Dar es Salaam is Tanzania's biggest and most economically important city with the current population size exceeding 2.5 million inhabitants and a total area of 1,400 km², corresponding to a mean human population density of 2,900 per km² 40. It is situated between latitude 6.0°-7.5° S and longitude 39.0°-39.6° E. The city is divided into three municipalities: Kinondoni, Temeke and Ilala and each of these municipalities is further divided into wards. The study site comprised the 15 wards with 614,000 residents 40 included in the Dar es Salaam UMCP, 7193237 covering an area of 55 km² with wards ranging in size from 0.96 to 15 km². All UMCP activities are coordinated by the City Medical Office of Health, and are fully integrated into the decentralized administrative system in Dar es Salaam, operating on all six administrative levels of the city: the city council, the three municipal councils it oversees, 15 wards chosen from those municipalities, containing 67 neighbourhoods referred to as mitaa in Kiswahili (singular mtaa, meaning literally street), and more than 3,000 housing clusters known as ten-cell-units (TCU) with each of them subdivided into a set of plots corresponding largely to housing compounds. The main tasks on the three upper levels are programme management and supervision, whereas mosquito larval surveillance and control is organized at ward level and implemented at the level of TCUs and their constituent plots. In principle, a TCU clusters ten houses with an elected representative known as an mjumbe, but typically comprises between 20-100 houses in practice 41. UMCP implements regular surveillance of mosquito breeding habitats as a means to monitor effective coverage of aquatic habitats with microbial larvicides. Surveillance is applied through a community-based 35 but vertically managed delivery system 37. The cross-sectional surveys described here to evaluate routine surveillance activities were conducted between end of June 2007 and January 2008. This period spanned a full dry season and was preceded by a typical rainfall pattern with a main rainy season from March to June and a much shorter rainy season from October to December.

Community owned resource persons (CORPs) were recruited through local administrative leaders including Street Health Committees and were remunerated at a rate of 3,000 Tanzanian shillings (US$ 2.45) per day through a casual labour system formulated by the municipal councils of Dar es Salaam for a variety of small-scale maintenance tasks such as road cleaning and garbage collection 3235. All essential standard operating procedures adopted by the recruited larval surveillance CORPs are described in detail elsewhere 37, but summarized as follows.

Over 90 larval surveillance CORPs were actively employed by the UMCP during the time of survey with each assigned to a defined area of responsibility, comprising a specific subset of TCUs within one neighbourhood. These lists of TCUs were initially allocated based on local knowledge of habitat abundance, difficulty of terrain and geographic scale and subsequently refined through detailed participatory mapping of the study area 41. On average, one CORP was responsible for an area of approximately 0.6 km². All CORPs worked under the oversight of a single ward-level supervisor. Each CORP followed a predefined schedule of TCUs, which they were expected to survey on each day of the week. In wards where larviciding was taking place, the schedule of TCUs visited by the surveillance CORPs followed one day after the application of microbial larvicides by a separate set of larval control CORPs 37 so that indicators of operational shortcoming, such as the presence of late-stage (3^rd or 4^th instar) mosquito larvae, could be reacted to in sufficient time to prevent emergence of adult mosquitoes. This system was designed for routine mosquito habitat surveillance and larviciding to allow timely interpretation and reaction to entomologic monitoring data.

The investigator (PPC) initially conducted three weeks of unscheduled guided walks with 23 of the surveillance CORPs nominated by the ward supervisor after the investigator reported to their office in the morning. The investigator did not pre-inform the CORPs nor did he reveal his role and independent status at any time before or during the visit. Both the investigator and the chosen CORPs would leave the ward office and survey TCUs that the CORPs were expected to survey according to their normal predefined schedule for that particular day 37, returning later to report to the ward supervisor. At this stage, the survey was led by the CORPs and the investigator followed passively, covertly observing and recording how CORPs conducted their routine larval habitat surveillance and prepared their daily reports for submission to the ward supervisor. Specifically, the following information was collected: did CORPs follow TCUs schedule correctly, were all TCUs and plots visited, whether fenced compounds were entered and if not, why not, how habitats were recorded, how habitats were searched for larvae, how CORPs interacted with residents. In cases of observed shortcomings in the operational practices of the CORPs or any additional opportunities for improved implementation of their duties, the CORPs were informally advised by the investigator. This approach was intended to maintain an open, non-authoritative relationship of the investigator with the CORPs, allowing him to observe and understand the operational challenges facing the CORPs and the program as a whole. A detailed formal analysis of these qualitative observations will be published elsewhere but informal appraisal of these observations was used to design a quantitative survey described as follows.

