Sensitivity Analysis for Unmeasured Confounder

[Gregorytus07]

Hi there,

I am trying to conduct a sensitivity analysis to assess the impact of an unmeasured confounder on the relationship between exposure (E) and outcome (Y). To estimate this effect, I utilized the proc genmod procedure in SAS, adjusting for covariates such as age, sex, and comorbidities. However, a recognized confounder, namely fluid volume (a continuous variable), has not been measured. I am seeking assistance with SAS code specifically tailored for performing the sensitivity analysis. The fluid volume can vary from 0 to 2000, with an average of 500. The outcome variable is binary - 0 or 1.

I appreciate any assistance in this matter.

[ballardw]

When you say "unmeasured" that makes me want to ask do you have a any variable in your data set with those values? If not, I am not sure how you expect SAS to analyze something that is not there at all.

[Gregorytus07]

@ballardw Thanks for your feedback. The variable is not measured in our dataset at all. However, I read about methods of using sensitivity analysis for unmeasured confounding by Lui et al (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3800481/). Also, I learned from literature that the method proposed by VanderWeele (https://journals.lww.com/epidem/fulltext/2016/05000/sensitivity_analysis_without_assumptions.11.aspx... using the E-value is widely recommended. Just trying to figure out how to actually do the analysis in SAS.

Thanks for your help.

[MelissaM]

In the manuscript you mentioned by Liu, et al, starting as the 3rd paragraph below "Setting, Assumption, and Notation", it lists the notation that will be used. Table 1 shows where each of those is necessary, as does most of the rest of the publication. There are five approaches to take, depending on which one you feel comfortable with in terms of obtaining the parameters related to fluid volume. The parameters are not just the range (0-2000) and the average (500). You'll also need to factor in prevalence in both those who had the event and those who did not, which is likely not readily available. The way I'm reading it, the assumptions don't work well. Also, keep in mind, fluid volume is almost definitely not the only confounder. I'll stop there because I think that's sufficient. Bottom line, I don't believe any of those methods - or any method - can estimate anything that was not measured in any of the groups. Realistically speaking, we cannot predict an event reliably using demographics.  

[Gregorytus07]

@MelissaM I appreciate your feedback. Your observation regarding the relevance of the assumptions in the article to our context is valid. While the article does propose a method to account for an unmeasured continuous variable, the procedure appears intricate and necessitates specifying more conditions than we are able to handle.