Survival Analysis I (CHL5209H)

Transcription

1 Survival Analysis Dalla Lana School of Public Health University of Toronto January 7,

2 Literature Clayton D & Hills M (1993): Statistical Models in Epidemiology. Not really useful as a reference text but interesting pedagogical approach. Kalbfleisch JD & Prentice RL (2002): The Statistical Analysis of Failure Time Data, Second Edition. Introductory, serves as a reference text. Klein JP & Moeschberger ML (2003): Survival Analysis - Techniques for Censored and Truncated Data, Second Edition. Introductory, serves as a reference text. Aalen OO, Borgan Ø, Gjessing H (2008): Survival and Event History Analysis - A Process Point of View. For those looking for something more theoretical. 31-2

3 31-3

4 Survival analysis? Survival analysis focuses on a single event per individual (say, first marriage, graduation, diagnosis of a disease, death). Analysis of multiple events would be referred to as event history analysis. Despite the name, a course in survival analysis is mostly not about survival. Rather, most of the time, we concentrate on estimating or modeling a more fundamental quantity, the rate parameter λ, or the time-dependent version, the hazard function λ(t). Survival probability in turn is determined by the hazard function. The rates can be for example mortality or incidence rates. 31-4

5 More about rates Rate parameter is the parameter of the distribution, characterizing the rate of occurrence of the events of interest. The expected number of events µ in a total of Y years of follow-up time and λ are connected by µ = λy. The observed number of events D in Y years of follow-up time is distributed as D (λy ). How to estimate the rate parameter λ? 31-5

6 An estimator for λ A possible estimator is suggested by µ = λy λ = µ Y. It would seem reasonable to replace here the expected number of events µ with the observed number of events D and take ˆλ = D Y. This is in fact the maximum likelihood estimator of λ. 31-6

7 Maximum likelihood criterion Maximum likelihood estimate is the parameter value that maximizes the probability of observing the data D. The probability of the observed data is given by the statistical model, which is now D (λy ). Probabilities under the distribution are given by (λy ) D e λy. D! We consider this probability as a function of λ, and call it the likelihood of λ. 31-7

8 Maximizing the likelihood We may ignore any multiplicative terms not depending on the parameter, and instead maximize the expression λ D e λy. Or, for mathematical convenience, its logarithm D log λ λy. How to find the argument value which maximizes a function? Set the first derivative to zero and solve w.r.t. λ: D λ Y = 0 λ = D Y. Check that the second derivative is negative: D λ 2 <

9 Follow-up data Clayton & Hills (1993, p. 41): 31-9

10 Estimate For the 7 subjects (individuals) there is a total of 36 time units of follow-up time/person-time, and 2 outcome events (for individuals 2 and 6). The follow-up of the other individuals was terminated by censoring (e.g. by events other than the outcome event of interest). In fact, censoring is the reason why we need to approach the estimation problem through rates, rather than looking at the event times directly. We arrive at ˆλ = D Y = To recap: Estimand/parameter/object of inference: λ Estimator: D Y Estimate:

11 Interpretation? We will define the rate parameter more precisely shortly, but let us first consider the interpretation. Unlike the risk parameter, the probability of an event occurring within a specific time period, the rate parameter does not correspond to a follow-up period of a fixed length. Rather, it characterizes the instantaneous occurrence of the outcome event at any given time. The rate parameter is not a probability, but it can be characterized in terms of the risk parameter when the follow-up period is very short

12 Time unit Suppose that each of the N = 36 time bins here is of length h = 0.05 years: In total there is Y = Nh = = 1.8 years of follow-up

13 From risk to rate The empirical rate is given by ˆλ = = 1.11 per person-year, or, say, 1110 per 1000 person-years. Per person-year, the empirical rate would be the same, had we instead split the person-time into 180 bins of length 0.01 years. Suppose that we have made the time bins short enough so that at most one event can occur in each bin. Whether an event occurred in a particular bin of length h is now a Bernoulli-distributed variable, with the expected number of events equal to the risk π. Thus, because rate is the expected count divided by person-time, when h is small, we have λ = π h π = λh This connection is important in understanding how rate is related to survival probability.

14 31-14 Survival probability One of the particular properties of the natural logarithm and its inverse is that when x is close to zero, e x 1 + x, and conversely, log(1 + x) x. Suppose that we are interested in the probability of surviving T years. By splitting the timescale so that N = T h, T = Nh. The probability of surviving through a single time bin of length h, conditional on surviving until the start of this interval, is 1 π = 1 λh. By the multiplicative rule, the T year survival probability is thus (1 λh) N. This motivates the well-known Kaplan-Meier estimator, to be encountered later. In turn, the logarithm of this is N log(1 λh) Nλh = λt.

