Parametric Survival Models

Survival models that name the shape of the hazard, fitted by maximum likelihood with censoring and delayed entry. Proves that under type II censoring twice the rate times the total time on test is exactly chi-squared, and compares four intervals for an exponential rate; proves that the Weibull's profile likelihood in the shape has exactly one maximum; proves that the Weibull is both a proportional hazards and an accelerated failure time model and that it is the only such family when the covariate takes a range of values, with a counterexample for a single binary covariate; shows that the log-logistic model is a proportional odds model, that Cox–Snell residuals are unit exponential and, in a proportional hazards model with a free constant multiplying the hazard, sum to the number of events, and that a piecewise exponential model is a Poisson regression whose limit is Cox's model. Fits exponential, Weibull, Gompertz, log-logistic and log-normal models to 4,114 NHANES adults followed for twenty years, on time since examination and on age, compares them by AIC and with Kaplan–Meier, and fits the Gompertz law to the U.S. life table of 2019. Simulations (30 seeds) measure what a correct model gains over Kaplan–Meier for the median and what a wrong one costs. Worked problems on an exact interval for a mean life, a Weibull fit by hand, hazard and time ratios, a median remaining life under the Gompertz law and a proportional odds ratio.

The 4,114 adults of the NHANES cohort of The Cox Proportional Hazards Model were examined in 1999 or 2000 (National Center for Health Statistics, 2002) and followed for deaths to the end of 2019, through the National Death Index (National Center for Health Statistics, 2022): 1,240 died during 70,389.5 person-years of follow-up. The Cox model compared their death rates by smoking, diabetes, age and sex without saying anything about how the death rate changes over time. Adult mortality has a well-known shape, though: over much of adult life, death rates rise exponentially with age (Kirkwood, 2015). What is gained by building that shape into the model, and what is lost if the chosen shape is wrong?

A parametric survival model specifies the distribution of the survival time $T$, or of $T$ given covariates, up to a few parameters, and fits them by maximum likelihood from censored data. The hazard, survival function and cumulative hazard are those of Censoring, Survival Functions and Hazards, whose Theorem 8 gives the likelihood of censored data used throughout: $L(\theta) = \prod_i h(y_i; \theta)^{\delta_i}S(y_i; \theta)$, with each term divided by $S(v_i; \theta)$ when person $i$ enters observation late, at time $v_i$.

First we treat the exponential model: its estimate, an exact interval under one censoring design, and four intervals compared. Then comes the Weibull model, with a proof that its likelihood has one maximum, and the two ways covariates enter a parametric model, proportional hazards and accelerated failure time, with a proof that the Weibull is the only family of both kinds. Next we fit five families to the cohort, change the time scale to age, where the Gompertz law appears, and fit that law to the U.S. life table. Then we check fitted models with Cox–Snell residuals, choose among them by AIC, and connect the piecewise exponential model to Poisson regression and to Cox's model. Finally a simulation measures the trade between efficiency and robustness. The article assumes Maximum Likelihood Estimation and the tests of The Likelihood Ratio, Wald and Score Tests, and it is part of Survival Analysis.

The exponential model

The simplest parametric model has a constant hazard $\lambda$, so $S(t) = e^{-\lambda t}$. Censoring, Survival Functions and Hazards derives its maximum likelihood estimate from censored data: with $D = \sum_i\delta_i$ events and total time at risk $T = \sum_i y_i$,

$$\hat\lambda = \frac{D}{T},$$

with log-likelihood $\ell(\lambda) = D\ln\lambda - \lambda T$ and observed information $D/\lambda^2$, so a standard error of $\hat\lambda/\sqrt{D}$. This is the crude rate of Crude Rates and Their Precision, events per unit of person-time, now as the estimate of a parameter. It is also the estimate of the rate of a Poisson process observed for a total time $T$, whose gaps are exponential.

On the cohort, the 1,240 deaths in 844,674 person-months give $\hat\lambda = 17.62$ deaths per 1,000 person-years, with standard error 0.50. Under this model the median survival time would be $\ln 2/\hat\lambda = 39.3$ years. The next section shows that the model is wrong for this cohort; first, how precise is $\hat\lambda$ if it is right?

Four intervals for a rate.

  • The Wald interval $\hat\lambda \pm 1.96\,\hat\lambda/\sqrt{D}$: 16.64 to 18.60 per 1,000 person-years.
  • The log-Wald interval $\hat\lambda\exp(\pm 1.96/\sqrt{D})$, from the standard error $1/\sqrt{D}$ of $\ln\hat\lambda$ (the observed information on the log scale is $D$): 16.66 to 18.62.
  • The likelihood ratio interval, the set of $\lambda$ with $2[\ell(\hat\lambda) - \ell(\lambda)] \le 3.84$: 16.65 to 18.62.
  • The exact Poisson interval of Crude Rates and Their Precision, which treats $D$ as Poisson with mean $\lambda T$ and $T$ as fixed. Its limits are the 2.5th percentile of the gamma distribution with shape $D$ and the 97.5th percentile of the one with shape $D + 1$, divided by $T$: 16.65 to 18.62.

With 1,240 deaths all four agree to within 0.03. They differ when deaths are few. The exact Poisson interval is exact only when $T$ is fixed in advance, as for a Poisson process watched for a fixed time. When people are followed until they die or are censored, $T$ depends on the death times, and the interval is an approximation. In 30 simulated sets of 2,000 studies (seeds 85,000,001 to 85,000,030), each of 20 people with exponential lifetimes of rate 1 and exponential censoring of rate 0.5, the median coverage was 0.946 for the Wald interval (interquartile range 0.944 to 0.949), 0.946 for the log-Wald (0.944 to 0.949), 0.948 for the likelihood ratio interval (0.944 to 0.950) and 0.961 for the exact Poisson interval (0.958 to 0.963). Studies with no death (no estimate) were left out. In this design the three large-sample intervals were close to 95%, and the exact Poisson interval was conservative.

An exact pivot under type II censoring. One design does give an exact interval. In type II censoring, $n$ items are put on test together and the test stops at the $r$th failure; the other $n - r$ are censored at that moment. The total time on test is $T = \sum_{i \le r}X_{(i)} + (n - r)X_{(r)}$, where $X_{(1)} < \cdots < X_{(n)}$ are the ordered lifetimes. Censoring, Survival Functions and Hazards shows that the likelihood is again $\lambda^re^{-\lambda T}$ (Theorem 8, type II case), so $\hat\lambda = r/T$.

