Owen Yang
Many of us are aware of the idea of confounding. When we find an association between two things, the apparent association can be confounded and so not necessarily causal.
If we find children taking inhaled corticosteroid have poor asthma control compared to children who do not take inhaled corticosteroid. This does not mean inhaled corticosteroid cause poor asthma control. In this case, it is probably that those who have poor asthma control require inhaled corticosteroid more often than others, and hence the apparent association.
Reverse causation is an example of strong confounding
In this case sometimes we would say it is reverse causation, but we can also see prior asthma control and current/future asthma control as two separate things, and the apparent association between inhaled corticosteroid and current/future asthma control is ‘confounded’ by prior asthma control. Prior poor asthma control had caused the prescription of inhaled corticosteroid, and had also caused current/future asthma control. If there is no effect between inhaled corticosteroid on subsequent asthma control, we can still see an apparent association between inhaled corticosteroid and poor subsequent asthma control, simply because of the confounding of the prior asthma control.
If using inhaled corticosteroid does improve asthma control, we may be able to see an association between inhaled corticosteroid and good current/future asthma control if the effect of inhaled corticosteroid is strong enough to overcome the existing confounding.
The misconception about adjustment
One misconception that I sometimes hear is that adjustment is a solution to confounding. This is partially true but perhaps misconstrued. The misconception goes like this: when there is confounding, the association should be adjustment for (or ‘controlled’ for) the confounding factor, and we can trust the finding after adjustment is unbiased or not confounded. Therefore, if we adjust for prior asthma control, then the adjusted association between inhaled corticosteroid and future asthma control should be ‘legit.’
This is probably true in an ideal world, but in reality we need to use professional judgement depending on the context. In this case, the link between prior and future asthma is so strong, and the quality of adjustment really needs to be very good to eliminate all the doubts for any residual confounding.
Adjustment is just weighted association
It is easy to give up technical details and let statisticians to lead, but hear me up: adjustment is just weighted association, and the rest details are really not that important. We believe an apparent association is because one group is different from another group based on a confounding factor, and in this case those who use inhaled corticosteroid have poor prior asthma control compared to who does not use inhaled corticosteroid. Instead of doing a raw association, we balance the weight between users and non-users based on prior asthma control, to look at what the association would be if the two groups have the same distribution of prior asthma control. This is basically what adjustment is. If the good and poor adjustment is 40:60 in one group and 60:40 in another group, they are re-weighted and we see what the association may look like if they were both 50:50.
Therefore, the adjustment should really be limited to ‘adjustment’ and should not go beyond reason. if it is not 40:60 vs 60:40, but 1:99 versus 99:1, a forced adjustment would be asking the 1% to represent 50% of the group, and the risk of misrepresentation would be really high. There could be something unpredicted for the 1% poorly controlled individuals with asthma not having been prescribed with inhaled corticosteroid.
There are many situations where adjustment could work, but in an extreme example of observational study like this, adjustment does not solve everything.
So tell me, should you be convinced with the result or worried about the result if you find the association changes after adjustment?
