Exposure to Schistosoma mansoni infection in a ... - Wiley Online Library

PDFlib PLOP: PDF Linearization, Optimization, Protection Page inserted by evaluation version www.pdflib.com – [email protected]

Tropical Medicine and International Health volume 6 no 2 pp 136±145 february 2001

Exposure to Schistosoma mansoni infection in a rural area in Brazil. II: Household risk factors Jeffrey Bethony1,2, Jeff T. Williams2, Helmut Kloos3, John Blangero2, Lucia Alves-Fraga4, Germaine Buck5, Arthur Michalek5, Sarah Williams-Blangero2, Philip T. LoVerde5, Rodrigo CorreÂa-Oliveira1 and Andrea Gazzinelli6 1 2 3 4 5 6

Centro de Pesquisas ReneÂ Rachou, FIOCRUZ, Belo Horizonte, Minas Gerais, Brazil Southwest Foundation for Biomedical Research, San Antonio, USA University of California at San Francisco, USA Universidade de Vale de Rio Doce, Governador Valadares, Minas Gerais, Brazil School of Medical and Biomedical Sciences, State University of New York at Buffalo, USA Escola de Enfermagem, Universidade Federal Minas Gerais, Minas Gerais, Brazil

Summary

A number of studies have pointed out the potential importance of the household in the transmission of schistosomiasis. The clustering of domestic activities associated with water collection, storage, and usage can result in the sharing of transmission sites and infective water contact behaviours. In this study, we employed a variance component method to estimate effects due to individual risk factors and shared residence on the variance in faecal egg counts during Schistosoma mansoni infection. A suite of covariates, which included demographic, socioeconomic, water supply, and water contact behaviour terms, contributed 15% to the variance in faecal egg counts. Shared residence alone accounted for 28% of the variance in faecal egg excretion. When both the suite of covariates and shared residence were considered in the same model, shared residence still contributed 22% to the variance in infection intensity. These results point to the importance of shared residence as a means of capturing the complex interrelationship between shared demographic, socioeconomic, physical environmental, and behavioural factors that in¯uence transmission of schistosomiasis at the household level. keywords Schistosoma mansoni, epidemiology, variance component methods, household, Brazil correspondence Dr Jeffrey Bethony, Centro de Pesquisas ReneÂ Rachou, Av. Augusto de Lima 1715, CEP 30190-002 Belo Horizonte, MG, Brazil. Email: jeff@cpqrr.®ocruz.br

Introduction Several important risk factors have been identi®ed that in¯uence intensity of infection with Schistosoma mansoni in rural communities in Brazil. These include demographic, socioeconomic and spatial factors, as well as environmental, sanitation, and water supply conditions (Gazzinelli et al. 1998; Kloos et al. 1998; Lima e Costa et al. 1998). Many studies have attempted to measure individual exposure to S. mansoni infection either by direct observation of behaviour at water sources (Fulford et al. 1996b) or by interview or questionnaire to characterize water contact behaviour (Sama & Ratard 1994; Lima e Costa et al. 1998). Watts et al. (1998) and Cairncross et al. (1996) have pointed out the potential importance of the household in

136

the transmission of schistosomiasis. An important aspect of the household is the clustering of domestic activities associated with water collection, storage and usage. Such activities can result in the sharing of infective sites and infective water contact behaviours, which in turn expose household members to similar risks of infection (Watts et al. 1998). Evidence for household clustering of infected individuals exists for other helminthic infections: e.g. Ascaris lumbricoides (Forrester et al. 1988), Trichuris trichiura (Forrester et al. 1988), and Strongyloides stercoralis (Conway et al. 1995; Lindo et al. 1995). Two methodological problems have constrained previous research on the household aggregation of helminthic infections. The ®rst is the de®nition of the `household', which is variously de®ned as a physical structure, a kinship locus, a social organization, an economic unit, or usually

