Baseline Hazard Flexibility And Association Identifiability

Why Baseline Flexibility Matters for Association

In a joint model, the survival submodel relates an individual's event hazard to a longitudinal marker, often through the marker's latent trajectory. The baseline hazard describes how event risk changes over time apart from that association and other covariates. If the baseline hazard is too restrictive, its misspecification may be attributed to the biomarker association. If it is very flexible, it can explain more of the time pattern in event risk, potentially leaving less information to estimate the association.

This is a trade-off, not a rule that flexibility always weakens association estimates. A common baseline hazard can account for risk changes shared across individuals at a given time. It cannot, by itself, explain systematic differences in risk between individuals who are at risk at that same time and have different biomarker values.

Definition: In a shared-random-effects joint model, let $m_i(t)$ be the latent longitudinal marker trajectory for individual $i$, and let $b_i$ denote that individual's random effects. A common survival submodel is $$ h_i(t\mid b_i)=h_0(t)\exp\{\gamma^\mathsf{T}w_i+\alpha m_i(t)\}, $$ where $h_0(t)$ is the baseline hazard, $w_i$ is a vector of survival covariates, $\gamma$ is its coefficient vector, and $\alpha$ is the biomarker association parameter. The value of $\alpha$ describes the log hazard ratio for a one-unit difference in $m_i(t)$, holding the other terms fixed. For fixed $b_i$, a one-unit increase multiplies the hazard by $\exp(\alpha)$. The full joint likelihood also uses the longitudinal model and the distribution of $b_i$.

To assess how baseline specification affects $\alpha$, fit the model under each baseline specification and examine the profile likelihood. If $\ell(\alpha,\eta)$ is the log likelihood, with $\eta$ collecting nuisance parameters such as baseline-hazard parameters, define

\ell_p(\alpha)=\max_{\eta}\ell(\alpha,\eta).

Each point on this profile is the maximized log likelihood with $\alpha$ held fixed and the nuisance parameters re-estimated. A profile that changes little over a wide range of $\alpha$ indicates that the data distinguish those values poorly under that model. In a joint model, the likelihood may involve integrating over the random effects; the same profiling principle applies to that likelihood.

What the Baseline Can and Cannot Absorb

At a given time, the log hazard can be written as

\log h_i(t\mid b_i)=\log h_0(t)+\gamma^\mathsf{T}w_i+\alpha m_i(t).

If the marker values at that time can be decomposed into a time-specific common component and individual deviations, $m_i(t)=c(t)+u_i(t)$, then

\log h_i(t\mid b_i)=\{\log h_0(t)+\alpha c(t)\}+\gamma^\mathsf{T}w_i+\alpha u_i(t).

The common component $\alpha c(t)$ has the same value for everyone at that time. A sufficiently flexible baseline can represent its contribution as part of the time-varying baseline. The deviations $u_i(t)$, however, vary between individuals. When there is adequate within-time variation and event information, those contrasts can still inform $\alpha$. This decomposition is a useful explanation, not a guarantee that the full likelihood separates the parameters well: random effects, censoring, sparse events, and the longitudinal data all affect the information available.

Worked Example: Aggregating to a Constant Baseline

Consider hypothetical grouped survival data with a binary marker $x$, where $x=0$ or $x=1$, and corresponding exposure totals of 100 person-years in each group. There are 25 events, 17 among individuals with $x=1$. For this calculation, use the exponential-hazard survival submodel $h(t\mid x)=\lambda\exp(\alpha x)$, with a common constant baseline rate $\lambda$. Treat the longitudinal marker values as fixed for this reduced calculation; a full joint model also estimates the longitudinal part and accounts for random effects.

Ignoring likelihood constants that do not depend on $\alpha$ or $\lambda$, the log likelihood is

\ell(\lambda,\alpha)=25\log\lambda+17\alpha-100\lambda-100\lambda e^\alpha.

For fixed $\alpha$, setting the derivative with respect to $\lambda$ to zero gives $25/\lambda-100(1+e^\alpha)=0$, so $\hat\lambda(\alpha)=25/\{100(1+e^\alpha)\}$. Substitution yields the profile difference

\ell_p(\alpha)-\ell_p(0)=17\alpha-25\log\left(\frac{1+e^\alpha}{2}\right).

