= Poisson surrogate for an uncensored proportional-hazards likelihood
Consider <independent> event times with <hazard function> $h_i(t)=\lambda(t)e^{\beta^Tx_i}$ and baseline <cumulative hazard function> $\Lambda(t)=\int_0^t\lambda(s)\,ds$. For optimization over $\beta$, first hold $\lambda$ and $\Lambda$ fixed and independent of $\beta$. Under ignorable <independent censoring>, an observation $(y_i,\delta_i)$ has log-<likelihood contribution>
$$
\ell_i(\beta)=\delta_i\{\log\lambda(y_i)+\beta^Tx_i\}
-\Lambda(y_i)e^{\beta^Tx_i}.
$$
<Set> $\mu_i=\Lambda(y_i)e^{\beta^Tx_i}$. A <Poisson regression> with response $\delta_i$ and <log link> has log-<likelihood contribution> $\delta_i\log\mu_i-\mu_i-\log(\delta_i!)$, which differs from $\ell_i$ only by terms independent of $\beta$. Thus the same $\beta$ maximizes both criteria. The linear predictor is $\beta^Tx_i+\log\Lambda(y_i)$, with $\log\Lambda(y_i)$ a known <offset> whenever $0<\Lambda(y_i)<\infty$.
For uncensored observations every response is one. This is an equivalence of objective functions: the event indicators need not actually be Poisson counts. A related split-interval construction handles piecewise-constant <hazard functions>.<Survival Models by Germán Rodríguez>[Section 7.4, likelihood equivalence with Poisson regression]{ref}
If the baseline is unspecified, its integrated values are not known offsets; fitting an ordinary <generalized linear model> as if they were known does not estimate an arbitrary <Cox proportional-hazards model>. One must estimate the baseline jointly, use a suitable baseline representation, or eliminate it through <Cox partial likelihood>. Zero exposure also needs separate handling rather than taking the <logarithm> of zero.
Back to article page