Year: 2022
Type: article
Abstract: A propensity score (PS) can be used to minimize confounding bias in observational studies, where the treatment assignment (or more broadly exposure) is not random1–3. In its most basic form, the PS is the estimated probability of being treated (as opposed to not treated) given the measured covariates. It can be estimated by fitting a regression model (typically logistic regression) with the treatment received (or exposure) as outcome. There are four common ways in which a PS is used to estimate the effect of the treatment: matching, stratification, adjustment, and inverse probability of treatment weighting. For example, it could be used to seek to compare use or not of oral antibiotic and mechanical bowel preparation following left-sided colonic rectal resections4. One of the most common implementations is PS matching where each treated subject is matched with one or more untreated (or unexposed) subjects with a similar value of PS. How ‘similar’ is defined varies, but is usually based upon the width of the so-called calliper (i.e. the difference in the PS between two subjects)5. In practice, matching often results in exclusion of study participants that have no match which in turn can result in bias. On occasion, the proportion excluded can be high which limits generalizability6. Recent reports suggest that matching on PS should not be used to adjust for confounding in observational studies, and may in fact increase confounder imbalance leading to estimates with greater bias7. Stratification is usually implemented by dividing the study sample into subsets based upon the PS; however, this can lead to large heterogeneity within strata. Adjustment is typically done by including the main confounders and the PS in the analysis model for estimation of the treatment–outcome association. However, this procedure may give little gain, as it generally leads to the same estimate as standard adjusted regression if the same confounders are used in the regression and in the estimation of the PS. Alternatively, PS can be used to derive weights, so-called inverse probability of treatment weights to adjust the analysis to account for differences in the propensity to receive treatment. The simplest approach is to derive weights such as 1/PS for the treated and 1/(1 − PS) for the untreated. These weights are then used in the statistical model for estimating the association between treatment and outcome8. Extreme weights can lead to bias and methods for stabilizing weights exists9. The inverse probability of treatment weighting can take into account time-dependent confounding and selection bias8,10. Using a PS approach does not necessarily mean that the estimated effect can be interpreted as causal. Using PS for causal inference, such determining the right treatment decisions (using a laparoscopic or open surgical approach) requires that certain assumptions are fulfilled. For example, that there are no unmeasured confounders and that each subject has a positive probability of receiving each treatment. It is important to remember that using the PS approach does not necessarily remove potential bias from unmeasured confounders. Further, the choice of variables to include in the model for estimation of PS is not a trivial one. Only true confounders (to remove bias) and covariates predictive of the outcome (to increase precision) are typically useful for inclusion in the PS model. However, not including relevant variables when estimating PS easily leads to bias. Including variables affecting treatment but unrelated to outcome or variables that are effects of exposure and outcome (so-called colliders) generally lead to bias. In surgery, the preoperative (such time from diagnosis until surgery and ASA grade) and operative variables (such as anaesthetic and surgical approach) can typically be safely used to assess operative (pre-operative only) and postoperative exposures (both pre-operative and operative), and confounders. Statistical hypothesis testing should not be used for selection of variables for inclusion in a model for estimating PS9. A situation when the use of PS can be particularly advantageous is when the outcome is rare, the exposure is common, and there are a large number of confounders to adjust for. Standard regression methods may not be suitable, as adjustment for many covariates with rare outcome is unreliable. The PS approach, where the confounders are included only in the model for treatment, allow for adjustment for more variables. Only the corresponding PS is used in the final model for treatment–outcome association11. Gorge Box stated that ‘all models are wrong but some are useful’ and this applies also to using PS. They can help in obtaining useful inferences when applied with care to suitable data. Disclosure. The authors declare no conflict of interest.
Source: British journal of surgery
DOI: 10.1093/bjs/znac032
Institutions: Lund University, University of Oxford
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4206193123
