This systematic review was conducted in accordance with the PRISMA statement using a pre-defined protocol (International Prospective Register of Systematic Reviews identification number: CRD42012002554; Text S1 and Protocol S1) [13],[14]. The PubMed database was searched on 4 March 2013, and Embase and WHO Global Index Medicus were systematically searched on 10 April 2012, without language, geographic, publication, date, or any other restrictions. In addition, the WHO International Clinical Trials Registry Platform, the Cochrane Central Register of Controlled Trials, the International Standard Randomised Controlled Trial Number Register, and ClinicalTrials.gov were systematically searched without language, publication, or date restrictions on 3 September 2012. Experts in the field were contacted to identify unpublished research and ongoing studies, and bibliographies of relevant studies were screened.

Community-based HTC was defined as HTC outside of health facilities. Facility-based HTC approaches were defined as those in healthcare sites (e.g., health facilities, hospitals, and fixed, stand-alone voluntary counselling and testing sites). Eleven different community-based HTC approaches were reviewed in this study: (1) door-to-door testing (systematically offering HTC to homes in a catchment area), (2) mobile testing for the general population (offering HTC via a mobile HTC service in areas visited by the general public, such as shopping centres, transport hubs, or roadside restaurants), (3) index testing (offering HTC to household members of people with HIV and persons who may have been exposed to HIV such as spouses, sexual partners, or children of people with HIV); (4) mobile testing for men who have sex with men (MSM), (5) mobile testing for people who inject drugs (PWID), (6) mobile testing for female sex workers (FSW), (7) mobile testing for adolescents, (8) self-testing, (9) workplace HTC, (10) church-based HTC, and (11) school-based HTC.

Several outcomes were analysed in this study. Uptake was calculated by dividing the number of individuals accepting HTC by the number of individuals offered HTC. The proportion of first-time testers was calculated by dividing the number of people reporting receiving their first HIV test by the total number of people tested. The proportion of participants with a CD4 count greater than 350 cells/µl was calculated among participants with HIV who had their CD4 count measured. Two steps of the retention continuum were assessed: (1) CD4 measurement (among all participants found to have HIV) and (2) initiation of ART (among participants eligible per national guidelines). In studies with a comparator arm, the HIV positivity rate was calculated by dividing the number of individuals found to be HIV positive by the number of individuals tested. HTC coverage was calculated by dividing the number of people tested by the total number of people living in the catchment area for the community-based HTC approach. HIV incidence was calculated by dividing the risk of infection in communities with access to community- and facility-based HTC by the risk of infection in communities with access to only facility-based HTC. Some of the outcomes were not independent. For example, the number of people tested was the denominator for the HIV positivity rate and first-time testers and also the numerator for HTC coverage. Moreover, the number of people living with HIV was the numerator for the HIV positivity rate and also the denominator for calculating the first step of the retention continuum (CD4 measurement). The cost per person tested was approximated by dividing the economic costs incurred during HTC in studies by the total number of people tested. Costs were adjusted for inflation from the year the costs were estimated to 2012 United States dollars using the US Bureau of Labor Statistics' inflation calculator [15].

The search strategies (Table S1) were designed with the assistance of a librarian to identify studies including community-based HTC. Following recommendations from PRISMA, eligibility criteria were based on key study characteristics: population, intervention, comparator, outcome, and design [13]. Specifically, studies were included when (1) the study population included people in generalised, concentrated, or low-level HIV epidemics; (2) the intervention was community-based HTC offered in combination with a background of facility-based HTC; (3) the comparator was facility-based HTC; (4) the outcome(s) included uptake, proportion of people reporting receiving their first HIV test, CD4 value at diagnosis, rates of linkage to care, HIV positivity rate, HTC coverage, HIV incidence, or cost per person tested; and (5) the study design was a randomised trial or observational cohort study. Given the lack of comparative studies for community-based HTC, studies without a comparator arm were also included if they met the remaining eligibility criteria.

A. B. S., N. F., and O. A. independently screened the abstracts of all articles identified via the literature database searches and then compared the full texts of all articles selected during screening against the inclusion criteria. Disagreements were resolved by discussion. J. S. R. and A. K. S. repeated the same process for the clinical trial registries.

A. B. S., J. S. R., and A. K. S. completed the data extraction of characteristics of study participants, community-based testing approaches, outcomes, and quality assessment using a standardised extraction form.

