Statistics from Altmetric.com
We read with interest the cohort study by Cheung et al 1 reporting that use of proton-pump inhibitors (PPIs) after Helicobacter pylori eradication is associated with an increased risk of gastric cancer (GC). We believe that the significantly elevated HR of GC with PPI use (HR 2.44; 95% CI 1.42 to 4.20) is a consequence of two time-related biases. Immortal time bias was introduced by misclassifying exposure,2–4 while latency bias was introduced by not incorporating latency in the exposure definition, a major issue for potential carcinogenic drug effects.5 6
Immortal time in a cohort study refers to a period of follow-up time when the outcome could not occur.3 It arises in this study from the definition of frequency of PPI use, ‘calculated by dividing the total treatment duration by the duration of follow-up’, namely the average use over the entire follow-up. Exposure to PPIs was then categorised ‘into non-regular use (<weekly use; reference group) and regular use (at least weekly use)’.1 Thus, the PPI ‘non-users’ included authentic non-users of PPIs and the non-regular PPI users (less than weekly use), together making up the reference group against which the risk of GC for the PPI users was compared. This explains the paucity of PPI ‘users’ in this study (3271 out of 63 397, ie 5%), which actually represent only the continuous weekly PPI use over the median of 7.6 years.
Consequently, a patient who used PPIs regularly for some time, but not for the entire follow-up, was considered a PPI non-user in this study because the overall average use is less than weekly, leading to exposure misclassification. Figure 1 depicts this pattern in the top two subjects. For the top subject, the initial period of regular use (in red) is called ‘immortal’ because they could not have developed GC; had they developed GC during that time, they would have been classified as PPI user, just like the bottom two subjects. This immortal time period was misclassified as non-use when it should have been classified as PPI use. The second subject is a regular PPI user for the second half of follow-up, which was misclassified as non-use. Such misclassified immortal time will artificially increase the time to cancer, thus falsely lowering the rate of GC for the reference group of ‘PPI non-users’.3
With respect to latency, Cheung et al did correctly exclude ‘prescriptions of these agents started within 6 months prior to the gastric cancer diagnosis’ to avoid protopathic bias.1 However, PPI exposure, defined as more than weekly regular use, had no requisite for a latent period. Figure 2 illustrates five typical subjects who all developed GC during follow-up. The top three GC cases have, by design, been classified as unexposed to PPIs when in fact they should have been classified as exposed for all or part of the follow-up, as the cancer can develop long after the exposure period has ceased.5
It would thus be important for the authors to reanalyse their cohort by classifying PPI exposure in a time-varying way over follow-up time and accounting for latency. First, a biologically plausible minimal duration of regular use would be needed to define exposure to PPIs— 1 year, for example. This analysis would use a Cox model with time-varying PPI exposure (same definition of ‘more than weekly use’, but for a year, say) computed and updated over follow-up time at each of the 153 risk sets, namely whenever a GC case occurs.7–9 Thus, the frequency of PPI use, calculated by dividing the treatment duration by the duration of exposure up to the point of the risk set, should be computed and updated in a time-varying manner. For example, the Cox model with time-varying PPI exposure would classify the top patient in figure 2 as a PPI non-user for the first year of use, and as a PPI user (1 year of regular use) subsequently, including the time they discontinued use, so as to account for latency. The second patient in figure 2 would be classified as a PPI non-user for the period before initiating PPIs as well as for the first year of use, and as a PPI user subsequently. This approach could be extended to different definitions of regular use.
In conclusion, whether PPI use is associated with an increased risk of GC is an important question for treating physicians and public health regulators. Alas, the study by Cheung et al used a design that introduces two time-related biases, immortal time bias, a recognised bias in gastroenterology,4 and bias from lack of latency. A proper reanalysis with time-varying PPI exposure to avoid immortal time bias and proper latency considerations will provide much reassurance on the validity of these important findings.
Competing interests None declared.
Provenance and peer review Not commissioned; internally peer reviewed.
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.