Why likelihood support matters
A likelihood calculation is not only a way to compare parameter values. It also reveals whether a model can generate the observed event history at all. If a path includes a transition at a time when the model assigns that transition intensity zero, its path likelihood is zero. No finite log-likelihood or score calculation can repair this conflict.
The distinction between a structural zero and a parameter-boundary zero is important. A structural zero comes from the model’s rules, such as a transition being prohibited in a state, and does not depend on the parameter value. A parameter-boundary zero occurs only at a particular boundary value: for example, a transition rate may become zero when a nonnegative rate parameter is set to zero. The two cases have different implications for model compatibility and estimation.
Path likelihood and its zero factors
Consider an observed follow-up interval \([0,\tau]\). Let \(N_k(t)\) count observed events of type \(k\) by time \(t\), and let \(\lambda_k(t;\theta)\) be the corresponding predictable intensity at parameter value \(\theta\). The intensity is conditional on the observed history immediately before time \(t\). It is zero when the model says that type \(k\) cannot occur then. Assume the observed initial state is conditioned upon, there are no simultaneous jumps, and any censoring mechanism is ignorable for the likelihood being used. The usual multiplicative-intensity path likelihood, up to factors not depending on \(\theta\), is
The exponential term accounts for the observed event-free exposure: larger integrated intensity makes a long period with no event less compatible with the model. Each event contributes its intensity at its observed time. The formula is understood as a density for the event history, with parameter-independent reference-measure factors omitted.
This formula gives a direct support check. If any observed event has \(\lambda_k(t_j;\theta)=0\), its event factor is zero and hence \(L(\theta)=0\). If every observed event has positive intensity and the integrated intensities are finite, the displayed likelihood is positive. An empty event history has an empty product equal to one, so its likelihood is the exponential term; it can be positive even when some intensities are zero throughout follow-up.
A useful log-likelihood expression is
but it is finite only when every event factor is positive. If even one observed event has zero intensity, the log-likelihood is \(-\infty\), not a finite value that can be handled by ordinary score equations.
Why an event at zero intensity gives zero likelihood
Over a short interval of length \(h\), a standard continuous-time intensity model assigns an event of type \(k\), conditional on the past, probability approximately \(\lambda_k(t;\theta)h\), with the probability of no event approximately \(1-\sum_k\lambda_k(t;\theta)h\). Across small intervals, the no-event contributions accumulate to the exponential of minus the integrated intensity. At an observed event time, the event contribution is the corresponding intensity. Multiplying these contributions yields the path-likelihood form above.
In particular, if the type-\(k\) intensity is zero at an observed event, the first-order event contribution is zero. This is the likelihood-level expression of a support violation. The argument assumes an ordinary continuous-time intensity model. If events can occur at discrete scheduled times with positive probability, their conditional event probabilities require an appropriate discrete-time or point-mass likelihood rather than this continuous-time density formula.
A pointwise zero requires care. Predictable intensities are defined up to sets that are negligible for integration with respect to time and probability. A zero on a time interval or on a relevant set of positive exposure is substantively informative; a zero assigned only at an isolated deterministic time may depend on the chosen version of the intensity and does not by itself establish a meaningful support restriction. Interpret a zero at an event in light of the model’s actual event mechanism and support, not merely a numerical value stored at one time point.
Worked Example: A Competing Event Type Has Zero Rate
Suppose a hypothetical subject begins in one state and remains at risk for two event types until the first event. On a period of length 2 time units, take constant intensities \(\lambda_1=0.4\) and \(\lambda_2=\beta\) per time unit. The observed path has no event before time 2 and then has a type-2 event. Evaluate its likelihood at \(\beta=0\).
The event-free exposure factor is \(\exp\{-(0.4+\beta)2\}\), and the observed type-2 event contributes \(\beta\). Thus
At \(\beta=0\), substitution gives
The observed path is unsupported at this parameter value: a type-2 event cannot occur when its intensity is zero. For any \(\beta>0\), the path has positive likelihood, since both the event factor and the exposure factor are positive. The calculation concerns likelihood support; it does not claim that a particular exact event time has positive point probability under a continuous-time model.
Structural zeros and parameter-boundary zeros
A structural zero is imposed by the model’s state space, transition rules, or a deterministic risk-set condition. For example, a transition may be forbidden from a particular state, or its intensity may be multiplied by a known indicator that is zero. If the model says this intensity is zero for every allowable parameter value but the data record the transition under that condition, the path is unsupported throughout the parameter space. This is a model-data or data-construction conflict, not evidence that an optimizer needs a better starting value.
A parameter-boundary zero arises when the intensity is positive for some parameter values but becomes zero at a boundary value. For a nonnegative rate \(\theta\), the model \(\lambda(t;\theta)=Y(t)\theta\) has zero event intensity at \(\theta=0\) whenever the subject is at risk. Whether that boundary value has zero likelihood depends on the observed events: an observed event during positive exposure rules it out, while no such event can leave it as a maximizing value.
Thus, a zero estimate is not automatically a support failure. Check whether the data contain events requiring positive intensity under that transition. A boundary estimate of zero can be a valid maximum likelihood estimate when the observed history contains no such event. Conversely, evaluating a candidate boundary value at which an observed event has zero intensity gives zero likelihood, even if other parameter values support the path.
Worked Example: A Structural Risk-Set Zero
In a hypothetical two-state model, a transition can occur only when a subject occupies state A. Its intensity is \(\lambda(t;\theta)=Y_A(t)Z(t)\exp(\theta)\), where \(Y_A(t)\) indicates occupancy of state A and a known gate \(Z(t)\) is either zero or one. Suppose the recorded history says the subject is in state B at time 3, so \(Y_A(3)=0\), but also records this transition at time 3. Take \(Z(3)=1\).
