Treatment Effects in the Presence of Unmeasured Confounding: Dealing With Observations in the Tails of the Propensity Score Distribution--A Simulation Study

Division of Pharmacoepidemiology and Pharmacoeconomics, Brigham and Women’s Hospital, Harvard Medical School, Boston, Massachusetts, USA.
American journal of epidemiology (Impact Factor: 5.23). 10/2010; 172(7):843-54. DOI: 10.1093/aje/kwq198
Source: PubMed


Frailty, a poorly measured confounder in older patients, can promote treatment in some situations and discourage it in others.
This can create unmeasured confounding and lead to nonuniform treatment effects over the propensity score (PS). The authors
compared bias and mean squared error for various PS implementations under PS trimming, thereby excluding persons treated contrary
to prediction. Cohort studies were simulated with a binary treatment T as a function of 8 covariates X. Two of the covariates were assumed to be unmeasured strong risk factors for the outcome and present in persons treated contrary
to prediction. The outcome Y was simulated as a Poisson function of T and all X’s. In analyses based on measured covariates only, the range of PS's was trimmed asymmetrically according to the percentile
of PS in treated patients at the lower end and in untreated patients at the upper end. PS trimming reduced bias due to unmeasured
confounders and mean squared error in most scenarios assessed. Treatment effect estimates based on PS range restrictions do
not correspond to a causal parameter but may be less biased by such unmeasured confounding. Increasing validity based on PS
trimming may be a unique advantage of PS's over conventional outcome models.

Full-text preview

Available from:
  • Source
    • "We examined the distributions of PS in both exposure groups to ensure that only patients with overlapping scores were included. Patients with scores in the extreme upper and lower tail of the PS distribution (outside the 5th to the 95th percentiles) were excluded, as the inclusion of people treated contrary to extreme scores can introduce bias from unmeasured confounding [29]. Sensitivity analyses were conducted without this exclusion. "
    [Show abstract] [Hide abstract]
    ABSTRACT: Background There is strong evidence of reductions in major vascular events from statins across all cardiovascular risk categories. However, trials of statin therapy have provided conflicting results regarding statin use and type 2 diabetes (T2DM). We aimed to assess the effect of statins on T2DM development. Method We carried out a population-based cohort study using the Clinical Practice Research Datalink (CPRD), a database of computerized clinical records. Every patient aged 30–85 years old starting a statin between 1989 and 2009 was matched with up to five non-statin users. The observation period in CPRD ended in 31 December 2011. Cox proportional hazard regression was used to compare rates of T2DM between statin users and non-users, using propensity score method to adjust for systematic differences between groups. Results The study basis comprised 2,016,094 individuals, including 430,890 people who received a statin, matched to 1,585,204 people not prescribed a statin. Mean follow-up time was 5.43 years for statin users and 3.89 years for nonusers. During follow-up 130,395 individuals developed T2DM. Statin use was associated with an increased risk of T2DM (HR 1.57; 95% CI 1.54-1.59), which increases with longer duration of use. The increased risk was smaller among people with hypertension or cardiovascular disease and was only apparent after 5 or more years treatment with statins in these groups. Conversely, age-specific risk ratios decreased in older people. Conclusions Statin use is associated with an increased risk of T2DM. Our results suggest that the relative risk is higher among people without diagnosed hypertension or cardiovascular disease. These findings should be considered in the context of the observational nature of the data which is prone to bias and unmeasured confounding.
    Full-text · Article · Jul 2014 · BMC Cardiovascular Disorders
  • Source
    • "– ajustement par modèle de régression permettant de prendre en compte simultanément plusieurs facteurs de confusion ou d'étudier l'effet de variables quantitatives. L'utilisation d'un score de propension dans les cas où les facteurs d'ajustement sont nombreux est encouragée.101102103104105106107108109110111112113114115D'autres méthodes permettent de comparer deux groupes de patients, comme l'utilisation d'une variable instrumentale,116117118119120121mais elles se basent sur des hypothèses difficilement vérifiables et leur utilisation ne peut être actuellement privilégiée.[122]Il "
    [Show abstract] [Hide abstract]
    ABSTRACT: The assessment of a health technology is frequently accompanied by uncertainty about its impact, at short or long terms, on the health of the population. The Health Authorities may request additional « post-registration » data that will allow a relevant reassessment of these technologies. The responsibility to collect this information lies with the industry and the HAS evaluates the methodology. This guideline provides practical benchmarks on methodological aspects of these studies. It describes the different types of studies to consider depending on the objectives, including the use of databases and cohorts and European studies. It emphasizes the importance of establishing a scientific committee, clearly defining the objectives of the study, justifying the methodological choices, documenting the representativeness or completeness of centers, investigators and patients, limiting the number of lost of follow-up patients and missing data, describing the statistical analysis methods, the bias and their possible impact on results. The publication of the results of these studies is strongly encouraged.
    Full-text · Article · Dec 2012 · Thérapie
  • Source
    • "To improve comparability of the 2 vaccine groups, we controlled for the propensity score in multivariate models, but first we excluded all outlier observations (“trimming”), defined as those below the lower 2.5% of the tail of the TIV observations (3,355 observations) and above the upper 2.5% tail of the ATIV observations (2,894 observations). These tails are outside the primary area of overlap of the propensity scores, and they increase residual confounding in any type of analysis (19). "
    [Show abstract] [Hide abstract]
    ABSTRACT: Although vaccination against influenza is recommended for elderly and high-risk patients in many countries, efficacy in the elderly has been suboptimal. The MF59 adjuvanted trivalent inactivated vaccine (ATIV) was developed to increase the immune response of elderly subjects to influenza vaccination, but its effectiveness has not yet been well documented. This prospective, observational study evaluated the relative effectiveness of ATIV versus nonadjuvanted trivalent inactivated vaccine (TIV) in individuals at least 65 years of age in Lombardy, northern Italy. Hospitalizations for influenza or pneumonia (International Classification of Diseases, Ninth Revision, Clinical Modification, codes 480-487) during the 2006-2007, 2007-2008, and 2008-2009 influenza seasons were identified from administrative databases. Stratified and regression analyses, including the propensity score to adjust for confounding, as well as generalized estimating equations to account for repeated vaccination, were used. Overall, 107,661 records were evaluated, contributing 170,988 person-seasons of observation. Since ATIV is preferentially recommended for more frail individuals, subjects vaccinated with ATIV were older and had more functional impairment and comorbidities. In the primary analysis, risk of hospitalization for influenza or pneumonia was 25% lower for ATIV relative to TIV (relative risk = 0.75, 95% confidence interval: 0.57, 0.98). To the extent that there is residual bias, ATIV is likely to be even more protective than this result suggests.
    Full-text · Article · Aug 2012 · American journal of epidemiology
Show more