Pre-incarceration police harassment, drug addiction and HIV risk behaviours among prisoners in Kyrgyzstan and Azerbaijan: results from a nationally representative cross-sectional study

Introduction The expanding HIV epidemic in Azerbaijan and Kyrgyzstan is concentrated among people who inject drugs (PWID), who comprise a third of prisoners there. Detention of PWID is common but its impact on health has not been previously studied in the region. We aimed to understand the relationship between official and unofficial (police harassment) detention of PWID and HIV risk behaviours. Methods In a nationally representative cross-sectional study, soon-to-be released prisoners in Kyrgyzstan (N=368) and Azerbaijan (N=510) completed standardized health assessment surveys. After identifying correlated variables through bivariate testing, we built multi-group path models with pre-incarceration official and unofficial detention as exogenous variables and pre-incarceration composite HIV risk as an endogenous variable, controlling for potential confounders and estimating indirect effects. Results Overall, 463 (51%) prisoners reported at least one detention in the year before incarceration with an average of 1.3 detentions in that period. Unofficial detentions (13%) were less common than official detentions (41%). Optimal model fit was achieved (X2=5.83, p=0.44; Goodness of Fit Index (GFI) GFI=0.99; Comparative Fit Index (CFI) CFI=1.00; Root Mean Square Error of Approximation (RMSEA) RMSEA=0.00; PCLOSE=0.98) when unofficial detention had an indirect effect on HIV risk, mediated by drug addiction severity, with more detentions associated with higher addiction severity, which in turn correlated with increased HIV risk. The final model explained 35% of the variance in the outcome. The effect was maintained for both countries, but stronger for Kyrgyzstan. The model also holds for Kyrgyzstan using unique data on within-prison drug injection as the outcome, which was frequent in prisoners there. Conclusions Detention by police is a strong correlate of addiction severity, which mediates its effect on HIV risk behaviour. This pattern suggests that police may target drug users and that such harassment may result in an increase in HIV risk-taking behaviours, primarily because of the continued drug use within prisons. These findings highlight the important negative role that police play in the HIV epidemic response and point to the urgent need for interventions to reduce police harassment, in parallel with interventions to reduce HIV transmission within and outside of prison.

Specifically, to explore the relative effects of detention on HIV risk-taking, and investigate potential mediating relationships among the variables identified as significant correlates through bivariate testing while also accounting for moderating impact of each country, we performed a multi-group path analysis with official and unofficial detention as exogenous variables, addiction severity as a mediator, and composite HIV risk as an endogenous variable. We controlled for depression, anxiety, social support, and the presence of alcohol use disorders, and estimated indirect effects via bootstrapping procedures, while step-wise eliminating insignificant paths and "hanging" variables.
As noted in the manuscript, our data is cross-sectional; nevertheless, we believe it is important to clarify the temporal ordering in the questionnaire. Detention history measures asked about experiences in the year before the current incarceration; while drug addiction and HIV risk behaviors were assessed by a set of items addressing behaviors in the 30-day period prior to the arrest that resulted in the current incarceration. Being able to measure the current within prison drug injection (WPDI) for the Kyrgyzstan sample has allowed us to further clarify and confirm temporal ordering in our cross-sectional data. Thus, the relationships between detention, addiction severity, and HIV-risk are correlational, although they follow the temporal ordering of the survey.

Measures of Model Fit
The chi-square value χ 2 (with degrees of freedom and a corresponding p value) can be used to assess whether a specified model fits the data; however, this statistic should be used with caution, as it is sensitive to sample size, and is likely to result in statistical significance with a large sample size, rejecting the null that the model fits the data. 3 Comparative fit index (CFI) 4 is a commonly used index that involves comparing the fit of a specified default model against the fit of the null independence model that assumes no relationships among the variables.
Root Mean Square Error of Approximation (RMSEA) is a widely used index of model fit that looks at the differences between observed and predicted covariances and corrects for model complexity. A value of the RMSEA of about .05 or less typically indicates a good fit of the model in relation to the degrees of freedom, but a value of .08 or less are also adequate. The RMSEA is reported with the 90% confidence interval: LO 90 and HI 90.
PCLOSE is a p value that tests the null hypothesis that the RMSEA is no greater than .05, and if p is greater than .05, the model fit is close.
The goodness of fit index (GFI) ranges from 0 to 1, where 1 indicates a perfect fit. It is important to note that this measure is strongly affected by sample size and is reported here conventionally, as it has not lost its popularity despite the documented bias (see Kline 5 ). AGFI is adjusted GFI for the degrees of freedom, bounded by 1 above but is not bounded below by zero.

Path Diagrams
SEM is a graphical modeling methodology, so the graphical representation is essential to defining and formalizing the specification of a model. As can be seen in the path diagram below (Model 1), rectangles represent observed variables, which are the measurement constructs described under the study measures section of the manuscript. These variables are exogenous (predictors) and endogenous (criterion).
As shown in the path diagram, small circles (labeled e) are error terms or disturbances, which are latent influences on the constructs and reflect both variance attributable to random processes as well as the processes specific to that construct, and are usually attached to all endogenous variables in the model. 6 Model 1. Base aggregate model with covariates for two countries combined without bootstrap procedure.
Note: variable description "Drugs" refers to addiction severity and "HIVoutcome" refers to HIV risk throughout.

RESULTS
Results are produced by the AMOS text output *** p < .05 Model 5. KYRGYZSTAN WPDI with two detention variables. The identical path diagram is also provided in the originally submitted appendixmanuscript (Model Figure 2), and the results are summarized in Table 7.
Note: WPDI model presented in the manuscript (Figure 2) combines Unofficial and Official detention.

Notes for Model
Number of distinct sample moments: 10 Number of distinct parameters to be estimated: