Almost all CPS workers experience client violence. Although accumulating evidence suggest its deleterious consequences, few longitudinal studies measure client violence and its subtypes over time. The current mixed effect analysis provides the opportunity to test components of both social dominance theory and the JD-R model by considering how workers’ demographic characteristics and their job attributes contribute to client violence risk and how these relationships may change as workers gain experience. Few have sought to understand how demographic and work characteristics shape violence trajectories by disaggregating changes both between and within workers over time. The current study used a statewide cohort of recently-hired CPS workers followed prospectively from 6-months to 3.5 years after hire to consider two research questions: (1) how do types of client violence evolve as workers learn their positions? and (2) how do demographic and job attributes contribute to its prevalence over early job tenure? Based on available theory and limited empirical evidence, our hypotheses are as follows:

To be included in the analysis, participants were employed in a child welfare position and carrying a caseload, as not all respondents were employed in in the field work with clients (e.g., quality assurance staff). Table 1 displays the total respondents for a given wave, those still employed in child welfare, and those carrying a caseload. The sample size declined over time largely due to turnover.

The analysis examined violence trajectories from wave 3 to wave 8 due to the timing and availability of measures. Although client violence items were collected at wave 2, workers had been out of training and engaging with clients for 3–4 months, as compared to the 6-month interval in the remaining waves. To have a common time interval for exposure to client violence, the analysis of trajectories begins at wave 3, or about one year on the job. In addition, the analysis employed time-lagged predictors. Job attributes were first collected at wave 2 after participants completed training and had responsibilities as case workers for about three months. Although data were collected beginning 6- and 12-months post-hire for independent and dependent variables, respectively, we consider workers’ early-tenure because responses reflect experiences up to 6 months prior to the current wave. In addition, the diversity of CPS worker responsibilities likely means that workers continued to learn their positions at the first included wave of data collection.

Workers were racially and ethnically diverse. Over one-third of workers (35.3%) identified as non-Hispanic Black, 17.5% identified as Hispanic, 3.8% identified as non-Hispanic of another race, and 43.4% identified as non-Hispanic White. About 87% of respondents identified as women, 13% as men, and one respondent as transgender. On average, workers were 31 years old at hire with most workers in their 20s (SD = 9.05). Sixty percent of workers served families with substantiated maltreatment as case managers and the remainder worked as child protective investigators (CPIs) assessing for child maltreatment. About 30% of workers were CPIs located outside of Sheriff offices, and 10% were CPIs located in Sherriff offices.

This study used data from the Florida Study of Professionals for Safe Families (FSPSF), a prospective, statewide longitudinal study of newly-hired CPS workers to learn about workers’ experiences on the job, their wellbeing, and their retention. During mandatory pre-service training, all workers hired between September 2015 and December 2016 were invited to participate and 84% of eligible workers completed the online baseline survey (N=1500). Workers received online follow-up surveys every six months and were followed for 3.5 years regardless of whether they exited child welfare (see Wilke et al., 2017 for additional study details). The Institutional Review Board at the authors’ university approved all study protocol.

The analysis consisted of a descriptive analysis followed by a multi-level mixed effects regression analysis. First, we conducted a descriptive analysis to examine client violence of recently-hired workers and provide a descriptive picture of work-related characteristics from 6 months to 3.5 years post-hire. Second, mixed-effects logistic models, including fixed and random effects (Singer & Willett, 2003), examined trajectories of violence over time and how demographic and job attributes shaped trajectories. The models incorporated all available data (e.g., data from workers who contributed data for any model) and accounted for right censoring (e.g., early exits from sample; Singer & Willett, 2003).

Mixed effects models use data in the person-year format in which workers contribute up to six waves of data from waves 3–8 for the violence indicators with time-lagged predictors from the previous wave. The models allow workers to vary in their interview schedules, an inevitability in a large longitudinal project, and we measured time using the number of months since the worker’s hire date. Models were estimated at two levels. The first level examined within-worker change (e.g., how much does client violence change over time?). The second level examined each worker’s unique change trajectory considering demographic and job characteristics (e.g., how do worker demographics and job attributes influence changes in violence across individuals?). The Level 2 model considers how workers differ in their changes over time to determine each worker’s trajectory net of demographic characteristics, job attributes, and the interrelationships among the variables (Singer & Willett, 2003). We tested whether the cluster-based variation of the effect of time improved model fit for each form of violence through log-likelihood tests and included time when appropriate (Bates et al., 2015; Sommet & Morselli, 2017).

