1. What Are Forward Addition and Backward Elimination?
Forward addition and backward elimination are stepwise approaches commonly used during covariate model development in population pharmacokinetics. They provide a structured way to evaluate candidate covariate relationships with PK parameters.
The basic strategy is to begin with a defined structural PK model and a set of plausible candidate covariates. Covariates are then added or removed according to prespecified criteria, often based on changes in the objective function value (OFV), likelihood-ratio testing, parameter estimates, diagnostics, and scientific plausibility.
A stepwise covariate strategy can use forward addition to identify candidate relationships and backward elimination to determine which relationships remain supported in the final model.
2. Why Use Forward Addition and Backward Elimination?
Population PK datasets may contain many potential covariates: body weight, age, renal function, sex, disease status, laboratory measurements, concomitant medications, and other patient characteristics.
Including every possible covariate can produce an unnecessarily complicated model. Conversely, excluding all covariates can leave important sources of variability unexplained.
A stepwise strategy provides a reproducible framework for moving between these extremes.
| Problem | Potential consequence | Role of stepwise development |
|---|---|---|
| Many candidate covariates | Large number of possible models | Provides an organized search strategy |
| Weak covariate relationships | Overly complicated model | Helps identify relationships that provide sufficient evidence |
| Correlated covariates | Unstable parameter estimates | Encourages evaluation of competing relationships |
| Multiple testing | Chance findings | Supports use of prespecified criteria and confirmation steps |
| Overparameterization | Poor precision or convergence problems | Provides a mechanism for simplifying the model |
3. Start With an Appropriate Base Model
Forward addition should not begin with an arbitrary model. The first step is to establish an adequate base structural and statistical model.
Depending on the analysis, the base model may specify the number of compartments, absorption process, elimination process, interindividual variability, and residual unexplained variability.
For example, suppose clearance is modeled as:
where \(\theta_{CL}\) represents typical clearance and \(\eta_{CL,i}\) represents interindividual variability for individual \(i\).
A candidate covariate can then be introduced to explain some of the between-subject variability in clearance.
4. Forward Addition
In forward addition, candidate covariates are introduced sequentially into the model. At each step, the candidate relationship that satisfies the prespecified inclusion criterion is considered for addition.
Step 1: Define candidate relationships
Candidate relationships should be based on biological knowledge, previous analyses, exploratory plots, literature, and the characteristics of the dataset.
For example, possible clearance relationships might include:
- Body weight → CL
- Creatinine clearance → CL
- Age → CL
- Sex → CL
- Concomitant medication → CL
Step 2: Fit each candidate relationship
Each candidate can be evaluated by adding it to the current model. The change in objective function value can then be examined along with parameter estimates and model diagnostics.
When the more complex model produces a lower OFV, the magnitude of the decrease can provide evidence that the additional covariate improves the likelihood-based fit.
Step 3: Select the next covariate
Among the candidate relationships that satisfy the inclusion criterion, the relationship producing the strongest improvement according to the predefined procedure can be added.
Step 4: Repeat
The process continues until no remaining candidate meets the forward-addition criterion or until the prespecified candidate set has been evaluated.
5. Objective Function Value and Likelihood-Ratio Testing
In likelihood-based population PK modeling, changes in OFV are often used to compare nested models.
For two nested models, the likelihood-ratio statistic can be expressed as:
where \(L_{\text{reduced}}\) and \(L_{\text{full}}\) are the likelihoods under the reduced and full models, respectively.
Under standard regularity conditions, the difference can be compared approximately with a chi-square distribution with degrees of freedom corresponding to the difference in the number of estimated parameters.
| Comparison | Typical interpretation |
|---|---|
| Large decrease in OFV | The added relationship substantially improves the likelihood under the model |
| Small decrease in OFV | The additional relationship provides limited statistical improvement |
| No meaningful improvement | The candidate may not justify the additional model complexity |
The exact OFV thresholds used in a covariate modeling workflow should be defined in advance. They should not be selected retrospectively to obtain a desired model.
6. Why Is the Forward Criterion Often Relatively Liberal?
A common strategy is to use a relatively liberal criterion during forward addition and a stricter criterion during backward elimination.
The reasoning is that forward addition is intended to identify potentially important relationships. A relationship that is modestly supported at this stage may become more clearly interpretable once other important covariates have been incorporated.
For example, a development procedure might specify a nominal inclusion threshold of:
for a one-parameter nested comparison, corresponding approximately to a likelihood-ratio test at the 0.05 level under the usual chi-square approximation.
This is only an illustrative threshold. Actual population PK workflows may use different criteria depending on the objective, estimation method, number of candidate tests, model structure, and analysis plan.
7. Backward Elimination
After forward addition has produced a relatively full covariate model, backward elimination evaluates whether the included relationships can be removed without losing sufficient model support.
The procedure works in the opposite direction from forward addition.
- Start with the full covariate model.
- Remove one covariate relationship at a time.
- Evaluate the resulting change in OFV and other diagnostics.
- Identify relationships that fail the prespecified retention criterion.
- Remove the least-supported relationship.
