Page 54 - Clinical Small Animal Internal Medicine

22 Section 1 Evaluation and Management of the Patient

be unlikely to have many cases of IMHA enrolled unless the cases, the distribution of the same risk factors is also
VetBooks.ir the population was very large, adversely affecting sta- measured in controls. Setting aside the issue of match-
ing, conceptually the distribution of risk factors in a con-
tistical precision and jeopardizing the ability to make
conclusions. Moreover, it would require comparing vac-
that same distribution in the source population. Matching
cination histories on a disproportionately large number trol group not subject to selection bias effectively mirrors
of dogs that never developed IMHA for comparison to is a technical violation of this principle, because match-
a disproportionately small number of dogs that did, ing on a confounder alters the distribution of the risk fac-
adversely affecting study efficiency. tors in the controls to be more aligned with that of the
If an adequate number of cases exist in a hospital reg- cases, but this intentional design feature that leads to an
istry (or those of multiple hospitals or diagnostic labora- unintentional selection bias can be remediated through a
tories, etc.), a far more efficient approach to studying proper matched statistical analysis.
factors affecting the incidence of a rare disease is to take The problem with this conception of a control group as
a census or sample of the cases from the registry, and for a representative sample of a source population of cases is
comparative purposes take a sample of dogs (controls) one of practicality: for a hospital that is not in a geo-
never diagnosed with IMHA (but theoretically at risk of graphically restricted region, it is not at all clear what the
getting it later in life) from the same registry. The ratio of source population is. There are many reasons, known
controls to cases is at the discretion of the investigator; if and unknown, why owners elect to have their pets treated
resources dictate that only a fixed number of cases and or not treated, and why they choose one hospital over
controls can be included in the study, a 1:1 ratio is most another. The reasons why owners make such choices
efficient. If the cost of obtaining risk factor information could be influenced in some way by the very risk factors
on the patients is not an important consideration, then a being studied, so without appreciating the subtleties of
ratio of up to 4–5 controls per case is usually sufficient selecting controls, it is possible to introduce selection
for statistical precision. bias into a case–control study.
Although controls can be taken at any time during the To illustrate this point, suppose that an investigator at
period of case accrual, there are distinct statistical and a community hospital wishes to determine if vaccina-
interpretive advantages to selecting them at the same tions lead to IMHA development, and observes that 50%
time that cases are diagnosed. Such matching on time of new cases had a history of vaccination in the prior two
can be extended to additional confounders, such as hos- months. The odds of exposure to the risk factor is there-
pital (in a multicenter study) and age, but most other fore 0.50 / 0.50 = 1.0. To obtain controls, a random sur-
confounders can be controlled for by an appropriate vey by the investigator of pet owners in the community
matched analysis. around the hospital reveals that 8.3% of their pets were
The advantages to matching controls to cases on time vaccinated in the prior two months; the odds of exposure
are several‐fold. First, if time itself was a confounder of in the community-based controls is 0.083 / 0.917 =
the risk factor–disease relationship (i.e., if disease inci- 0.091. The odds ratio relating two‐month prior vaccina-
dence and the distribution of the risk factor in the popu- tion to IMHA incidence is 1.0 / 0.091 = 11.0.
lation change over time), then effective analytic control The assumption behind using a random sample of
of it would almost certainly be impossible without dogs in the community as controls is that they represent
matching if the period of case accrual was long. Second, the source population of cases of IMHA seen at the hos-
the odds ratio calculated from the case–control study pital; namely, had those dogs in the community devel-
can be interpreted as a risk ratio when the incidence of oped IMHA, they would have been brought to the
disease is believed to be rare in all categories of the risk hospital and undergone the same diagnostic work‐up as
factors (and confounders). Third, if the disease incidence other cases of IMHA. In cases when only a single hospi-
is not rare, the odds ratio can still be interpreted as an tal serves a well‐circumscribed region, this may be a ten-
incidence rate ratio, described earlier for cohort studies. able assumption. However, if there are multiple hospitals
However, analyses of matched case–control data are that serve communities, the reasons why some owners
complex when more than one control is matched to a select one hospital over another may be related to factors
case, and require a computer with software capable of associated with the study exposure. Moreover, the pro-
performing conditional logistic regression. portion of owners who are willing to bear the costs of
The most difficult design issue in case–control studies seeking medical diagnosis and care for their pets for seri-
is that of proper control selection. Cases can be thought ous illnesses is likely to be lower than the proportion of
of as originating from an underlying source population people who would seek medical care for their own seri-
of individuals at risk of developing the disease, and con- ous illnesses.
trols are, simplistically, a sample of that source popula- Now suppose that the investigator’s hospital is known
tion. Just as the distribution of risk factors is measured in for higher medical costs relative to other hospitals that