At $\alpha=\log 2$, the substituted calculation is $17\log 2-25\log(1.5)$, or approximately $1.6469$ using the rounded profile calculation. The constant-baseline profile is therefore higher at $\log 2$ than at zero. Its maximum occurs where the derivative $17-25e^\alpha/(1+e^\alpha)$ is zero: $e^\alpha=17/8$, so $\hat\alpha=\log(17/8)\approx0.7538$. The profile at $\log 2$ is close to this maximum. This is evidence about the fitted profile under this particular baseline specification, not proof that the marker causes higher risk.

Comparing Profiles When the Baseline Is More Flexible

A piecewise-constant baseline allows the underlying hazard to differ across time intervals. In a grouped exponential representation, let $Y_{0j}$ and $Y_{1j}$ be the person-time exposures for marker groups 0 and 1 in interval $j$, and let $d_{1j}$ be the events in group 1. With $D_j$ total events in that interval, profiling the interval-specific baseline rate gives a contribution to the profile difference of

d_{1j}\alpha-D_j\log\left(\frac{Y_{0j}+e^\alpha Y_{1j}}{Y_{0j}+Y_{1j}}\right).

The reason is that the interval-specific rate is estimated separately, using total interval exposure $Y_{0j}+e^\alpha Y_{1j}$ after accounting for the relative hazard multiplier. The profile sums these contributions over intervals. Compared with a constant baseline, the flexible model can explain changes in the overall event rate from interval to interval without requiring $\alpha$ to explain them.

Worked Example: The Same Totals Under a Piecewise Baseline

Now use hypothetical data with two intervals. In interval 1, $Y_{0,1}=80$, $Y_{1,1}=20$, $D_1=5$, and $d_{1,1}=1$. In interval 2, $Y_{0,2}=20$, $Y_{1,2}=80$, $D_2=20$, and $d_{1,2}=16$. Thus the aggregate totals remain 100 person-years in each marker group, 25 events overall, and 17 group-1 events, as in the preceding example.

With a separate baseline rate for each interval, the profile difference is

\ell_p(\alpha)-\ell_p(0) =17\alpha -5\log\left(\frac{80+20e^\alpha}{100}\right) -20\log\left(\frac{20+80e^\alpha}{100}\right).

At $\alpha=0$, the derivative is $17-5(0.2)-20(0.8)=0$. Each interval's contribution is concave in $\alpha$, so the profile is maximized at zero. At $\alpha=\log 2$, the calculation is

17\log 2-5\log(1.2)-20\log(1.8)\approx -0.8838.

Thus the likelihood at $\log 2$ is $\exp(-0.8838)\approx0.4132$ times the profile maximum, with constants omitted identically from both values. In this hypothetical data set, the piecewise-baseline profile does not favor $\log 2$ over zero as the constant-baseline profile did. The event proportions across intervals align with different marker-group exposure mixes; the separate baseline rates accommodate the interval pattern, changing what the profile attributes to $\alpha$. The comparison illustrates sensitivity to baseline specification, not a general direction of change.

Worked Example: Exact Nonidentifiability in a Limiting Case

Suppose every individual at risk in each time interval has the same marker value $c_j$, and the baseline is allowed to vary freely by interval. In a hypothetical two-interval case, take $c_1=2$, $c_2=3$, and baseline rates $0.04$ and $0.02$ per year. At $\alpha=\log 2$, the corresponding hazards are $0.04e^{2\log 2}=0.16$ and $0.02e^{3\log 2}=0.16$ per year.

For any other value $\alpha'$, define new interval baseline rates $h'_{0j}=h_{0j}\exp\{(\alpha-\alpha')c_j\}$. Then $h'_{0j}\exp(\alpha'c_j)=h_{0j}\exp(\alpha c_j)$ in each interval. The hazards, and hence the survival likelihood, are unchanged. Because the marker has no between-individual variation within an interval, the data cannot distinguish its common time pattern from the baseline. In this limiting case, $\alpha$ is not identified by the survival component.