At the recorded event time, the model intensity is
for every finite allowable \(\theta\). The event factor in the path likelihood is zero, so the recorded path has zero likelihood for every parameter value. This is structural: changing \(\theta\) cannot make the transition possible while the subject is not at risk in state A. The appropriate response is to check the state history, event coding, risk indicator, and transition specification. It is not appropriate to interpret the result as an extreme but finite parameter estimate.
Worked Example: No Events and a Boundary Estimate
Suppose a hypothetical transition has intensity \(\lambda(t;\theta)=Y(t)\theta\), where \(\theta\geq 0\) is a constant rate per year. A subject contributes 10 years at risk and has no transition of this type. The likelihood contribution is
For \(\theta\geq0\), \(L(\theta)\leq1\), with equality at \(\hat{\theta}=0\). Therefore the maximum likelihood estimate is zero events per year, a boundary value. At that estimate, the path remains supported: there was no observed transition whose event factor would be zero. If instead one transition had occurred during those 10 years, the likelihood would be \(L(\theta)=\theta\exp(-10\theta)\), which is zero at \(\theta=0\). For \(\theta>0\), its log-likelihood is \(\log\theta-10\theta\), and setting its derivative \(1/\theta-10\) equal to zero gives \(\hat{\theta}=0.1\) events per year. The observed event excludes the zero-rate boundary as a likelihood maximizer.
Diagnostics and common failure modes
Before maximizing or differentiating a path likelihood, inspect its event factors. This check can identify a model incompatibility that would otherwise appear as an optimizer failure, a log of zero, or a seemingly enormous negative objective. When a computed likelihood is zero, identify which observed event factor caused it; do not infer the cause from the total alone.
- Check the event and risk records. At each observed transition, verify that the subject is in an eligible state, the event type is allowed, and the risk indicator is one where required.
- Separate exact zeros from small positive values. A tiny positive intensity gives a very small likelihood contribution, not a zero. Numerical underflow can nevertheless make a positive likelihood appear to be zero; use a log-likelihood only after checking that all event intensities are positive.
- Check parameter constraints. A nonnegative rate equal to zero may be an allowed boundary value. A log-rate parameter can approach a zero rate only as it tends to \(-\infty\); a finite log-rate does not produce an exact zero.
- Check the observation model. The displayed likelihood conditions on the initial state and assumes an ignorable or parameter-free observation mechanism. If the initial state or censoring process is modeled jointly, its likelihood contribution may also matter.
A common mistake is to replace a zero event intensity by a small numerical constant, such as \(10^{-8}\), and proceed as if the data had positive likelihood under the stated model. That changes the model’s support and can conceal a prohibited transition or an incorrectly constructed risk set. A small positive replacement is defensible only as an explicitly chosen approximation in a revised model, not as a silent fix to a zero likelihood.
How to use the distinction in inference
The likelihood-support check comes before score equations, information calculations, or optimization. If all observed events have positive intensities, ordinary likelihood calculations may proceed subject to the model’s other assumptions. If an event has zero intensity at a particular parameter value, exclude that value from any maximization over supported parameter values. If the zero is structural for all parameter values, reassess the data or the model before inference.
This is different from merely observing few events. Sparse data can yield a flat likelihood, a large standard error, or a boundary estimate without making the observed path impossible. In the no-event rate example, the maximum occurs at zero while the observed path still has positive likelihood. By contrast, one event under a zero rate makes the likelihood at that rate exactly zero.
The main alternative is not a competing formula for the same support question, but a different observation model when the continuous-time intensity assumptions do not apply. For example, if event opportunities occur only at scheduled visits, conditional event probabilities at those visits may be more appropriate. Within a continuous-time model, the path likelihood is the natural way to assess support: its event factors expose prohibited observed events, while its integrated intensities account for event-free time.
Check Your Understanding
For each question, decide what the likelihood implies. Concise solutions follow the questions.
- A path has one observed type-1 event, and \(\lambda_1(t_j;\theta)=0\) at that event time for every allowable \(\theta\). What should be concluded?
- A nonnegative rate \(\theta\) has likelihood \(L(\theta)=\exp(-6\theta)\), with no observed events and 6 years of exposure. Find the maximum likelihood estimate and say whether the path is supported there.
- For the likelihood in Question 2, suppose one event is observed instead, so \(L(\theta)=\theta\exp(-6\theta)\). Find the maximum likelihood estimate for \(\theta\geq0\).
- Why is replacing an exact zero event intensity by a small positive number not a neutral numerical step?
- A fitted model gives a very small but positive event intensity at an observed transition. Does that alone imply zero likelihood?
Solutions. (1) The path has zero likelihood for every allowable parameter value. Because the zero is structural, check the state history, event coding, and model’s permitted transitions rather than treating this as a boundary estimate. (2) The likelihood is maximized at \(\hat{\theta}=0\), since \(\exp(-6\theta)\leq1\) for \(\theta\geq0\). The no-event path is supported there: its likelihood is \(1\). (3) For \(\theta>0\), the log-likelihood is \(\log\theta-6\theta\), with derivative \(1/\theta-6\). Setting this to zero gives \(\hat{\theta}=1/6\), or approximately \(0.167\) events per year. At \(\theta=0\), the event factor is zero, so that boundary value cannot maximize the likelihood. (4) The replacement changes the model’s support and may hide a prohibited event or data error; it must be treated as an explicit model modification, not a harmless calculation trick. (5) No. A positive intensity gives a positive event factor, although a very small one can make the likelihood small. Underflow in a computation should be distinguished from a mathematical zero.