- Repeat until all remaining relationships satisfy the retention criterion.
The final model therefore does not necessarily contain every covariate that entered during forward addition.
8. Why Use a Stricter Criterion for Backward Elimination?
The purpose of the backward step is different from the purpose of the forward step. Forward addition searches for potentially useful relationships, whereas backward elimination asks whether relationships remain sufficiently supported when considered within a model containing multiple covariates.
For that reason, a stricter criterion is often used for retention.
As an illustration, suppose the backward-elimination threshold is:
for a one-parameter comparison. This corresponds approximately to a nominal likelihood-ratio threshold of 0.01 under the usual chi-square approximation.
The precise criterion should be prespecified and interpreted in the context of the complete modeling strategy.
9. The Complete Forward–Backward Workflow
A typical stepwise covariate-development sequence can be summarized as follows.
| Stage | Action | Decision |
|---|---|---|
| 1 | Develop the base PK model | Establish an adequate structural and statistical model |
| 2 | Define candidate covariate relationships | Use scientific and exploratory evidence |
| 3 | Perform forward addition | Add relationships satisfying the inclusion criterion |
| 4 | Repeat forward testing | Continue until no additional candidate qualifies |
| 5 | Begin backward elimination | Test removal of included relationships |
| 6 | Apply stricter retention criterion | Remove relationships that fail the criterion |
| 7 | Repeat backward elimination | Continue until all remaining relationships qualify |
| 8 | Evaluate final model | Assess diagnostics, precision, plausibility, and predictive performance |
10. Worked Example: Selecting Covariates on Clearance
Suppose a population PK model has typical clearance of 10 L/h. Five candidate covariates are being considered:
- Body weight
- Creatinine clearance
- Age
- Sex
- Concomitant medication
Step 1: Base model
The base model has an OFV of 1,250.
Step 2: Forward addition
Each candidate is tested individually.
| Candidate | OFV | ΔOFV |
|---|---|---|
| Body weight | 1,243 | 7 |
| Creatinine clearance | 1,230 | 20 |
| Age | 1,247 | 3 |
| Sex | 1,248 | 2 |
| Concomitant medication | 1,245 | 5 |
If the prespecified forward criterion is \(\Delta OFV > 3.84\), several relationships qualify. The candidate producing the largest improvement would be added first.
Step 3: Add creatinine clearance
Creatinine clearance produces the largest initial OFV reduction, so it becomes the first covariate in the model.
Step 4: Re-test remaining candidates
The remaining candidates are now evaluated on top of the creatinine-clearance relationship. The results may change because covariates can explain overlapping portions of between-subject variability.
Step 5: Continue forward addition
Suppose body weight produces an additional \(\Delta OFV=8\). It is added. Age, sex, and medication are then evaluated again in the updated model.
Step 6: Begin backward elimination
Suppose the resulting model contains creatinine clearance, body weight, and medication.
The relationships are now removed one at a time. If medication produces only a small OFV increase when removed, while creatinine clearance and body weight produce large increases, medication may fail the stricter retention criterion and be removed.
11. The Covariate Relationship Matters
Forward addition and backward elimination determine whether candidate relationships are supported, but the relationship itself must also be specified appropriately.
For a continuous covariate such as body weight, several forms might be considered.
Linear relationship
Power relationship
These relationships make different assumptions about how the covariate changes the PK parameter. A stepwise procedure should therefore evaluate scientifically reasonable functional forms rather than treating the covariate name alone as the model.
12. Forward and Backward Selection With Categorical Covariates
Categorical covariates can be incorporated using indicator variables.
For a binary covariate such as sex, a simple model might be:
where \(SEX_i=0\) represents the reference category and \(SEX_i=1\) represents the other category.
For multi-level categorical variables, multiple indicator parameters may be required. The number of additional parameters should therefore be considered when defining the appropriate model-comparison criterion.
13. What Happens When Covariates Are Correlated?
One of the most important limitations of automated covariate selection is that candidate covariates may be strongly correlated.
For example, body weight and body surface area may convey overlapping information. Similarly, renal biomarkers may be mathematically or biologically related.
If two covariates explain much of the same variability, the first covariate added to the model can reduce the apparent contribution of the second.
14. Why Stepwise Selection Does Not Eliminate Overfitting
A common misconception is that forward addition followed by backward elimination automatically produces an unbiased or optimal covariate model.
It does not.
Stepwise procedures repeatedly evaluate candidate relationships using the same dataset. Consequently, the final model reflects the search process as well as the underlying data-generating relationships.
Potential consequences include:
- Selection of relationships that are partly driven by random variation.
- Unstable parameter estimates.
- Dependence of the final model on the candidate set.
- Reduced reproducibility in another dataset.
- Overoptimistic assessment of model performance when evaluated only on the development dataset.
This is one reason why biological plausibility, parameter precision, diagnostics, external evaluation, and predictive performance remain important after stepwise selection.
15. Do Not Select Covariates Using OFV Alone
A covariate can produce an attractive OFV reduction while still creating an implausible or poorly behaved model.
After each important model-development step, consider:
- Magnitude and precision of the covariate effect.
