Page 39 - Clinical Small Animal Internal Medicine

7

VetBooks.ir

2

Statistical Interpretation for Practitioners

Philip H. Kass, BS, DVM, MPVM, MS, PhD

Department of Population Health and Reproduction, School of Veterinary Medicine, University of California, Davis, CA, USA

The need for an understanding of how to conduct statisti- key assumptions. For example, the initial epidemiological
cal analyses and, more importantly, how to interpret them research into the association between tobacco smoking and
derives from a natural tension between aspiration and lung cancer was performed in the 1950s in a cohort of male
reality: the desire to make encompassing statements con- British doctors [1]. Technically, the findings of this research
cerning the characteristics and causal properties about strictly applied to the entire population of male British doc-
populations of animals, juxtaposed with the inability to tors who were contemporaries of the study subjects. The
study more than a small sample of them. Statistical infer- choice to generalize these findings – namely, that the inci-
ence therefore provides the necessary linkage between dence of lung cancer was many‐fold higher among smokers
using samples to make inferences about populations. than nonsmokers – to other populations rested on key
assumptions motivated by scientific reasons independent of
the actual research. These assumptions included that the
effect of tobacco smoking on lung cancer incidence should
External Validity not meaningfully vary by gender, occupation, country of ori-
gin, ethnic identity, and birth cohort. Nothing in the original
It is often taken as a matter of faith that studies conducted research could have provided evidence to support these
in relatively circumscribed subpopulations (such as a cohort assumptions; nevertheless, they helped create the basis for
of patients seen at an individual hospital in a defined period the landmark 1964 report in the US officially affirming a
of time) can have relatively generalizable findings. For causal link between tobacco smoking and lung cancer [2].
example, a hospital‐based study examining the relative clin- A more recent example of the dangers of extrapolating
ical efficacy of two or more chemotherapeutic regimens to study results to nonstudy populations can be found in an
treat newly diagnosed canine lymphoma by inducing remis- article on the association between mitotic index (MI)
sion may motivate the authors to make recommendations and survival in dogs diagnosed with mast cell tumors [3].
for adoption well beyond the hospital’s patient catchment The authors found a substantially lower survival in dogs
area. When are such generalizations justified? seen at a California veterinary medical teaching hospital
Sampling of populations is required to scientifically whose MI was 5 or fewer versus those whose MI was 6 or
justify extrapolating results from sample‐based studies greater. In contrast, Elston et al. performed similar anal-
to target populations. Thus, to make scientific inferences yses on dogs from Brazil and recommended somewhat
about a population, it is necessary to study a representa- different MI cut‐off values [4]. The original authors
tive sample. In many cases, the sample size need not be responded by underscoring the difficulty of externally
particularly large, and can be obtained through random validating studies:
(or more complex) sampling schemes. The process of
random sampling ensures representative selection, and This underscores the fallibility of classification
that in turn provides the key link between a study sample schemes in clinical veterinary medicine: they may
and a target population. work well in a study population and its corre-
Because few studies are actually conducted using true sponding reference population but may not per-
sampling of a target population, the ability to generalize form nearly as well in a target population of
study findings (that is, having “external validity”) typically inherently different animals or where measure-
depends on prior medical belief and knowledge, as well as ment standards may not necessarily be completely

Clinical Small Animal Internal Medicine Volume I, First Edition. Edited by David S. Bruyette.
© 2020 John Wiley & Sons, Inc. Published 2020 by John Wiley & Sons, Inc.
Companion website: www.wiley.com/go/bruyette/clinical