Baseline Hazard Flexibility and Association Estimates

Why the Baseline Hazard Matters

In a joint model, the longitudinal submodel describes a subject’s biomarker trajectory, while the event submodel relates that trajectory to the hazard of the event. The baseline hazard describes how the event rate changes over time for a subject at the reference covariate values and biomarker level. If its shape is constrained incorrectly, the association estimate can shift: the model may use the biomarker link to compensate for time patterns that the baseline hazard cannot represent.

The purpose of this sensitivity analysis is not to choose a baseline specification simply because it produces a preferred association estimate. It is to ask whether the estimated biomarker link and its scientific interpretation are stable when the baseline hazard is represented in a different, defensible way. The comparison is meaningful only when the other important parts of the joint model are held fixed.

Definition: Let $T_i$ be subject $i$’s event time, $C_i$ the censoring time, and $Y_i=\min(T_i,C_i)$ the observed time. Let $\delta_i=1$ when the event is observed and $0$ when the observation is censored. A shared-random-effects joint model may specify the event hazard, conditional on the subject’s random effects $b_i$, as $$ h_i(t\mid b_i)=h_0(t)\exp\{\gamma^\top w_i+\alpha m_i(t)\}, $$ where $h_0(t)$ is the baseline hazard, $w_i$ is a vector of event-submodel covariates, $m_i(t)$ is the latent biomarker value from the longitudinal model, $\gamma$ is a vector of covariate effects, and $\alpha$ is the current-value association parameter. With this convention, a one-unit increase in $m_i(t)$, holding $w_i$ and the rest of the model fixed, multiplies the hazard by $\exp(\alpha)$.

A parametric baseline specifies a particular function for $h_0(t)$ or its cumulative hazard. For example, a Weibull baseline can be written $H_0(t)=\lambda t^\kappa$, with $\lambda>0$ and $\kappa>0$, so that $h_0(t)=\lambda\kappa t^{\kappa-1}$. A semiparametric specification leaves the baseline cumulative hazard $H_0(t)$ unspecified apart from being nondecreasing; it estimates the baseline flexibly from event-time and risk-set information rather than imposing a named smooth shape.

Both specifications retain assumptions. The comparison assumes that the longitudinal model, the association form, the covariate effects, and the event and censoring definitions are appropriate. It also relies on independent subjects and on censoring being independent of the event process conditional on the variables and random effects included in the model. A flexible baseline does not repair a misspecified biomarker trajectory, an incorrect association link, or informative censoring that the model has not represented.

What Must Stay Fixed in a Sensitivity Comparison

For a focused comparison, fit models to the same subjects and use the same longitudinal submodel, random-effects structure, event covariates, association link, time origin, and biomarker units. Change the baseline-hazard specification alone. In particular, retain the same definition of $m_i(t)$: changing from current value to slope or cumulative exposure at the same time as changing the baseline would make it impossible to attribute an estimate change to either modeling choice.

The estimate $\hat{\alpha}$ is on the log-hazard scale. Report it together with $\exp(\hat{\alpha})$, the hazard ratio for a one-unit increase in the biomarker, and state the biomarker unit. Compare uncertainty as well as point estimates. A confidence interval that crosses zero for $\alpha$ corresponds to a hazard-ratio interval that crosses one. A change in whether an interval crosses its null value is a change in the strength of statistical evidence, not proof that the underlying association changed between models.

Joint-model estimation must account for the latent trajectories and random effects. In general, one cannot treat fitted biomarker values as known covariates and then assume an ordinary survival-model estimate is the joint-model estimate. Full likelihood methods integrate over the random effects, using numerical integration or related computation. In a semiparametric fit, the baseline is estimated alongside the other model components; its flexibility does not remove the need to account for uncertainty and latent trajectories.

Worked Example: Comparing Association Estimates on the Same Scale

Suppose two hypothetical joint-model fits use the same longitudinal model and current-value link. The only planned difference is a Weibull baseline versus an unspecified semiparametric baseline. The biomarker is measured in units of 10 mg/L. The fitted association estimates are $\hat{\alpha}=0.48$, $SE=0.18$ for the Weibull model and $\hat{\alpha}=0.31$, $SE=0.17$ for the semiparametric model.

For the Weibull model, the estimated hazard ratio per 10 mg/L is $\exp(0.48)=1.616$, or about 1.62. A Wald 95% confidence interval on the log-hazard scale is $$ 0.48\pm 1.96(0.18)=0.48\pm0.3528=(0.1272,\ 0.8328). $$ Exponentiating the endpoints gives $(\exp(0.1272),\exp(0.8328))=(1.136,\ 2.300)$, approximately $(1.14,\ 2.30)$.

