Cox regression is not very suitable in the analysis of huge data sets with a lot of events (e.g., deaths). For instance, consider analyzing the mortality of the Swedish population aged 60–110 during the years 1968-2019, where we can count to more than four million deaths.
The obvious way to handle that situation is by tabulation and applying a piecewise constant hazard function, because it is a well-known fact that any continuous function can arbitrary well be approximated by a step function, simply by taking small enough steps.
The data sets swepop and swedeaths in eha contain age and sex
specific yearly information on population size and number of deaths,
respectively. They both cover the full Swedish population the years 1968–2019.
The first few rows of each:
head(swepop)##          age   sex year   pop  id
## 102.1969   0 women 1969 52673 102
## 1.1969     0   men 1969 55728   1
## 103.1969   1 women 1969 56831 103
## 2.1969     1   men 1969 59924   2
## 104.1969   2 women 1969 58994 104
## 3.1969     2   men 1969 62502   3head(swedeaths)##        age   sex year deaths id
## 2.1969   0 women 1969    491  2
## 1.1969   0   men 1969    773  1
## 4.1969   1 women 1969     33  4
## 3.1969   1   men 1969     45  3
## 6.1969   2 women 1969     22  6
## 5.1969   2   men 1969     43  5The funny rownames and the column id are created by the function reshape,
which was used to transform the original tables, given in wide format, to
long format. In the original data, downloaded from
Statistics Sweden, the population size refers to the last
day, December 31, of the given year, but here it refers to an average of that
value and the corresponding one the previous year. In that way we get an estimate
of the number of person years, which allows us to consider number of
occurrences and exposure time in each cell of the data. This information will
allow us to fit proportional hazards survival models. So we start by joining
the two data sets and remove irrelevant stuff:
dat <- swepop[, c("age", "sex", "year", "pop")]
dat$deaths <- swedeaths$deaths
rownames(dat) <- 1:NROW(dat) # Simplify rownames.
head(dat)##   age   sex year   pop deaths
## 1   0 women 1969 52673    491
## 2   0   men 1969 55728    773
## 3   1 women 1969 56831     33
## 4   1   men 1969 59924     45
## 5   2 women 1969 58994     22
## 6   2   men 1969 62502     43tail(dat)##       age   sex year    pop deaths
## 10499  98 women 2020 2033.5    764
## 10500  98   men 2020  578.5    301
## 10501  99 women 2020 1445.0    581
## 10502  99   men 2020  367.0    178
## 10503 100 women 2020 1933.5   1004
## 10504 100   men 2020  394.5    243We note that the age column ends with age == 100, which in fact means
age >= 100. There are in total 4745063 observed deaths during the years
1968–2019, or 91251 deaths per
year on average. There are 101 age groups, two sexes, and 52 years, in all 10504
cells (rows in the data frame).
Assuming a piecewise constant hazards model on the 101 age groups, we can fit a
proportional hazards model by Poisson regression, utilizing the fact that two
likelihood functions in fact are identical. In R, we use glm.
fit.glm <- glm(deaths ~ offset(log(pop)) + I(year - 2000) + sex + 
                 factor(age), data = dat, family = poisson)
summary(fit.glm)$coefficients[2:3, ]##                Estimate Std. Error z value Pr(>|z|)
## I(year - 2000) -0.01497  3.136e-05  -477.5        0
## sexmen          0.43997  9.304e-04   472.9        0The 101 coefficients corresponding to the intercept and the age factor can be used to estimate the hazard function: The intercept, -5.6181, is the log of the hazard in the age interval 0-1, and the rest are differences to that value, so we can reconstruct the baseline hazard by
lhaz <- coefficients(fit.glm)[-(2:3)]
n <- length(lhaz)
lhaz[-1] <- lhaz[-1] + lhaz[1]
haz <- exp(lhaz)and plot the result, see Figure 1.1.
oldpar <- par(las = 1, lwd = 1.5, mfrow = c(1, 2))
plot(0:(n-1), haz, type = "s", main = "log(hazards)", 
     xlab = "Age", ylab = "", log = "y")
plot(0:(n-1), haz, type = "s", main = "hazards", 
     xlab = "Age", ylab = "Deaths / Year")Figure 1.1: Age-specific mortality, Sweden 1968-2019. Poisson regression.
