Use of Point‐in‐Time or Window Approach in the Case‐Crossover Design, Implications for Pharmacoepidemiologic Research Using Registries

The case-crossover design and similar outcome-anchored case-only designs compare exposure frequency in a window immediately or shortly before an outcome (“focal window”) with the frequency at one or more control times (“referent windows”) selected from the same person [1]. Such within-person between-time comparison is the defining characteristic of “self-controlled” designs, which are increasingly popular in pharmacoepidemiology because self-matching eliminates confounding by characteristics that are stable over time [1]. There are two alternative traditions for implementing the case-crossover design in a pharmacoepidemiologic study: Using windows and using discrete points in time. In the window approach, one focal and one or more referent windows of equal length are placed backward in time from the outcome's occurrence. Each of these windows is considered exposed if there is a prescription fill within it. The point-in-time approach is to have one focal (outcome) point-in-time and one or more referent points-in-time before the focal time. Each of these points-in-time, whether focal or referent, is considered exposed if a prescription's exposure period (e.g., determined by the quantity of tablets) covers it. If the same duration is assigned to all prescriptions' exposure periods and to all windows, then the two approaches are in effect identical (Figure 1). Both approaches allow for washout. In the window approach, it is generally recommended to have a washout window immediately before the focal (outcome) window whose exposure status is not considered. Its purpose is to guard against bias by carry-over between referent and focal windows. This carry-over could be either biological, that is, caused by a lingering effect, or it could be a carry-over of drug intake, caused by minor non-adherence. In the point-in-time approach, a similar washout is achieved by having twice the distance between the latest of the referent times and the focal time, than between neighboring referent times (see Figure 1). Under some circumstances, the point-in-time approach may have an advantage over the window approach. First, it allows for flexibility when assigning durations to single prescriptions, for example, if one has qualifying information on how long it is supposed to last, for example, the number of tablets or a prescribed daily dose. Such information cannot be incorporated into the window approach. In addition, the point-in-time approach offers a more intuitive exposure definition, since by the window approach, a drug package that is dispensed on the last day of a window will mostly be consumed during the following window but will formally define exposure in the window in which it occurred. Finally, the point-in-time approach allows the researcher to have short intervals between these points, for example, if there is a short look-back prior to the outcomes and the researcher wants to have the added statistical precision offered by using multiple reference points in time [2]. With short distances between reference times, some prescriptions' assigned exposure period might cover more than one referent or focal point-in-time, thereby qualifying these as exposed. Apart from exposure autocorrelation, there is no obvious source of bias inherent in this. The autocorrelation bias can be mitigated by the Mantel–Haenszel procedure or a newly developed weighting technique [3]. There is no principled difference in statistical precision since these two approaches will contribute with the same number of observations if the count of referent times equals the count of referent windows. Both approaches can be implemented in case–time–control studies [4] or case–case–time–control studies [5] as well. The width of exposure assessment windows is no trivial matter, and it may not have had the attention it deserves. If, for example, the windows are set to a width that is narrower than the typical distance between consecutive prescriptions, then some windows will falsely appear to be unexposed, since two neighboring prescriptions might fall right before and right after a window. Thereby, patients who should have been excluded from the analysis as being exposed at all times will have some spuriously unexposed windows. Instead, they will be kept in the analysis and show a false discordance. With abundant chronic use, as is seen surprising often in case-crossover studies [6], this could confer a strong bias towards the null [7]. To minimize such misclassification bias, the width of the windows should correspond to a high percentile (90 or 95) of the distances between consecutive prescriptions belonging to the same episode. In the point-in-time approach, there would be no bias by selecting referent dates with intervals that are either longer or narrower than the typical intervals between consecutive prescriptions, since the selected points in time represent a sample of the look-back of interest before the outcome. If the prescription durations are specified correctly in a chronic user, then the focal and all the referent times will be categorized as exposed, and the subject will—correctly—be excluded from further analysis. With correct exposure assignment, there is no inherent bias from having very wide intervals in the point-in-time approach, since these points-in-time will still represent an unbiased sample of the individual's exposure history. If too wide windows are assigned in the window approach, there is a risk of overlooking true treatment gaps, and if this affects focal and referent windows differentially, there will be a bias. The optimal width for windows is thus unknown and may be subject to future research. The self-controlled case series, another self-controlled design, does not require that the researcher specifies points in time or windows. It is, however, dependent on an accurate assignment of exposure based on the prescription data, and owing to its bidirectional nature, it may be vulnerable to bias incurred by the outcome affecting future drug exposure [8]. It is our impression that in current pharmacoepidemiologic practice, window-based analyses are much more common than point-in-time based analyses, that the width of windows, although usually quite important, is rarely informed by exploratory analyses of prescription renewals, and that data that could inform the choice of prescription duration are too rarely used. We would like to see the point-in-time approach used more often in pharmacoepidemiologic publication. We would also like to see more researchers let their analysis be informed by the observed intervals between prescriptions—before using either of the approaches. The authors declare no conflicts of interest. We thank Lars Christian Lund for valuable input and critical reading.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.