Important, the measured job attributes overlapped. For example, conceptually, work pressure can include role overload, an aspect of role demands. Initial correlational analyses also indicated overlap among many job attributes. Therefore, we ran separate multilevel regressions examining each job attribute. To test whether the relationship between each job attribute was time contingent, we interacted each job attribute with time. We compared model statistics (i.e., the Akaike Information Criterion (AIC), Baysian Information Criterion (BIC), and deviance) and added additional variables based on lower scores indicating better model fit. No work attribute-time interaction was significant or improved model fit. We present the best fitting models. All analyses were conducted in Stata 17.0.

Longitudinal analysis includes missing data due to attrition, non-response to a survey, and non-response to individual survey items. All FSPSF participants received surveys at every wave unless they notified us of their withdrawal from the study. About 2% (n = 23) withdrew from the study and nearly 5% (n = 67) did not complete any surveys after baseline. Survey response rates ranged from 77% to 87% (W2: 8%7; W3: 85%; W4: 80%; W5: 79%; W6: 77%; W7: 78%; W8: 78%). Across all waves, missing data on survey items ranged from 0% to 5% on all included variables. To retain as many workers as possible, we used person mean substitution to impute scale scores (Dodeen, 2003). For respondents with fewer than 50% missing values on a scale, we computed mean scores and multiplied by the number of items for an overall imputed sum score; otherwise, the scale score was left as missing. Missing data analysis indicated that those with missing data did not differ from those included in the analysis with three exceptions: those without missing data were older (Wave 3), more likely to be White (Waves 4, 5 ,7, and 8), and more likely to have experienced yelling (Wave 7).

The literature lacks consensus about appropriate strategies to estimate power within multilevel models (Bolger et al., 2011). A widely used, respected method noted for its flexibility uses Monte Carlo simulation (Arend & Schäfer, 2019). Monte Carlo simulation suggests that 80% power at α = .05 with a medium-sized intraclass correlation coefficient (ICC) and Level 1 units = 7, Level 2 samples of at least 200 can detect a .1 minimum effect size for Level 1 direct effects (i.e., changes over time), a .23 minimum effect size for Level 2 direct effects, and .32 for cross-level interactions. The current study (N = 7, n = 837–859) allows for the detection of effects in the small to medium range (Arend & Schäfer, 2019).

Table 2 displays the percentage of workers that experienced violent acts in the previous six months. The experience of being yelled at declined over time from 80% in the 6- to 12-month period to 64% in the 36- to 42-month period. The experience of being threatened declined from nearly 48% in the first period to nearly 39% in the last period. Physical violence did not significantly change over time. Additional analyses examining within-worker experiences over time suggest that change is more common for being yelled at or threatened compared to experiencing physical violence. For example, 55% of workers had waves in which they experienced yelling and waves in which they did not; over 60% of workers had waves in which they were threatened and waves in which they were not. In large part due to the low incidence of physical violence, only 15% of workers changed over time in their report of such behavior, with only two individuals reporting a physical incident at every wave.

Table 3 presents percentage/mean distributions, scale reliabilities, and sample sizes for the measures of job demands and resources. Cronbach’s alpha scores were at least acceptable at each wave, all exceeding .8. Table 4 provides correlations, and tests of significance with Bonferroni correction for all job attributes at Wave 2, the first wave of analysis for time-lagged variables. On average, correlations were high demonstrating high overlap. The number of hours worked and organizational characteristics significantly correlated with almost all other measures.