A total of 173 TCUs from neighbourhoods distributed across all 15 wards were randomly selected from the list of TCUs in the UMCP study area. A total of 64 CORPs were responsible for these selected TCUs. The investigator accompanied the relevant CORP in guided walks through each TCU one day after their scheduled routine surveillance of that TCU and implemented his own larval habitat surveys following the standard operating procedures 37. Results of the investigator were compared with the CORP's datasheet of the previous day. Every potential habitat found by the CORP in each plot, and any additional habitats identified by the investigator that had not been detected by the surveillance CORPs, were distinguished and recorded using standardized forms (Additional file 1). Habitats were further classified into three habitat categories and constituent 11 habitat types 35 as follows: (1) natural habitats comprising (i) marshy or swampy areas, (ii) river-beds and (iii) springs or seepages; (2) agricultural artificial habitats comprising (i) rice paddies, (ii) ridge and furrow agriculture (matuta) and (iii) other habitats associated with agriculture; (3) non-agricultural artificial habitats comprising (i) drains and ditches, (ii) construction pits, foundations and other excavations (iii) water storage containers, (iv) tyre tracks and puddles and (v) ponds or pools. Additional information was collected regarding the presence or absence of a fence around a plot and whether or not a particular TCU was targeted with larvicide application at the time that it was surveyed. Lastly, records were taken regarding evidence of lack of familiarity of a CORP with the specific TCU and plots. Unfamiliarity was assumed if the CORP was not readily able to find his or her way around the TCU or plot, when plot boundaries could not be clearly defined and/or when residents of the plot were unable to recognise him/her as a regular visitor to the area.

All the data were entered in coded numeric form and analysed using SPSS 15.0. Any association between the occupancy of different mosquito habitat categories and types by Anopheles and Culicine larvae was analysed using multivariate binary logistic regression 42. Specifically, generalized estimating equations (GEE) were fitted to determine the influence of lack of familiarity of the CORP with the area, presence of a fence around the plot and whether larviciding was operational in that time and place upon the proportion of wet habitats (detection coverage) reported by CORPs and the proportion of habitats which contained larvae that were reported to be occupied by the CORP (detection sensitivity) for different habitat categories or types. While all observed habitats were included in the model fits to assess detection coverage, only those found to contain larvae by the investigator were considered in the denominator of models to assess detection sensitivity. The detection of the wet habitat or larval occupancy by the CORP was treated as the binary outcome variable and was fitted to a binomial distribution with a logit link function. CORP identity was treated as the subject variable and an exchangeable correlation matrix chosen for the repeated measurements distinguished by plot identity as the within subject variable. Differences between frequency distributions were assessed using likelihood ratio χ² analysis.

A total of 8,395 plots were visited during the cross-sectional surveys, 60.0% (5,039) of which were from larviciding areas (Figure 1). Approximately one quarter of these plots (26.8%; 2,253) was behind fences. There was an unequal distribution of fenced plots between the visited larviciding and non-larviciding areas with the majority of the fenced plots (69.7%; 1,571) recorded in areas where larviciding was taking place. Overall 3,997 potential mosquito breeding habitats were recorded. Ofthese, 2,965 (74.2%) contained water at the time of survey. The vast majority of these wet habitats were non-agricultural artificial habitats (90.0%), such as drains, ditches, construction sites, foundations, man-made holes and tire tracks. The remainder was composed of a small number of natural habitats (7.4%), such as swampy areas with high groundwater level, riverbeds, seepages and springs, and a few agricultural artificial habitats (2.6%) mainly associated with rice and sweet potato cultivation; crops grown in ridge and furrow systems known as matuta (Table 1).

