Techniques to adjust for confounding in observational studies
Confounding
Observational data are often used for research in critical care. Unlike randomized controlled trials, where randomization theoretically balances confounding factors, studies involving observational data pose the challenge of how to adjust appropriately for the bias and confounding that are inherent when comparing two or more groups of patients.
Table Techniques to adjust for confounding in observational studies
| Technique | Strengths | Weaknesses |
|---|---|---|
| Matching | Simple Balances confounding factors |
Difficulty finding matches Possibility of overmatching Requires strong understanding of confounders involved Inability to examine effect of confounders used for matching |
| Stratification | Simple Ability to see effect modification |
Difficult to interpret with many subgroups Requires strong understanding of confounders involved |
| Multivariable adjustment | Can include many confounders Can examine effects of individual confounders Ability to examine multilevel effects |
More complicated analysis Potentially poor fit of model Possibility of missing effect modification |
| Propensity scores | Single number generated for simpler matching Ability to assess for bias between groups |
Potentially matching very different patients with similar scores |
| Instrumental variables | Only single variable needed Ability to look at questions where other types of adjustment can not be easily accomplished |
Difficult to ensure variabl is not at all associated with the outcome |
Matching
A relatively simple technique (although not so commonly used in critical care research) is matching to account for bias and confounding. This involves identifying the variables that constitute confounders and then matching patients in the exposed and unexposed groups based on these variables so that the two groups are “the same” with regard to all of these factors. This is an artificial way of forcing a balance of confounding factors, which would otherwise have been taken care of by randomization. Patients can be matched on as simple a confounder as age or on many individual factors (ventilated on admission to ICU, receiving antibiotics, etc).
Stratification
The concept of stratification is similar to matching and provides another simple approach to controlling for confounding. Confounding variables are identified and subgroups created using these variables. For example, if age is a confounder, patients may be grouped into categories of age less than 60, age 60 to 80, and age greater than 80.
Multivariable adjustment
The most common method used for adjustment of confounding in critical care research is multivariable adjustment. After initially examining the relationship between the exposure of interest and the outcome (giving a “crude” or “unadjusted” result), variables that are known confounders are then added to the model to provide an effect that is “adjusted” for these known confounders.
The exact model used depends on the research question being asked and specifically whether the outcome variable of interest is binary (eg, hospital mortality), continuous (eg, number of blood transfusions), or involves an outcome associated with time (eg, survival time). If the outcome of interest is binary, then logistic regression modeling is most commonly used.
Propensity Scores
Propensity scores are similar to severity of illness scores, as they provide a single number to represent a large quantity of variables. Rather than estimating severity of illness and probability of hospital mortality, a propensity score is constructed to represent the probability of a patient encountering the exposure of interest by taking into account many variables.