2.5. Sensitivity analyses

We performed multiple sensitivity analyses to assess whether our results were robust under alternative inclusion criteria and methodological specifications. First, we ran our primary models as unadjusted models. Second, we ran all models using stabilized, truncated inverse probability of treatment weights (IPTWs) as to calculate average treatment effects, 20 , 21 , 22 , 23 based on propensity scores generated from a multiple logistic regression model, where the covariates that were included in the regression adjustment of the primary analysis were included in the propensity score. The use of propensity scores served to balance on observable characteristics between those with higher vs lower levels of unmet social need while reducing the issue of overfitting under the direct adjustment approach. 24 Third, rather than limiting our sample to low‐income adults only, we re‐ran our primary analyses to assess effects across adults of all income levels; these results are more generalizable to all US adults, but also introduce a greater degree of confounding due to unobservable factors across income strata, as income is highly associated with both unmet social needs and access to and quality of care. Finally, we examined the relationship between each of the seven individual social needs measures and each of our four primary outcome measures to better understand which social needs were most strongly associated with quality and access.