Comparison between VD model and CAU model
T, Chi-square and various kinds of regression tests will be used to compare DCR (binary variable), DCS (continuous variable), and hospitalization (continuous variable) and incidence of risky behavior (binary variable) rates between the VD and CAU models. Repeated analyses of variance will be used to compare psychiatric symptoms and social function (continuous variable) scores between the two groups.
We will use a difference-in-difference regression model with one treatment variable and additional regressors as control variables (see below). Insofar as perfect randomization and program execution are rarely achieved, we make use of an equation with a difference-in-difference Ordinary Least Square (OLS) estimator with additional regressors to estimate population true coefficients. The example below uses DCS as the dependent variable.
∆DCS: This is the value of DCSi for the ith individual at the end of the experiment minus the value of DCSi at the start, or the change. DCS is the Drug Compliance Score, a continuous variable ranging from 0-10, with 0 indicating zero compliance and 10 full compliance.
Treatment (VDmodel): This is a binary variable indicating whether an individual is enrolled in the treatment model (VD) or the control group (CAU). Treatment = 1 if an individual is in the treatment group and 0 for individuals in the control.
Social EconDemog: This refers to social, economic and demographic characteristics that can be plugged into the model as additional control variables for analysis.
Illness Severity: This is a composite of variables relating to illness severity such as number of years an individual has been diagnosed with schizophrenia before the start of our program, and days of hospitalization due to schizophrenia over the two years before the start of the program.
β4FamilyCare i: This variable indicates the level of care an individual receives from his or her family.
ui: This is the residual of all other determinants of DCS.
In the above equation, β1 indicates program effect, i.e., the effectiveness of village doctors in managing villagers with schizophrenia as measured by the average change in DCS for those in the treatment group (VD) minus the average change in DCS for those in the control group (CAU).
There are several advantages to using this model. First, the coefficient in which we are most interested, β1, is more efficient than other estimators as it has a smaller variance or standard error than a simple difference estimator when unobserved determinants of DCS persist over time for a given individual. Second, this model eliminates possible pre-program differences in DCS between the treatment and control groups; for example, while using cluster randomization, we may not achieve true randomization at the individual level since it is possible (though unlikely) that all eligible individuals within a cluster could be assigned to a treatment or to a control group. Finally, the difference-in-difference equation allows for additional variables to be added to the treatment variable in the VD group. As these added regressors are possible determinants of the dependent variable DCS, including them improves the efficiency of the OLS estimator, β1, which is the coefficient for program effect, thus addressing: potential bias introduced by imperfect randomization; partial compliance wherein patients with more severe illness in the VD group may at times elect not to comply with village doctors’ care, while patients with milder disease or more supportive family members in the CAU group may, of their own initiative, seek care from their village doctors; and attrition as severely ill patients are lost to follow-up (depending on the data collected, we will seek instrumental variables to address bias introduced by drop-outs). That said, we anticipate a high take-up rate by patients in rural Liuyang because the free antipsychotics program has been operating in the municipality for several years and has been well accepted by patients and their families. Given past experience in Liuyang, we anticipate a drop-out rate of 10% over the course of the study.
Cost effectiveness analysis
Addressing whether the program is cost-effective is also important. Our cost-utility analysis will assess the cost per quality-adjusted life-years (QALY) added with the program. The analysis is from the societal perspective such that any cost charged to any stakeholder is counted, including volunteer time. The lower the cost per QALY, the more cost-effective the program is. In the present study, costs for treating schizophrenia and controlling symptoms, e.g. VD and family members’ time as well as medication costs, will be weighed against averted costs, e.g. medical costs of hospitalization and the social costs incurred by the disruption of family and village life by untreated patients. The QUALY costs – and savings – of the proposed study can then be weighed relative to other medical interventions.
We will estimate QALY using a Markov Chain approach, largely modifying and following Garcia-Ruiz
et al. [
22]. As in Garcia-Ruiz, our model will allow for relapses into schizophrenia while taking medication, discontinuing medication due to side effects, or discontinuing medication due to other reasons.
Methods for distributing antipsychotics
In village A, the VD will distribute antipsychotics to family members on a weekly basis for two months. In village B, patients will receive their medicine in the village clinic every day for DOT. These medication distribution methods will be alternated in villages A and B after two months, so that the total time during which medication distribution is facilitated by VDs will be four months. Feedback on these distribution methods will be collected from participating patients, family members, and VDs in structured focus group discussions. The more feasible distribution method will be determined based on these structured discussions.