3. Risk of bias ACTG‐A5164, 2009.

Bias	Authors' judgement	Support for judgement
Random sequence generation	Low	Random sequence was generated by central computer using permuted blocks within strata. Neither block size nor treatment assignments to other sites were public.
Allocation concealment	Unclear	No details provided in protocol or included study.
Blinding of participants and personnel	High	Protocol stated that for arm B (deferred ART), no study‐provided drugs were to be provided initially hence blinding of participants and personnel was not possible.
Blinding of outcome assessment	Low	Primary outcome was a composite endpoint of survival and viral load. Detection bias was unlikely.
Incomplete outcome data	Low	Equal numbers withdrew without primary endpoint data in each study arm. Details provided for these participants.
Selective reporting	Low	Reported outcomes were consistent with protocol.