Theorem 1 (type II censoring). If $X_1, \ldots, X_n$ are independent exponential lifetimes with rate $\lambda$ and the test stops at the $r$th failure, then $2\lambda T$ has the chi-squared distribution with $2r$ degrees of freedom.

Here is the idea. While $m$ items are on test, the time to the next failure is the minimum of $m$ independent exponentials, which is exponential with rate $m\lambda$, and memorylessness makes the successive gaps independent. The total time on test adds up each gap multiplied by the number of items running during it, and each such product is exponential with rate $\lambda$.

Proof. The ordered lifetimes have joint density $n!\,\lambda^n\exp(-\lambda\sum_{i=1}^n x_{(i)})$ on $0 < x_{(1)} < \cdots < x_{(n)}$: each of the $n!$ assignments of the items to the ordered values contributes $\prod_i\lambda e^{-\lambda x_{(i)}}$. Define the spacings $s_i = x_{(i)} - x_{(i-1)}$ (with $x_{(0)} = 0$) and $e_i = (n - i + 1)s_i$. The map from the ordered values to the spacings is linear and triangular with ones on the diagonal, so its Jacobian is 1. The map from spacings to the $e_i$ multiplies each by $n - i + 1$, so the density picks up the factor $1/n!$. Since $\sum_i x_{(i)} = \sum_i (n - i + 1)s_i = \sum_i e_i$ (the spacing $s_i$ is part of the $n - i + 1$ lifetimes that have not ended before it), the density of $(e_1, \ldots, e_n)$ is $\prod_i\lambda e^{-\lambda e_i}$ on $e_i > 0$. So the $e_i$ are independent exponentials with rate $\lambda$. During the $i$th spacing, $n - i + 1$ items are on test, so $T = \sum_{i \le r}(n - i + 1)s_i = \sum_{i \le r}e_i$. Then $2\lambda e_i$ is exponential with mean 2, which is the chi-squared distribution with 2 degrees of freedom, and a sum of $r$ independent such variables is chi-squared with $2r$ degrees of freedom (The Chi-Squared and t Distributions derives the density; with $\nu$ degrees of freedom it is the gamma with shape $\nu/2$ and scale 2, and gammas with a common scale add their shapes). $\square$

The exact 95% interval for $\lambda$ is therefore $(\chi^2_{2r,\,0.025}/(2T),\ \chi^2_{2r,\,0.975}/(2T))$, where $\chi^2_{\nu, p}$ is the $p$ quantile. The script checked the pivot by simulation: for $n = 10$, $r = 7$, $\lambda = 0.2$ and 20,000 tests per seed, the Kolmogorov–Smirnov test of $2\lambda T$ against the chi-squared distribution with 14 degrees of freedom gave p-values with median 0.52 (interquartile range 0.13 to 0.67) over the 30 seeds 82,000,001 to 82,000,030, as uniform p-values should. Problem 1 uses the interval.

The Weibull model

Is the death rate of the cohort constant over the twenty years? Clearly not: the cohort ages, and the death rate of every member rises. A model whose hazard can rise or fall is needed, and the Weibull is the standard first choice.

The Weibull distribution with shape $k > 0$ and rate-type scale $\lambda > 0$ has

$$S(t) = \exp\{-(\lambda t)^k\}, \qquad H(t) = (\lambda t)^k, \qquad h(t) = k\lambda^kt^{k-1}.$$

The hazard rises with $t$ when $k > 1$, falls when $k < 1$, and is constant, the exponential, when $k = 1$. The distribution is named for Weibull (1951), whose paper presented it under the title "A statistical distribution function of wide applicability".

With censoring the log-likelihood is

$$\ell(k, \lambda) = D\ln k + Dk\ln\lambda + (k - 1)\sum_i\delta_i\ln y_i - \lambda^k\sum_i y_i^k.$$

For fixed $k$ it is maximized over $\lambda$ in closed form, which leaves a one-dimensional problem in the shape.

Theorem 2 (the Weibull estimate exists and is unique). Suppose all observed times $y_i$ are positive, $D \ge 1$, and some event time is smaller than the largest observed time $y_{\max}$. For fixed $k$ the likelihood is maximized at $\hat\lambda(k)^k = D/\sum_i y_i^k$. The profile log-likelihood $\ell_p(k) = \ell(k, \hat\lambda(k))$ is strictly concave, and its derivative

$$\ell_p'(k) = \frac{D}{k} + \sum_i\delta_i\ln y_i - D\,\frac{\sum_i y_i^k\ln y_i}{\sum_i y_i^k}$$

has exactly one root $\hat k$. So the maximum likelihood estimate $(\hat k, \hat\lambda(\hat k))$ exists and is unique.

Proof. Write $\rho = \lambda^k$. For fixed $k$, $\ell = D\ln k + D\ln\rho + (k - 1)\sum\delta_i\ln y_i - \rho\sum y_i^k$, a strictly concave function of $\rho > 0$ maximized where $D/\rho = \sum y_i^k$. Substituting gives $\ell_p(k) = D\ln k + D\ln D - D\ln\sum_i y_i^k + (k - 1)\sum_i\delta_i\ln y_i - D$, whose derivative is the stated one. Put $w_i = y_i^k/\sum_l y_l^k$. The derivative of $\sum_i y_i^k\ln y_i/\sum_i y_i^k = \sum_i w_i\ln y_i$ with respect to $k$ is $\sum_i w_i(\ln y_i)^2 - (\sum_i w_i\ln y_i)^2$, the variance of $\ln y$ under the weights $w$, which is nonnegative. So $\ell_p''(k) = -D/k^2 - D\operatorname{Var}_w(\ln y) < 0$, and $\ell_p'$ is strictly decreasing. As $k \to 0^+$, $D/k \to \infty$ while the other terms stay bounded, so $\ell_p' \to +\infty$. As $k \to \infty$, the weights concentrate on the observations with $y_i = y_{\max}$, so $\sum_i w_i\ln y_i \to \ln y_{\max}$ and $\ell_p'(k) \to \sum_i\delta_i(\ln y_i - \ln y_{\max})$, which is negative because some event time is below $y_{\max}$. A continuous, strictly decreasing function that goes from $+\infty$ to a negative limit has exactly one root, where $\ell_p$ has its unique maximum. $\square$

The script finds the root by bisection and computes standard errors from the observed information of $(\ln k, \ln\lambda)$. On the cohort, with time in months since the examination, $\hat k = 1.285$ (standard error 0.035, 95% interval 1.218 to 1.356 from the log scale) and $1/\hat\lambda = 44.2$ years. The log-likelihood rises from $-9{,}329.6$ for the exponential to $-9{,}290.6$ for the Weibull, a likelihood ratio statistic of 77.9 for the hypothesis $k = 1$ on 1 degree of freedom. The death rate rose over follow-up, as an aging cohort's must.