While it straightforward to use glm and Poisson regression to fit the model, it
takes some efforts to get it right. That is the reason for the creation of
the function tpchreg (“Tabular Piecewise Constant Hazards REGression”), and
with it, the “Poisson analysis” is performed by
fit <- tpchreg(oe(deaths, pop) ~ I(year - 2000) + sex, 
               time = age, last = 101, data = dat)Note:
The function oe (“occurrence/exposure”) takes two arguments, the first
is the number of events (deaths in our example), and the second is exposure time,
or person years.
The argument time is the defining time intervals variable. It can be either character, like
c(“0-1”, “1-2”, …, “100-101”) or numeric (as here). If numeric, the value refers
to the start of the corresponding interval, and the next start is the end of the
previous interval. This leaves the last interval’s endpoint undefined, and if not given
by the last argument (see below), it is chosen so that the length of the last interval is
one.
The argument last closes the last interval, if is not already closed, see
above. The exact value of last is only important for plotting and for the calculation
of the restricted mean survival time, (RMST) see the summary result below.
summary(fit)## Covariate             Mean       Coef     Rel.Risk   S.E.    LR p
## I(year - 2000)       -4.480    -0.015     0.985     0.000    0.000 
## sex                                                          0.000 
##            women      0.503     0         1 (reference)
##              men      0.497     0.440     1.553     0.001
## 
## Events                    4745063 
## Total time at risk        459651727 
## Max. log. likelihood      -19283067 
## LR test statistic         443018.82 
## Degrees of freedom        2 
## Overall p-value           0
## 
## Restricted mean survival:  81.58 in (0, 101]The restricted mean survival time is defined as the integral of the survivor function over the given time interval. Note that if the lower limit of the interval is larger than zero, it gives the conditional restricted mean survival time, given survival to the lower endpoint.
Graphs of the hazards and the log(hazards) functions are shown in Figure 1.2.
oldpar <- par(mfrow = c(1, 2), las = 1, lwd = 1.5)
plot(fit, fn = "haz", log = "y", main = "log(hazards)", 
     xlab = "Age")
plot(fit, fn = "haz", log = "", main = "hazards", 
     xlab = "Age", ylab = "Deaths / Year")Figure 1.2: Age-specific mortality, Sweden 1968-2019. ‘tpch’ regression. Baseline refers to women and the year 2000.
Same results as with glm and Poisson regression, but a lot simpler.
Sometimes you have a large data file in classical, individual form, suitable for
Cox regression with coxreg, but the mere size makes it impractical, or even
impossible. Then help is close by tabulating and assuming a piecewise constant
hazard function, returning to the method in the previous section, that is, using
tpchreg.
The helper function is toTpch, and we illustrate its use on the oldmort data
frame:
head(oldmort[, c("enter", "exit", "event", "sex", "civ", "birthdate")])##   enter  exit event    sex       civ birthdate
## 1 94.51 95.81  TRUE female     widow      1765
## 2 94.27 95.76  TRUE female unmarried      1766
## 3 91.09 91.95  TRUE female     widow      1769
## 4 89.01 89.59  TRUE female     widow      1771
## 5 90.00 90.21  TRUE female     widow      1770
## 6 88.43 89.76  TRUE female     widow      1772oldmort$birthyear <- floor(oldmort$birthdate) - 1800
om <- toTpch(Surv(enter, exit, event) ~ sex + civ + birthyear, 
             cuts = seq(60, 100, by = 2), data = oldmort)
head(om)##      sex       civ birthyear   age event exposure
## 1   male unmarried        -2 60-62     0    0.578
## 2 female unmarried        -2 60-62     0    4.109
## 3   male   married        -2 60-62     0   17.366
## 4 female   married        -2 60-62     0   14.129
## 5   male     widow        -2 60-62     0    0.148
## 6 female     widow        -2 60-62     0    5.805Note two things:
The creation of a new variable, birthyear. The original birthdate is
given with precision days and contains 3570
unique values, which will contribute to creating a very large table, so the
transformation gives birth year with 52
unique values. Further, the new variable is given a
reference value of 1800 by subtraction, necessary so that the baseline does
not coincide with the birth of Christ. Will foremost affect plotting of the
estimated survivor function. However, regression parameter estimates are
unaffected, as long as no interaction effect including birthyear is present.
