Updating the Longitudinal-Survival Association Parameter

Why the Association Parameter Needs Its Own Update

In a shared-random-effects joint model, the longitudinal trajectory links the repeated measurements to the event hazard. A current-value association parameter controls how the subject’s latent longitudinal value at time $t$ enters that hazard. Updating this parameter is not generally a regression calculation on a single estimated trajectory: in the EM algorithm, the M-step maximizes an expectation over the posterior distribution of the latent random effects.

The focus here is the survival contribution to that maximization, with the longitudinal parameters and the other survival parameters held fixed at their current M-step values. The result is a one-dimensional optimization problem. Its score has a useful form, and under the standard current-value proportional-hazards specification its objective is concave. But a finite maximum is not guaranteed, and replacing the posterior expectation by a hazard evaluated at a posterior mean can change the update.

Model and M-Step Objective

Definition: Let $T_i$ be subject $i$'s observed follow-up time, $\delta_i$ equal 1 for an observed event and 0 for right censoring, and $b_i$ denote the subject-specific latent random effects. Write the latent longitudinal mean as $m_i(t;b_i)$. In a current-value proportional-hazards model, the event hazard conditional on $b_i$ is $$ h_i(t\mid b_i)=h_0(t)\exp\{w_i(t)^\mathsf{T}\eta+\alpha m_i(t;b_i)\}, $$ where $h_0(t)$ is the baseline hazard, $w_i(t)$ is a vector of survival covariates, $\eta$ is their coefficient vector, and $\alpha$ is the current-value association parameter. A one-unit increase in the latent current value multiplies the hazard by $\exp(\alpha)$, holding the other terms fixed.

At an EM iteration, let $\theta^{\mathrm{old}}$ denote the parameter values used to form the E-step posterior distribution of $b_i$, conditional on the observed longitudinal and survival data. For the association update, write $E_i^{\mathrm{old}}[\cdot]$ for expectation with respect to that posterior. The survival log-likelihood contribution conditional on $b_i$, omitting terms unrelated to $\alpha$, is

\ell_{S,i}(\alpha\mid b_i) = \delta_i\alpha m_i(T_i;b_i) - \int_0^{T_i}h_0(t)\exp\{w_i(t)^\mathsf{T}\eta+\alpha m_i(t;b_i)\}\,dt.

Thus, with the other parameters fixed, the part of the EM auxiliary function that depends on $\alpha$ is

Q_S(\alpha) = \sum_i E_i^{\mathrm{old}}\left[ \delta_i\alpha m_i(T_i;b_i) - \int_0^{T_i}h_0(t)\exp\{w_i(t)^\mathsf{T}\eta+\alpha m_i(t;b_i)\}\,dt \right].

This expression assumes the usual right-censored survival likelihood, a correctly specified conditional hazard, and censoring that is noninformative conditional on the model variables. The longitudinal mean $m_i(t;b_i)$, baseline hazard, and covariate term are treated as fixed functions while updating $\alpha$. The posterior distribution in the expectation is also held at its E-step value; it is not recomputed as a function of the candidate $\alpha$ during this maximization.

Score, Curvature, and the Maximization

If differentiation can be interchanged with the posterior expectation and the time integral—for example, when the relevant exponential moments are finite—the score is

U_S(\alpha) = \frac{\partial Q_S(\alpha)}{\partial\alpha} = \sum_i E_i^{\mathrm{old}}\left[ \delta_i m_i(T_i;b_i) - \int_0^{T_i} h_0(t)\exp\{w_i(t)^\mathsf{T}\eta+\alpha m_i(t;b_i)\} m_i(t;b_i)\,dt \right].

The first term is the posterior-expected event contribution. The second term accounts for the change in cumulative hazard as $\alpha$ changes. It is not enough to match the observed-event values of the marker: censored subjects and the full at-risk follow-up contribute through the cumulative-hazard term.

Differentiating once more gives

\frac{\partial^2 Q_S(\alpha)}{\partial\alpha^2} = -\sum_i E_i^{\mathrm{old}}\left[ \int_0^{T_i} h_0(t)\exp\{w_i(t)^\mathsf{T}\eta+\alpha m_i(t;b_i)\} m_i(t;b_i)^2\,dt \right]\leq 0.

The inequality follows because the hazard is nonnegative and the squared marker value is nonnegative. Therefore $Q_S$ is concave in $\alpha$. It is strictly concave if the expected integrated hazard-weighted squared marker is positive. Strict concavity means there can be at most one finite maximizer, but it does not ensure that one exists.