For the semiparametric model, the hazard ratio is $\exp(0.31)=1.363$, or about 1.36. Its Wald interval is $$ 0.31\pm 1.96(0.17)=0.31\pm0.3332=(-0.0232,\ 0.6432), $$ which becomes $(\exp(-0.0232),\exp(0.6432))=(0.977,\ 1.902)$, approximately $(0.98,\ 1.90)$.

In these hypothetical results, both point estimates indicate a higher hazard with higher biomarker values, but the magnitude is smaller under the semiparametric baseline. The Weibull interval excludes $\alpha=0$, whereas the semiparametric interval includes it. This is a meaningful sensitivity in the estimated evidence. It does not establish that the semiparametric model is correct, nor is the overlap or nonoverlap of these separate intervals a formal test of the difference between estimates. The two fitted coefficients describe the same one-unit biomarker contrast, so their change is interpretable on a common scale.

Why a Flexible Baseline Can Change the Link Estimate

The event model assigns time-pattern information between the baseline hazard and the covariate-dependent multiplier. A restrictive baseline may fit poorly if the underlying event rate rises, falls, or changes curvature in a way its chosen form cannot represent. If biomarker values also change systematically over time, the association parameter can partly absorb that mismatch. A more flexible baseline can reduce this pressure, although it may also increase uncertainty because fewer restrictions are imposed on the baseline shape.

The mechanism is visible in a simple conditional derivation. Suppose the random effects and the linear predictor $\eta_i(t)=\gamma^\top w_i+\alpha m_i(t)$ are treated as fixed, and the baseline cumulative hazard is represented by jumps at distinct observed event times. For event time $t_j$, let $\Delta H_j$ be the baseline cumulative-hazard increment and let $R_j=\sum_{i\in\mathcal{R}(t_j)}\exp\{\eta_i(t_j)\}$, where $\mathcal{R}(t_j)$ is the risk set just before $t_j$. For a single event at that distinct time, the event contribution contains $\log(\Delta H_j)+\eta_{\text{event}}(t_j)$, while the survival contribution for the risk set contains $-R_j\Delta H_j$. Terms not involving $\Delta H_j$ can be left out when maximizing with respect to that increment.

\ell_j(\Delta H_j)=\log(\Delta H_j)-R_j\Delta H_j,\qquad \frac{\partial\ell_j}{\partial\Delta H_j}=\frac{1}{\Delta H_j}-R_j.

Setting the derivative to zero gives $\Delta H_j=1/R_j$. The second derivative is $-1/(\Delta H_j)^2<0$, so this stationary point is a maximum. This derivation assumes one event at a distinct event time and fixed linear predictors. It illustrates how an unconstrained baseline increment can adapt to the risk set; it is not, by itself, a complete estimation algorithm for a joint model with latent random effects.

For tied events, this expression is not the exact tied-event likelihood result. Using $d_j/R_j$ for $d_j$ tied events corresponds to the Breslow approximation in the usual risk-set construction; the exact tied-event likelihood depends on the tie-handling model. A report should identify the method used rather than presenting that ratio as an exact result for ties.

Worked Example: A Baseline Increment at a Distinct Event Time

Consider a hypothetical conditional calculation at time $t_j$. There is exactly one event at this distinct event time. Three subjects are at risk immediately before it, with fixed relative hazard weights $\exp\{\eta_i(t_j)\}$ equal to 2.0, 1.0, and 0.5. Thus $R_j=2.0+1.0+0.5=3.5$.

The conditional log-likelihood contribution involving the baseline increment is $\ell_j(\Delta H_j)=\log(\Delta H_j)-3.5\Delta H_j$. Its derivative is $1/\Delta H_j-3.5$. Equating this to zero gives $$ \widehat{\Delta H}_j=\frac{1}{3.5}=0.2857. $$

The estimated increment is dimensionless, as a cumulative hazard is. This calculation shows the local baseline flexibility for one distinct event time when the linear predictors are held fixed. In an actual joint fit, the event-time contributions are combined with the longitudinal likelihood and the random-effects distribution, and the latent predictors are not simply known constants.

Reading the Comparison and Checking Fit

A useful comparison includes the association estimate, its standard error or interval, the corresponding hazard ratio per stated biomarker unit, and evidence about the baseline fit. If the association changes substantially, inspect whether the parametric baseline has an implausible time pattern or whether the flexible fit has unstable increments in sparse risk sets. A baseline estimate that changes abruptly where few subjects remain at risk should be interpreted cautiously: flexibility cannot create information that the data do not contain.