Survival of the 4,114 NHANES adults against years since the examination: the Kaplan–Meier estimate (steps) with the fitted exponential and Weibull curves. Notice that the exponential lies below the data for most of the follow-up and crosses it only after 18.8 years, because it forces a constant death rate on a cohort whose death rate rises, while the Weibull follows the steps closely.
Survival of the 4,114 NHANES adults against years since the examination: the Kaplan–Meier estimate (steps) with the fitted exponential and Weibull curves. Notice that the exponential lies below the data for most of the follow-up and crosses it only after 18.8 years, because it forces a constant death rate on a cohort whose death rate rises, while the Weibull follows the steps closely.

The two curves differ where it matters. At 5, 10 and 20 years the Kaplan–Meier estimates are 0.941, 0.859 and 0.697. The exponential gives 0.916, 0.838 and 0.703, too low early and slightly high late. The Weibull gives 0.941, 0.862 and 0.697.

Example (the residuals of the exponential sum to the number of deaths). At the exponential estimate, $\sum_i\hat\lambda y_i = \hat\lambda T = D$. The Weibull estimate has the same property, $\sum_i(\hat\lambda y_i)^{\hat k} = D$, because $\hat\lambda^{\hat k} = D/\sum y_i^{\hat k}$. The script confirms it on the cohort. The section on model checking explains why it holds for every proportional hazards model with a free constant multiplying the hazard (Proposition 2(b)). It fails for the log-normal and log-logistic models, whose free parameter rescales time rather than the hazard: with the covariates of the regression section, their residuals sum to 1,180.6 and 1,174.5, against 1,240 deaths.

Proportional hazards and accelerated failure time

Covariates can enter a parametric model in two natural ways. In a proportional hazards (PH) model they multiply the hazard, $h(t \mid \mathbf{x}) = h_0(t)\exp(\mathbf{x}^{\mathsf T}\boldsymbol\beta)$, as in the Cox model, so that $S(t \mid \mathbf{x}) = S_0(t)^{\exp(\mathbf{x}^{\mathsf T}\boldsymbol\beta)}$. In an accelerated failure time (AFT) model they stretch or shrink time,

$$S(t \mid \mathbf{x}) = S_0\big(t\,e^{-\mathbf{x}^{\mathsf T}\boldsymbol\gamma}\big), \qquad\text{equivalently}\qquad \ln T = \mu + \mathbf{x}^{\mathsf T}\boldsymbol\gamma + \sigma W,$$

where $W$ is a random error with a fixed distribution, $\mu$ an intercept and $\sigma > 0$ a scale. The equivalence: $P(T > t \mid \mathbf{x}) = P(\mu + \sigma W > \ln t - \mathbf{x}^{\mathsf T}\boldsymbol\gamma)$, which depends on $t$ and $\mathbf{x}$ only through $te^{-\mathbf{x}^{\mathsf T}\boldsymbol\gamma}$. Under an AFT model, $e^{\gamma_j}$ is a time ratio: a unit increase in $x_j$ multiplies every quantile of $T$ by $e^{\gamma_j}$. The AFT model is a linear regression for $\ln T$, with censored responses.

Theorem 3 (the Weibull is both, and the only family that is). (a) If $S_0(t) = \exp\{-(\lambda t)^k\}$, then the AFT model with coefficients $\boldsymbol\gamma$ is the PH model with $\boldsymbol\beta = -k\boldsymbol\gamma$, and $\ln T = -\ln\lambda + \sigma W$ with $\sigma = 1/k$ and $P(W > w) = \exp(-e^w)$, the minimum extreme value distribution. (b) Conversely, let $S_0$ be continuous with $0 < S_0(t) < 1$ for $t > 0$. Suppose that for every $c$ in an open interval $J$ containing 1 there is a number $\psi(c) > 0$ with $S_0(ct) = S_0(t)^{\psi(c)}$ for all $t > 0$, that is, stretching time by $c$ is the same as multiplying the hazard by $\psi(c)$. Then $S_0(t) = \exp(-\rho t^k)$ for some $\rho > 0$ and $k > 0$: the Weibull.

The covariate in (b) acts through $c = e^{-\mathbf{x}^{\mathsf T}\boldsymbol\gamma}$, and $c$ fills an interval around 1 when $\mathbf{x}^{\mathsf T}\boldsymbol\gamma$ takes all values near 0, as it does for a continuous covariate such as age. The idea of the proof is to take logarithms twice, which turns the condition into Cauchy's functional equation, whose only monotone solutions are linear.

Proof. (a) $S_0(te^{-\mathbf{x}^{\mathsf T}\boldsymbol\gamma}) = \exp\{-(\lambda t)^ke^{-k\mathbf{x}^{\mathsf T}\boldsymbol\gamma}\} = S_0(t)^{\exp(-k\mathbf{x}^{\mathsf T}\boldsymbol\gamma)}$, a PH model with $\boldsymbol\beta = -k\boldsymbol\gamma$. For the error form, $P(\ln T > s) = \exp\{-(\lambda e^s)^k\} = \exp\{-e^{k(s + \ln\lambda)}\}$, so $W = k(\ln T + \ln\lambda)$ has $P(W > w) = \exp(-e^w)$, and $\ln T = -\ln\lambda + W/k$.

(b) Let $H_0 = -\ln S_0$, which is continuous, nondecreasing (as $S_0$ is nonincreasing) and positive on $(0, \infty)$, and let $G(u) = \ln H_0(e^u)$ for real $u$, also continuous and nondecreasing. Taking logarithms of $-\ln S_0(ct) = \psi(c)\{-\ln S_0(t)\}$ at $t = e^u$ and $c = e^a$ gives

$$G(u + a) - G(u) = \ln\psi(e^a)\qquad\text{for all real }u\text{ and all }a \in A = \{\ln c : c \in J\},$$

