Skip to main content

Table 2 Main bias-control strategies in observational studies of pharmacoepidemiologic databases

From: Bias in pharmacoepidemiologic studies using secondary health care databases: a scoping review

Category

Control strategies

Confounding

 Measured confounding

- Multivariate analysis

- Restriction*

- Stratification

- Matching

- New-user design

- Propensity score

- Large-scale, simple randomized trials

- Meta-analysis of clinical trials

* Confounding by indication: Restricting the untreated group to a population with the same indication, or limiting participation to patients without a risk factor for the effect that could have determined the treatment

 Time-dependent confounding

- G–estimation

- Marginal structural models

 Unmeasured confounding

- Crossover design

- Asymmetric exclusion of patients with extreme propensity-score values

- Instrumental variables

- Proxy measures

- Restriction (active comparison group)

- Sensitivity analysis

- Validation study + external adjustment

Selection bias

 Protopathic bias

- Restriction (e.g. restricting the untreated group to a population with the same indication, or restricting the treated group to a population with an indication that is not a subclinical stage of the disease)

- Excluding a specific period of time prior to the date of diagnosis of the disease (lag-time) from the etiologic window

 Losses to follow-up (informative censoring)

- Inclusion of variables that affect censoring and event times in the multivariate regression model

- Inverse probability of censoring weighting

- Sensitivity analysis

 Depletion of susceptibles (prevalent user bias)

- New-user design

- Meta-analysis of clinical trials

 Missing data

- Replacing each absent observation with a mean value based on observed values of the variable or the predicted value based on a regression model

- Imputation methods (e.g. multiple imputation)

- Likelihood-based methods

- Inverse probability weighting

Measurement bias

 Misclassification bias

- Validation study (exposure/outcome/confounders) + (sensitivity analysis/misclassification control techniques using multivariate regression)

Time-related bias

 Immortal time bias

- Data analysis with procedures that take into account time-dependent exposure in a cohort

- Transferring the start of treatment to the end of the immortal time period in both groups

 Immeasurable time bias

- Data analysis accounting for the time-varying exposable period

 Time-window bias

- Accounting for duration of treatment in the selection of controls

- Time-dependent analysis

 Time-lag bias

- Comparing patients at the same stage of disease