Diagnostics should focus on whether the models represent the observed event-time pattern and whether conclusions depend on a small number of times or subjects. Examine fitted cumulative baseline hazards and compare their broad time shapes, especially where events and risk-set information are plentiful. Assess event-model residual patterns using diagnostics appropriate to the fitted joint model, and check that conclusions are not driven by a small number of late events. Also verify that both fits converged and that reported standard errors are credible; numerical instability can masquerade as model sensitivity.

The main alternatives answer related but different questions. A parametric baseline is parsimonious and can be efficient when its time shape is well supported, but it risks bias if that shape is wrong. A semiparametric baseline reduces commitment to a particular shape, but can be less precise and still depends on the event model, link, censoring assumptions, and computation. A highly flexible parametric family or a piecewise baseline can provide intermediate flexibility, but it introduces its own choices about shape or intervals. Comparing these options is a robustness analysis, not a substitute for checking assumptions.

Worked Example: How Baseline Shape Affects Predicted Risk

Suppose a hypothetical semiparametric fit estimates $H_0(2)=0.12$ by two years, while a fitted Weibull baseline is $H_0(t)=0.04t^{1.4}$, with time measured in years. At two years the Weibull cumulative baseline hazard is $0.04(2^{1.4})$. Since $2^{1.4}\approx2.639$, this is approximately $0.1056$.

For a subject with a constant biomarker value $m=2$ in the stated units and no other covariates, use the semiparametric model’s hypothetical $\hat{\alpha}=0.31$. The hazard multiplier is $\exp(0.31\times2)=\exp(0.62)\approx1.859$. The corresponding cumulative hazard at two years is $0.12(1.859)=0.2231$, giving event risk $1-\exp(-0.2231)\approx0.200$, or 20.0%.

Using the Weibull baseline value at two years with that same association coefficient gives cumulative hazard $0.1056(1.859)=0.1963$ and risk $1-\exp(-0.1963)\approx0.178$, or 17.8%. The difference reflects the baseline cumulative hazards at that time, not a different biomarker multiplier. In a full comparison, each model would use its own jointly estimated association and baseline; this calculation isolates how different baseline levels affect a prediction when the link coefficient is held fixed.

Common Pitfall and Practical Limits

Common mistake: “The semiparametric model gives a smaller association estimate, so the parametric model must have overstated the true biomarker effect.” This reasoning treats one fit as truth merely because it is more flexible. The correction is to report the numerical sensitivity and examine baseline fit, uncertainty, and the shared assumptions. A flexible baseline can change an estimate for legitimate reasons, but it does not guarantee an unbiased association estimate.

Another pitfall is changing the baseline and the link at the same time, then attributing the difference to the baseline. Keep the link, biomarker scale, and other model components fixed for this comparison. Also distinguish a coefficient shift from a change in clinical interpretation: a hazard ratio is conditional on the specified model and is not a causal effect merely because it is expressed as a ratio.

The comparison has limits. Similar estimates do not prove that either baseline model is correct, and different estimates do not identify which model is wrong. Estimates may be imprecise when event counts are low or follow-up is sparse. Moreover, conclusions apply to the analyzed event process, time scale, and study population; changing these may change the baseline and the association estimate. The central question is whether the biomarker-link conclusion is robust enough to support the intended interpretation under plausible baseline representations.

Key takeaway: Compare baseline-hazard specifications while holding the longitudinal model and association link fixed. Interpret $\hat{\alpha}$ on the same biomarker scale, report its uncertainty and hazard-ratio translation, and use fit diagnostics to understand—not simply rank—any change.

Check Your Understanding

Work each question, then check the concise solutions below.

A current-value association estimate is $0.20$ per 5 mg/L. What hazard ratio does it imply for a 5 mg/L increase?
With one event at a distinct event time and risk-set sum $R_j=4$, what is the conditional estimate of the baseline increment under the fixed-linear-predictor derivation?
A parametric-baseline interval for $\alpha$ excludes zero, but a semiparametric-baseline interval includes zero. What can and cannot be concluded?
Why is it not a focused baseline-hazard sensitivity analysis to change from a current-value link to a slope link at the same time?

Solutions. (1) $\exp(0.20)=1.221$, so the fitted hazard is about 1.22 times as high per 5 mg/L increase, conditional on the model. (2) With one event, $\widehat{\Delta H}_j=1/R_j=1/4=0.25$. (3) The statistical evidence is sensitive to baseline specification; neither interval alone establishes which model is correct, and the interval change is not a formal test that the coefficients differ. (4) Both the baseline and the association definition have changed, so the estimate difference cannot be attributed to the baseline representation alone.