Table 5 presents the logistic mixed effects linear regression models of yelling, threats, and physical acts of violence. Examining yelling first, the empty model confirmed sufficient variation between workers to support the appropriateness for mixed effects models. The between-person constant approximates the average odds of reported yelling across workers and time. The constant indicates that the average probability of being yelled at during the 6-wave period was 0.853 (P = (5.79)/(1+5.79)). Its statistical significance suggests the probability of yelling was not constant across workers and time. The interclass correlation coefficient (r = 0.44) indicated a high degree of residual autocorrelation and that differences between workers accounted for 44% of the total variance in the probability of a yelling incident, with the remainder (56%) due to external circumstances. This significant variance suggested substantial residual variation between workers over time that may be explainable with additional variables.

The unconditional growth model added the time indicator (i.e., # of months since hire) to capture the rate of change in the odds of a yelling incident, and a random time component. The likelihood ratio test was significant (χ² (2) = 51.45, p = .000), indicating that allowing the specific effects of the independent variables to vary over time improved the model fit. The statistical significance of the within-worker constant means the odds of reporting a yelling incident decreased over time. Coupled with the significant between-worker constant, there is variation in both their initial odds of experiencing a yelling incident and the probability of experiencing such an incident over time.

The empty and unconditional growth model results suggest that workers’ experiences of yelling decreased over time and that between-worker differences partially accounted for this decline. Following, the third model examining yelling reflects the best fitting model. Worker demographic characteristics contributed to the odds that a worker experienced yelling. Women had twice the odds of being yelled at compared to men. Compared to Whites, Black and Hispanic workers had 78% (1– .22) and 66% (1 – .34) lower odds of yelling incidents, respectively. CPIs reported higher odds of being yelled at than case managers. Select job attributes also predicted yelling. Workers who classified a higher percentage of their cases as difficult and those who experienced higher levels of time pressure had higher odds of yelling incidents.

The middle columns of Table 5 report parallel results for threats. The constant indicates that the average probability of being yelled at during the 6-wave period is 0.42 (P = (.73)/(1+.73)). Similar to yelling, the odds of being threatened was not constant across workers and time. The interclass correlation coefficient of .40 indicated a high degree of residual autocorrelation such that worker characteristics accounted for 40% of the variation in experiencing a threat. The unconditional growth model for threats also indicated that the odds of an experience decreased slightly over time. The likelihood ratio test was significant (χ² (3) = 23.06, p = 0.000), indicating an improved model fit. The failure of the between-worker constant to achieve statistical significance suggested that the average probability of change did not vary over time. In other words, although workers varied in their initial odds of experiencing a threat, the typical worker did not change in the probability of experiencing a threat over the observed period. However, as indicated in the best-fitting model, variables related to important between-worker differences in change trajectories. Black and older workers experienced lower odds of experiencing a threat compared to their younger and White counterparts. Also, perceiving additional time pressure and more work demands increased the odds of experiencing a threat.

For the third series of models, the rightmost columns of Table 5 present the findings for physical violence. The empty model for physical violence confirmed sufficient variation between workers making mixed effects models appropriate. The between-worker constant suggested that the probability of experiencing physical violence over the study period was 0.03 (P = ((.03)/(1+.03)). The much smaller interclass correlation coefficient (.14) indicates a little autocorrelation, and worker characteristics account for a smaller proportion of the variation in experiencing physical violence. Put simply, physical violence is a more random event than yelling or threats. Also, unlike the earlier models, the introduction of time did not improve model fit. The likelihood ratio test was not significant (χ² (3) = 1.57, p = .666), and, thus, the best fitting model did not include time as a random effect. Reflecting the low correlation coefficient, no worker demographic characteristic was significant in predicting physical violence. However, CPIs located in sheriff’s offices had 69% (1–.31) lower odds of experiencing physical violence than case managers. A post-hoc comparison test indicated that the comparison between CPIs in sheriff offices to those in social services offices approached significance (OR = .36, p = .06). Time pressure was also significant: congruent with yelling and threats, perceiving higher levels of time pressure related to higher odds of exposure to physical violence.

