Chapter 4. Cancer Epidemiology and Statistics
Jeff W. Tyler and Jose Armando Villamil
This chapter reviews some of the epidemiological and statistical concepts related to clinical oncology. Because veterinarians should base their clinical decisions upon advances presented as published, peer-reviewed evidence, an understanding of basic concepts of experimental design and data analysis is essential. For those unfamiliar with statistical terminology, a glossary of terms is provided in Box 4-1 .
BOX 4-1
GLOSSARY OF TERMS 1-4
Bias —A systematic deviation of measured endpoints from true or representative values.
Case-control study —Exposure to risk factors is determined by history or survey data, and differences in exposure to hypothesized risk factors are assessed between cases and controls.
Cohort study —Subjects are placed in groups on the basis of the presence or absence of postulated risk factors or treatments, and disease outcomes are followed over time.
Confounding —An association between a postulated risk factor and a measured outcome, which is the result of some other variable that has not been considered.
Normal —Symmetrically dispersed around a measure of central tendency following a specific pattern of distribution whereby approximately two thirds of observations are located with a single standard deviation of the mean and approximately 95% of observations are within 2 standard deviations of the same mean.
P -value —The likelihood that a difference of the observed magnitude could occur because of chance.
Power —The likelihood of detecting a significant difference among groups or exposures at a given P -value, sample size, group means, and standard deviation.
Sensitivity —The likelihood of a positive test result given that a patient has the disease in question.
Specificity —The likelihood of a negative test result given that a patient does not have the disease in question.
Type I error —see P -value
AVAILABLE STUDY DESIGNS AND QUALITY OF EVIDENCE
Two broad classifications of study design exist: descriptive and analytical. Descriptive studies typically report disease manifestations in a defined group, without comparisons of disease incidence or outcome based on exposure to risk factors. The classic example of a descriptive report is a case series. Descriptive reports provide weaker evidence than analytical studies because no attempt is made to link exposure to a risk factor with causality.
Analytical studies attempt to draw associations between exposure to risk factors and either the incidence of disease or a disease outcome. Among analytical studies, randomized, controlled clinical trials provide the strongest evidence, followed by cohort studies and case-control studies.
Analytical studies may be either experimental or observational. The optimal evidence is provided by experimental studies that are randomized, blind, controlled, clinical trials (see Chapter 5 for more detail). Although this approach provides the strongest evidence, this design is problematic if exposure to a risk factor has either a small or delayed effect on disease outcome. For example, if exposure to a pesticide is postulated to cause an increased incidence of bladder cancer 5 to 10 years after exposure, an experimental study exploring this relationship would entail years of study and effort to answer a single clinical question.
Less rigorous analytical studies include cohort studies and case-control studies. In cohort studies, subjects are placed in groups on the basis of the presence or absence of postulated risk factors or treatments, and disease outcomes are followed over time. This design, like the classic experimental study, is prospective in nature; however, the group assignment and standardization of outcome monitoring is less rigorous. Studies of this type typically assess the impact of a limited number of risk factors, usually treated vs. control, on disease outcome.
In contrast, case-control studies are retrospective. Patients with a specific diagnosis are defined as cases. Controls are drawn from a similar population during the same time interval, but do not have the specific diagnosis. Exposure to risk factors is determined by history or survey data, and differences in exposure are assessed between case and control groups. Case-control studies have a number of characteristics which make them attractive. First, because these studies are retrospective, they are relatively inexpensive and can be performed quickly. Secondly, exposure to multiple risk factors can be compared among cases and controls in a single study. These studies provide weaker evidence because owners may have imperfect memory and the owners of case animals may have recollections of risk factor exposures because they have either consciously or unconsciously attempted to determine the cause of their pets’ disease.
CONTROLS AND RANDOMIZATION
Clinical observations without controlled comparison groups may be interesting and thought provoking; unfortunately, observations of this type that are erroneous often become accepted, but unproven, dogmas. Examples of publications that lack controls include individual case reports and case series. The limitations of case reports are obvious. Case reports present novel diagnoses, diagnostic approaches and treatments. However, these reports are singular observations, and consequently, lack a control or reference group and fail to provide direct evidence. In some ways, case series are even more problematic. The number of observations in a case series is substantially greater than in a case report, and these manuscripts often include statistics. Thus, they may outwardly appear to carry greater scientific weight, and conclusions may be accepted by the casual reader without appropriate skepticism. Although clinical outcomes are often compared among subjects with differing historical risk factors within a case series, these comparisons are often invalid because historical controls have different parameters and their own set limitations. Consequently, these studies, although more rigorous than case reports, have limited value as direct evidence of diagnostic or therapeutic efficacy. The factor missing from both case reports and case series is the identification of a valid and appropriate comparison or control group. Controls are the single most important factor determining the validity of a study .
For studies undertaken to determine the efficacy of a new treatment or intervention, the control group should be comprised of animals that were either left untreated or, preferably, received the current standard therapy. 1 The control group should be as identical to the exposure group or treatment group as possible, except for the particular factor or variable being investigated. 2 Ideally, assignment to either the treatment or control group should be random. Methods of random assignment include coin toss, drawing of a card, or use of a random number table. Regardless of the method used, each subject should have an equal likelihood of being placed in the treatment or control groups, ensuring that treatment and control groups have a similar composition. Examples of non-random and unacceptable group assignment strategies include clinician preference, evolutionary changes in treatment protocols, financial constraints, and owners’ preferences. Consequently, owner and clinician willingness to participate in a clinical trial should be determined prior to group assignment. Willingness to participate in a trial that hinges upon group assignment outcome cannot be viewed as a random event. Owners may choose to participate in a trial if given subsidized access to novel treatment that may have the potential to improve clinical outcomes. Likewise, clinicians are likely motivated in a similar manner. These potential biases in group assignment may be accentuated in unsubsidized trials because clients with fewer financial resources may be shuttled into the less-expensive treatment arms of a clinical trial. Another common mistake is the use of a historical population or a population from a previous published study as a control group. The problem with this control group is that diagnostic procedures, standard of care, client expectations, clinician expertise, available drugs, and the definition of the disease in question may have changed over time. Differences over time between groups create the potential for temporal effects in all the identified variables to be misidentified as differences in response to treatment.
Random group assignment will also minimize the impact of extraneous factors associated with the variable in question. Confounding has been described as the mixing together of the effect of two or more factors. Thus, when confounding is present, practitioners might think they are measuring the association of an exposure factor with an outcome, but the association measure also includes the effect of one or more extraneous factors. 3 For example, suppose a new chemotherapy agent has become available for treatment of osteosarcoma in dogs but is very expensive. In an effort to maximize the number of patients enrolled, small dogs are given the new, expensive agent and larger dogs receive an older, less-expensive treatment. A significant between-group difference is observed. Unbeknownst to the researcher, the course of disease differs between large and small dogs. The investigator erroneously attributes improved survival in the smaller dogs to improved treatment efficacy when, in fact, the difference is related to patient size.
SAMPLE SIZE
Factors that impact sample size requirements include designated significance threshold P -value, power, the rate of the observed event in the control population, and the anticipated reduction in mortality associated with treatment. In general, common events and dramatic treatment effects result in smaller sample size requirements for a study.
< div class='tao-gold-member'>
Only gold members can continue reading. Log In or Register to continue
Stay updated, free articles. Join our Telegram channel
Full access? Get Clinical Tree
