Understanding Statistical Approach

#1
I am currently enrolled in a summer course that is A Review of Statistics for Evidence Based Healthcare. My task was to superficially understand an applied statistical approach in a research paper. Here is the research paper I chose: https://www.bmj.com/content/348/bmj.g1251. This article focuses on prolonged-opioid use after major surgery in opioid-naive patients. As a pharmacy student, this is obviously a very important thing to focus on as the opioid crisis is still strong and present in society as a whole.

They utilized multiple statistical tools to develop their patient characteristics as well as come to a significant statistical outcome. They first utilized descriptive statistics to determine characteristics of the cohort for any patterns post-opioid testing. This was to figure out what characteristics they want to focus on when coming to conclusions in the future about associations. They utilized a bivariate test to determine the relationship of the characteristics to whether or not they experienced prolonged-opioid use. This was to determine associations between the outcome of prolonged-opioid use and each of the patient characteristics. Each of these characteristics were studied to see if they were statistically significant based on if they were found to have p-values lower than 0.05, which is the alpha value chosen by the researchers. They then use a multivariable logistic regression model for adjusted association of patients after discharge in regards to the specific characteristics that they identified as being relevant. They use this because they are focusing on one outcome, prolonged-opioid use after discharge, with regards to multiple variables, the different characteristics. They utilize this to get more reliable and accurate data. Statistical significance is determined by the p-values found from the adjusted analysis, with the same alpha value. With unadjusted analyses, only three characteristics came out to be significant, but after adjustments, four characteristics were found to be significant.

Is this correct or am I still missing important details or explanations?

@hlsmith
 
Last edited:

hlsmith

Less is more. Stay pure. Stay poor.
#2
@allyklein8

The analyses present an interesting phenomenon. Ideally, researchers draw out a causal graph to reveal the relationships between variables and then models them within a single model (e.g., multiple logistic regression). It is uncommon for bivariate analyses not to show an association and then the multiple model to reveal an association. This usually goes the other direction (effect in bivariate and no association beyond chance in multiple model) and is why it can be misleading to conduct bivariate statistics in a complex observational or non-randomized setting. Variable may have a more attenuated association in the multiple model, because two variables may be representing a similar effect or system effect (so they are explaining the same thing) which become apparent in the multiple model. An example is height and age explaining weight. If you know one of them the other becomes less important.
 

noetsi

Fortran must die
#3
I am trying to imagine how a variable could have no bivariate impact, but in the presence of other variables have a real impact. That seems really strange to me. :p As hlsmith says its normally the other way around. Bivariate effects are large, marginal effects in a multiple regression being very small.

Hopefully you build your model based on theory not descriptive or bivariate analysis. But in practice you may have no theory (I think this is more true in medical research than in others because so little is known for instance with new drugs). But you always look for past studies to anchor your analysis in - thus the required literature review. :)

"They utilize this to get more reliable and accurate data." You don't get better data this way - the data is what it is. Other than transformations there is only so much you can do to it. Some methods lead to better results than others, they deal better with data limits for example.