ã 2001 Blackwell Science Ltd

Tropical Medicine and International Health

volume 6 no 2 pp 136±145 february 2001

J. Bethony et al. Exposure to S. mansoni. II: Household risk factors

some combination of these characteristics (Kroeger 1985; Berman et al. 1994). The simplicity of these de®nitions conceals the complexity of what a household represents, especially in rural areas where extended families share residences, dwell in compounds, or employ temporary `live-in' labourers (Berman et al. 1994; Kroeger 1985). Another limitation of previous studies has been the use of a dichotomous dependent variable (ìnfected' vs. `non-infected'), to reduce the analysis to a test of independence in 1 2 ´ K contingency tables (Forrester et al. 1988; Conway et al. 1995). The observed distribution of cases between households is then compared with that expected 2 from a binomial distribution using a v2 goodness-of-®t procedure. The use of a dichotomous dependent variable has several drawbacks for a household study of S. mansoni. The most widely used diagnostic technique, the Kato±Katz faecal thick smear technique (Katz et al. 1972), would require an impractically large number of stool examinations from participants in order to achieve the sensitivity suf®cient to accurately separate egg-positive from egg-negative individuals (Engels et al. 1996, 1997). The faecal thick smear technique is much better for detecting differences in the magnitude of faecal egg excretion, i.e., in differentiating lightly, moderately, and heavily infected individuals (de Vlas 1996). Further, the identi®cation of household clustering based on infection status seems of secondary importance to identifying factors that determine the intensity of faecal egg excretion. Heavy faecal egg excretion may play an important role in the development of severe clinical forms of the disease (hepatosplenism) (Butterworth et al. 1996) and in the transmission of the parasite in the community (Etard & Borel 1992; Woolhouse 2000). We did not attempt to address these limitations by employing a more comprehensive set of sampling rules for the household or by using a more sensitive diagnostic method to distinguish ìnfected' from `noninfected'. Instead, we used a simple inclusion rule to determine household membership (i.e. individuals who sleep and eat in the same dwelling and share an income), and examined data on speci®c individual demographic, socioeconomic, physical environmental, and water contact behaviours. We applied a variance com3 ponent methodology (Searle et al. 1992) to estimate on the proportion of variation in faecal egg excretion due to shared residence and individual risk factors. Variance component methods have a long and successful history in quantitative genetic research (Searle et al. 1992), but their utility for disentangling the environmental and behavioural complexities of infectious diseases remains largely unevaluated. ã 2001 Blackwell Science Ltd

Methods Study sample The rural area of MelquõÂades is located in the state of Minas Gerais, Brazil and has been described in detail in 4 Gazzinelli et al (2001). The house was the main sampling unit, and the survey was conducted in three phases. In phase I, the research team visited dwelling structures within the 100-km2 area of the study site to obtain informed consent using a verbal version of the standard consent form approved by the National Ethics Committee of Brazil and the Internal Review Board of the University at Buffalo, State University of New York. At this time, each house was assigned a unique household identi®cation number and the residents within each household were given a personal identity number (PID). There were three exclusion criteria: (i) having left the study area for a continuous period of ³ 60 days within the previous 2 years; (ii) residence in the endemic area only on weekends or once a month; (iii) prior treatment for S. mansoni infection during the previous two years (n 46). Those who claimed to have received treatment were examined and treated for current helminthic infections, but were not included in the analysis of the data. In phase II, three plastic containers for faecal samples were distributed to each participant. The methods for sample retrieval, storage, and slide reading are detailed in Gazzinelli et al. (2001). In phase III of the survey, a questionnaire was administered to either the head male or female of the household. The questionnaire elicited data on household membership, demographics, socioeconomics, sanitation, house construction, and occupation. A member of a household was de®ned as an individual who slept, took meals, and shared (either by contributing or using) in the income of the household over the last 2 years. Genealogical and demographic information including age, sex, parental names, and residences for all individuals were also collected. The genealogical data allowed construction of multi-household extended pedigrees using PEDSYS, a pedigree-based data management system (Dyke 1989). Individuals were de®ned as belonging to the same pedigree if they were related biologically to anyone else in the pedigree or if they were married to anyone else in the pedigree. Age was taken in 1999 and con®rmed by birth certi®cate. A person was classi®ed as having been born in the endemic area if both parents resided in the area at the time of the individual's birth. Income was determined for the household as a whole on a monthly basis and then divided by the number of individuals in each household. Due to the dif®culty in de®ning social and economic status in 137




