¹ No serious inconsistency. All four studies report a decrease in attrition at 12 months.
² Not downgraded for indirectness. The studies included adults (two studies), children (1 study) or both (1 study); and were conducted in sub‐Saharan Africa (South Africa, Malawi).
³ Upgraded by 1 for large effect size. The effect estimate indicated a 54% decrease in attrition in the decentralised group.
⁴ Adjusted rates for Brennan 2011, Chan 2010 and Fatti 2010 are consistent with the crude proportions reported here. In Brennan 2011, the adjusted hazard ratio was 0.3 (95% CI 0.2 to 0.6)/ 100 person years indicating better outcomes at the health centre. Chan 2010 reported an adjusted odds ratio of 0.48 (95% CI 0.4 to 0.58) indicating better outcomes at the health centre. Fatti 2010 presented the results inversing the site of risk, the adjusted hazard ratio was 2.19 (1.94 to 2.24) indicating greater problems with patients failing to attend the hospital.
⁵ No serious inconsistency. Three of the four studies show benefit with varied effect sizes (39%. 51% and 66% reduction in patients lost to care), the smallest study reports no difference in clinic follow‐up at 12 months.
⁶ Adjusted rates for Brennan 2011, Chan 2010 and Fatti 2010 are consistent with the crude proportions reported here. In Brennan 2011, the adjusted hazard ratio was 0.2 (95% CI 0.04 to 0.8)/ 100 person years indicating better outcomes at the health centre. Chan 2010 reported an adjusted odds ratio of 0.19 (95% CI 0.15 to 0.25) indicating better outcomes at the health centre. Fatti 2010 presented the results inversing the site of risk, the adjusted hazard ratio was 1.6 (95% CI 1.3 to 1.99) indicating relatively increased risk of death in patients attending the hospital.
⁷ Not downgraded for methodological limitations. For one included study (Fatti 2010), the health centre group had balanced CD4 cell counts, but more severe illness ‐ 79% had WHO clinical stage III or IV disease compared with 58% in the hospital group. However, this would tend to favour the hospital group so we did not downgrade on baseline imbalance.
⁸ No serious inconsistency. All four studies show decrease in death at 12 months with varied effect sizes (10%, 74%, 77% and 81% reductions).
⁹ Not upgraded for large effect size, despite large effect size and narrow confidence interval, this review is not aiming to explore whether decentralisation decreases death, rather excluding that it increases death. The model of care down refers healthier patients for maintenance therapy, generally sicker patients remain at the hospital setting, this therefore favours decentralisation.

¹ Downgraded by 1 for methodological limitations. Bedelu 2008, McGuire 2013 and Massaquoi 2009 included sicker patients at the hospital setting, Assefa has unknown baseline risk as other baseline characteristics were not reported. This bias would tend to favour therapy provided at the health centre.
² Not downgraded for inconsistency. Three studies report significantly reduced attrition with decentralisation (13%, 42% and 52%), while one study reported no difference.
³ Not downgraded for indirectness. The studies included adults (3 studies) or adults and children (1 study); and were conducted in sub‐Saharan Africa (South Africa, Malawi and Ethiopia). This model of care is probably applicable in better resourced settings where basic levels of healthcare are likely to be better resourced, favouring decentralisation.
⁴ Downgraded by 1 for imprecision. Although the sample sizes are large and event rates are high, the confidence interval is wide including both appreciable benefit and the null effect.
⁵ Not downgraded for risk of bias. Four retrospective cohorts provided data. Although there were differences in their baseline health status (Bedelu 2008, Massaquoi 2009 and McGuire 2012 included sicker patients at the hospital), this study limitation is not expected to impact on the attendance at the clinic.
⁶ Not downgraded for inconsistency. All four studies showed substantially better clinic attendance with decentralisation, however, the effect sizes varied, 24%, 63%, 80% and 89% reductions.
⁷ Upgraded by 1 for large effect size . The effect size indicates a 70% lower rate of failure to attend clinic follow‐up at the health centre compared to hospital.
⁸ Downgraded for inconsistency.There is qualitative heterogeneity, Bedelu 2008, Massaquoi 2009 and McGuire 2013 include sicker patients at the hospital, yet only McGuire showed increased death in that setting. Therefore the inconsistency is unexplained.
⁹ Downgraded by 1 for imprecision. Although the sample sizes are large and event rates are high, the confidence interval is wide including both appreciable benefit and harm.

¹ Not downgraded for indirectness. Note that the trials were conducted in Kenya and Uganda in adult populations.
² Downgraded by 1 for imprecision. These two cluster trials have been pooled after adjusting for the design effect. The intra‐cluster co‐efficient was assumed, as it was not provided in the trial reports.The included studies have small sample sizes and wide confidence intervals which include appreciable harm and benefit.
³ The cluster randomised controlled trials Selke 2010 and Jaffar 2009 are included in this pooled analysis. Selke 2010 reports the adjusted incidence rate ratio for patients lost to care as IRR 1.15 (95% CI 0.24 to 3.03), P = 1.0
⁴ The cluster randomised controlled trials Selke 2010 and Jaffar 2009 are included in this pooled analysis. Jaffar 2009 reports the adjusted rate ratio for death, RR 0.95(95% CI 0.71 to 1.28); Selke 2010 did not provide adjusted rates for this outcome.