an open interval containing 0, say containing $(-\varepsilon, \varepsilon)$. Define $\Delta(a, u) = G(u + a) - G(u)$. For $|a| < \varepsilon$ it does not depend on $u$. For any real $a$, choose a whole number $N$ with $|a|/N < \varepsilon$; then $\Delta(a, u) = \sum_{m=0}^{N-1}\Delta(a/N, u + ma/N)$, a sum of $N$ terms each equal to $\Delta(a/N, 0)$, so $\Delta(a, u)$ does not depend on $u$ for any $a$. Write it $\Delta(a)$. Then $\Delta(a + b) = G(u + a + b) - G(u + b) + G(u + b) - G(u) = \Delta(a) + \Delta(b)$, and $\Delta$ is nondecreasing because $G$ is. Additivity gives $\Delta(m) = m\Delta(1)$ for whole numbers $m$ and $\Delta(p/q) = (p/q)\Delta(1)$ for rationals, since $q\Delta(p/q) = \Delta(p) = p\Delta(1)$. For real $a$, rationals $p_1 < a < p_2$ give $p_1\Delta(1) \le \Delta(a) \le p_2\Delta(1)$ by monotonicity, and letting $p_1, p_2 \to a$ gives $\Delta(a) = ka$ with $k = \Delta(1) \ge 0$. So $G(u) = G(0) + ku$, which means $H_0(t) = \rho t^k$ with $\rho = H_0(1) > 0$. Finally $k > 0$: if $k = 0$, $H_0$ would equal $\rho > 0$ on all of $(0, \infty)$, but continuity of $S_0$ at 0, where $S_0(0) = 1$, forces $H_0(t) \to 0$ as $t \to 0$. $\square$

Example (one binary covariate is not enough). Condition (b) for a single factor $c$ does not force the Weibull. Take $H_0(t) = t\exp\{\varepsilon\sin(2\pi\ln t/\ln 2)\}$ with $\varepsilon = 0.1$. Then $H_0(2t) = 2H_0(t)$ exactly, since the sine has period $\ln 2$ in $\ln t$, so for a binary covariate that doubles time the AFT model and the PH model with hazard ratio 2 coincide. The function is increasing because $d\ln H_0/d\ln t = 1 + \varepsilon(2\pi/\ln 2)\cos(\cdot) > 0$ whenever $\varepsilon < \ln 2/(2\pi) = 0.110$. It is not a Weibull: the ratio $H_0(3t)/H_0(t)$ ranges from 2.47 to 3.64 as $t$ varies, so stretching time by 3 is not a proportional change of hazard. Theorem 3 needs the stretch factors to fill an interval.

So the Gompertz model below, which is PH but not Weibull, is not an AFT model; and the log-normal and log-logistic AFT models are not PH. Two AFT families are in common use besides the Weibull:

  • Log-logistic: $W$ has the logistic distribution, $S(t \mid \mathbf{x}) = 1/(1 + e^{z})$ with $z = (\ln t - \mu - \mathbf{x}^{\mathsf T}\boldsymbol\gamma)/\sigma$. For $\sigma < 1$ its hazard rises and then falls: with $k = 1/\sigma$ and $\lambda = e^{-\mu - \mathbf{x}^{\mathsf T}\boldsymbol\gamma}$ the hazard is $k\lambda^kt^{k-1}/\{1 + (\lambda t)^k\}$, and $\frac{d}{dt}\ln h = \{(k - 1) - (\lambda t)^k\}/[t\{1 + (\lambda t)^k\}]$ is positive for small $t$ and negative once $(\lambda t)^k > k - 1$.
  • Log-normal: $W$ is standard normal, $S(t \mid \mathbf{x}) = 1 - \Phi(z)$.

Proposition 1 (the log-logistic model has proportional odds). In the log-logistic AFT model the odds of failure by time $t$, $(1 - S)/S$, equal $e^{(\ln t - \mu)/\sigma}\exp(-\mathbf{x}^{\mathsf T}\boldsymbol\gamma/\sigma)$, so two covariate values have a constant odds ratio $\exp\{-(\mathbf{x} - \mathbf{x}')^{\mathsf T}\boldsymbol\gamma/\sigma\}$ at every $t$.

Proof. $(1 - S)/S = e^z = e^{(\ln t - \mu)/\sigma}e^{-\mathbf{x}^{\mathsf T}\boldsymbol\gamma/\sigma}$, and the factor depending on $t$ cancels in the ratio. $\square$

Regression: the cohort with covariates

Back to the cohort, now with the covariates of the Cox article: age in decades from 60, male sex, current and former smoking, and diabetes. With time in years since the examination, the Weibull PH model has shape $\hat k = 1.587$ and these coefficients, set against the Cox model's (Efron's ties):

Covariate Weibull hazard ratio (SE of $\hat\beta$) Cox hazard ratio (SE) Weibull time ratio
Age, per 10 years 2.38 (0.023) 2.43 (0.024) 0.58
Male 1.30 (0.059) 1.31 (0.059) 0.85
Current smoker 2.26 (0.078) 2.27 (0.078) 0.60
Former smoker 1.11 (0.067) 1.12 (0.067) 0.93
Diabetes 1.85 (0.069) 1.86 (0.069) 0.68

The two models agree closely on the hazard ratios, and their standard errors differ by at most 0.001. Here the parametric baseline bought no visible precision for the coefficients. That fits Efron's (1977) finding that the partial likelihood loses little information under realistic conditions, cited in the Cox article. The shape is where a parametric model adds something: a smooth baseline hazard with three or four parameters instead of a step at every death, and a survival curve that can be extrapolated, at the model's risk.

The last column reads the same fit as an AFT model by Theorem 3, with time ratio $e^{\gamma_j} = e^{-\beta_j/k}$. A current smoker's survival time is 0.60 times that of an otherwise similar never smoker, at every quantile. The hazard ratio and the time ratio are two descriptions of one fitted model, not two findings.

Age as the time scale: the Gompertz law

Time since the examination mixes people of all ages. With age as the time scale, as recommended for survey cohorts by Korn, Graubard and Midthune (1997), each person enters observation at the age of examination and leaves at death or at the end of 2019. The likelihood divides each person's contribution by $S(v_i \mid \mathbf{x}_i)$ at the entry age $v_i$ (delayed entry), so the log-likelihood is $\sum_i\{\delta_i\ln h(y_i \mid \mathbf{x}_i) - H(y_i \mid \mathbf{x}_i) + H(v_i \mid \mathbf{x}_i)\}$. The examination age in months (RIDAGEEX) is missing for 12 people, leaving 4,102.

The Gompertz model has hazard

$$h(t) = a\,e^{bt}, \qquad H(t) = \frac{a}{b}\left(e^{bt} - 1\right),$$

a death rate that grows exponentially with age at rate $b$, doubling every $\ln 2/b$ years. Gompertz (1825) proposed the law in a paper titled "On the nature of the function expressive of the law of human mortality", writing the number living at age $x$ as $d\,g^{q^x}$, which is the same law (its log-hazard is linear in $x$); the hazard form above came later (Kirkwood, 2015). Kirkwood summarizes that the paper showed age-specific death rates rising exponentially over much of adult life, and that other patterns are now recognized in advanced old age.