- Clinical or pharmacological plausibility.
- Parameter correlations and identifiability.
- Changes in interindividual variability.
- Changes in residual unexplained variability.
- Goodness-of-fit diagnostics.
- Visual predictive checks or other simulation-based diagnostics.
- Influence of extreme observations.
- Model stability and successful estimation.
- Predictive performance in relevant populations.
16. Common Mistakes in Forward–Backward Selection
| Mistake | Why it is problematic |
|---|---|
| Adding every statistically significant covariate | Can produce an unnecessarily complex model and ignore biological plausibility. |
| Using the same threshold for addition and elimination automatically | Does not distinguish the different purposes of the two stages. |
| Changing thresholds after seeing the results | Makes the procedure less reproducible and can introduce selection bias. |
| Ignoring correlated covariates | Can make selection order and parameter interpretation unstable. |
| Ignoring functional form | A real covariate effect may appear weak if modeled using an inappropriate relationship. |
| Relying exclusively on OFV | Statistical improvement does not guarantee a useful or scientifically credible model. |
| Reporting only the final model | Hides the model-development process and makes the analysis harder to reproduce. |
| Assuming selection proves causality | A covariate association does not establish a causal mechanism. |
17. How Should the Procedure Be Reported?
A transparent population PK analysis should describe the covariate-selection procedure sufficiently for another analyst to understand how the final model was obtained.
Important details include:
- The candidate covariates considered.
- The PK parameters to which each covariate could be related.
- The functional forms evaluated.
- The forward-addition criterion.
- The backward-elimination criterion.
- The model-comparison method.
- How correlated covariates were handled.
- Any prespecified biological or clinical restrictions.
- The diagnostic and predictive assessments performed after selection.
A model-development table can be particularly useful because it shows which relationships entered, which were removed, and why.
18. What Does the Final Covariate Model Mean?
Suppose the final model for clearance is:
The model states that typical clearance changes with body weight according to a power relationship centered at 70 kg.
For a 70-kg individual, the weight term equals one:
so the typical clearance predicted by the covariate component is 10 L/h.
For an 80-kg individual:
The important point is that the covariate model provides a quantitative relationship between an individual characteristic and a PK parameter. The stepwise selection process determines whether the relationship is retained; the final parameter estimate describes its magnitude.
19. Limitations of Stepwise Covariate Selection
Forward addition and backward elimination are useful organizational frameworks, but they have important limitations.
- Selection uncertainty: small changes in the dataset can sometimes change which covariates are selected.
- Multiple testing: many candidate relationships may be evaluated during the search.
- Correlation: related covariates may compete to explain the same variability.
- Model dependence: covariate effects depend on the underlying structural and statistical model.
- Data dependence: a selected relationship may reflect characteristics of the development dataset.
- Functional-form dependence: failure to evaluate a suitable functional form can obscure a real relationship.
- Causal ambiguity: statistical association does not establish mechanism.
20. A Practical Forward–Backward Covariate Workflow
- Define the scientific objectives. Decide which PK parameters and covariate relationships are scientifically relevant.
- Establish the base model. Develop the structural model, interindividual variability model, and residual error model before beginning systematic covariate selection.
- Explore candidate relationships. Examine plots, summaries, biological knowledge, and prior evidence.
- Prespecify the selection criteria. Define the forward and backward thresholds before reviewing the final results.
- Perform forward addition. Add supported relationships sequentially.
- Perform backward elimination. Remove relationships that fail the stricter retention criterion.
- Evaluate the final model. Examine parameter estimates, precision, diagnostics, variability, and predictive performance.
- Assess robustness. Consider whether important conclusions depend strongly on the exact selection path.
- Interpret pharmacologically. Explain what the retained covariates imply for PK variability and prediction.
21. Key Takeaways
- Forward addition and backward elimination are stepwise approaches for population PK covariate model development.
- Forward addition generally begins with a base model and sequentially adds candidate covariate relationships.
- Backward elimination starts from the fuller covariate model and evaluates whether individual relationships can be removed.
- The forward criterion is often more liberal than the backward criterion because the two stages serve different purposes.
- Changes in objective function value can be used for likelihood-based model comparisons when the relevant assumptions are appropriate.
- A covariate that enters during forward addition can later be removed during backward elimination.
- Correlated covariates can make stepwise selection dependent on the order in which relationships are evaluated.
- The functional form of a covariate relationship is as important as the identity of the covariate itself.
- OFV should not be the sole basis for deciding whether a covariate belongs in the final model.
- Biological plausibility, parameter precision, diagnostics, stability, and predictive performance remain important after statistical selection.
- Stepwise selection does not eliminate overfitting or selection uncertainty.
- The final model should be interpreted as a model-based description of covariate–PK relationships, not automatically as evidence of causality.
Where to Go Next
A natural progression is to study covariate model building strategies in greater depth, including exploratory covariate screening, allometric relationships, categorical covariates, continuous covariates, correlation among candidate predictors, and full-model approaches.
The next tutorial can build directly on this framework by examining how to choose between competing covariate relationships and how to distinguish statistical association from a clinically meaningful PK effect.