| Covariates reported with justification |
These covariates represent the variables which are included in the PSM model. Omitting a true confounder (if available) may bias results. Justification for these covariates provides rationale to readers regarding each selection and allows readers to independently assess if important variables were omitted from the propensity score model. |
“Eighty-four pairs of patients were successfully matched using 14 covariates: sex, age, affected side, body mass index (BMI), concomitant symptoms such as vertigo and tinnitus, lifestyle factors such as drinking and smoking, systemic disease such as hypertension and diabetes, audiometric curves, the average of pure tone audiometry evaluations pre and post treatment, and time to treatment initiation” [28] |
| “We selected covariates known to affect treatment selection. These primarily included sociodemographic characteristics (age, highest education level, and marital status). These variables have demonstrated associations with the ability to travel large distances to specific medical centers, or to affect how severe a patient’s disease was at presentation Additionally, we included variables believed to be related to the outcome but not necessarily the treatment to reduce bias…[etc]” [29] |
| Summary Statistics |
Summary statistics including baseline numbers and percentages for the overall study sample and the post-matched sample. |
Table 3 [30] |
| Covariate Balance |
Covariate balance is used in order to assess whether the two matched groups (treatment and control) differ substantially based on the covariates described above. If a large difference remains, this indicates the two groups have not been ideally matched and that confounding is still present. |
“To test the covariate balance after propensity-score matching, we calculated standardized differences to compare the baseline characteristics of patients between the cetuximab-based RT and CCRT groups for both unmatched and propensity score–matched groups. A standardized difference of >10% was defined as out of balance.” [31] |
| Estimation of Propensity Score |
Specifications regarding the type of regression model used to generate propensity scores. |
“…[P]ropensity scores were estimated for each patient using a multivariable logistic regression adjusting for all covariates.20” [32] |
| Sensitivity Analysis |
Sensitivity analyses can be used to assess for residual confounding, particularly due to bias that was unaccounted for during matching. Specifically, it determines the extent to which an omitted covariate could impact the treatment effect. |
“Formal sensitivity analysis was performed as described elsewhere.” [33] |
| Matching Algorithm |
The method through which patients in each group are matched based on their calculated propensity scores (e.g. Greedy algorithm vs Optimal algorithm). |
“The second step was matching patients 1:1 via the nearest-neighbor matching strategy without replacement, with 0.2 SD of the logit of the propensity score as the caliper value.” [34] |
| Matching Ratio |
This describes the ratio of untreated subjects that are matched to treated subjects (e.g. 1:1, 2:1, 3:1, etc.). 1:1 is the most commonly used ratio. |
| Caliper Specifications |
The caliper is the maximum distance or value that propensity scores between matched subjects is allowed differ. |
| Replacement |
This describes whether a single untreated subject is allowed to be matched with more than 1 treated subject (matching with replacement) or to only 1 treated subject (matching without replacement) |
| Paired Statistical Methods |
The statistical method used to assess treatment effects. Statistical tests can either assume samples are independent or paired. |
“We used Kaplan-Meier methods and multivariable Cox proportional hazards regression models to evaluate OS…The Cox models were then stratified by matched pair, and CIs were calculated using robust SEs to account for correlated observations.” [35] |
