio test (LRT), the larger and reduced models must use the same observations, response model, random-effects structure, residual model, and likelihood definition. They should be fitted using the same estimation method and settings. Maximum likelihood fits are compared; likelihoods from different datasets or incompatible approximations do not form a valid LRT. In addition, the models must be nested, adequately identifiable, and fitted successfully. Under standard regularity conditions, the null parameter values are not on a boundary, and the sample is sufficiently informative, the LRT has an approximate chi-square distribution with degrees of freedom equal to the number of independent parameters constrained by the deletion.

A full model that is too complex for the available information may fail to converge or produce unstable estimates. Check covariate support, missingness, and collinearity before beginning; these affect whether the proposed full model can be fitted meaningfully. The candidate model and elimination rule should be specified in advance. In particular, retain lower-order terms when required by a model hierarchy—for example, do not remove a main effect while retaining an interaction that includes it.

The Likelihood-Ratio Calculation

Write \(\ell(M)\) for the maximized log likelihood of model \(M\). Since the larger model can reproduce the reduced model by satisfying its constraints, its maximized log likelihood cannot be smaller than that of the reduced model, apart from numerical fitting error. The likelihood ratio compares the likelihoods of the two fits:

The statistic is large when the reduced model fits substantially worse. Under the null hypothesis that the deleted parameter or block is unnecessary, its approximate reference distribution is \(\chi^2_{df}\), where \(df\) is the number of constraints. A small p-value is evidence against deleting that effect under the specified test; a larger p-value means the data did not provide strong evidence against the reduced model. It does not establish that the effect is exactly zero.

The derivation follows directly from the likelihood ratio. For a nested pair, the null model is the reduced model and the alternative is the larger model. The likelihood-ratio statistic is \(-2\log\{L(M_R)/L(M_F)\}\). Substituting \(\log L(M)=\ell(M)\) gives \(2\{\ell(M_F)-\ell(M_R)\}\). Under the regularity conditions above, this statistic is approximately chi-square under the null, with one degree of freedom for one constrained coefficient, or more for a block of coefficients.

Worked Search: One Effect Removed at a Time

Deleting Effects as Blocks

A covariate effect may require more than one coefficient. If a categorical predictor has several levels, for example, its representation may use multiple coefficients relative to a reference level. If the scientific question concerns the predictor as a whole, its deletion is a block constraint: remove all its associated coefficients together and use the corresponding difference in parameter count for \(df\). Testing or removing only one coefficient at a time answers a different question and can leave an incoherent representation of the effect.

Audit the Comparisons and the Search Record

A useful search record includes the current model, the effect tested for deletion, the reduced-model log likelihood, the LRT statistic, degrees of freedom, p-value, the decision, and the estimates after refitting. Recording estimates is not optional bookkeeping: it reveals whether the remaining effects are stable or shift substantially as other terms disappear.

For each fit, check convergence status, parameter estimates and standard errors, and whether the likelihood is plausible relative to the nested fit. A reduced model with a larger maximized likelihood than its parent usually signals an optimization problem, mismatched settings, or a coding error; it is not a reason to reverse the LRT formula. If a model does not converge or an estimate is effectively unidentifiable, treat the deletion comparison as unresolved until the cause is addressed. Do not turn a failed fit into evidence for removing the term.

How Backward Elimination Differs—and What It Cannot Establish

Forward inclusion starts with a simpler model and adds candidate effects; backward elimination starts with a covariate-rich model and considers deletions. Both can use nested likelihood comparisons, but they may reach different final models because each step conditions on the effects currently present. Backward elimination also depends on being able to fit its starting model. If the full model is unsupported by the data, a forward strategy may be more feasible, though it has its own limitations. A prespecified full model and a scientifically justified candidate set matter more than treating either search direction as an automatic discovery procedure.

The deletion p-values are conditional on the current model and the path taken to reach it. Repeatedly selecting and testing on the same data does not preserve the simple interpretation of a single prespecified test at its nominal error rate. The final model should therefore be reported as selected by a search, with the candidate set, rule, and sequence documented. A retained effect is not necessarily important in every scientific sense; a removed effect is not proven irrelevant. External knowledge, parameter uncertainty, model adequacy, and predictive performance remain part of model evaluation.

A common mistake is to test every deletion once in the full model and then remove all terms with large p-values simultaneously. This skips refitting and can produce a different sequence and final model, because the effect estimates and likelihood contrasts change after each deletion. The correction is to remove the selected effect, refit, and recompute every remaining comparison against that updated model. Another pitfall is to read a nonsignificant deletion test as proof that the covariate has no effect. It means only that the current data and model did not provide sufficiently strong evidence against the reduced model under the chosen rule.