Group Lasso for Related Transition Effects

Why Group Selection Across Transitions?

A covariate may matter for several transitions in a multi-state model, but its association need not have the same size—or even the same direction—for each transition. Fitting a separate regression for every transition allows this flexibility, but can leave a large collection of terms to estimate and interpret. Group lasso addresses a particular selection question: does this covariate contribute to the model across the specified set of transitions?

Instead of penalizing every transition coefficient independently, group lasso penalizes the length of the coefficient vector for a covariate. This encourages the vector to be either entirely zero or retained as a group. It does not impose equality of the transition effects. In a proportional-hazards formulation, for example, a retained covariate can have a positive coefficient for one transition and a negative coefficient for another.

For transition-specific proportional hazards, one possible model is \(\lambda_r(t\mid \mathbf{x})=\lambda_{0r}(t)\exp\{\sum_j x_j(t)\beta_{jr}\}\), where \(\lambda_r\) is the hazard for transition \(r\), \(\lambda_{0r}\) is its baseline hazard, and \(x_j(t)\) is the covariate value used at time \(t\). The group penalty acts on the regression coefficients, not on the baseline hazards. The covariate values and risk sets must be defined for the transitions being modeled, as in transition regression.

The likelihood or loss must correspond to the chosen transition model. The penalty can be applied to a joint likelihood, a sum of transition-specific contributions, or an appropriately scaled objective. Subjects—not transition records treated as independent observations—are the independent sampling units when the model’s likelihood assumptions require independent subjects. If a composite objective is used, its construction and variance calculations must respect within-subject dependence.

Objective and What Selection Means

Let \(\ell(\boldsymbol{\beta})\) denote the log-likelihood, or the log partial likelihood when that is the fitted model’s likelihood contribution. A common objective is the negative log-likelihood divided by the number of subjects, with unpenalized nuisance parameters such as baseline-hazard parameters included as appropriate. For clarity, write the penalized objective as

The factor \(1/n\) is a convention, not a mathematical requirement. If the loss is instead an unscaled negative log-likelihood, the numerical value of \(\lambda\) must be on the corresponding scale. A group weight \(w_j=\sqrt{R_j}\), where \(R_j\) is the number of transition coefficients in group \(j\), is a common choice to account for group size; other weights are possible and should be stated.

When the fitted loss is convex in the coefficients, the group-lasso objective is convex because each norm penalty is convex. Thus a global minimizer can be found, although it may not be unique if the design is rank deficient or coefficients are weakly identified. Convexity does not make the model or the selected groups automatically correct. Interpretation still depends on a meaningful transition specification, appropriate covariate coding, and a defensible likelihood.

A nonzero group vector does not mathematically require every component to be nonzero. Group lasso encourages joint inclusion or exclusion, but an individual transition coefficient may still equal zero, for example because of the data structure or other constraints. “Retained across transitions” means that the covariate group is not removed wholesale; it does not guarantee a nonzero estimated effect for each transition.

Why the Penalty Can Remove a Whole Group

The thresholding mechanism is easiest to see in a quadratic problem for one coefficient group. Suppose the loss, after fixing other groups or under an orthogonal quadratic approximation, has the form \(\frac12\|\mathbf{b}-\mathbf{z}\|_2^2\). Here \(\mathbf{z}\) is the unpenalized estimate in this simplified problem. The group-lasso estimate minimizes

If \(\mathbf{z}=\mathbf{0}\), the objective is minimized at \(\widehat{\mathbf{b}}=\mathbf{0}\). For \(\mathbf{z}\neq\mathbf{0}\), consider a nonzero minimizer. Its gradient must vanish:

This equation makes \(\mathbf{b}\) point in the same direction as \(\mathbf{z}\). Write \(\mathbf{b}=c\mathbf{z}\) with \(c>0\). Taking norms in the gradient equation gives \(\|\mathbf{z}\|_2=c\|\mathbf{z}\|_2+\lambda w\), so \(c=1-\lambda w/\|\mathbf{z}\|_2\). This is possible for a nonzero minimizer only when \(\|\mathbf{z}\|_2>\lambda w\). At zero, the subgradient condition is \(\|\mathbf{z}\|_2\leq\lambda w\). Therefore the solution is

The explicit zero case matters: the second expression divides by \(\|\mathbf{z}\|_2\), so it cannot be used when \(\mathbf{z}=\mathbf{0}\). This derivation gives the exact update for the stated quadratic problem. In a general transition-regression likelihood, the full optimizer usually requires an iterative algorithm; the threshold formula explains the groupwise shrinkage rather than replacing that optimization.

Worked Selection and Interpretation

Assumptions, Tuning, and Diagnostics

The group-lasso calculation requires a specified grouping of coefficients and a defined loss. Statistical interpretation also requires the transition model and covariate coding to be suitable for the question. For example, if effects are modeled as constant over time, that proportional-effects assumption must be defensible; if covariates vary over time, their values must be assigned consistently to the risk intervals and transitions. Group lasso does not repair misspecified transition hazards, inappropriate risk sets, or dependent observations incorrectly treated as independent.

The penalty parameter \(\lambda\) controls the trade-off between fit and shrinkage. At \(\lambda=0\), the penalty contributes nothing; as \(\lambda\) grows, more groups tend to be set to zero. A practical analysis can choose \(\lambda\) using cross-validation or another prespecified criterion, but the choice should match the goal. Prediction-oriented tuning need not identify the scientifically “true” set of covariates. The scaling of the loss and the group weights must be held consistent when comparing candidate values.

Before interpreting a selected set, check that continuous predictors are scaled sensibly and that the transition coefficient groups are constructed as intended. Otherwise, the penalty can reflect units or arbitrary coding rather than a meaningful notion of group size. Examine the selected groups and their coefficient paths over a range of \(\lambda\), and assess whether small changes in tuning or data handling alter the conclusions. A group that appears only at a narrow range of penalties should not be presented as a stable finding.

Also inspect model fit and transition-specific diagnostics using appropriate methods. A group-lasso penalty is a selection and shrinkage device, not a diagnostic for time-varying effects, functional form, or influential observations. If one transition has little information, its coefficient may be unstable even when the covariate group is retained through the other transitions. That limitation should be reported rather than hidden by the group-level decision.

How Group Lasso Differs From Alternatives

These methods answer different modeling questions. Separate lasso is appropriate when transition-by-transition sparsity is wanted. A shared-effect model is appropriate when equal transition effects are justified or are the intended restriction. Group lasso is useful when the covariate should enter or leave as a set, while preserving transition-specific estimates. None of these penalties supplies ordinary unpenalized standard errors or p-values for a selected model without further justification; selection itself affects uncertainty.

Common Mistakes and Limitations

• Confusing selection with equality. A selected group does not mean its transition coefficients are equal. Group lasso penalizes their joint norm, not their differences.
• Reading a zero group as evidence of no effect. A zero estimate is a result of the chosen data, model, penalty, and tuning parameter; it is not proof that the covariate has no association.
• Applying the quadratic threshold to a general model. The closed-form update requires the stated quadratic loss. Correlated groups and general likelihoods require optimization of the full objective.
• Ignoring units and group definitions. Rescaling covariates or changing which transitions belong to a group can change the penalty’s practical effect and the selected result.
• Treating selection as inference. A nonzero penalized coefficient is not automatically statistically significant, and a zero coefficient is not a hypothesis test result.

Group lasso is most valuable when the scientific or predictive aim genuinely concerns covariate inclusion across a set of transitions. Its limitations are equally important: the result depends on the transition model, penalty weights, tuning, and data support, and it does not by itself quantify post-selection uncertainty or establish causal effects.