15 Survival and cumulative hazard The quantity λt is known as the cumulative hazard. We have (approximately, without calculus) obtained a fundamental relationship of survival analysis, namely that the T year survival probability is (1 λh) N e λt. Let us test whether this approximation actually works. Now ˆλ = If T = 1 and h = 0.05, N = 20 and we get ( ) The exact one year survival probability is e We should get a better approximation through a finer split of the time scale. If h = 0.01, N = 100 and ( )

16 Regression models Recall the relationship µ = λy. If γ = log λ, uquivalently we have a log-linear form log µ = γ + log Y, where we call log Y an offset term. γ is an unknown parameter, which we could estimate in an obvious way. (How?) Such a one-parameter model is not very interesting, but serves as a starting point to modeling. Consider now the expected number of events µ 1 in Y 1 years of exposed person-time and the expected number of events µ 0 in Y 0 years of unexposed person-time. The corresponding probability models are now D 1 (µ 1 ) and D 0 (µ 0 ), where µ 1 = λ 1 Y 1 and µ 0 = λ 0 Y 0. Here µ 1 + µ 0 = µ and D 1 + D 0 = D.

17 Combining the two models We may now parametrize the two log-rates in terms of an intercept term α and a coefficient β as log λ 0 = α and log λ 1 = α + β. By introducing an exposure variable Z, with Z = 1 (Z = 0) indicating the exposed (unexposed) person-time, we can express these definitions as a equation log λ Z = α + βz λ Z = e α+βz. This results in a single statistical model, namely ( ) D Z Y Z e α+βz. What is the interpretation of the coefficient? Now we have λ 1 = eα+β λ 0 e α = eα e β e α = eβ, or β = log ( λ1 λ 0 ), that is, the log rate ratio.

18 Reparametrizing rates The model can be easily extended to accommodate more than one determinant. For example, unadjusted comparisons of rates are susceptible to confounding; we can move on to consider confounder-adjusted rate ratios. Consider the following dataset: Introduce an exposure variable taking values Z = 1 (energy intake < 2750 kcals) and Z = 0 ( 2750 kcals), and an age group indicator taking values X = 0 (40 49), X = 1 (50 59) and X = 2 (60 69).

19 The original parameters There are now six rate parameters λ ZX, corresponding to each exposure-age combination: Z = 0 Z = 1 X = 0 λ 00 λ 10 X = 1 λ 01 λ 11 X = 2 λ 02 λ 12 The corresponding statistical distributions are Z = 0 Z = 1 X = 0 D 00 (Y 00 λ 00 ) D 10 (Y 10 λ 10 ) X = 1 D 01 (Y 01 λ 01 ) D 11 (Y 11 λ 11 ) X = 2 D 02 (Y 02 λ 02 ) D 12 (Y 12 λ 12 ) 31-19

20 Transformed parameters Now, we are not primarily interested in estimating six rates; rather, we are interested in the rate ratio between the exposure categories, adjusting for age. We could parametrize the rates w.r.t. the baseline, or reference, rate λ 00 which is then modified by the exposure and age (cf. Clayton & Hills 1993, p. 220). Define Z = 0 Z = 1 X = 0 λ 00 = λ 00 λ 10 = λ 00 θ X = 1 λ 01 = λ 00 φ 1 λ 11 = λ 00 θφ 1 X = 2 λ 02 = λ 00 φ 2 λ 12 = λ 00 θφ 2 Now θ is the rate ratio within each age group (verify)

21 Regression parameters As before, we can specify the reparametrization in terms of a link function and a linear predictor as Since log λ ZX = α + βz + γ 1 1 {X =1} + γ 2 1 {X =2}. λ ZX = e α+βz+γ 11 {X =1} +γ 2 1 {X =2}, we have that λ 00 = e α, θ = e β, φ 1 = e γ 1 and φ 2 = e γ 2. The rates are now given by the equation as Z = 0 Z = 1 X = 0 λ 00 = e α λ 10 = e α+β X = 1 λ 01 = e α+γ 1 λ 11 = e α+β+γ 1 X = 2 λ 02 = e α+γ 2 λ 12 = e α+β+γ 2 The number of parameters has been reduced from six to four.