Almost half (45.6%; 1,351/2,965) of all aquatic habitats were located within fenced plots. One fifth (20.5%; 608/2,965) of all aquatic habitats were recorded in plots with which CORPs clearly appeared to be unfamiliar and 91.9% (539/608) of these were located behind fences. A large number of wet habitats were surveyed in both larviciding areas (1,895) and in non-larviciding areas (1,070) and the proportion of habitats within fenced plots was higher in areas with larviciding than those without (50.8% (962) versus 36.4% (389), respectively; χ² = 57.3, df = 1, P < 0.001).

Only 7.8% of all the surveyed habitats contained any aquatic stages of Anopheles larvae (Table 1) so there were relatively few habitats in which the sensitivity with which CORPs detected these key indicators of malaria vector proliferation could be assessed. Unexpectedly, three quarters (74.8%, 172/230) of anopheline-occupied habitats were found in larviciding areas and anopheline larval occupancy was twice as high in wards where larviciding took place as those without (9.1% (172/1,895) versus 5.5% (59/1,070); Odds Ratio [95% Confidence Interval] = 2.11 [1.20-3.67], P = 0.009). Overall, 7.0% (207/2,965) of wet aquatic habitats contained early-stage (1^st and 2^nd instars) Anopheles larvae, whereas 5.2% (155/2,965) of aquatic habitats were inhabited by late-stage Anopheles larvae (3^rd and 4^th instars), with 71.6% (111/155) of the latter recorded in areas with larviciding.

The probability of a habitat containing Anopheline larvae depended on category and habitat type (Table 1). Agricultural sites were twice as likely to contain Anopheline larvae than natural habitats, whilst the chance of finding larvae in artificial non-agricultural habitats was much lower. Nevertheless, non-agricultural artificial habitats were the most abundant (90%) and, therefore, constituted 58% (135/231) of all Anopheles-occupied habitats (Table 1).

Over one quarter of wet habitats contained culicine larvae (Table 1), with 25.9% (767) and 22.1% (656) inhabited by early-stage and late-stages respectively. Natural and agricultural habitats were equally likely to harbour culicine larvae whilst the probability of their presence was significantly higher in artificial, non-agricultural habitats (Table 1).

CORPs recorded 1963 wet habitats during their routine surveillance. Seven of these habitats were confirmed to be non-existent by the investigator, suggesting these CORPs had filled the surveillance forms without visiting the relevant plots so these were excluded from the analyses. Therefore, CORPs correctly recorded two thirds of wet habitats (Table 2). Detection coverage varied significantly between individual CORPs and between different habitat types (P < 0.001 for both as determined by logistic regression). CORPs were unfamiliar with 20.5% (608) of wet habitats and 92% (539) of these were located behind fences. Furthermore, the majority of wet habitats that the CORPs failed to record (61.1%; 619/1009) were located within fenced plots.

Detection coverage differed significantly for different habitat types (χ² = 432.8, df = 10, p < 0.001) and categories (Table 3) with artificial non-agricultural habitats 1.6 times more likely to be recorded than others (Table 3). Consistent with the baseline evaluation 35 conducted before the introduction of current procedures for mapping 41, surveillance and larvicide application 37, most conspicuous habitat types like ponds, rivers, seepages, springs and drains were more readily recorded, whereas water receptacles were poorly detected (Table 2). Furthermore, the type of habitats that CORPs did not find was significantly different between the fenced and unfenced plots: the majority of water storage containers, tyre tracks and artificial pits were located behind fences (Figure 2).

