Page 51 - Clinical Small Animal Internal Medicine

3 Using Data for Clinical Decision Making 19

Cohort Studies Suppose that the patient was lost to follow‐up after
VetBooks.ir Assemblages of patients that share one or more defining four years, but was known to be alive up to that time. Did
death from IMHA occur? The answer at the four‐year
conditions, typically diseases or syndromes, are often
referred to as a “case series” but can be more expansively time point is no, because the patient was not observed
long enough for death to occur and be recorded.
thought of as a patient cohort. In a hospital setting, such However, such an individual who did not die from IMHA
patients present over a period of time; if the condition is after four years should not statistically be treated as
common then the period of time for sufficient patient equivalent to a patient still alive after an even longer
accrual may be brief, but if the condition is rare, it may period of observation. Because most patients are fol-
take many years for an adequate number of patients to lowed for different periods of time after disease diagno-
be enrolled unless a decision is made to extend patient sis, an investigator would provide a distorted impression
enrollment to multiple hospitals. While population‐ of the natural course of the disease by responding to the
based (epidemiologic) cohort studies typically study risk simplistic question, “What proportion of patients died?”
factors for disease incidence, the goal of hospital‐based A legitimate analysis would instead have to be limited to
cohort studies is different: to study the outcomes of specifying an identical follow‐up period for each patient,
patients already diagnosed with diseases. Such out- such as “What proportion of patients succumbed to
comes are contextual to the specific diseases, but often IMHA within one year of diagnosis?” Of course, patients
include the occurrence of or time to remission, relapse, followed for less than one year would necessarily be
development of another disease condition, recovery, or excluded even from this analysis.
death. By subdividing cohort members into groups Because follow‐up times vary so much in patients
defined by factors thought to influence the respective recruited over long periods of time, a preferable analytic
outcomes, it becomes possible to quantify and compare approach takes into account not only if a patient eventu-
their effects. ally died from the disease, but also how long they were
There are two main types of outcomes of interest in
the study of factors affecting cohorts of patients that alive prior to death. The apposite question instead
becomes: “What is the probability of patients dying from
share a disease in common. In this example, death will be IMHA after accounting for different periods of time
the illustrative outcome, and sex will be the hypothesized under observation following diagnosis?” The answer, a
determinant of death. The duality of outcomes is:
probability that depends on the time interval after diag-
whether or not death from the disease occurs follow-
● nosis and characteristics of patients that influence sur-
ing its occurrence in patients vival, will fall between 0 and 1. This kind of analysis is
how long it takes death from the disease to occur (if it
● generically known as a “survival analysis.” One approach
does) following its occurrence in patients. commonly used to estimate these probabilities is called
The first outcome is not as straightforward to evaluate as the “product‐limit method of survival/failure function
might initially appear, because it depends on two related estimation” or, equivalently and eponymously, a “Kaplan–
factors: how long a patient is diseased, and how long a Meier analysis.”
patient is followed up. To illustrate the distinction between counting the pro-
To illustrate these points, consider a patient diagnosed portion of patients who die and the probability of a
with immune‐mediated hemolytic anemia (IMHA) on patient dying as a function of time, consider a hypotheti-
January 1, 2013, and that succumbed to the disease on cal cohort of 12 patients. Four patients each have either
January 1, 2019. From the perspective of the investigator, five, six, or seven years of follow‐up, and within each
did death occur? The answer is no if the follow‐up time is group of four patients, two die and two either do not die
less than six years (although death occurred after six or are lost to follow‐up at the end of their follow‐up
years, the investigator would not have been aware of it); period.
the answer is yes if the time period evaluated is longer The crude proportion of these 12 patients dying is
than six years. Therefore, the qualified answer to the 50%, but that figure is entirely misleading because of
question depends on specifying the time of follow‐up unequal follow‐up times. In contrast, a statistical analy-
after the diagnosis is made. Clearly, the longer a patient sis projects that the probability of dying after five years is
is followed up, the greater the probability they will die 17%, after six years is 38%, and after seven years is 69%.
(although not necessarily from IMHA). By convention, The latter analysis is valid because it accounts for not
patients who do not experience the health outcome at only whether or not death occurred, but also how long a
the point that they are no longer observed at risk of it, patient was at risk of dying (by including censored
because they were lost to follow‐up, withdrew from the patients).
study, recovered, or died from a competing cause, are It is also possible to estimate the time it takes to reach
called “censored.” a certain probability of death. For example, the time it