The length of the time pieces is set to two years, in order to avoid empty intervals in the upper age range. This choice has only a marginal effect on the final results of the analyses. You can try it out yourself. Note that it is not necessary to use equidistant cut points, it is chosen here only for convenience.
Now we can run tpchreg as before
fit3 <- tpchreg(oe(event, exposure) ~ sex + civ + 
                  birthyear, time = age, data = om)
summary(fit3)## Covariate             Mean       Coef     Rel.Risk   S.E.    LR p
## sex                                                          0.000 
##             male      0.406     0         1 (reference)
##           female      0.594    -0.245     0.783     0.047
## civ                                                          0.000 
##        unmarried      0.080     0         1 (reference)
##          married      0.530    -0.397     0.672     0.081
##            widow      0.390    -0.258     0.773     0.079
## birthyear             2.114    -0.006     0.994     0.004    0.150 
## 
## Events                    1971 
## Total time at risk         37824 
## Max. log. likelihood      -7265.5 
## LR test statistic         43.45 
## Degrees of freedom        4 
## Overall p-value           8.34423e-09
## 
## Restricted mean survival:  12.65 in (60, 100]And the hazards graphs are shown in Figure 2.1.
oldpar <- par(mfrow = c(1, 2), las = 1, lwd = 1.5)
plot(fit3, fn = "haz", log = "y", main = "log(hazards)", 
     xlab = "Age", ylab = "log(Deaths / Year)", col = "blue")
plot(fit3, fn = "haz", log = "", main = "hazards", 
     xlab = "Age", ylab = "Deaths / Year", col = "blue")Figure 2.1: Old age mortality, Skellefteå 1860-1880.
The plots of the survivor and cumulative hazards functions are “smoother”, see Figure 2.2.
oldpar <- par(mfrow = c(1, 2), las = 1, lwd = 1.5)
plot(fit3, fn = "cum", log = "y", main = "Cum. hazards", 
     xlab = "Age", col = "blue")
plot(fit3, fn = "sur", log = "", main = "Survivor function", 
     xlab = "Age", col = "blue")Figure 2.2: Old age mortality, Skellefteå 1860-1880. Cumulative hazards and survivor functions.
par(oldpar)A comparison with Cox regression on the original data.
fit4 <- coxreg(Surv(enter, exit, event) ~ sex + civ + I(birthdate - 1800), 
               data = oldmort)
summary(fit4)## Covariate             Mean       Coef     Rel.Risk   S.E.    LR p
## sex                                                          0.000 
##             male      0.406     0         1 (reference)
##           female      0.594    -0.244     0.783     0.047
## civ                                                          0.000 
##        unmarried      0.080     0         1 (reference)
##          married      0.530    -0.397     0.673     0.081
##            widow      0.390    -0.259     0.772     0.079
## I(birthdate - 18      2.602    -0.005     0.995     0.004    0.212 
## 
## Events                    1971 
## Total time at risk         37824 
## Max. log. likelihood      -13557 
## LR test statistic         42.79 
## Degrees of freedom        4 
## Overall p-value           1.14378e-08And the graphs, see Figure 2.3.
oldpar <- par(mfrow = c(1, 2), lwd = 1.5, las = 1)
plot(fit4, main = "Cumulative hazards", xlab = "Age", 
     col = "blue")
plot(fit4, main = "Survivor function", xlab = "Age", 
     fn = "surv", col = "blue")Figure 2.3: Old age mortality, Skellefteå 1860-1880. Cox regression with original data.