200 likes | 311 Views
Impact Evaluation in the Real World. One non-experimental design for evaluating behavioral HIV prevention campaigns . Implementation realities. BCC program: H as already started Builds on the previous campaign (not the first one addressing behaviour )
E N D
Impact Evaluation in the Real World One non-experimental design for evaluating behavioral HIV prevention campaigns
Implementation realities • BCC program: • Has already started • Builds on the previous campaign (not the first one addressing behaviour) • Is being rolled out in communities that have other HIV prevention interventions • There are endogenous ‘interventions’ (e.g. conversations on the way to school, or in the waiting line at the clinic) • Diffusion is a good thing • Cannot (and does not want to) control implementation
Effect = 3.47 – 11.13 = - 7.66 Participants 66.37 – 62.90 = 3.47 57.50 - 46.37 = 11.13 Non-participants
Effect = 8.87 – 16.53 = - 7.66 Before 66.37 – 57.50 = 8.87 62.90 – 46.37 = 16.53 After
Counterfactual assumption: Without intervention participants and nonparticipants’ pregnancy rates follow same trends
74.0 16.5
74.0 -7.6
Implementation realities • BCC program: • Has already started • Builds on the previous campaign (not the first one addressing behaviour) • Is being rolled out in communities that have other HIV prevention interventions • There are endogenous ‘interventions’ (e.g. conversations on the way to school, or in the waiting line at the clinic) • Diffusion is a good thing • Cannot (and does not want to) control implementation
What do we need to know? • Can a specific set of communication messages manipulate a specific set of sexual behaviors? • What magnitude of behaviour change will give what magnitude of changes in incidence?
Approach decided on • NON-intervention approach: • We are NOT trying to prove that one campaign works….. BUT we are trying to see whether a specific set of messages work, irrespective of the method of delivery or transmission of the method • Observation approach • We are not trying to force one intervention to work; not focusing on implementation of one intervention
So what WILL we do? • Non-experimental design • Researcher does not manipulate the independent variable (message exposure) • No control group in the community; create the control group statistically through matching • Collection of exposure, behavioural and biological data from random sample of individuals and their sexual partners • Develop a measurement of intensity of exposure (‘doses’ of exposure) • Determine the probability of having a specific dose of exposure • Match individuals with similar covariates, but different doses of exposure • Compare biological and behavioural outcomes
So what WILL we do? • Survey to measure demographic covariates (or use pop survey data) • Measure type and intensity of exposure to messages • Different doses of exposure to MCP campaign messages among the population • Detailed measurement of method of exposure to messages during surveys: Direct channels (# times heard messages on radio…); AND indirect channels (conversation with friends, relatives, etc.; as shown to be important in accounting for HIV declines in Uganda) • Construct message exposure scale (low vs. high, or more detailed) using statistical techniques (e.g., principal components analysis) • Every individual has a single score for message exposure
So what WILL we do? • Survey to measure exposure, behavioural outcomes, couple and social network norms and HIV incidence amongst random selection of individuals • Nested sub-study to trace partners of those who reported one or more sexual partners, and collect same data from them • Parallel measurement of ‘social norms’ – hearsay ethnography or other methods
So what WILL we do? • Analyses • Use covariates to calculate an individual’s propensity (scalar summary of all covariates) to receive a specific ‘dose of treatment’ (message exposure scale) • Match pairs of participants (index cases and their sexual partners) with similar propensity scores and different doses of treatment (control and treatment groups) • Now, we can calculate impact (behavioural and biological outcomes) by comparing the means of outcomes across participants and their matched pairs • Modeling • Has the density of the sexual network changed over time, and to what extent has it changed? • How 'much' behaviour change is needed, over what period of time in how many individuals, to bring about what levels of reductions in new infections • What are the individual and the combined effects of MC, ART, increased condom use, and MCP reductions, respectively, on the number of new infections • What is the ideal 'mix' of interventions to implement?
Density of scores for high exposure ‘Low exposure’ Density of scores for low exposure ‘High exposure’ Low probability of exposure given X High probability of exposure given X
What we will know • Can a specific set of communication messages (delivered in different ways) manipulate a specific set of sexual behaviors? • What magnitude of behaviour change will give what magnitude of changes in incidence?