This study has several limitations. First, the study relied on perceived violence. Subjectivity is important in defining violence and its consequences (Barth et al., 2013), yet multiple workers may classify a single incident differently. Second, the sample was a statewide sample of CPS workers hired during one period in one state. Although theory and prior research (e.g., Bakker & Demerouti, 2007; Bride et al., 2015) suggest similarities between CPS workers and others in vulnerable positions (e.g., home health, substance abuse, healthcare), results may not generalize to other types of frontline workers or to CPS workers in other states. Third, due to available data, analyses did not include measures of client violence during workers’ first six months on the job, including roughly the first three months of casework. Findings do not contribute to understating trajectories during this time. Fourth, although the dataset provided a unique statewide sample, we collapsed races and ethnicities other than non-Hispanic Black, Hispanic, and non-Hispanic White into one category due to small sample sizes. Data did not allow for analyses of races and ethnicities other than the three groups. In terms of gender, the options for gender did not follow best practices; respondents were restricted to three inaccurate options (i.e., male, female, transgender) that did not cover all gender identities. The sample included only one self-identified transgender person; therefore, findings cannot inform how gender operates for transgender or other gender diverse groups. In addition, minoritized workers were more frequently excluded from the sample due to missing data, a common problem in longitudinal research (Yancy et al., 2006). Missing data also related to age and yelling at single waves, potentially overrepresenting older workers and those who experienced yelling. Fifth, the study did not test reciprocal associations between job attributes and client violence. Considering potential bi-directionality is important for future work.

Future studies can build upon the current one in several ways. First, the consistency of job demands in predicting non-physical violence suggests examining how these demands relate to conflictual client engagement, given the connection between job stress, emotional dissonance, and burnout in other helping fields (Bakker & Heuven, 2006). Next, current findings indicate that high rates declined over time for yelling and threats, yet a similar proportion of workers experienced physical violence regardless of tenure. Future work can examine why this is the case. Workers may become more effective at dissipating situations that could result in yelling or threats or workers’ thresholds for what constitutes these forms of non-physical violence may change over time. Further, the current study’s results suggest that worker demographic characteristics matter for violence. Worker-client dyad data could provide insight into the mechanisms through which worker demographic characteristics operate. Perhaps, race, age, or gender worker-client differences increase the likelihood of non-physical violence. Alternatively, perhaps cultural definitions of violence contribute to disparate rates. Minoritized workers, for example, may hold higher thresholds for yelling or verbal aggression (Phelps et al., 1991) or value more direct approaches to dealing with client conflict (Shuter & Turner, 1997). Finally, the persistent rates of physical violence, coupled with few significant predictors, call for more research to understand the context in which these incidents occur to inform prevention strategies. Also, understanding why sheriff office affiliation may be one of few protectors of physical violence is important.

The current study provides important insight into trajectories of client violence among CPS workers, one vulnerable group in the social services employment sector, that has implications for both policy and prevention. First, child welfare services are a state responsibility overseen by federal mandates. State or federal agencies can provide organizations guidance to promote worker safety and system level changes. For example, the Department of Labor’s Occupational Safety and Health Administration’s (2016) voluntary guidelines provide a starting point for protection (e.g., management commitment to worker safety, hazard identification protocol) and model policy development.

Second, organizations can support their staff through developing policies that offer protocol to guide reporting of violent incidents and proscribe effective responses. Developing responsive, transparent procedures for social service workers at all levels to guide client interactions can nourish a culture of safety. The lack of protection that job resources offered workers suggests that articulating clear steps for workers and agencies to prevent violence and handle the aftermath of unavoidable incidents, particularly given the high unpredictability of physical violence, is also important.

Finally, it is notable that rates of non-physical violence are highest during early careers as case workers learn their positions. This argues for the need to provide effective training and skill development to assess for the potential for client violence and de-escalate situations when possible. Moreover, this study’s finding that job attributes, particularly demands, continue to contribute to client violence in spite of experience suggest the importance of workplace supports that address significant job demands. Child welfare and related professionals susceptible to client violence operate in fast-paced, unpredictable, volatile, and stressful environments (NIOSH, 2021). Organizational policies or practices that allow supervisors or other team members to step in and assist workers with time-sensitive tasks during highly stressful times can promote worker safety and wellbeing.