Fitted with sex, smoking and diabetes as PH covariates, the Gompertz model on age gives $\hat b = 0.0895$ per year (standard error 0.0023): the death rate doubled every 7.75 years of age. The hazard ratios are 1.31 for men, 2.27 for current smokers, 1.12 for former smokers and 1.86 for diabetes. A Cox model on the same age scale, with the same delayed entry, gives 1.32, 2.30, 1.13 and 1.90. The Weibull on age needs a shape of 7.07 to mimic the steep rise, and fits worse (next sections).

Death rates by age in the U.S. life table of 2019 (points, log scale) with the fitted Gompertz line, and the Gompertz baseline fitted to the NHANES cohort (women who never smoked and have no diagnosed diabetes). Notice that the life-table rates follow a straight line on the log scale from about 50 to 90, rise above it at younger ages, and that the NHANES line runs parallel to it, lower because its baseline group is at low risk.
Death rates by age in the U.S. life table of 2019 (points, log scale) with the fitted Gompertz line, and the Gompertz baseline fitted to the NHANES cohort (women who never smoked and have no diagnosed diabetes). Notice that the life-table rates follow a straight line on the log scale from about 50 to 90, rise above it at younger ages, and that the NHANES line runs parallel to it, lower because its baseline group is at low risk.

The life table. The Gompertz law can also be fitted to the U.S. life table of 2019 (Arias and Xu, 2022), whose construction is the subject of Life Tables. For ages 30 to 95, the table gives the deaths ${}_1d_x$ and the person-years ${}_1L_x$ of its hypothetical cohort, and their ratio is the death rate ${}_1m_x$. Fitting $\ln m_x = \alpha + b(x - 60)$ by the Poisson likelihood with offset $\ln L_x$ (the form of Poisson Regression) gives $\hat b = 0.0910$, a doubling every 7.62 years, close to the NHANES estimate. The fitted rate at 60 is 8.64 per 1,000 against the table's 9.13. The ratio of the table's rate to the fitted one is 2.33 at age 30, 1.42 at 40, 1.12 at 50, 0.85 at 70, 0.89 at 80, 1.13 at 90 and 1.21 at 95. From 50 to 90 the table's rates stay within 17% of the line. The law describes the middle of adult life, misses young adults badly (accidents, overdoses and other causes of death that do not rise with age), and curves upward at the oldest ages. The life-table "deaths" are not observed counts, so this fit has no standard errors; it describes the shape of the table's rates. The NHANES baseline group has a fitted death rate of 5.64 per 1,000 at 60, below the table's, as expected for women who never smoked and do not have diagnosed diabetes.

Checking a fitted model

A fitted parametric model claims a whole distribution for every person. One check transforms each observed time by the person's own fitted cumulative hazard.

Proposition 2 (Cox–Snell residuals). (a) If $T$ has a continuous distribution with cumulative hazard $H$, and $H(t) \to \infty$ as $t \to \infty$ (no mass at infinity, true for every family on this page), then $H(T)$ is exponential with rate 1. Given $T > v$, $H(T) - H(v)$ is exponential with rate 1. (b) In a parametric PH model $H(t \mid \mathbf{x}) = \exp(\beta_0 + \mathbf{x}^{\mathsf T}\boldsymbol\beta)H_*(t)$ with a free intercept $\beta_0$, the residuals $\hat r_i = \hat H(y_i \mid \mathbf{x}_i) - \hat H(v_i \mid \mathbf{x}_i)$ at the maximum likelihood estimate sum to the number of events $D$.

Proof. (a) For $s \ge 0$, $P(H(T) > s) = P(T > H^{-1}(s)) = S(H^{-1}(s)) = e^{-s}$, where $H^{-1}(s) = \inf\{t : H(t) \ge s\}$ is finite because $H(t) \to \infty$, and the second equality uses the continuity of $H$. The first equality holds up to an event of probability zero: $H(T) > s$ implies $T > H^{-1}(s)$, and the converse fails only if $T$ falls in a stretch beyond $H^{-1}(s)$ where $H$ stays equal to $s$, on which $S$ is constant, so $T$ falls there with probability zero; this is the probability integral transform of Random Variables and Distribution Functions applied to $S$. Given $T > v$, $P(H(T) - H(v) > s \mid T > v) = S(H^{-1}(H(v) + s))/S(v) = e^{-H(v) - s}/e^{-H(v)} = e^{-s}$. (b) The log-likelihood with delayed entry is $\sum_i\{\delta_i\ln h(y_i \mid \mathbf{x}_i) - H(y_i \mid \mathbf{x}_i) + H(v_i \mid \mathbf{x}_i)\}$. The intercept enters $\ln h$ additively and multiplies $H$ by $e^{\beta_0}$, so the derivative in $\beta_0$ is $D - \sum_i\{H(y_i \mid \mathbf{x}_i) - H(v_i \mid \mathbf{x}_i)\}$, which vanishes at the estimate. $\square$

So if the model is right, the pairs $(\hat r_i, \delta_i)$ should look like a censored sample from the unit exponential, apart from the error in the estimates (the observed $\hat r_i$ are not exactly exponential, since the parameters were fitted to the same data). Residuals of the kind that Cox and Snell (1968) defined in general, applied in this way, are called Cox–Snell residuals. The check computes the Nelson–Aalen estimate of the cumulative hazard of the residuals, treating each residual as a censored or uncensored time (The Kaplan–Meier Estimator), and compares it with the line of slope 1. For delayed entry, person $i$ is at risk from $\hat H(v_i \mid \mathbf{x}_i)$ to $\hat H(y_i \mid \mathbf{x}_i)$.

Nelson–Aalen estimates of the cumulative hazard of Cox–Snell residuals for three fitted models, against the line of slope 1 that a correct model would follow. Notice that all three follow the line up to a residual of about 1 and then run a little above it, where few people remain, and that the exponential model on age is the first to leave the line, from about 0.4.
Nelson–Aalen estimates of the cumulative hazard of Cox–Snell residuals for three fitted models, against the line of slope 1 that a correct model would follow. Notice that all three follow the line up to a residual of about 1 and then run a little above it, where few people remain, and that the exponential model on age is the first to leave the line, from about 0.4.