developing countries (Kroeger 1985), we supplemented questions on income with questions on the ownership of 17 valuable items (e.g. car, microwave, refrigerator). These items were weighted proportionate to their value in Brazilian `reais', with the weighted values summed for each household and termed `household worth index'. For questions concerning possession of water supply or sanitation facilities, the interviewer assessed the adequacy of each response by visual con®rmation. Each participant was interviewed individually as to his or her water contact behaviour. Questions concerning the frequency of water contact were derived from a water contact observation study in the same area (Gazzinelli et al. 2001). Questions consisted of the frequency per week for each water contact behaviour and the source of the water (e.g. well, bamboo conduit). Parents were asked about their children's water contact behaviour only when the children were younger than 6 years; no proxy responses for water contact behaviour for participants older than age 6 years were accepted. All water contacts were then strati®ed according to whether or not they took place in potentially infective water contact sites. The sites were then divided into possibly infected (streams, canals, ®shponds, `bicas') or potentially safe (springs, wells, water hoses, and household taps, wash basins and showers sites). The determination of infected or safe was based on the presence (infective) or absence (safe) of snail intermediate hosts by malacological surveys undertaken at each contact site. Water contact frequencies that occurred at potentially infective sites were then multiplied by standardized duration and body immersion values from Gazzinelli et al. (2001). The residence of the survey team in the study area made it possible to check the quality of data collection and correct data at once by revisiting the household concerned. Statistical methods The distribution of eggs per gram of faeces (epg) and its natural log transformed version (lnepg) were tested for normality using a suite of empirical distribution function (edf) statistics (Stephens 1974). Because the individuals in our sample were related, their epg values do not comprise a set of independent observations, which violates one of the main assumptions of the tests. Partly as a consequence of this non-independence, naive application of EDF statistics, even to properly transformed data, will almost certainly lead to rejection of the null hypothesis of normality. We therefore applied tests for normality to epg and lnepg values for the founding members in our sample, since these comprised a set of assumed independent individuals. A founder was de®ned as an individual whose parents are 138

not in the pedigree. Founders are assumed to be non-inbred and unrelated (Lange 1997). Student's t-test was used to test for signi®cant differences in lnepg between genders. Analysis of variance (A N O V A ) was used to test the mean differences of egg counts by 10-year age intervals for both sexes. Bonferroni post hoc tests, with a signi®cance level of 0.05, were used to determine which 10-year age intervals differed signi®cantly from others. A Pearson product moment correlation was used for all correlations. Variance component analysis was used to test the effects of covariates and shared residence on faecal egg counts during S. mansoni infection. Detailed development of variance component methods in various contexts is 5 available elsewhere (Lange et al. 1976; Hopper & Mathews 1982; Searle et al. 1992). Brie¯y, let y (y1,¼,yn) denote the vector of quantitative trait values for a collection of n individuals. Under a model of 6 multivariate normality (Lange et al. 1976) for the trait vector y the ln-likelihood l of the sample is given by n 1 1 l ÿ ln2p ÿ lnjRj ÿ y ÿ l0 Rÿ1 y ÿ l 2 2 2 where l and R are, respectively, the mean and variancecovariance matrix for y. Covariate effects are introduced by modelling the trait mean vector as l y + Ab where y is the overall trait mean; A is an n ´ m design matrix for the trait whose n rows contain m covariates such as age, sex, water contact, and any other demographic, socioeconomic, physical environmental, or behavioural indices for each individual; and b is an m-vector of regression coef®cients for the selected covariates. The covariance matrix R depends on the speci®c effects and interactions to be modelled. In general R takes the form R RjXjr2j where the parameter rj2 is a scalar variance component and Xj is the corresponding structuring matrix representing the expected pattern of covariation. Minimally, R must contain a term re¯ecting random, individual-speci®c environmental variation. This `sporadic' model can be written R Inre2 where In is the n ´ n matrix and re2 is the variance due to random, individual-speci®c environmental effects. Additional environmental factors are expected to be important, however. For example, the contribution of shared household to variation in the dependent variable can be modelled by introducing the term H, where H is a matrix whose ijth element hij 1 if individuals i, j live in the same house, and hij 0 otherwise. The associated parameter rh2 represents the variance in the dependent variable that is attributable to the effect speci®ed by H. The total trait covariance for this `household' model is then R Hrh2 + Ire2.





The statistical signi®cance of any modelled effect is tested by comparing the likelihood L0 of the null model in which the effect is absent with the likelihood L1 of a model in which the effect is estimated. Under standard conditions the likelihood ratio statistic k )2ln(L0 ) L1) is asymptotically distributed as a chi-square variable with degrees of freedom equal to the difference in number of free parameters between the two models (Kendall & 7 Stuart 1967). With variance component models, however, the likelihood-ratio test is often made under non-standard conditions ± some parameters may fail to exist under the null hypothesis, or some parameters may have their true value under the null hypothesis on a boundary of the parameter space de®ned by the alternative hypothesis. In these cases the distribution of the likelihood ratio statistic under the null hypothesis is a complex mixture of chisquared distributions, with mixing proportions determined by the geometry of the parameter space (Self & Liang 1987). For our analyses we used the `sporadic' and `household' covariance models described above, with covariates as speci®ed in the text. Maximum-likelihood methods were used to jointly estimate the regression coef®cients bi and the variance components r2h. All analyses were performed using the variance component analysis package Sequential Oligogenic Linkage Analysis Routines (SOLAR) (Blangero & Almasy 1997; Almasy & Blangero 1998).