22 Specification in terms of expected counts A model is always specified in terms of the expected event count: D ZX (µ ZX ). The model for the expected count is specified by µ ZX = Y ZX λ ZX = Y ZX e α+βz+γ 11 {X =1} +γ 2 1 {X =2} = e α+βz+γ 11 {X =1} +γ 2 1 {X =2} +log Y ZX. We have obtained the model ) D XZ (e α+βz+γ 11 {X =1} +γ 2 1 {X =2} +log Y ZX. When fitting the model, log-person years has to be included in the linear predictor as an offset variable

23 Fitting the model The data as frequency records are entered into R as: d <- c(4,5,8,2,12,14) y <- c(607.9,1272.1,888.9,311.9,878.1,667.5) z <- c(0,0,0,1,1,1) x <- c(0,1,2,0,1,2) The model is specified as model <- glm(d ~ z + as.factor(x) + offset(log(y)), family=poisson(link="log")) The as.factor(x) term specifies that we want to estimate separate age group effects (rather than assume that the X -variable modifies the log-rate additively)

24 Results Call: glm(formula = d ~ z + as.factor(x) + offset(log(y)), family = poisson(link = "log")) Deviance Residuals: Coefficients: Estimate Std. Error z value Pr(> z ) (Intercept) < 2e-16 *** z ** as.factor(x) as.factor(x) Signif. codes: 0 *** ** 0.01 * (Dispersion parameter for poisson family taken to be 1) Null deviance: on 5 degrees of freedom Residual deviance: on 2 degrees of freedom AIC: Number of Fisher Scoring iterations: 4

25 Proportional hazards From the model output, we may calculate estimates for the original rate parameters (per 1000 person-years) as Z = 0 Z = 1 X = 0 ˆλ 00 = 4.44 ˆλ 10 = X = 1 ˆλ 01 = 5.05 ˆλ 11 = X = 2 ˆλ 02 = 8.86 ˆλ 12 = Note that the rate ratio stays constant across the age groups. This is forced by the earlier model specification. This is a modeling assumption, namely the proportional hazards assumption. Compare these estimates to the corresponding six empirical rates. Is assuming proportionality of the hazard rates justified? How could one test this? Or relax this assumption? 31-25

26 31-26

27 Counting process In survival analysis, the outcome is generally a pair of random variables (T i, E i ), where T i represents the event time, and E i the type of the event that occurred at T i. Usually, we have to consider at least two types of events, namely the outcome event of interest (say, E i = 1), and censoring (say, E i = 0), that is, termination of the follow-up due to some other reason than the outcome event of interest. Equivalently, we can characterize the outcomes in terms of the cause-specific counting processes N ij (t) = 1 {Ti t,e i =j}, j = 0, (What is a process?) We can also introduce an overall counting process N i (t) = 1 j=0 N ij(t) for any type of event. If there is no censoring, this is all we need.

28 History The observed histories of such counting processes just before time t are denoted as F t = σ({n ij (u) : i = 1,..., n; j = 0, 1; 0 u < t}). The mathematical characterization of this as a generated σ-algebra is not important at this point, we only need to understand F t as all observed information before t. Consider now the conditional probability for an event (of any type) occurring at t, that is, P(dN i (t) = 1 F t ), where dn i (t) N i (t + dt) N i (t)

29 Hazard function Alternatively, this may be characterized as P(dN i (t) = 1 F t ) = E[dN i (t) F t ] = P(t T i < t + dt F t ) λ(t) dt, or λ(t) = P(t T i < t + dt F t ). dt When T i t and the outcome of individual i does not depend on the outcomes of other individuals, this is equivalent to the familiar definition λ(t) = lim t 0 P(t T i < t + t T i t). t 31-29

30 Connection to the survival function Now P(t T i < t + dt T i t) = P(t T i < t + dt) P(T i t) λ(t) = f (t) S(t), where f (t) is the density function and S(t) the survival function of the event time distribution. Note that S(t) = 1 F (t) and f (t) = Further, d[log F (t)] dt df (t) dt. = f (t) d[log S(t)] F (t) and dt This gives us again the fundamental relationship { t } S(t) = exp λ(u) du, 0 = f (t) S(t) where t 0 λ(u) du Λ(t) is the cumulative hazard.

31 Cause-specific hazards In the presence of censoring, we are not interested in the overall hazard rate, but rather, the cause-specific hazard rate P(dN ij (t) = 1 F t ) = E[dN ij (t) F t ] = P(t T i < t + dt, E i = j F t ) λ j (t) dt for event type j. It is easy to see that these are related to the overall hazard by λ(t) = 1 j=0 λ j(t). The sub-density function corresponding to event type j is given by f j (t) P(t T i < t + dt, E i = j) dt = λ j (t)s(t), where S(t) is the overall survival function.