At a residual of 0.5 the Nelson–Aalen curve is 0.479 for the Weibull on time since examination, 0.501 for the Gompertz on age and 0.544 for the exponential on age, against 0.5 for a perfect fit. The plot is a weak check: the exponential on age, which AIC rejects by more than 1,900 units against the Gompertz below, still lies near the line. One plausible reason is that with delayed entry each residual covers only the stretch of age over which the person was observed, about twenty years at most, and over such a stretch a constant hazard is a rough but not absurd description. The plot can miss a misfit that the likelihood detects. The Weibull residuals sum to 1,240.0, the number of deaths, as part (b) requires. A comparison of a fitted survival curve with the Kaplan–Meier estimate, as in the first figure, is the other standard check; it works only for groups large enough to have their own Kaplan–Meier curve.

Choosing among the models

Several families can pass a residual check. The Akaike Information Criterion compares models fitted by maximum likelihood by $\mathrm{AIC} = -2\hat\ell + 2k$, where $k$ counts the parameters: smaller is better, and differences of a few units matter. It estimates the expected predictive divergence, the quantity that article derives, and does not need the models to be nested.

With the five covariates on time since examination (each model has the covariates plus one or two shape and scale parameters):

Model Log-likelihood AIC
Exponential (PH) −5,270.8 10,553.6
Weibull (PH) −5,128.9 10,271.7
Gompertz (PH) −5,112.6 10,239.1
Log-logistic (AFT) −5,175.4 10,364.9
Log-normal (AFT) −5,284.2 10,582.3

On age as the time scale, with sex, smoking and diabetes, the AIC values are 12,116.5 for the exponential, 10,247.2 for the Weibull, 10,189.0 for the Gompertz, 10,411.5 for the log-logistic and 10,600.0 for the log-normal.

On both scales the Gompertz is best and the Weibull second. On time since examination the Gompertz rate is $\hat b = 0.0924$ per year, close to its value on age: the cohort's death rate rises over follow-up mainly because its members age. The AFT families do worse. The log-normal is the worst of the five on time since examination, with $\hat\sigma = 1.066$, and the log-logistic, with $\hat\sigma = 0.523$, is in between. Their time ratios for current smoking, 0.54 for the log-normal and 0.55 for the log-logistic, differ from the Weibull's 0.60, a reminder that a time ratio depends on the error distribution assumed. The values on the two scales cannot be compared with each other: the age-scale likelihood is conditional on the entry ages and is a different function of different data.

Piecewise exponential models are Poisson regressions

Between the single-shape families and Cox's model lies the piecewise exponential model. Cut time into pieces $I_m = (\tau_{m-1}, \tau_m]$ and let the baseline hazard be constant on each, $h(t \mid \mathbf{x}) = \lambda_m\exp(\mathbf{x}^{\mathsf T}\boldsymbol\beta)$ for $t \in I_m$.

Proposition 3 (the piecewise exponential model as a Poisson regression). (a) Let $T_{im}$ be the time person $i$ spends at risk in piece $m$ and $d_{im}$ the indicator that person $i$ dies in piece $m$. Under independent, noninformative censoring, the likelihood of the piecewise exponential PH model is proportional to the likelihood of independent Poisson counts $d_{im}$ with means $\mu_{im} = T_{im}\exp(\alpha_m + \mathbf{x}_i^{\mathsf T}\boldsymbol\beta)$, $\alpha_m = \ln\lambda_m$: a Poisson regression with offset $\ln T_{im}$, a factor for the piece and the person's covariates. (b) If the pieces are the intervals between consecutive distinct event times, $(t_{(m-1)}, t_{(m)}]$, and each member of the risk set $R(t_{(m)})$ is credited with the whole piece, of length $\ell_m$, while no one else is at risk in it, then the profile likelihood of $\boldsymbol\beta$ is Breslow's partial likelihood up to a constant.

Proof. (a) This is Theorem 1 of Poisson Regression with each (person, piece) pair as a cell: that proof shows that a person who spends time $\tau_c$ in cells with constant rates $\lambda_c$ contributes $\prod_c\lambda_c^{\delta_c}e^{-\lambda_c\tau_c}$, and nothing in it requires the rate of a cell to be shared by everyone in it. With the rate $\lambda_me^{\mathbf{x}_i^{\mathsf T}\boldsymbol\beta}$ for person $i$ in piece $m$, the likelihood is $\prod_{i,m}(\lambda_me^{\mathbf{x}_i^{\mathsf T}\boldsymbol\beta})^{d_{im}}e^{-\lambda_me^{\mathbf{x}_i^{\mathsf T}\boldsymbol\beta}T_{im}}$, which differs from the product of Poisson probabilities $e^{-\mu_{im}}\mu_{im}^{d_{im}}/d_{im}!$ only by the factor $\prod T_{im}^{d_{im}}/d_{im}!$, free of the parameters. (b) Under the stated crediting, $T_{im} = \ell_m\mathbf{1}\{i \in R(t_{(m)})\}$. With $D_m$ deaths in piece $m$, the log-likelihood is $\sum_i\delta_i\mathbf{x}_i^{\mathsf T}\boldsymbol\beta + \sum_m\{D_m\ln\lambda_m - \lambda_m\ell_m\sum_{l \in R(t_{(m)})}r_l\}$ with $r_l = e^{\mathbf{x}_l^{\mathsf T}\boldsymbol\beta}$. Maximizing each term over $\lambda_m$ gives $\hat\lambda_m = D_m/(\ell_m\sum_{R(t_{(m)})}r_l)$, and substituting leaves $\sum_i\delta_i\mathbf{x}_i^{\mathsf T}\boldsymbol\beta - \sum_mD_m\ln\sum_{R(t_{(m)})}r_l$ plus terms free of $\boldsymbol\beta$: Breslow's log partial likelihood. $\square$

Part (b) is the profile argument of The Cox Proportional Hazards Model (Theorem 2) in the clothing of a Poisson regression. Holford (1980) showed that rates log-linear in covariates lead to the same estimating equations for data from a Poisson process, an exponential distribution, a survival model or a log-linear model, and Laird and Olivier (1981) that approaches including Cox's "can be handled by using existing computer packages developed for the log-linear analysis of contingency table data" (both abstracts). Censoring, Survival Functions and Hazards gives the case without covariates, where the estimate of each $\lambda_m$ is the life table's death rate.

On the cohort, with yearly pieces (21 of them, 72,459 person-year rows, each person's exposure in months), the Poisson regression gives coefficients 0.885, 0.266, 0.819, 0.109 and 0.622 for age, sex, current and former smoking and diabetes, within 0.002 of the Cox estimates. With one piece for each of the 243 months in which someone died, and exposure credited as in part (b), a general-purpose optimizer applied to the Poisson likelihood with 248 parameters reproduced Breslow's Cox estimates to five decimals.