Table 1 Percentage distribution of gender by 10-year age intervals Age (years)

Females n

Males (%)

n

%

Total

1±9 10±19 20±29 30±39 40±49 50±59 60 +

63 79 38 44 32 24 30

(53) (49) (55) (65) (47) (52) (45)

56 81 31 24 36 22 37

(47) (51) (45) (35) (53) (48) (55)

119 160 69 68 68 46 67

Total

310

(52)

287

(48)

597

Table 2 Distribution of individuals by household Size of household 1 2 3 4 5 6 7 8 9 10 12

Number of households

Number of individuals

15 20 26 28 26 14 4 4 4 3 1

15 40 78 112 130 84 28 32 36 30 12

145

597

Results

Total

A total of 197 dwellings were counted in MelquõÂades, but only 145 of these structures were permanent residences; 676 individuals were assigned PIDs and 26 (4%) of these individuals failed to hand in complete stool samples, while 79 (12%) did not complete the household survey. These omissions reduced the sample to 597 individuals or 88% of the original sample size. The sample comprised more females (n 310) than males (n 287). The relationship between 10-year age intervals and gender is presented in Table 1, where the greatest disparity in gender can be seen in the 30±39-year age interval, in which females outnumber males. Table 2 presents the numbers of individuals sampled per household. The households ranged in size from one to 12 inhabitants. The mean ( SD) for the number of individuals per household was 4.12 (2.19). Of the households 84 (58%) were represented by four or more inhabitants. Figure 1 shows that faecal egg excretion was overdispersed in the sample, with marked leptokurtosis (63.95) and skewness (6.96). Results of EDF tests for normality are summarized in Table 3. By any measure the assumption of normality is grossly violated for epg, but lnepg + 1 is

effective at normalizing the distribution and was used in all subsequent analyses. The prevalence of S. mansoni infection in the study sample was 59.1% (95% CI 55.1, 63.1%) and the mean epg was 128 (95% CI 97, 161). Overall, males (174; 95% CI 97, 161) had a signi®cantly higher (P 0.008) intensity of infection than females (88; 95% CI 58, 117). Figure 2 presents the distribution of lnepg for both genders by 10-year age intervals. The highest mean intensity of infection occurred in the 1±19-year age intervals for both males and females. Table 4 presents the mean, standard deviation, and range for continuous covariates. The mean ( SD) age for the sample was 29.95 (21.53) years and included individuals from 1 to 96 years of age. The mean ( SD) of washing limbs was 24.08 (27.07) and that for crossing streams was 17.09 (16.00), making these the most common water contact activities. These two behaviours, along with washing utensils and ®shing, also showed the greatest range in frequency per week. The mean ( SD) total water


139




Figure 1 Distribution of S. mansoni eggs in the study sample expressed in eggs per gram of faeces (n = 597). Table 3 Results of EDF tests for normality of epg and lnepg in founders (n = 32) epg

ln-transformed epg

EDF statistic

t

P

Kolmogorov D Kuiper V CrameÂr-von Mises W2 Watson U2

2.17 4.37 1.69 1.61

< < <
> > >

0.15 0.15 0.15 0.15

contacts per week was 118.34 (56.37) and ranged from 0 to 504 contacts per week. Table 5 shows the distribution of dichotomous covariates in the sample. The largest occupational group was `domestics', followed by `secondary school student' and then àgricultural worker'. Only 19% claimed to have contact with potentially infective water sources while performing agricultural activities, whether as a part of their regular work or as an extra activity (e.g. gardening). However, 68% claimed to contact potentially infective water sources at least once a week. More than half of the individuals lived in houses with a tap, basin, or latrine. Table 6 presents the results of multiple regression on covariates signi®cant at the 0.10 level. This suite of covariates contributed 15% to the variation in lnepg and comprised a near equal distribution of demographic, socioeconomic, physical environmental, and behavioural factors, with demographic and socioeconomic factors being the most signi®cant. Gender, birth in the study area, 140

