Researchers conduct an RCT comparing a new tuberculosis regimen (6-month) versus standard therapy (6-month) in 400 newly diagnosed patients in Delhi. Randomization is done using a computer-generated sequence, and both clinicians and patients are blinded to assignment. At 6 months, 320 patients complete the trial; 40 are lost to follow-up (20 from each arm), and 40 withdraw due to adverse events (15 from new regimen, 25 from standard). The investigators perform per-protocol analysis on the 320 completers only. Which bias is most likely introduced, and why is intention-to-treat (ITT) analysis the preferred approach here?

Explanation

## The Problem: Per-Protocol Analysis on Incomplete Data **Key Point:** Excluding 80 participants (20% of the cohort) from analysis violates the principle of intention-to-treat. The differential dropout (15 vs. 25 adverse events) suggests that the new regimen may be better tolerated—but per-protocol analysis excludes those who withdrew, creating a biased sample. ## Bias Introduced: Attrition Bias **High-Yield:** Attrition bias (a form of selection bias) occurs when dropouts and losses to follow-up are *not random* and are *related to the outcome*. | Characteristic | Impact | |---|---| | **Differential dropout** | 15 vs. 25 adverse events suggests regimen tolerability differs | | **Outcome-related loss** | Patients withdrawing due to AEs may have worse outcomes; excluding them biases efficacy estimates | | **Loss of baseline balance** | Per-protocol analysis compares non-random subsets, not the original randomized groups | | **Direction of bias** | New regimen may appear *more* efficacious if sicker/intolerant patients are excluded | ## Why ITT Is the Gold Standard 1. **Preserves randomization:** All 400 randomized participants are analyzed in their assigned groups, regardless of adherence, dropout, or outcome. 2. **Maintains baseline comparability:** The groups remain balanced on known and unknown confounders at baseline. 3. **Reflects real-world effectiveness:** ITT estimates the effect of *assignment* to treatment, not the effect of *receiving* treatment—more clinically relevant for policy decisions. 4. **Prevents attrition bias:** Dropouts are included, so differential loss-to-follow-up does not distort the estimate. **Clinical Pearl:** ITT is conservative—it dilutes the observed treatment effect by including non-adherent and lost-to-follow-up participants. But this conservatism is *appropriate* because it reflects the true intent-to-treat effect in a real population where not everyone adheres or completes follow-up. **Mnemonic: ITT = Integrity of Randomization** — ITT preserves the causal structure created by randomization, even when participants don't follow the protocol. ## Why Other Options Are Incorrect - **Detection bias (Option B):** Detection bias arises from unblinded outcome assessment. The question states both clinicians and patients are blinded; detection bias is not the issue. ITT does not reduce observer bias; it prevents attrition bias. - **Performance bias (Option C):** Performance bias occurs when treatment allocation influences how participants are managed. Blinding addresses this. ITT does not prevent clinicians from knowing who completed the protocol; it includes all participants in analysis regardless. - **Recall bias (Option D):** Recall bias is a measurement error in self-reported data. ITT does not minimize recall bias; it addresses selection bias from dropout. The timing of analysis relative to randomization does not reduce memory errors.