Efficiency and robustness: a simulation

When the model is right, a parametric estimate uses the shape of the distribution as information; when it is wrong, it reports a quantity that does not exist. The median survival time is a convenient target, because the Kaplan–Meier estimate gives it without any model (The Kaplan–Meier Estimator, section "The median and its interval"). The parametric median is $\hat\lambda^{-1}(\ln 2)^{1/\hat k}$ for the Weibull and $\ln 2/\hat\lambda$ for the exponential.

Each of four designs drew 30 seeds of 1,000 samples, each sample of 200 lifetimes with independent exponential censoring:

  • A, Weibull lifetimes with shape 1.5 and scale 1, censoring rate 0.3 (23% censored), estimated by Kaplan–Meier and the Weibull (seeds 81,000,001 to 81,000,030);
  • B, exponential lifetimes with rate 1, censoring rate 0.3 (23% censored), estimated by Kaplan–Meier, the Weibull and the exponential (seeds 81,100,001 to 81,100,030);
  • C, log-normal lifetimes ($\ln T$ standard normal), censoring rate 0.15 (19% censored), estimated by Kaplan–Meier and a Weibull, which is wrong (seeds 83,000,001 to 83,000,030);
  • D, Weibull lifetimes with shape 2, censoring rate 0.3 (23% censored), estimated by Kaplan–Meier and an exponential, which is wrong (seeds 84,000,001 to 84,000,030).

For each seed the script computed the relative bias of each estimate of the median (mean over the 1,000 samples, as a percentage of the true median) and the efficiency of the parametric estimate relative to Kaplan–Meier, the ratio of mean squared errors $\mathrm{MSE}_{\mathrm{KM}}/\mathrm{MSE}_{\mathrm{param}}$. Values above 1 favor the parametric model. Medians over the 30 seeds, with interquartile ranges:

Design Kaplan–Meier bias Parametric bias Efficiency (MSE ratio)
A, Weibull right 0.3% (0.1 to 0.4) Weibull 0.2% (0.1 to 0.3) 1.56 (1.51 to 1.58)
B, exponential right 0.4% (0.3 to 0.6) Weibull 0.3% (0.1 to 0.5) 1.54 (1.49 to 1.58)
B, exponential right exponential 0.1% (0.0 to 0.2) 1.81 (1.75 to 1.86)
C, log-normal, Weibull wrong 0.6% (0.2 to 0.7) Weibull 15.3% (15.2 to 15.4) 0.29 (0.28 to 0.30)
D, Weibull, exponential wrong 0.1% (−0.1 to 0.2) exponential −18.8% (−18.9 to −18.7) 0.09 (0.08 to 0.09)

In these designs a correct model estimated the median with a mean squared error about 1.5 to 1.8 times smaller than Kaplan–Meier's, the equivalent of 50% to 80% more data. The correct one-parameter model did better than the two-parameter Weibull that contains it. A wrong model was biased by 15% to 19%, and its mean squared error was 3.4 (design C) and 12 (design D) times Kaplan–Meier's. The experiment used one sample size, 200, so it does not itself show what happens as $n$ grows. Under the regularity conditions for misspecified maximum likelihood, a fact this site does not prove, a wrong model's estimates converge to the parameter that maximizes the expected log-likelihood of the censored data, and its median then converges to that member's median, not to the true one. The limit depends on the censoring distribution as well as on the lifetimes. The script computes these limits by numerical integration: a bias of −18.86% for design D and +14.85% for design C, close to the −18.8% and 15.3% observed at $n = 200$. If that convergence holds, the bias does not shrink with $n$: the gain from a right model is a constant factor, and a wrong model's mean squared error relative to Kaplan–Meier's grows in proportion to $n$, since its squared bias stays while Kaplan–Meier's variance falls. That asymmetry is the case for checking a parametric model before using it, and for Kaplan–Meier and Cox's model where no model can be checked well.

What is proved here, what is cited, and what the experiments show

Proved on this page: the exact chi-squared pivot under type II censoring (Theorem 1); the existence and uniqueness of the Weibull estimate through the concave profile likelihood (Theorem 2); the Weibull as both a PH and an AFT model, and the only such family when the time-stretch factors fill an interval, with a counterexample for a single factor (Theorem 3); the proportional odds property of the log-logistic model (Proposition 1); the unit exponential distribution of the cumulative hazard at the lifetime, also under delayed entry, and the sum of Cox–Snell residuals (Proposition 2); the piecewise exponential model as a Poisson regression and its profile as Breslow's partial likelihood (Proposition 3).

Cited: the censored-data likelihood, the exponential estimate and its standard error, and the piecewise-constant estimate without covariates (Censoring, Survival Functions and Hazards); the person-time likelihood (Theorem 1 of Poisson Regression); the exact Poisson limits (Crude Rates and Their Precision); the large-sample theory of maximum likelihood estimates and tests (Asymptotics of the Maximum Likelihood Estimator and the tests article), applied here to likelihoods of censored data without a separate proof; the AIC (The Akaike Information Criterion); the Gompertz law as Kirkwood (2015) summarizes Gompertz (1825); the time scale (Korn, Graubard and Midthune, 1997); the log-linear treatment of survival data (Holford, 1980; Laird and Olivier, 1981), at the strength of their abstracts. Not proved and not cited: the convergence of a misspecified maximum likelihood estimate to the maximizer of the expected censored log-likelihood, used only to read the computed limits of the simulation biases.

What the experiments show: on the NHANES cohort, the Weibull fits time since examination much better than the exponential, the Gompertz fits best on both time scales by AIC, and the parametric hazard ratios agree with Cox's to within 0.05. The Gompertz slope of the cohort, 0.0895 per year, is close to that of the U.S. life table, 0.0910. In the simulations, correct models estimated the median more efficiently than Kaplan–Meier by a factor of 1.5 to 1.8, and wrong models were biased by 15% to 19%; the four interval methods for a rate covered at 0.946 to 0.961 in one small design. These are observations on one cohort and four simulation designs.

Worked problems

The problems use this article's NHANES fits and two small datasets made up for hand computation (one of them the machine data of The EM Algorithm, Problem 2); they come from no course assignment. The script tools/experiments/statistics/parametric-survival-models.py checks every number.

Problem 1 (an exact interval for a mean life). Ten machines are tested together until the seventh fails, which happens at 5 thousand hours; the seven failure times total 21 thousand hours. Assuming exponential lifetimes, estimate the mean life with an exact 95% interval, and compare with the log-Wald interval.