The probability of a CORP detecting and recording a wet habitat was similar in larviciding and non-larviciding areas but was 84% less likely if he or she was unfamiliar with the area (Table 3). As mentioned earlier, the vast majority of the sites with which the CORPs were unfamiliar were within fenced plots. The covariance between these two variables (Pearson correlation, r² = 0.40, P < 0.001) implies that the presence of fences around plots contributed to the lack of familiarity with plots among CORPs. Although excluded from the selected model presented in Table 3 because of this covariance, fenced plots, selected when familiarity was excluded, reduced the detection coverage by half (OR [95%CI] = 0.49 [0.37-0.65], P < 0.001).

Overall detection sensitivity of mosquito larvae was very low among CORPs. They found only 29 of 230 anopheline-positive habitats and 263 out of 830 culicine-positive habitats, corresponding to under-reporting rates of 87.4% and 68.4%, respectively. CORPs reported a higher proportion of larva-containing habitats in non-larviciding areas (anophelines: 27.6% (16/58) and culicines: 44.4% (138/311)) than larviciding areas (anophelines: 7.6% (13/172) and culicines: 24.1% (125/519)). Detection sensitivity was twice as high for early instars 13.5% (28/207) than late instars 6.5% (10/155) of anopheline larvae (χ² = 4.72, df = 1, P = 0.029). Detection sensitivity for early and late-stage culicine larvae did not differ (32.2%, (247/767) and 30.0%, (196/653) respectively (χ² = 0.787, df = 1, P = 0.375). Not only did most habitats (71.6%; 111/155) that contained late-stage anophelines during the investigator's survey occur in the larviciding areas, CORPs had reported this indicator of mosquito proliferation in only 3 of these cases (2.7%). Detection sensitivity of late stage Anopheles in non-larviciding areas was also very low (15.9%; 7/44) and did not differ significantly (P = 0.124) from larviciding areas.

Failures to detect mosquito larvae can be attributed to two distinct causes: (1) the aquatic habitat was not found and therefore no larval search took place or (2) the larvae were not detected during the inspection of that habitat. More than half of the anopheline (57.8%; 133/230) and culicine-positive (56.0%, 465/830) habitats were not recorded as wet by CORPs. In 60.9% and 95.7% of these non-reported anopheline and culicine-occupied habitats, respectively, the CORPs was either unfamiliar (anophelines; 45.1%, (60/133), culicines; 52.5%, (244/465)), the plot was fenced (anophelines; 45.9%, (61/133), culicines; 64.3%, (299/465)) or both (anophelines; 30.1%, (40/133), culicines; 21.1% (98/465)).

Anopheline larvae were identified by CORPs in only 29 of the 97 occupied habitats which they recorded as wet so overall detection sensitivity was 29.9%. More importantly they appeared unfamiliar with very few of both the anopheline-positive habitats which they reported as wet (5.2%, 5/97) and those which they did not (7.4%, 5/68). It therefore appears likely that not reporting larvae is due to insufficient dipping, examination or training in mosquito identification rather than not visiting the site. Notably, the detection sensitivity for culicine larvae in habitats that were reported as wet was much higher with almost three quarters of habitats containing these more obvious larvae being successfully identified (Table 4).

Larval detection sensitivity was different for different habitat types for anophelines (χ² = 28.9, df = 10, P = 0.001) and culicines (χ² = 21.6, df = 10, P = 0.016). CORPs more readily detected anopheline larvae in larger, more obvious habitats like drains, riverbeds, ponds and matuta (Table 4). To enable fitting of a logistic model, these types had to be pooled into categories which had no significant effect. However, the probability of CORPs reporting larval anophelines occupying a habitat was drastically reduced if the habitat was located in an area where larviciding was ongoing (Table 5).

Late-stage Anopheles occupancy was reduced by over 70% in habitats in the intervention areas where the surveillance CORPs actually found and reported the wet habitats (Table 6). Note that no significant reduction of late-stage Anopheles occupancy was revealed for habitats in areas not covered by the intervention, regardless of whether the surveillance CORPs found them or not (Table 6).