monthly income, and level of education had P-values < 0.01. Less statistically signi®cant were the four types of water supply (possession of a shower, tap, basin, latrine) and the three water contact behaviour terms (frequency of ®shing, hand irrigation of plants, and bathing per week). It is notable that terms involving age (age, age2, and the interaction term age ´ sex) were not included in the ®nal model. Table 7 shows that strong and signi®cant bivariate correlations were found between water supply covariates from Table 6 and between these covariates and monthly income. Formal likelihood-ratio tests for the effect of shared household on faecal egg excretion are outlined in Table 8. The `sporadic' model of variation incorporates only the contribution of random individual-speci®c environmental effects (r2ej). The `household' model retains this effect, and also introduces an effect due to shared household (r2hj). Twice the difference between the ln-likelihood for the sporadic model and that for the household model yields a test statistic that is distributed as a : mixture of a v2 variable with 1 df and a unit point mass at the origin (Self & Liang 1987). Tests are shown with and without the covariates modelled in Table 6. While covariate effects account for 15% of the variance in the sporadic model, covariate effects account for only 12% in the household model. Irrespective of covariate inclusion, however, the test for a shared household effect is highly signi®cant, with 22 to 28% of the observed variation in faecal egg excretion accounted for by shared residence alone. ã 2001 Blackwell Science Ltd




Figure 2 The relationship between ln(epg + 1) and 10 year age intervals by gender. The bars indicate 95% CI for the mean of each 10 year age interval.

Discussion The most important result from this study was the role of shared residence on variation in faecal egg counts (lnepg). Shared residence accounted for between 22 and 28% of the total variance in lnepg. When both shared residence and the covariates from Table 6 were included in the same analysis, shared residence still accounted for 22% of the variance in faecal egg excretion. The proportion of variance of faecal egg excretion accounted for by shared residence may be the result of two factors: (i) the de®nition of shared residence overlaps with kinship, which is a factor not taken into account in the current study, and (ii) the generality of the term `shared residence' incorporates a large number of unknown elements. In either case, these results point to the importance of shared residence as a means of capturing the complex interrelationship of shared genetic, demographic, socioeconomic, physical environmental, and behavioural factors that in¯uence transmission at the household level. These results also support the ideas of Cairncross et al. (1996) and Watts et al. (1998) on the importance of the household in the transmission of S. mansoni infection. Both observe that, even when the de®nition of the household is problematic, it remains a focal point for bringing individuals into contact with disease agents. In schistosomiasis, it is the clustering of domestic activities associated with water collection, storage, and usage (Watts et al. 1998). Such ã 2001 Blackwell Science Ltd

activities can result in the sharing of infective sites and infective water contact behaviours, which in turn expose household members to similar risks of schistosomiasis infection. Similar results were reported by Etard & Borel (1992) for Schistosoma haematobium in a Mauritanean village, where domestic contacts, mostly by females, represented 62% of all observations. Kloos et al. (1998) noted the importance of the household in determining water contact activities for Schistosoma mansoni in Brazil due to the greater privatization of water resources. A strength of the study was the variance component method (Searle et al. 1992), which enabled us to measure the contribution of both ®xed covariate and random household effects on the variance of faecal egg counts. The use of a variance component method also allowed us to model the household using a simple àggregate' measure of shared residence in a single dwelling. More speci®c kinds of shared physical environment might also be investigated using the variance component method; for example see Astemborski et al. (1985), Hopper (1983), Clifford et al. (1984), and Lange (1986). Another strength of the study was the survey instrument. It was derived from extensive preliminary observations on water usage and accessibility in the study area (Gazzinelli et al. 2001). This preliminary work gave us insights into water contact behaviours and resources that would not be yielded by conventional surveys, which are often structured around questions imposed by outside researchers and, hence, dislocated from 141




Table 4 The distribution of continuous demographic, socioeconomic, household, and water contact behaviour variables (n = 597) Covariate Demographic Age Years in study area

Mean

SD*

Range

29.95 30.83

21.53 81.87

1±96 1±96

Socioeconomic Income per month ($R) Household worth index Number of possessions Level of education (years)

286.97 19.62 4.26 7.02

326.49 17.10 2.25 5.99

50±2500 0±98 0±12 1±23

Household Number of rooms in house Age of household members Individuals per household

6.5 31.61 4.16

2.3 14.45 4.82

1±14 1±96 1±12

Water contacts frequency (weekly) Bathing 9.15 Washing limbs 24.08 Crossing streams 17.09 Washing utensils 8.55 Hand irrigation of plants 3.18 Washing clothes 1.40 Fishing 3.68 Fetching water 5.63 Maintaining canals 0.51 Playing in streams 1.51 12 Total water contacts 118.34

3.98 27.07 16.00 12.70 4.86 2.60 9.98 13.13 2.44 4.92 56.37

0±28 0±150 0±140 0±140 0±28 0±21 0±120 0±70 0±48 0±63 0±504

*SD, standard deviation; Total income of household divided by number of household members.