The Newcastle-Ottawa Quality Assessment Scale was used to assess bias in studies with a comparator arm included in pooled analyses [16]. This scale rates studies based on eight criteria in three sources of bias. We modified this scale to remove one criterion, demonstration that the outcome of interest was not present at the start of study, since a previous HIV test may not affect all the outcomes analysed in this article. The Cochrane Collaboration's “risk of bias” tool was used to assess bias in randomised trials with a comparator arm [17].

Outcome proportions from studies meeting inclusion criteria were stabilised using the Freeman-Tukey-type arcsine square-root transformation and then pooled to summarise the proportion of participants who (1) accepted different community-based HTC approaches, (2) reported receiving their first HIV test, (3) had CD4 counts measured after diagnosis, (4) were diagnosed with HIV with a CD4 count above 350 cells/µl, and (5) initiated ART after their CD4 count indicated they were eligible for treatment [18],[19]. Pooled relative risks (RRs) were used to compare participants of community- and facility-based HTC with respect to uptake, proportion of first-time testers, the HIV positivity rate, proportion with CD4 counts above 350 cells/µl, and HTC coverage. Random-effects models were used for all analyses. Given the differences in HIV epidemiology, sexual mixing patterns, transmission factors, and healthcare utilisation rates for key populations, key population outcome data were reported individually and not pooled. I² statistics were used to measure heterogeneity [20]. I² statistics near 25% indicate low heterogeneity, values near 50% indicate moderate heterogeneity, and those above 75% indicate high heterogeneity [21]. All analyses were completed in STATA version 12.0.

108 articles, describing studies conducted from 1987 to 2012 and including 864,651 participants completing HTC, met the eligibility criteria (Table 1; Figure 1). Two articles were randomised trials [22],[23], and the rest were observational in design. Data from one multi-centre cluster-randomised trial were stratified into three studies (based on the country where the testing was offered) [23], data from three articles were stratified based on the year community-based HTC was offered [24]–[26], and data from four articles were stratified based on the community-based HTC approach used [27]–[30]. Given that 108 articles provided data from 108 studies and there were nine additional studies after stratification, there were a total of 117 studies included (Table S4). 76 studies were from Africa, 28 were from North America (excluding Central America), six were from Asia, four were from Central and South America, three were from Europe, and one was from Australia. The clinical trial registers identified ten ongoing trials: one on index testing [31], one on mobile testing [32], five on door-to-door testing [33]–[37], one on self-testing [38], and two on community-based testing for key populations [39],[40].

The percentage of participants who were male was 45.3% for index testing, 45.9% for door-to-door testing, 44.9% for mobile testing, 62.6% for self-testing, 67.0% for workplace testing, and 42.2% for school-based testing (Table 1). Excluding studies including only MSM or only FSW, 62.9% of testers were male in mobile testing for key populations (Table 1). Population-level HTC efforts found that implementation of community-based HTC increases the number of couples receiving testing (Table 2).

61 studies reported uptake of different community-based testing approaches among 713,632 participants: seven studies evaluated index testing among 12,052 participants [29],[30],[41]–[46], three evaluated self-testing among 1,839 participants [47]–[49], 14 evaluated mobile HTC among 79,475 participants [26],[28],[50]–[59], 28 evaluated door-to-door testing among 555,267 participants [24],[25],[28],[30],[60]–[81], six evaluated workplace HTC among 62,406 participants [22],[82]–[86], and three evaluated school-based HTC among 2,593 participants [87]–[89] (Figure 2). The percentage of participants accepting HTC was 88.2% for index testing (95% confidence interval [CI] 80.5%–95.9%; I² 99.7%, 95% CI 99.7%–99.8%; Figure 3), 87.1% for self-testing (95% CI 85.1%–89.0%; I² 28.8%, 95% CI 0%–92.6%; Figure 4), 86.8% for mobile HTC (95% CI 85.6%–88.1%; I² 99.9%, 95% CI 99.9%–99.9%; Figure 5), 80.0% for door-to-door HTC (95% CI 76.9%–83.1%; I² 99.9%, 95% CI 99.9%–99.9%; Figure 6), 67.4% for workplace HTC (95% CI 32.8%–100.0%; I² 100%, 100.0%–100.0%; Figure 7), and 62.1% for school-based HTC (95% CI 39.6%–84.5%; I² 99.0%, 95% CI 98.5%–99.4%; Figure 8). Uptake was higher in community-based HTC compared to providing vouchers to participants for facility-based HTC (RR 10.65, 95% CI 6.27–18.08; I² 96.1%; Figure 9) [22],[43].