Key result: A finite interior M-step update solves $U_S(\alpha)=0$. Under the stated model, the score is nonincreasing in $\alpha$, so a root is a global maximizer. Check the score and the objective at the reported solution; if no root exists on the real line, the objective may have only a limiting supremum or the update may require a scientifically justified parameter constraint.

A short justification links the score root to maximization. For a differentiable concave function, its derivative is nonincreasing. If $U_S(\widehat{\alpha})=0$, then the derivative is nonnegative to the left of $\widehat{\alpha}$ and nonpositive to its right. Consequently, the objective cannot increase by moving away from $\widehat{\alpha}$, so $\widehat{\alpha}$ is a global maximum. If the curvature is strictly negative in the relevant region, this maximum is unique.

In practice, one can solve the scalar score equation with a bracketed root-finding method, or maximize $Q_S$ directly with a safeguarded one-dimensional optimizer. Newton’s method uses $\alpha_{\mathrm{new}}=\alpha-U_S(\alpha)/Q_S''(\alpha)$, but should be safeguarded: a small curvature can produce an excessively large step, and a numerical approximation can make an unguarded step decrease the objective.

Worked Example: A Closed-Form Update From Discrete Posterior Support

Consider three hypothetical subjects, each followed for one time unit. Two have an event at time 1 and one is censored at time 1. Let the baseline hazard be constant at 1, let $w_i(t)^\mathsf{T}\eta=0$, and let the latent marker be $m_i(t;b_i)=b_i$ throughout follow-up. Suppose the E-step posterior for each subject assigns probability $1/2$ to $b_i=0$ and probability $1/2$ to $b_i=1$. These assumptions make the posterior expectations explicit; the example is an illustration, not observed study data.

For either posterior, $E[b_i]=1/2$ and $E[\exp(\alpha b_i)]=(1+\exp(\alpha))/2$. Each event contributes $\alpha E[b_i]=\alpha/2$, while each subject contributes cumulative hazard $E[\exp(\alpha b_i)]=(1+\exp(\alpha))/2$. There are two events and three subjects, so

Q_S(\alpha) =2\left(\frac{\alpha}{2}\right) -3\left(\frac{1+\exp(\alpha)}{2}\right) =\alpha-\frac{3}{2}\{1+\exp(\alpha)\}.

The score and second derivative are

U_S(\alpha)=1-\frac{3}{2}\exp(\alpha), \qquad Q_S''(\alpha)=-\frac{3}{2}\exp(\alpha)<0.

Setting the score equal to zero gives $\exp(\widehat{\alpha})=2/3$, hence $\widehat{\alpha}=\log(2/3)\approx-0.4055$. At this value, $Q_S(\widehat{\alpha})=\log(2/3)-\frac{3}{2}(1+2/3)\approx-2.9055$. At $\alpha=0$, $Q_S(0)=-3$, so the update increases the objective by approximately $0.0945$. The fitted association corresponds to a hazard ratio $\exp(\widehat{\alpha})=2/3$ for a one-unit higher latent marker value, conditional on the other model terms.

Posterior Expectations Are Not Plug-In Values

The E-step expectation must be applied to the whole survival contribution. In particular, the cumulative-hazard term involves $E_i^{\mathrm{old}}[\exp\{\alpha m_i(t;b_i)\}]$, not $\exp\{\alpha E_i^{\mathrm{old}}[m_i(t;b_i)]\}$. Because the exponential is nonlinear, these quantities generally differ. This distinction is central when the posterior random-effects distribution is diffuse or skewed.

Worked Example: Posterior Averaging Changes the Update

Use two hypothetical subjects with constant baseline hazard $0.4$, no survival covariates, and $m_i(t;b_i)=b_i$. Both are followed for one time unit; one has an event at time 1 and the other is censored at time 1. For each subject, suppose the E-step posterior assigns probability $1/4$ to $b_i=0$ and $3/4$ to $b_i=1$. Then $E[b_i]=3/4$, while $E[\exp(\alpha b_i)]=1/4+(3/4)\exp(\alpha)$.

The one event contributes $3\alpha/4$. The total cumulative-hazard coefficient is $0.4(1+1)=0.8$, so

Q_S(\alpha)=\frac{3}{4}\alpha -0.8\left\{\frac{1}{4}+\frac{3}{4}\exp(\alpha)\right\} =\frac{3}{4}\alpha-0.2-0.6\exp(\alpha).

Its score is $U_S(\alpha)=3/4-0.6\exp(\alpha)$, giving $\exp(\widehat{\alpha})=5/4$ and $\widehat{\alpha}=\log(5/4)\approx0.2231$. The second derivative is $-0.6\exp(\alpha)<0$, verifying a unique maximum. The posterior-averaged hazard term at the solution uses $E[\exp(\widehat{\alpha}b_i)]=1/4+(3/4)(5/4)=1.1875$, not $\exp(\widehat{\alpha}E[b_i])=(5/4)^{3/4}$. The score equation and update follow from the former quantity, as required by the E-step objective.

