What drives treatment response?
Response tracks adherence in the trial data. Does adherence cause response, or are both driven by comorbidity — and how much does dose matter on top?
The decision
Whether to invest in an adherence program, and whether dose escalation is worth it alongside. The action is adherence; the outcome is response; the unit is a participant.
- Action
- adherence
- Adherence
- Outcome
- response
- Unit
- participants
Why it resists a spreadsheet
A confounded adherence effect is the classic way a trial team concludes that an adherence program is the lever, when what actually moved response was the illness mix of the enrolled population. The table shows adherence and response moving together and cannot distinguish the two stories.
The stakes are asymmetric. Concluding that adherence matters when it does not buys an expensive program; concluding it does not matter when it does abandons the cheapest available improvement. Both mistakes are made from the same correlation.
The worked example
Run against the shipped trial-response scenario: four binary variables, 1,500 rows, screened at an uncorrected threshold of 0.01 over 18 tests.
Same declared-truth scoring. All 12 of 12 observable rows fitted from counts. Fewer tests than the other scenarios because there are fewer variables, which is why the recovered graph here is the cleanest.
What the screen found
- Precision
- 1.000
- of the edges found, the share that are true
- Recall
- 1.000
- of the true edges, the share that were found
All four true edges recovered and nothing spurious. This is the scenario where the classical confounder is present and found: comorbidity drives both adherence and response, exactly as the vignette says.
The effect
Effect
+32.60 percentage points
risk difference, on the probability of a good response
risk ratio 2.181
Band
+24.72 to +40.49 percentage points
sensitivity_envelope, basis table_sensitivity@0.05
High adherence moves a good response from 27.61% to 60.21%, adjusting for comorbidity. This is the only positive effect of the four, and it is the only one whose band is wide relative to the effect — which is a better sign than the opposite.
What to adjust for, and what not to
Two of the four variables carry advice. This is the cleanest of the four examples and the one where the textbook answer is right.
- adherencen/a
treatment
The variable being intervened on.
- responsen/a
outcome
The variable whose response is being measured.
- comorbidityrequired
confounder
Leaving it out leaves a backdoor path open, and the estimate absorbs the difference. An ancestor of both, with a direct route to the outcome that does not pass through the treatment — which is what distinguishes it from the manufacturing page's version of this row.
- doseoptional
cause of outcome
It cannot confound the effect, so adjusting for it changes precision rather than the answer.
The minimal adjustment set is comorbidity alone. Dose is in the full backdoor set and not in the minimal one: it is upstream of the outcome only, so it cannot confound the estimate, and including it changes precision rather than the answer.
One question this scenario cannot answer
If an unmeasured clinical factor drives both adherence and response, what is the effect then?
Adjustment cannot help against an unmeasured common cause, and there is no front-door route because the treatment has no child leading to the outcome. Identification is not established; both failed routes are named.
This is the question that decides whether the adherence effect is real, and the honest answer is that it cannot be settled from observational data without further assumptions. The E-value says an unmeasured confounder would have to be associated with both adherence and response by a factor of 3.79 to explain the effect away. In a trial population, a factor of 3.79 is not a high bar.
What this does not establish: This is adherence in trial data, not a randomised adherence program, and nothing here controls for the fact that adherence is itself affected by side effects and tolerability. The scenario also has no dropout, so the people who stop taking the treatment simply are not in it — which in a real trial is exactly the population at issue.
What the record keeps
The adherence intervention as framed, who approved it against which stored analysis, which participants were assigned versus actually adherent, and the measured response reconciled against the design-time claim.
- The analysis as run, with its engine version
- The adjustment set and the roles table, so the comorbidity control is on the record
- Assigned versus reached — which in adherence studies is where most of the interesting variation lives
- The measured response against the design-time claim
Book a walkthrough
Bring the response question. The measurement worth having is the one that says whether the adherence effect survives an unmeasured clinical factor.