19 studies reported uptake among 41,110 participants in key populations, including 16,725 MSM [90]–[97], 4,681 PWID [92],[98]–[100], 81 FSW [101], 13,240 adolescents [102],[103], and 6,383 individuals from combinations of key populations. The percentage accepting HTC was 99.7% among FSW, ranged from 13.7% to 94.5% among PWID, ranged from 9.4% to 95.0% among MSM, and ranged from 33.9% to 96.6% among adolescents (Figure 10). One study reported an uptake percentage of 95.2% among PWID and FSW [104], another reported an uptake percentage of 75.1% among PWID, FSW, and MSM [105], and another reported an uptake percentage of 60.0% among PWID and MSM [106]. Uptake was higher for community-based testing than for facility-based testing among FSW (RR 1.10, 95% CI 1.03–1.17) [101] and MSM (RR 1.53, 95% CI 1.42–1.65) [96] (Figure 11).

33 studies reported the HTC history among 597,016 participants in community-based HTC approaches [23],[25],[27]–[29],[44],[47],[51],[53]–[55],[58]–[61],[63],[64],[69],[70],[72],[74],[83],[107]–[112]. 62.2% (95% CI 58.0%–66.4%; I² 99.9%, 95% CI 99.9%–99.9%; Figure 12) of participants at community-based HTC sites reported receiving their first HIV test. In the nine studies with a facility-based comparator arm, a larger proportion of participants reported receiving their first HIV test at community-based HTC than at facility-based HTC (RR 1.23, 95% CI 1.06–1.42; I² 99.8%, 95% CI 99.8%–99.9%; Figure 9) [23],[27],[29],[107],[108],[111].

17 studies reported the HTC history of 25,311 participants from key populations receiving community-based HTC [27],[90],[92],[94],[96],[98],[105],[108],[113]–[118]. 9% to 79% of participants reported receiving their first HIV test (Figure 13). Five of these studies included a facility-based comparator arm (Figure 11). There were more first-time testers in community-based HTC than facility-based HTC for two study populations of MSM (RR 2.24, 95% CI 1.27–3.93 [113] and RR 1.37, 95% CI 1.18–1.59 [108]); however, there were fewer first-time testers in community-based HTC for a different study population of MSM (RR 0.33, 95% CI 0.26–0.43 [96]). There were more first-time testers in community-based HTC than facility-based HTC for a study population including PWID and MSM (RR 1.10, 95% CI 1.08–1.13 [116]); however, there was no difference in the proportion of first-time testers for a study population of FSW (RR 0.97, 95% CI 0.72–1.30 [108]).

14 studies included data on the HIV positivity rate among people testing in community-based approaches relative to people testing in facility-based approaches [22],[23],[27],[29],[43],[62],[87],[108],[119]–[121]. Overall, the HIV positivity rate was lower in community-based approaches relative to facility-based approaches (RR 0.59, 95% CI 0.37–0.96; I² 99.6%, 95% CI 99.6%–99.7%; Figure 9). The median number needed to screen to identify one person with HIV in community- and facility-based HTC was 17 (range 3–86) and 6 (range 4–154), respectively (Table 3). The number needed to screen with community-based testing was highest in settings with a low national HIV prevalence: 54 in Thailand and 86 in Guatemala [23],[108].

Six community-based testing studies for key populations included a facility-based comparator arm (Figure 11). Studies including FSW and a combination of PWID and MSM found no difference in the positivity rate for community- versus facility-based approaches (FSW, RR 0.62, 95% CI 0.29–1.34 [108] and RR 1.39, 95% CI 0.85–2.29 [101]; PWID and MSM, RR 1.13, 95% CI 0.91–1.39 [116]). There was a lower positivity rate among a study population of MSM (RR 0.09, 95% CI 0.03–0.33 [108]) and among a study population including PWID, FSW, and MSM (RR 0.51, 95% CI 0.28–0.94 [122]). There was also a higher positivity rate among a study population of MSM (RR 2.37, 95% CI 1.35–4.15 [123]). The number needed to screen to identify one person with HIV varied depending on the key population and study setting (Table 4).

18 studies reported the CD4 counts of 8,993 participants found to be HIV-positive using point-of-care or standard lab diagnostics [29],[30],[51],[55],[56],[60],[61],[74],[76],[82],[110],[121],[124]–[127]. 56.7% (95% CI 49.6%–63.9%; I² 97.6%, 95% CI 97.0%–98.1%; Figure 14) of participants testing positive had CD4 counts above 350 cells/µl. In the five studies with a facility-based HTC comparator arm, more participants in community-based HTC approaches had CD4 counts above 350 cells/µl than in facility-based approaches (RR 1.42, 95% CI 1.16–1.74; I² 94.8%, 95% CI 90.5%–97.1%; Figure 9) [29],[121],[125],[127].

