Reader Comments
Post a new comment on this article
Post Your Discussion Comment
Please follow our guidelines for comments and review our competing interests policy. Comments that do not conform to our guidelines will be promptly removed and the user account disabled. The following must be avoided:
- Remarks that could be interpreted as allegations of misconduct
- Unsupported assertions or statements
- Inflammatory or insulting language
Thank You!
Thank you for taking the time to flag this posting; we review flagged postings on a regular basis.
closeTIV, H1N1 and the perils of confounding by indication
Posted by Smahmud on 07 Apr 2010 at 23:25 GMT
Observational studies of the effects of medical interventions (e.g., vaccines and other pharmaceuticals) are often difficult to interpret, because of the lack of randomizations which makes them susceptible to many subtle and not so subtle biases, and because of the complex and hard to measure factors that influence both the receipt of the intervention and the development and detection of the disease under study. For instance, the findings of Skowronski et al could be readily explained by any factor that would both increase the likelihood of receiving the TIV and the chances of the acquisition and detection of pH1N1.
Some of these factors were considered by the authors, e.g., age, gender, presence of certain chronic conditions, increased utilization of health care services, occupation and delay in testing, although none of the described studies adjusted for all these variables simultaneously, and when adjustment was attempted it was often made based on crude measurements and broad categorizations that likely limited the ability of the authors to adequately adjust for confounding. For instance, instead of adjusting for specific chronic conditions that would increase the risk of acquisition of pH1N1, adjustment was made for the mere presence or absence of any chronic condition. The small number of exposed cases makes it difficult to perform fine adjustment for confounders (even for age) without running into colinearity and sparse data issues.
Other factors were not considered, e.g., use of steroids and other immunosuppressants, ethnicity, socio-economic status, access to health care (e.g., having a regular family physician), travel history, residence in a long-term care facility (e.g., for youth with developmental and mental health issues), and place of residence. All these factors could influence the receipt of TIV, but could also influence the risk of acquiring pH1N1, the likelihood of presenting for medical care and the chance of getting tested for pH1N1.
So it is possible that people who are more likely to receive the TIV are on average more likely to be tested for pH1N1 or more likely to acquire pH1N1 due to other associated illnesses or medication use, and that could readily explain the findings of Skowronski et al. For example, the observed associations were strongest among younger people, because young people are usually not offered vaccination unless they have a chronic illness or meet other eligibility criteria. Similarly, the associations were much weaker in the Ontario data, because of universal access to vaccination, which tends to weaken the confounding effects of chronic illness and health care access factors (a form of imperfect randomization).
One test of this alternate hypothesis is to repeat the same analysis for the pneumococcal vaccine, which has similar indications and patterns of utilization as the TIV. Since the stipulated biological mechanisms for the effect of TIV seem to be unique to influenza vaccines, a causal interpretation of the findings of Skowronski et al will be less likely if the pneumococcal vaccine was also found to be associated with increased risk of pH1N1 detection.
RE: TIV, H1N1 and the perils of confounding by indication
DanutaS replied to Smahmud on 15 Apr 2010 at 22:08 GMT
We thank Dr. Mahmud for his thoughtful comments.
We agree that bias and confounding cannot be ruled out in observational designs.
Randomized controlled trials were not possible to assess observations during the spring-summer 2009 in Canada. Our challenge has thus been to identify the methodological flaw in our several observational designs that can explain the increased risk we observed for pandemic H1N1 (pH1N1) but still allow the significant vaccine protection we observed against seasonal strains.
In all studies that we present, combinations of the following possible confounders on influenza vaccine effect were assessed: age, sex, comorbidity and delay between symptom onset and testing. The further follow-up studies conducted in Ontario and Quebec allowed assessment of additional covariates. This included exploration of occupation (notably health care worker status), prior health care utilization (as proxy for health care access or health care seeking behaviours), and number of children in the household (as proxy for exposure opportunities). In multivariable analyses, none of these additional covariates was influential. Additional analytical approaches undertaken in the Ontario dataset such as the application of propensity scores also did not alter findings.
Further analyses that were explored (but not presented) from the Ontario and/or Quebec data included select categories of comorbidity, travel history, smoking and ethnicity – again none of which was influential. Importantly (but often over-looked in interpreting our results), the effect persisted with restriction to healthy participants without comorbidity. It thus seems unlikely that differences in immunization programs in Ontario (universal) versus Quebec (high-risk targeted), can explain variation in provincial point estimates of vaccine effect as suggested by Dr. Mahmud. In any case, the studies in both provinces showed increased pH1N1 risk associated with prior seasonal vaccination. Differences in precision may be explained by obvious differences in study design and power. We have also considered that the domestically-produced vaccine comprised a greater proportion of vaccine distributed in Quebec than in Ontario.
Other methodological influences proposed by Dr. Mahmud such as residence in long term care are not anticipated to have played a major role in our community-based studies. Long term care residence is uncommon and may be associated with greater immunization, but it is also generally associated with advanced age and thus anticipated to be associated with lower rather than higher risk of pH1N1. We had few elderly participants in our studies overall. We do not consider the indications for pneumococcal vaccine to be the same as that of influenza (different target groups and scheduling) but we nevertheless explored pneumococcal vaccine effect on the risk of pH1N1 in the Ontario data set as also suggested by Dr. Mahmud. During the same period and with the same adjustments we found this effect to be null.
We do not believe our findings can be explained by differential health care access or health care seeking behaviour. With its active and systematic sampling of all household members, the prospective cohort study fully addressed this potential bias. Because both cases and controls presented to a clinician within a comparable interval since onset of influenza-like illness, the sentinel test-negative study similarly addressed this issue. Canada has a publicly-funded health care system that is free to patients at use, thus removing restrictions on access that may apply elsewhere.
In order that confounding may explain our observations, vaccinated participants must have had some other quality which resulted not in their being tested more often, but rather in their being more likely to test positive for pandemic H1N1 (but not seasonal influenza) than unvaccinated participants. None of the covariates proposed to date has served in this role. Instead, analytical approaches to address proposed confounders have generally resulted in similar or slightly increased odds ratios. It may be that bias or residual confounding by some other unrecognized factor is responsible for increasing odds ratios in our analyses but if that is true, it is unclear why this should be operating in Canada alone. Furthermore, over the several years of the sentinel study’s implementation in Canada, and even in 2008-09 prior to pH1N1 circulation, we have consistently identified vaccine protection. The methodological crux of the conundrum is thus primarily to understand why suggested biases and confounders, if operating, should act differently between years, and even within a single season, and why only in Canada.
Our findings thus raise important scientific questions – either in terms of the immunological mechanisms to explain our findings (if real) or else the methodological issues inherent to current systems for routinely monitoring vaccine effectiveness. Since it is not possible to resolve either of these issues in the current set of studies, we hope our findings will instead inspire others within the broader scientific community to conduct further empirical and theoretical explorations.
Danuta M. Skowronski MD, MHSc, FRCPC
BC Centre for Disease Control
On behalf of the Canadian SAVOIR Team