How to Compare Models and Diagnose Trouble

A meaningful comparison holds the longitudinal specification, association structure, covariates, and estimation settings fixed while changing the baseline representation. For each baseline choice, compute $\ell_p(\alpha)$ over a useful range, re-estimating all nuisance parameters at each value. Compare the maximizing value, the curvature and width of the profile, and whether it is asymmetric or nearly flat. The introductory material on structural and practical identifiability distinguishes exact nonidentifiability from weak information in finite data; both can appear as profiles that do not sharply distinguish values.

Baseline choice	Strength	Risk for association assessment
Parametric, such as constant or Weibull	Economical and often precise when its time pattern is adequate	Misspecified time pattern can distort the estimated association
Piecewise constant	Allows interval-specific risk levels and has a transparent likelihood	Many intervals with few events can leave weakly determined baseline rates and association profiles
Spline or other smooth baseline	Represents gradual changes while controlling roughness	Results may depend on knot placement or smoothing; excess flexibility can absorb time-pattern information
Stratified baseline	Allows distinct baseline patterns across prespecified strata	Uses information only within strata for contrasts and may reduce precision

Several checks help reveal whether the trade-off is affecting the result:

Plot the profile for each baseline choice on the same $\alpha$ scale. A large shift in its maximum or a marked widening merits investigation.
Check event counts and risk-set marker variation over time. If most intervals have few events or little within-interval variation in $m_i(t)$, the baseline may leave little information for the association.
Check whether the fitted baseline is implausibly rough or whether its estimated shape changes substantially across reasonable smoothness settings. This is a warning about baseline estimation, not a direct test of $\alpha$.
Keep the longitudinal model and random-effects distribution consistent during the comparison. Otherwise, a changed profile cannot be attributed to baseline flexibility alone.

Comparing maximized likelihood values from models with different numbers of baseline parameters is not, by itself, a fair measure of which baseline is preferable: a model with more parameters can fit at least as well. Information criteria or a principled penalty can help compare overall model fit, but they do not replace examining the association profile. Likewise, a sharply peaked profile under one restrictive baseline does not establish that the baseline is adequate. The conditional grouped calculations above explain the mechanism; in a fitted joint model, uncertainty in the latent trajectories and integration over random effects can further alter the profile.

Common Pitfall and Limitations

Common pitfall: “The flexible baseline changed the estimated association, so it must have absorbed the biomarker effect.” That conclusion is too strong. A common baseline absorbs shared time variation, not automatically the within-time differences between individuals. A changed estimate can also reflect sparse events, limited marker variation, random-effects uncertainty, or numerical instability. Compare the profiles, inspect within-time information, and verify that the rest of the model and estimation settings were held fixed.

Baseline flexibility is therefore best treated as a sensitivity dimension. A highly flexible baseline does not solve every form of misspecification, and a flatter profile does not say which model is scientifically correct. When estimates differ materially, report the dependence on baseline assumptions and consider whether the data contain enough event and within-time marker information to support the desired association estimate. The result is conditional on the joint model's assumptions, including the chosen marker trajectory, association form, and random-effects distribution.

Check Your Understanding

Answer each question, then compare your reasoning with the brief solutions.

What is profiled out when computing $\ell_p(\alpha)$, and why must those nuisance parameters be re-estimated at each fixed value of $\alpha$?
Why can a flexible baseline absorb a common marker pattern over time but not necessarily a marker contrast between individuals at the same time?
In the two-interval example, what is the profile derivative at $\alpha=0$, and what does its value imply together with concavity?
What feature of the limiting example makes $\alpha$ exactly unidentified from the survival component?

Solutions. (1) The nuisance parameters include the baseline-hazard parameters and other non-target parameters. Maximizing them at each fixed $\alpha$ gives the likelihood available to that value, rather than comparing values fitted with arbitrary nuisance settings. (2) A common marker component can be represented as a function of time in the baseline; individual departures differ within the risk set and cannot generally be represented by one common baseline. (3) It is $17-5(0.2)-20(0.8)=0$. Since the profile is concave, zero is its maximum. (4) Everyone at risk in an interval has the same marker value, so changing $\alpha$ can be offset exactly by changing that interval's baseline rate.