Lymphatic filariasis is endemic to parts of Haiti, including the town of Leogane and its surrounding communities. The descriptions of the treatment and prevention programs in the valuation exercises are modeled after projects conducted in the Leogane area. The treatment intervention includes the management of lymphedema and prevention of adenolymphangitis (ADLs) through improved hygiene, skin care, and physical measures, such as elevation and movement. Two prevention strategies can be used to interrupt transmission of LF, either regular use of table salt fortified with diethylcarbamazine (DEC) [8] or mass administration with a 2-drug combination: DEC plus albendazole or ivermectin plus albendazole [9]. Both kill the microfilaria produced by adult parasites, preventing further transmission. For this study DEC-fortified salt was the intervention used in the valuation exercise.
Survey Instrument
Individual WTP was elicited using an in-person survey conducted between November and December 1997, prior to the implementation of a DEC-fortified salt distribution program [8]. The survey was translated into Haitian Creole and was developed and pre-tested with the help of bilingual (Haitian Creole and English) residents of Leogane. The survey's four sections included information on demographics, knowledge and attitudes about LF, and two dichotomous choice WTP exercises (one for a prevention intervention and the other for a lymphedema treatment intervention). Demographic information included sex, age, household size, monthly household cash income, and type of house construction (cement, wood, or thatch). The latter was used as an index of wealth. The questions regarding knowledge and attitudes towards filariasis were based on earlier anthropological work on community perceptions of the disease in Leogane [7]. The section included questions on familiarity with the different disease-related conditions, perceived risk of being infected with the filarial parasite, and the respondents' beliefs about transmission of the disease.
In the prevention exercise, individuals were asked about WTP for the prevention of infection in the entire household. Respondents were asked whether they would be willing to pay a predetermined amount to purchase DEC-fortified salt for their family, in order to prevent new filarial infections among family members. The respondents replied "yes" or "no" to the questions. Dichotomous choice questions are recommended over open-ended questions that ask individuals directly how much they are willing to pay [10]. Although the latter are statistically easier to analyze, they are often difficult for respondents to answer and may not result in individuals reporting their true maximum willingness to pay.
The valuation exercise for the treatment intervention followed the same basic format as that used to value prevention. Respondents were asked to imagine that they had lymphedema. A hypothetical condition was described, along with a course of treatment and expected improvement in health status from the treatment. Improvement was described as including cessation of acute attacks, reduced swelling, and improvement in mobility. The amount of time each day required for treatment was also described. Respondents were then asked whether they would pay different predetermined amounts for the supplies required for the lymphedema treatment (e.g., soap, towel, basin, antiseptics). The treatment exercise elicited WTP for treatment of one individual. In both exercises, bid amounts were presented in Haitian gourdes. They are converted to U.S. dollars in the results presented here.
For both exercises, each respondent was offered a single bid amount. The bids for each respondent were chosen randomly from 5 and 4 possible values for the prevention and treatment interventions, respectively. Following the general approach outlined by Alberini [11], bids were selected using limited data from pilot testing. Bid amounts were intended to cover the expected range of true underlying WTP values of respondents based on data from pilot testing with 25 individuals. Using a single bid value for all respondents can result in inefficient and biased parameter estimates [12].
The study area has an estimated population of approximately 20,000 individuals. The survey was administered to a random sample of adults in the town of Leogane and its immediate surroundings. Sampling was based on the division of area into four geographically defined regions: two in central Leogane (east and west) and two immediately adjacent communities (one to the north and one to the south). Within each area houses were selected by starting at a random corner and choosing every third house on the street. When the end of the street was reached, the process was begun on an adjacent street. Sampling within each area continued until 150 sample households were identified and one adult household member was interviewed. The study protocol and data collection instrument were approved by the human subjects protection committees of the Centers for Disease Control and Prevention and the Hospital Ste. Croix (Leogane, Haiti).
Data analysis
The data were analyzed from three perspectives: the distribution of household WTP within the community, the community's total WTP, and the expected participation of households in the interventions. Data were analyzed in two stages. In the first stage, data from the survey were used in two regression models to estimate household WTP for prevention of infection in the entire household and treatment of one individual, based on household characteristics. Willingness to pay for treatment (WTPt) and prevention (WTPp) interventions were each estimated with a single bound probit model using the responses to the valuation question. The probit model estimates the likelihood that the household would be willing to pay different amounts for the interventions, rather than estimating a single WTP for each household. The shape and location of the likelihood function depend on several household characteristics.
The model for WTPt included sex, age, income, whether someone in the household had lymphedema or hydrocele, accurate knowledge of the causes of filariasis, household construction type (cement vs. wood/thatch) as a proxy for wealth, and the starting bid. The model for WTPp included these same variables plus two additional ones: household size and perceived risk. The model was specified as
Pr {'yes'} = Φ(α + βA),
where Φ is the cumulative distribution function of the standard normal distribution, α and β are the estimated regression parameters, and A is the vector of household characteristics described above. The LIMDEP econometric software package was used to obtain the maximum likelihood estimates for the regression parameters [13]. The order in which the WTP questions were presented did not influence valuation responses in the initial model and this was not included in the final model.
In the second stage of the data analysis the results of the regression model were used to estimate the distribution of household WTP for prevention and treatment within the community (rather than that of individual households). Several approaches can be used to estimate population mean willingness to pay from the estimated household likelihood functions. The traditional approach of estimating WTP for each respondent in the sample to estimate the mean and distribution was not used. This approach was not used in effort to adjust for potential selection bias in our sample. Because a disproportionate number of respondents were women and approximately 35% did not provide their income, to adjust for potential selection bias in our sample a Monte Carlo simulation was used to estimate the mean and the distribution of household WTP, using Crystal Ball software [14]. The simulation model captures the variability within the population by repeatedly calculating the function estimated in the regression, using different combinations of household characteristics (income, number of members, etc.). This approach allowed us to use the expected gender distribution in the community and an estimated income distribution (based on those who responded), in order to adjust for potential biases within our sample. The model also incorporates uncertainty in the individual regression coefficient estimates.
Variability in respondent and household attributes was characterized by fitting distributions to the different independent variables (income, household members, age, etc.), using the survey data. Continuous distributions were fit for income, age, and number of household members using Crystal Ball's distribution fitting procedures. Discrete distributions were used for categorical data (e.g. sex, knowledge, house type). Uncertainty in the regression coefficients was characterized based on the estimated variances and covariances of the parameters estimated in the regression model.
A 2-dimensional Monte Carlo simulation model was used to separate variability in household characteristics from uncertainty in the estimated coefficients of the determinants of WTP. For both prevention and treatment, the model operates by first selecting a random set of regression coefficients from their respective distributions. Holding these coefficient values constant, the model then randomly selects 1,000 combinations of household characteristics, based on the distribution of each. These are used to describe a likelihood distribution (based on the probit regression results) from which each household's WTP is estimated. A new set of regression coefficients is then selected and the process is repeated (250 iterations). The result provides a description of the distribution of household WTP for prevention and treatment within the community, and confidence intervals for these estimates. The model was also run allowing all variables to change simultaneously to develop an overall estimate of the distribution (using 100,000 iterations), including variability within the population and uncertainty regarding the fitted distribution.
Participation rates for both interventions were then estimated by calculating the percentage of the population willing to pay more than a series of different amounts. For both interventions, mean household WTP was estimated among the entire population and those willing to pay a positive amount. For the prevention intervention, community willingness to pay was estimated based on estimates of population, household size, percentage of the population willing to participate, and mean WTP.