Brief Communication
Pooling of Confounders Did Not Induce Residual Confounding in Influenza Vaccination Studies

https://doi.org/10.1016/j.annepidem.2009.02.001Get rights and content

Purpose

In observational studies on influenza vaccine effectiveness, confounding variables such as individual chronic diseases often are pooled before inclusion into a multivariable regression model. It has been suggested, however, that the pooling of confounders induces residual confounding, although empirical evidence is scarce. We set out to study the effects of combining several confounders into classes of co-morbidity.

Methods

In a retrospective cohort study on the association between influenza vaccination and mortality, the effect of pooling of 20 individual diagnoses into three dichotomous co-morbidity variables indicating the presence of at least one of a range of diagnoses was studied. The sample size allowed for adjustments for 22 confounders (age, sex, and 20 individual cardiovascular, pulmonary, or oncologic diagnoses).

Results

After adjustment for age and sex, further adjustment for 20 separate confounders or the three pooled co-morbidity variables resulted in comparable estimates of influenza vaccine effectiveness: odds ratio 0.78 (95% confidence interval, 0.62–0.98) and odds ratio 0.74 (95% confidence interval, 0.59–0.93), respectively.

Conclusion

We conclude that pooling of several (related) confounders in influenza vaccine effectiveness studies in health care databases is unlikely to induce important residual confounding.

Introduction

In observational studies on influenza vaccine effectiveness, several confounding variables are often pooled into a single dichotomous confounder 1, 2, 3. For example, several diagnoses of chronic diseases are combined into a single variable, indicating the presence of at least one of the included diagnoses. However, concerns have been raised recently that such pooling of confounders may result in residual confounding (4). Nevertheless, empirical studies quantifying this potential bias are scarce. We set out to study the effects of combining several confounders into classes of co-morbidity in a study on the association between influenza vaccination and mortality risk among community-dwelling elderly.

An important reason to combine several confounders is the limited number of cases in a dataset; hence, precision may be at stake when the number of covariates adjusted for in multivariable models is too large 5, 6. We used a large dataset from a retrospective cohort study with enough cases to include all confounders as separate covariates in a multivariable model according to the rule of thumb (10 cases per variable) 5, 6. Effect estimates derived from the dataset using different confounding variable approaches could, therefore, be adequately compared.

Section snippets

Methods

We used data from the computerized medical database of the University Medical Center Utrecht General Practitioner Research Network. This database complies with Dutch guidelines on the use of medical data for research purposes and has shown to be valid in influenza vaccine effectiveness studies (2). Diagnoses are coded according to the International Classification of Primary Care (ICPC) coding system. Information on all elderly (age ≥65 years) from seven influenza epidemic periods (1995/1996,

Results

The prevalence of different types of cardiovascular co-morbidity ranged from 0.3% (chronic ischemic heart disease) to 3.3% (congestive heart failure). The prevalence of pulmonary co-morbidity ranged from 0.2% (lung cancer) to 4.4% (COPD). The prevalence of different types of cancer ranged from 0.04% (Hodgkin's disease) to 0.8% (skin cancer). In total, 379 persons died during 44,418 influenza epidemic periods of observation. During 32,388 periods of observation (72.9%) subjects received

Discussion

Estimates of the effect of influenza vaccination on mortality obtained after adjustment for all individual diagnoses and after adjustment for three pooled co-morbidity variables were similar.

Effects of influenza vaccination on serious, infrequent outcomes such as mortality have only been studied in observational studies 1, 9. Obviously, such study designs are prone to confounding bias and which variables are important confounders in these studies has been fiercely debated 4, 9, 10, 11. These

References (16)

There are more references available in the full text version of this article.

Cited by (0)

This study is part of a personal grant of Dr. E. Hak to study confounding in observational intervention studies by the Netherlands Scientific Organization (VENI no. 916.56.109). There are no conflicts of interest.

View full text