Worked Example: A Finite Maximizer Need Not Exist

For one hypothetical censored subject, let $T=1$, $h_0(t)=1$, $w(t)^\mathsf{T}\eta=0$, and suppose the posterior places all its mass at $b=1$, so $m(t;b)=1$. Since $\delta=0$, there is no event contribution. The objective is

Q_S(\alpha)=-\int_0^1\exp(\alpha)\,dt=-\exp(\alpha), \qquad U_S(\alpha)=-\exp(\alpha)<0.

The objective increases as $\alpha$ decreases, approaching 0 as $\alpha$ tends to negative infinity, but it never reaches 0 at a finite value. Thus there is no finite maximizing update. A numerical optimizer that returns a very large negative number has not found a well-determined finite estimate; it is revealing the lack of a finite maximum in this example.

Practical Checks, Alternatives, and Limitations

Before accepting an update, verify that the E-step posterior and the current values of the other model parameters were held fixed throughout the M-step calculation. Evaluate the score at the proposed estimate and confirm it is close to zero for an interior solution. Check that the objective is no smaller than at the previous $\alpha$; the EM M-step is intended to maximize the fixed auxiliary function. For a bracketed score root, the score should change sign across the bracket. If the score is nearly flat, the update can be numerically unstable and the parameter may be weakly informed by the data.

The concavity result depends on the current-value term entering the log hazard linearly as $\alpha m_i(t;b_i)$, with the trajectory and other model components fixed. It does not imply that all parameters of a joint model can be updated independently or that the full observed-data likelihood is concave. If the model uses a different association structure, such as a nonlinear function of $\alpha$, the curvature argument must be checked anew.

A direct observed-likelihood optimization treats the random effects as integrated out and optimizes the marginal likelihood, rather than maximizing the posterior-expected complete-data contribution. The EM update described here is a conditional maximization of one component of the auxiliary function; it is not generally the same numerical step as direct optimization. A plug-in alternative that substitutes posterior mean trajectories is simpler, but it is not the EM M-step unless the relevant posterior distribution is degenerate or the required expectations happen to agree. Within the M-step, a score-root method and direct one-dimensional optimization target the same objective when both are implemented accurately; the latter can be preferable when a root cannot be bracketed.

Finally, the scale of $\alpha$ depends on the scale of the longitudinal marker. If the marker is measured in different units, the numerical value of $\alpha$ changes accordingly, while the hazard ratio must be stated for a specified marker increment. Weak variation in the latent marker over at-risk time, few events, or posterior uncertainty that leaves little effective information can make the association update imprecise or poorly conditioned. The M-step objective alone does not provide an observed-data standard error or a test of the association.

Common pitfall: Replacing $E[\exp\{\alpha m_i(t;b_i)\}]$ by $\exp\{\alpha E[m_i(t;b_i)]\}$ and calling the result the EM update. The correction is to take the posterior expectation of the complete survival contribution, including the exponential hazard, before maximizing over $\alpha$.

Check Your Understanding

Try each question, then compare your reasoning with the solutions below.

In the current-value hazard, what hazard ratio corresponds to a one-unit increase in the latent marker when $\alpha=-0.3$?
Why does the cumulative-hazard term contribute to the score for $\alpha$ even for a censored subject?
Suppose the computed score is strictly decreasing and changes from positive to negative across an interval. What does this imply about a root in that interval?
For one censored subject with constant marker $m=2$, follow-up time 1, and constant baseline hazard 1, write $Q_S(\alpha)$. Does it have a finite maximizer?
Why is posterior averaging of the hazard term generally different from evaluating the hazard at the posterior mean marker?

Solutions. (1) The hazard ratio is $\exp(-0.3)\approx0.7408$ per one-unit increase in the marker. (2) Censoring contributes the probability of remaining event-free through observed follow-up, represented by the negative cumulative hazard; that term depends on $\alpha$. (3) Continuity gives a score root between the two points, and concavity makes any such root a global maximizer; strict decrease makes it unique. (4) The event indicator is zero, so $Q_S(\alpha)=-\exp(2\alpha)$. Its derivative is $-2\exp(2\alpha)<0$, and its supremum 0 is approached as $\alpha$ tends to negative infinity, so no finite maximizer exists. (5) The exponential is nonlinear, so in general $E[\exp(\alpha m)]\ne\exp(\alpha E[m])$; the E-step requires the expectation of the full conditional survival contribution.