Thread by @rheum_cat

Biostatician @MikeLaValley8 @BU_BMC_Rheum presents a quick overview on how to appropriately interpret results from studies #ACR22

Of course this is not an exhaustive deep dive, but covers some common errors of interpretation

@MikeLaValley8 @BU_BMC_Rheum First, what is your research question? What is the question you are trying to answer with t he study?

👉Does this exposure CAUSE this outcome?

#ACR22

@MikeLaValley8 @BU_BMC_Rheum What are common observational study designs and what are big issues to watch out for?

These are issues that can affect the validity of their results.

There are more potential issues within these study designs that is beyond the scope of this table

#ACR22

@MikeLaValley8 @BU_BMC_Rheum We care about confounding in observational studies, because individuals are not randomized to each exposure of interest.

But first, what is a confounder?


#ACR22

@MikeLaValley8 @BU_BMC_Rheum Some basic DAGs (directed acyclic graphs) that illustrate exposures, outcomes, and confounders

When thinking about adjustment or other ways to deal with confounding, it can be useful to draw out the DAGs to show what you think the relationships between variables are

#ACR22

@MikeLaValley8 @BU_BMC_Rheum Example A: a common cause creates a non-causal association between the exposure and outcome

⭐Adjustment: tries to remove the effect of a common cause that could create an apparent relationship between exposure and outcome

#ACR22

@MikeLaValley8 @BU_BMC_Rheum Example B: Age confounds the relationship between CKD and mortality

Example C: Adjusting for age removes that confounding

Example D: If ethnicity is not measured well in a study, this leads to residual confounding even after adjustment

#ACR22

@MikeLaValley8 @BU_BMC_Rheum ⭐Restriction: Another way to try to handle confounding

It creates a homogenous group who are more likely to show the effect of the exposure or treatment

But caution! 🛑This can lead to regression to the mean

#ACR22

@MikeLaValley8 @BU_BMC_Rheum Example of regression to the mean with restriction using a BP threshold

🔎No restriction (full plot): BP is strongly correlated with time

🔎Restrict to high BP at baseline: there appears to be a reduction in mean DBP over time when there really is no change

#ACR22

@MikeLaValley8 @BU_BMC_Rheum A word on using p-values to estimate effect size: DON'T DO IT. That's not what this means at all.

#ACR22

@MikeLaValley8 @BU_BMC_Rheum What is effect size?

👉provides the measure of difference between groups or the magnitude of association between exposure and outcome
👉gives the clinical significance of the exposure for the outcome – is the observed effect large enough to be important?

#ACR22

@MikeLaValley8 @BU_BMC_Rheum A bit about confidence intervals

The confidence interval range can be evaluated to see:
🔹If values of clinical relevance lie in the range (clinical significance)
🔹If the null value lies in the range (statistical significance)

#ACR22

@MikeLaValley8 @BU_BMC_Rheum Finally, #Table2Fallacy

Table 2 is often used to show the adjusted results of the exposure

Variable types are highlighted in this Table 1:
🔹Exposure of interest in blue
🍎Outcomes in red
🍏Confounders (that were adjusted for) in green

#ACR22

@MikeLaValley8 @BU_BMC_Rheum #Table2Fallacy

The effect estimates for the blue (exposure of interest) are interpretable ✅

The effect estimates for the red (confounders that were also adjusted for) may not be interpretable 🛑

#ACR22

@MikeLaValley8 @BU_BMC_Rheum More on #Table2Fallacy here: academic.oup.com/aje/article/17…

#ACR22