Data were collected through a survey of CYCWs conducted from April through June 2015 as part of a broader formative evaluation of the Isibindi programme, at approximately the halfway point in a major expansion designed to train and deploy 10 000 additional CYCWs at new sites across the country [12]. A cross-sectional design was employed with individuals in the sample selected to represent all 4834 CYCWs employed at 318 Isibindi sites in total. The sample frame was based on a list of project sites with CYCW population size estimates provided by NACCW. Sites were stratified by pre-post expansion establishment status, with 83 sites established before 2012 and 235 sites established in 2012 or later. Within each time-of-establishment substratum, 40 sites were chosen with probability proportionate to size, with the size being the number of CYCWs and distribution across all provinces. All 1342 CYCWs employed at the selected sites were invited to complete an anonymous, self-administered, paper-based questionnaire in English. A total of 1158 CYCWs took part (86% response rate). Two selected project sites failed to submit any questionnaires, yielding 78 sites in the analytic sample.
Outcome measure
Retention
CYCWs were asked, “Within the next 12 months, do you think you will look for a new job with another employer?” Those who answered “no” (as opposed to “yes” or “maybe”) were categorised as intending to stay.
Predictor measures
Emotional exhaustion
Participants were asked to complete a modified version of the Maslach Burnout Inventory Human Services Scale (MBI-HSS) that included measures adapted from the emotional exhaustion subscale, such as “My work tires me emotionally” and “I feel frustrated by my work” [13]. The MBI-HSS has been used in South Africa among registered nurses with good internal and external validity [14]. Response options were coded 1 for “rarely,” 2 for “sometimes,” and 3 for “often”—and summed to create the total score (Cronbach’s alpha = 0.61).
Mentorship
CYCWs were asked, “How often does your mentor’s help and advice influence your day to day work?” “How often does your mentor treat you with respect?” and “In general, how comfortable do you feel asking your mentor for help with work-related concerns?” Responses of “always or often” (versus “sometimes” or “rarely or never”) on the first two items and “very comfortable” on the third item (versus “somewhat” or “not at all”) were considered indicative of high-quality mentorship and used to generate a binary variable.
Supervision
Supervision was measured using four items addressing the CYCW’s experience during the last 3 months: “How often did you meet with your [mentor]/[team leader] in a group with other CYCWs to discuss your work?” and “How often did you and your [mentor]/[team leader] meet in person—only the two of you?” Response options included “never” (0), “less than once a month” (0.5), “once a month” (1), and “more than once a month” (2). Coded values were summed to estimate the monthly number of supervisory interactions.
Motivation
Respondents were asked to select the “most important reason that [they] became a CYCW” from a list of six structured response options (plus other/specify). Those who responded that they became a CYCW to help their community or to work with children and youth, instead of for personal gain, were dichotomously categorised as altruistically motivated.
Perceived workload burden
Respondents were asked to indicate whether they felt their workload was “…too high, I have too much work to do a good job with all of it,” “…okay, I am busy but I can meet all my responsibilities,” or “…too small, I can take on more work and still do a good job.” Those who responded that their workload was too high were categorised as having a high burden.
Unsafe conditions
Respondents were asked to indicate if they had ever felt threatened or unsafe while doing their job as a CYCW (yes/no).
Compensation problems
CYCWs were asked how much they are supposed be paid per month, how much pay they received the previous month, and whether they had received that payment on time. Those who reported receiving no pay, less than their expected pay in the previous month, or having been paid late were categorised as having compensation problems.
Training modules completed
Respondents were asked to indicate every core training module they had completed, using a list. This measure reflects the number of core modules completed (out of 14 maximum).
Statistical analyses
Nationally representative sampling weights were calculated from the sample frame data provided by NACCW and the numbers of completed questionnaires received. Regression models were estimated at the individual CYCW level with clustering at the project/site and mentor levels, with CYCWs nested within project sites and project sites nested within mentors. Records with a single missing response on the emotional exhaustion scale items had its value imputed (n = 79), whereas those with more than one missing response were excluded (n = 33). Records without information on the CYCWs’ duration of employment were dropped from the model (n = 75). Records with missing data for dichotomous variables were coded to reflect the converse of the condition indicated.
Descriptive statistics were estimated using SAS v9.3 (Carey, NC) and adjusted to account for sampling weights. Mixed effects logistic regression using Stata IC 14 (College Station, TX) was utilised to estimate odds ratios associated with factors predictive of intention to remain employed as a CYCW. Respondents’ age, gender, length of time spent working as a CYCW, and high school completion status were included in the model. Random intercepts for project site within mentor were included to correct standard errors for clustering at those levels. Intraclass correlations were estimated at the sampling strata levels of province and established/new project site and found to be negligible after accounting for clustering by project site and, therefore, not accounted for in the multilevel model. Although the intraclass correlation for mentors was below 10% in the model, mentor-level random intercepts were included to enable comparisons of the amount of variance explained at the mentor as well as project site levels.