Table 5 The distribution of dichotomous variables (n = 597) Factor Demographic Born in study area Occupation* Agriculture Domestic Non-agricultural Retired or handicapped School student Pre-schooler Water contact Performing agricultural activities Infective water source Water supplyà Shower Faucet Basin Bamboo conduit (Bica) Latrine Cistern

n

%

438

73

146 172 27 24 166 62

24 29 5 4 29 10

116 408

19 68

235 308 322 339 409 200

39 52 56 57 69 33

*Occupational categories are mutually exclusive. Domestic includes work as housewife, domestic servant, or washerwoman; non-agricultural includes work as herder, driver, bricklayer, teacher, carpenter, or vendor; The determination of infected or safe was based on the presence (infective) or absence (safe) of the snail intermediate host by malacological survey; àWater supply sources are not mutually exclusive. Table 6 Best-®tting regression model of demographic, socioeconomic, physical environmental, and behavioural covariates on lnepg with signi®cance at 0.10 level (n = 597)

the social and cultural context in which these activities 13 Factor b SE P-value occur. The suite of covariates shown in Table 6 contributed Gender )0.6357 0.191 0.0018 Born in study area 0.6239 0.213 0.0002 15% to the variation in faecal egg excretion, and was Monthly income )0.0001 3.0 ´ 10)4 0.0002 evenly distributed among demographic, physical environLevel of education 0.9328 0.016 0.0002 mental, socioeconomic, and water contact factors. The Shower )0.7016 0.296 0.0499 importance of gender as a covariate is re¯ected by its level Faucet 0.8191 0.368 0.0934 of signi®cance (P 0.017) and con®rms our previous Basin )0.6123 0.321 0.0644 observations from the study area (Gazzinelli et al. 2001), Latrine 0.5635 0.230 0.0451 8 as well as those of Watts et al. (1998), who noted that Fishing 0.0217 9.3 ´ 10)4 0.0944 Irrigation of plants )0.0150 0.020 0.0640 water contact space and activities are often gendered and Bathing 0.0159 0.023 0.0886 that this plays an important role in infection intensity. The presence of water supply covariates in the ®nal model b indicates slope of regression; SE, standard error of the con®rms the importance of shared water resources at the b coef®cient. household level. However, the relationship between water supply covariates and faecal egg counts may be confounded by a third factor: monthly income. This is possible for two each other and with monthly income. It may be that the reasons: (i) the signi®cance of monthly income in the possession of a shower, basin, tap or latrine is a better multivariate model (less than 0.01) compared to the three predictor of socioeconomic status than the other indicators water supply covariates (all greater than 0.05) and (ii) the used in the study: e.g. household worth index, number of four water supply covariates were highly correlated with rooms per house, or the ownership of land. Finally, the 142





Table 7 Bivariate correlations between covariates shown in Table 6 (n = 597) Factor

Gender

Birth

Income

Education

Shower

Faucet

Basin

Latrine

Fishing

Irrigate

14 Gender Birth Income Education Shower Faucet Basin Latrine Fishing Irrigating Bathing

± 0.042 0.027 0.072 )0.028 0.000 0.024 0.026 )0.176à 0.255à 0.023

± )0.095 0.101 0.020 0.061 0.072 0.008 0.091à )0.171à 0.136à

± 0.161à 0.377à 0.316à 0.291à 0.232à 0.010 0.041 0.099

± 0.099 0.137à 0.133à 0.052 0.166à 0.006 0.066

± 0.732à 0.616à 0.509à 0.076 0.072 0.169à

± 0.821à 0.447à 0.027 0.110à 0.053

± 0.439à 0.032 0.076 0.072

± )0.028 0.055 0.014

± )0.097 0.066

± )0.076

Correlation is signi®cant at 0.05 (two-tailed); àCorrelation is signi®cant at 0.01 (two-tailed). Table 8 Maximum likelihood estimates of 15 Model shared household on the proportion of variance in lnepg in adjusted and 16;17 Sporadic unadjusted models (n = 597) Household Adjusted* Sporadic Household

Household

Random

LnL

v2

P-value

± 0.28 0.04

1.0 0.72 0.04

)879.472 )842.155

± 74.63

±