Appropriate control selection in case-control studies is key to generating valid results.
Sampling controls
There are three methods for sampling controls in case-control studies (represented pictorially in Figure 1).128
Cumulative sampling: Traditionally employed by case-control studies, where non-cases are selected at the end of follow-up from “survivors” or those who are disease free. Case-cohort sampling: Where controls are selected from the entire cohort at baseline,
in other words the base population. This method is also known as case-base sampling and is the least common method.
Incidence density sampling: where controls are selected from the “risk set” of persons at risk at the time that each case occurs. Also called concurrent sampling.
Figure 1: Pictorial guide to the three different sampling methods used in case-control studies.
Note: the horizontal lines represents individual patients and their follow-up ends when the line stops.129
The matched case-control study in this chapter utilises incidence density sampling to match cases to controls; so for each case, controls were selected from all those “at risk” of zoster on the day the zoster case was diagnosed. It can be shown that when this method of control
sampling is used, the resultant odds ratios (ORs) will estimate rate ratios for the outcome.128,130
Cumulative sampling END OF FOLLOW- UP START OF FOLLOW- UP END OF FOLLOW- UP START OF FOLLOW- UP END OF FOLLOW- UP START OF FOLLOW- UP Incidence density sampling Case-cohort sampling
Case: patient diagnosed with zoster Potential control: Non-zoster patient lost to follow-up
94
Please note that although this method does provide rate ratio estimates, ORs were reported in the paper, as this is still technically correct and reporting rate ratios may have led to confusion.
In incidence density sampling, a single control could be matched to more than one case.131 In
other words, even after being selected as a control, individuals should remain eligible to be
controls for other cases. Furthermore, cases can be controls prior to their zoster diagnosis.131
These features ensure that the resulting OR will provide an unbiased estimate of the underlying rate ratio, by sampling controls from all those at risk of zoster at the time of sampling.
This sampling method was chosen as it is seen as the gold-standard sampling method for case- control studies and provides usefully interpretable effect estimates. The key disadvantages of
cumulative sampling are: 1) cases cannot be in the control group, therefore the control group
will not truly represent the entire base population at risk; and 2) by sampling controls at the end of follow-up, factors influencing loss to follow-up will influence the selection of controls. Finally, time-window bias, where the observation time during which exposures could have been recorded differs between cases and controls, is less of a problem if incidence density
sampling is used.132 By selecting controls from the all person-moments instead of the end of
follow-up (as in cumulative sampling), the resulting exposure measurement for controls and cases are based on a more similar time span.
Matching cases and controls
Individual-level matching in case-control studies is the process of selecting a specified number of controls for each case, who are similar in terms of certain pre-specified characteristics, in
order to reduce the confounding effects of those characteristics.131 If the matching
characteristic is a very strong confounder, then matching increases the efficiency of the study. Without matching, adjusting for these confounders could lead to multiple strata with few data. Forcing the cases and controls to have a similar distribution of the matching variables results in
gains in precision and narrower confidence intervals (CIs).133 Matching may also control for
unmeasured confounders.
4.4.2.1.
Matching factors
The matching factors were; age (within 1 year), gender and general practice. Calendar time is matched on by default when using incidence density sampling, as the controls have to be “active” (in this case, currently registered) when the case was diagnosed with zoster.
95
Age is a very strong risk factor for zoster, therefore an important confounder to control for as tightly as possible. Gender is another well-reported risk factor for zoster, and associated with a number of the potential risk factors. Matching on practice controlled for practice level SES helping to account for some unmeasured socioeconomic differences as well as allow heterogeneity between patterns of prescribing or recording to be accounted for.
4.4.2.2.
Ratio of cases to controls
Cases were matched to up to four controls, as any more controls is not deemed to increase efficiency.134
Matching programme
With the help of a statistician, an algorithm was developed to individually match cases to controls, as follows:
1. All potential controls for each case were identified: as cases were eligible to be controls up until their zoster diagnosis, they were included in the pool of potential controls, with their end date redefined as their zoster date, minus one day. 2. Up to four controls were randomly selected for each case at random without
replacement. Priority was given to potential controls closest in age to the case.
96