Objective Text messages significantly improve uptake of infant HIV testing in clinical trial contexts. Women who were excluded from a randomized trial in Kenya were followed to create a comparison between women who were enrolled and did not receive the study SMS intervention and women who were screened but not enrolled. Design Parallel-cohort randomized controlled trial analysis. Methods We compared time to infant HIV testing between women in three groups: the Trial SMS group, the Trial Control group, and the Comparison Cohort comprised of women who were screened but not enrolled. Results Of the 1,115 women screened, 388 (35%) were eligible for trial enrollment, and were randomized to receive either intervention text messages (Trial SMS; N = 195) or continue usual care (Trial Control; N = 193). Among 727 women not enrolled in the study (Comparison Cohort), we obtained infant HIV testing data from clinic records for 510 (70%). The cumulative probability of infant HIV testing was highest in the Trial SMS group (92.0%; 95% CI 87.5–95.3), followed by the Trial Control group (85.1%; 95% CI 79.5–89.8), and lowest among women in the Comparison Cohort (43.4%; 95% CI 39.2–47.8). Conclusions Both the Trial SMS group and the Trial Control group were significantly more likely to have their infants tested for HIV compared to the Comparison Cohort, providing evidence of a “clinical trial effect.” This analysis suggests that SMS interventions should be implemented as an adjunct to consistent and engaged delivery of basic health services.
In this “parallel-cohort RCT” analysis [5], the primary exposure was participation in the randomized efficacy trial of the TextIT strategy (ClinicalTrials.gov {“type”:”clinical-trial”,”attrs”:{“text”:”NCT01433185″,”term_id”:”NCT01433185″}}NCT01433185). The design, analysis, and results of the trial are presented elsewhere [3, 4]. Briefly, HIV-infected pregnant women in prevention of mother-to-child HIV transmission (PMTCT) programs were randomly assigned to receive either theory-based two-way text messages during pregnancy and the postpartum period (intervention arm) or usual care (control arm). Eligible women were 18 years of age or older, pregnant at a gestational age of 28 weeks or greater (or had delivered on the day of enrollment), enrolled in the PMTCT program, planning to remain in the study area, and had access to a mobile phone plus reported ability to read or have someone who read text messages on their behalf. Women who reported sharing phones were ineligible unless they had disclosed their HIV status to the person sharing the phone. The intervention consisted of 14 text messages sent at weeks 28, 30, 32, 34, 36, 38, 39, and 40 during pregnancy, and weeks 1, 2, 3, 4, 5, and 6 after delivery. Messages were tailored by inserting the mother’s and infant’s name, the infant’s sex, and allowing participants to choose their preferred language and time for receiving messages. The primary trial outcome was the proportion of infants undergoing HIV virologic testing within 8 weeks after birth, assessed using an intention-to-treat analysis. Ethical review committees of the Kenya Medical Research Institute, the University of Washington, and the University of California San Francisco approved the trial. All participants provided written informed consent. This novel “parallel-cohort RCT” design aims to maximize inclusion of the population at risk of disease to make information from the original TextIT trial more applicable during potential scale up of the intervention [5]. Women who were recruited and screened for the randomized trial but did not meet trial inclusion criteria were selected as the comparison group for this secondary analysis. The primary outcome was time to infant HIV testing, defined as obtaining a dried blood spot sample for HIV virologic testing within 8 weeks after birth. Baseline demographic and clinical characteristics for trial participants were collected using a questionnaire administered prior to randomization, while those for non-trial participants were extracted from trial screening logs, patient charts, antenatal care clinic registers, and electronic medical records. Follow-up and primary outcome information for both trial and non-trial participants were extracted from the HIV-exposed infants (HEI) register, HEI patient chart, health facility maternity register, postnatal clinic register, and PMTCT clinic patient charts. Data extraction was done approximately six months after analysis and publication of the initial randomized trial. Baseline characteristics were summarized using descriptive statistics, and differences across the three study groups were compared using chi-square tests. The three study groups included women receiving the SMS intervention (Trial SMS), women enrolled in the trial control group (Trial Control), and women screened but not enrolled (Trial Ineligible). Time to infant HIV testing in the three study groups was compared using the Kaplan-Meier method and log-rank tests. Hazard ratios and associated 95% confidence intervals (CI) were estimated using a Cox proportional hazards regression model. Time to infant HIV testing was measured in days from the date of birth to the date when a dried blood spot (DBS) sample was collected for HIV virologic testing. In routine practice, DBS samples are collected six weeks after birth to coincide with infant immunization clinic schedules. However, DBS collection before or after six weeks would still be valid for determining infant HIV status. Survival time was censored at eight weeks or at the time of maternal or infant death. The proportional hazards assumption was assessed using tests and graphs based on scaled Schoenfeld residuals [6]. Potential confounding factors were included in the final regression model using a sequential approach by first applying background knowledge using directed acyclic graphs (DAGs), and then using statistics-based methods to determine adjustment variables [7, 8]. In the first step, variables were selected for model inclusion based on a priori importance of being potential confounders of the association between trial participation and infant HIV testing [9]. These variables included trial eligibility criteria, age and education level of the mother, receiving ART (because of CD4 count <350 or stage 3 or 4 disease, which were indications for ART, independent of PMTCT, at the time of the trial [10]), year of maternal enrollment into PMTCT, maternal knowledge about PMTCT, and receiving ART prophylaxis for PMTCT (AZT during pregnancy, AZT+3TC+NVP at delivery, AZT+3TC after delivery, and infant NVP; as opposed to receiving ART because of CD4 count <350 or stage 3 or 4 disease). The relationships between these variables were summarized and analyzed qualitatively using a DAG (S1 Fig) [11]. Based on this DAG, age, education, and enrollment in the PMTCT program were determined to be sufficient for adjustment to control for confounding. In the second multivariable model-building step, a Cox proportional hazards model with a backward stepwise model building approach was used. All three variables determined in the first step were included in this second step. Variables that changed the estimated hazard ratio of the main effect by ≥10% were retained [12]. In the end, both age and education were dropped from the model. The final model adjusted only for enrollment in the PMTCT program (versus pregnant women living with HIV who were not enrolled to receive PMTCT services at the time of screening). To investigate whether results were biased by the high proportion of non-trial participants with missing outcome information, sensitivity analyses were performed that considered them all first as failures (no infant HIV testing by 8 weeks), and then as successes (infant HIV testing completed by 8 weeks). The multivariable regression models for these sensitivity analyses were unadjusted, given that non-trial participants without outcome information were also missing baseline demographic and clinical characteristics. We report unadjusted and adjusted hazard ratios and corresponding 95% CIs. All tests were two-sided and a significance level of P < 0.05 was used. Stata v13 was used to perform all statistical analyses (StataCorp, College Station, TX).