Two studies reported the CD4 counts of participants found to be HIV-positive in a key population. Using standard lab diagnostics these studies reported a median CD4 count of 550 cells/µl among MSM [115] and 385 cells/µl among MSM, PWID, and FSW [105].

17 studies, including 5,852 participants with HIV, reported linkage to care from HIV diagnosis to CD4 measurement [30],[51],[55],[60],[61],[76],[82],[86],[89],[110],[124]–[126],[128],[129]. Overall, 80.1% of participants had their CD4 count measured after HIV diagnosis (95% CI 74.8%–85.4%; I² 99.5%, 95% CI 99.4%–99.5%; Figure 15). Nine studies with 527 participants reported linkage to care from being eligible to ART to initiating ART [30],[61],[76],[82],[89],[124],[129]. Overall, 73.1% of participants initiated ART after their CD4 count indicated that they were eligible (95% CI 61.3%–84.9%; I² 96.9%, 95% CI 95.6%–97.9%; Figure 15).

Two studies, including 52 participants with HIV, reported linkage to care from HIV diagnosis to CD4 measurement in key populations. 12 of 15 MSM had their CD4 count measured after HIV diagnosis [115]. 26 of 37 MSM and/or PWID had their CD4 count measured after HIV diagnosis [105]. No studies reported linkage to care from being eligible for ART to initiating ART in key populations.

14 studies summarised HTC coverage among all people living in the testing site's catchment area [23],[51],[56],[63],[70],[71],[74],[81]. Coverage of HTC ranged from 5% to 93% depending on the type of approach used (Table 5). Mobile HTC available as part of multi-disease health campaigns achieved high coverage in the shortest period of time. Community-based HTC increased coverage of HTC relative to facility-based approaches (RR 7.07, 95% CI 3.52–14.22; I² 99.7%, 95% CI 99.7%–99.8%; Figure 9).

One study reported HIV incidence [130]. There was a decreased risk of HIV infection in communities randomised to community-based testing relative to communities randomised to facility-based testing, although this estimate lacked statistical significance (RR 0.86, 95% CI 0.73–1.02; Figure 9).

The cost per person tested ranged from US$2.45 to US$881.63 using different community-based testing approaches (Table 6) [29],[45],[71],[74],[107],[117],[122],[131]–[136]. The cost per person tested was US$2.45 to US$14.37 for door-to-door testing, US$3.26 to US$33.54 for mobile testing, US$12.91 for hospital testing, US$15.30 to US$203.04 for index testing, US$21.28 to US$29.56 for testing at a fixed HTC site, US$126.48 for church-based testing, US$92.83 to US$881.63 for community-based testing for key populations, and US$93.73 for testing at an HIV clinic. Due to the heterogeneity in health systems and HIV prevalence within and between countries, the cost per person identified with HIV was not included.

No studies reported harm arising as a result of having been tested. 18 studies gave a description of the testers' experiences or listed reasons for tester refusal [47]–[49],[52],[58],[64],[71],[81],[82],[91],[102],[109],[113],[115],[137]–[139]. The studies discussed both the clients' positive testing experiences and their fears. Eight studies (including one targeting key populations) reported instances where participants refused HTC because of fear of status disclosure or stigma [52],[58],[64],[71],[81],[82],[102],[109]. In contrast, 12 studies (including three studies targeting key populations) specifically reported either no evidence of harm [47]–[49],[55],[91],[113],[115] or benefit through improved privacy or reduced stigma and fear [52],[82],[137]–[139].

There was concern of selection bias in nine of the studies included in pooled analyses [22],[23],[62],[111],[119],[125], concern of confounding in five studies [27],[62],[108],[111],[119], and concern of measurement bias in five studies [81],[107],[111],[121],[127] (Table S2). The randomised trials appeared to have limited selection, attrition, and reporting bias; however, their lack of blinding made them susceptible to performance and detection bias [22],[23] (Table S3).

While there was high uptake for community-based approaches in most studies, there were several outliers with low uptake. To gauge whether these outliers influenced pooled uptake estimates and increased heterogeneity we conducted sensitivity analyses without them (Table 7). Although the pooled estimates increased and the CIs tightened without the outliers, there was still high heterogeneity using the I² statistic. There was potential for selection bias, confounding, and measurement bias in several of the observational studies identified (Table S2). To determine whether these studies introduced bias into our results we ran sensitivity analyses without them and found the results to be similar (Table 8).