Solution. The total time on test is $T = 21 + 3 \times 5 = 36$ thousand hours and $r = 7$, so the mean life $\theta = 1/\lambda$ is estimated by $T/r = 5.143$ thousand hours. By Theorem 1, $2T/\theta$ is chi-squared with 14 degrees of freedom, whose 2.5% and 97.5% quantiles are 5.629 and 26.119. Then $P(5.629 \le 2T/\theta \le 26.119) = 0.95$, and solving for $\theta$ gives the exact interval $(72/26.119,\ 72/5.629) = (2.76, 12.79)$ thousand hours. The log-Wald interval $5.143\exp(\pm 1.96/\sqrt 7)$ is $(2.45, 10.79)$, too far to the left: with seven failures the distribution of $\hat\theta$ is still skewed. The same numbers appear in The EM Algorithm under type I censoring at 5 thousand hours, where the estimate is the same but Theorem 1 does not apply, because the number of failures is random there.

Problem 2 (a Weibull fit with five observations). Fit exponential and Weibull models to the times 2, 3, 5+, 7, 8+ (a plus marks censoring), and test $k = 1$.

Solution. $D = 3$ and $T = 25$, so the exponential gives $\hat\lambda = 3/25 = 0.12$. For the Weibull, $\sum\delta_i\ln y_i = \ln 2 + \ln 3 + \ln 7 = \ln 42 = 3.7377$. At $k = 1$ the profile score of Theorem 2 is $3 + 3.7377 - 3\sum y_i\ln y_i/\sum y_i = 1.579 > 0$, so by the theorem's monotonicity $\hat k > 1$. Bisection on the score gives $\hat k = 1.711$, and then $\hat\lambda = (3/\sum y_i^{\hat k})^{1/\hat k} = 0.1383$. The likelihood ratio statistic for $k = 1$ is 0.950, well below 3.84: five observations cannot distinguish a rising hazard from a constant one.

Problem 3 (from a hazard ratio to a time ratio). In the Weibull PH fit to the cohort, the coefficient of current smoking is 0.814 and the shape 1.587. Give the hazard ratio and the time ratio, and say what each means.

Solution. The hazard ratio is $e^{0.814} = 2.26$: at every time, a current smoker's death rate is 2.26 times that of an otherwise similar never smoker. By Theorem 3(a), $\gamma = -\beta/k = -0.814/1.587$, so the time ratio is $e^{-0.814/1.587} = 0.60$: every quantile of a current smoker's survival time is 0.60 times the corresponding quantile for the never smoker. Both statements describe the same fitted distribution; the time ratio is smaller in magnitude because the hazard rises over time ($k > 1$), so a modest shortening of time produces a large multiplication of the hazard.

Problem 4 (median remaining life under the Gompertz law). Using the NHANES Gompertz fit on age ($\hat b = 0.0895$ per year, baseline hazard 5.64 per 1,000 at 60 for a woman who never smoked and has no diagnosed diabetes, hazard ratio 2.27 for current smoking), find the median remaining life at 60 for the baseline woman and for a current smoker.

Solution. Write $h_{60}$ for the hazard at 60. The cumulative hazard from 60 to $60 + s$ is $\int_0^s h_{60}e^{bu}\,du = (h_{60}/b)(e^{bs} - 1)$, and the median remaining life solves $(h_{60}/b)(e^{bs} - 1) = \ln 2$, so $s = b^{-1}\ln(1 + b\ln 2/h_{60})$. For the baseline woman, $h_{60} = 0.00564$, which gives 27.8 years. For a current smoker, $h_{60} = 2.27 \times 0.00564 = 0.01282$ (12.82 per 1,000), which gives 19.7 years. Under a Gompertz PH model the hazard ratio of 2.27 is a shift of age by $\ln 2.27/b$ years, about nine years, which is why the median remaining life falls by about eight years: the smoker ages, in hazard terms, nine years ahead.

Problem 5 (a proportional odds ratio). The log-logistic AFT fit on time since examination has $\hat\sigma = 0.523$ and a coefficient $-0.596$ for current smoking. Find the odds ratio of death by any given time for current against never smokers.

Solution. By Proposition 1 the odds ratio is $\exp(-\gamma/\sigma) = \exp(0.596/0.523) = 3.13$, the same at every time. The time ratio is $e^{-0.596} = 0.55$. Under this model, at every $t$ the odds that a current smoker has died by $t$ are 3.13 times those of a never smoker. The model's AIC is 93.1 units worse than the Weibull's on these data, so this description fits less well than the proportional hazards one.

Summary

Result Statement
Exponential $\hat\lambda = D/T$, the crude rate; intervals Wald, log-Wald, likelihood ratio, exact Poisson
Type II censoring $2\lambda T \sim \chi^2_{2r}$ exactly (Theorem 1)
Weibull $S = e^{-(\lambda t)^k}$; profile in $k$ strictly concave, unique estimate (Theorem 2)
PH and AFT Weibull is both, with $\boldsymbol\beta = -k\boldsymbol\gamma$; the only family when the stretch factors fill an interval (Theorem 3)
Log-logistic proportional odds, odds ratio $e^{-\gamma/\sigma}$ (Proposition 1)
Cox–Snell residuals $H(T) \sim$ unit exponential; in a PH model with a free hazard multiplier, residuals sum to $D$ (Proposition 2(b))
Piecewise exponential Poisson regression with offset $\ln T_{im}$; profile is Breslow's likelihood (Proposition 3)
NHANES Gompertz best by AIC; death rate doubles every 7.75 years of age; hazard ratios as Cox's
Simulation correct model 1.5 to 1.8 times more efficient for the median; wrong model biased 15% to 19%

What we have learned

A parametric survival model is a likelihood for censored data with the hazard's shape named in advance. The exponential gives the crude rate, and under type II censoring an exact interval. The Weibull adds a shape whose estimate always exists and is unique, and it is the one family in which a multiplied hazard and a stretched time are the same thing. Mortality in the NHANES cohort is described better by the Gompertz law than by the other four families tried, with a death rate doubling about every eight years of age, as in the national life table. Its hazard ratios are those of the Cox model. In the simulations a correct model estimated a median with 55% to 65% of the mean squared error of Kaplan–Meier; a wrong one missed by a fixed fraction that, if the usual convergence of a misspecified maximum likelihood estimate holds, no sample size cures.

This article closes Survival Analysis. The models here assume a single cause of failure and a hazard that depends only on the covariates measured; the unit's topic page lists what lies beyond, and The Cox Proportional Hazards Model treats the semiparametric alternative used when no parametric shape can be trusted.

References