Mediation analysis of an adolescent HIV/STI/pregnancy prevention intervention

Abstract

Most interventions designed to prevent HIV/STI/pregnancy risk behaviors in youth have multiple components based on psychosocial theories (e.g., social cognitive theory) dictating sets of mediating variables to influence to achieve desired changes in behaviors. Mediation analysis is a method for investigating the extent to which a variable X (e.g., intervention indicator) influences an outcome variable Y (e.g., unprotected sex) by first influencing an intermediate variable M (e.g., self-efficacy to use a condom) and provides a way for empirically validating theoretical hypothesized mediation. In this way, mediation analysis is a critical tool for suggesting which components of complex interventions should be the focus of future more efficient and effective interventions. The present study applied multilevel mediation analysis to outcome data from the All4You2! study to begin to examine the relationships between a theory-based HIV/STI/pregnancy prevention curriculum for alternative school youth at risk for education failure, targeted psychosocial mediating variables, and the primary outcome unprotected sex in the past three months. Results suggest helping alternative school youth identify and avoid exposure to risky situations and improving their self-efficacy to refuse sex should be focal points of future interventions with these youth.

Introduction

The vast majority of HIV/STI/pregnancy prevention interventions for young people and adolescents seek to reduce sexual risk-taking behaviours, such as frequency of vaginal intercourse without a condom, initially by impacting on psychosocial mediating factors, such as normative beliefs regarding those behaviours (Kirby, Laris, and Rolleri 2007; Bell et al. 2007; DiClemente et al. 2008). Many such programmes are guided by cognitive behavioural theories, such as social cognitive theory (Bandura 1986) and social influence theory (McGuire 1972; Fisher 1988), which identify the psychosocial variables to target to produce changes in behaviours.

Unfortunately, there is little empirical evidence showing which psychosocial variables provide the actual mediating causal mechanisms through which existing HIV/STI/pregnancy prevention interventions affect sexual risk-taking behaviours for adolescent populations (Kirby, Laris, and Rolleri 2007; Hardnett et al. 2009). This key issue is something Doug Kirby highlighted in his work, as well as in his encouraging personal communications about the potential for different analytic techniques – in particular, mediation analysis explore here - to contribute needed information about intervention mechanisms in the field of sexual and reproductive health.

Mediation analysis is a statistical technique that can be used to empirically investigate these mechanisms by examining the degree to which an intervention influences an outcome variable (e.g., unprotected sex) by influencing first a mediating variable (e.g., self-efficacy to refuse sex) posited to be related to the outcome (Baron and Kenny 1986; MacKinnon 2008; Gelfand, Mensinger, and Tenhave 2009). In particular, the careful application of mediation analysis to data from intervention studies can provide valuable information about which psychosocial variables provide the strongest pathways to potential behaviour changes, thereby providing insight into how intervention components might be prioritised and strengthened and whether psychosocial theories underlying interventions need to be modified (Hardnett et al. 2009; Flay, Petraitis and Hu 1999). This information is particularly important given the recent movement towards simplifying complex behavioural interventions for more efficient and effective real-world implementation (Rotheram-Borus, Swendaman, and Chorpita 2012; Collins, MacKinnon, and Reeve 2013).

Operationally, a mediation analysis involves performing a set of regression analyses to obtain estimates of 1) the strength of the relationship between an independent variable X (e.g., intervention indicator) and a hypothesised mediator variable M (e.g, self-efficacy), and 2) the strength of the relationship between the mediator variable M and the dependent variable Y (e.g., unprotected sex outcome), adjusting for the effects of X; combining these estimates provides a measure of the strength of the mediated, or indirect effect of X on Y through mediator variable M. A key assumption of mediation hypotheses is that there is a causal chain of events in which X occurs and produces a subsequent change in M, which in turn causes a subsequent change in Y (MacKinnon, 2008; Baron & Kenny, 1986). Of special interest is the amount of change in Y caused by X indirectly transmitting change through M – i.e., the indirect, or mediated effect. This indirect effect may be whole or partial, accounting for a proportion of the total effect of X on Y.

Figure 1 shows a schematic path diagram for assessing effects within a mediation process. The effect, a, of the intervention on the mediator reflects the “action theory” of an intervention, which suggests how the intervention should be designed to affect a given psychosocial variable – e.g., inclusion of a component to increase students’ self-efficacy to refuse sex. The effect, b, of the mediator on the outcome, adjusted for the intervention, reflects the “conceptual theory” of an intervention, which is the psychosocial theory (e.g., social cognitive theory) from which the targeted mediators were selected. The indirect (“mediated”) effect of the intervention on the outcome transmitted through the mediator is ab; the effect of the intervention on the outcome, adjusted for the mediator, is c’. Thus, the total effect, c, (not shown) of X on Y is decomposed by the mediation process into a direct effect, c′, and an indirect effect, ab, which is transmitted through the mediator M.

Figure 1.

Single mediator process.

Given that mediators may work in complex ways and exert opposite indirect influences on the outcome, a significant intervention impact on the outcome is not a prerequisite to performing mediation analysis (MacKinnon 2008). It has been argued that a significant intervention effect on the hypothesised psychosocial mediator also is not a necessary condition to conducting a test of whether the psychosocial variable was a significant mediator of the intervention effect (Hayes 2009). Indeed, mediation analysis involves not just a single statistical test of whether the mediated (“indirect”) effect of a psychosocial variable is significant at a preset level (e.g., p<0.05), but examination of the magnitude and direction of all the individual pathways involved in a mediation process. For example, a substantial b but small a might suggest that the mediator being targeted is correct but the intervention needs to be strengthened in its strategies for influencing the mediator, and/or the mediator needs to be measured more accurately. A large a effect but small b would suggest that although the intervention is successful in influencing the targeted mediator, the psychosocial theory dictating which mediator to target needs some refinement (MacKinnon 2008). Thus, mediation analysis can provide rich information about the strengths and weaknesses of an intervention’s underlying conceptual model, critical information to help intervention developers tailor their interventions to maximise resources and impact (Bell et al. 2007; DiClemente et al. 2008).

Only a handful of studies to date have used mediation analysis to validate the hypothesised psychosocial mediators for HIV/STI/pregnancy prevention interventions (Stigler et al. 2006; Dilorio, McCarty, and Denzmore 2006; Villaruel et al. 2008; Soper et al. 2010; Li et al. 2011; Baumler et al. 2012), and no mediation analyses of HIV/STI/pregnancy prevention interventions for US high school populations were found in the literature. The purpose of this study is to apply mediation analysis to data from the All4You2! outcome study, on which Doug was a co-investigator. This study evaluated the impact of HIV/STI/pregnancy prevention education and service learning on psychosocial mediators and sexual risk-taking behaviours among students attending alternative high schools for youth at risk of education failure. The outcome study found a modest statistically significant (at p<.05) effect of the HIV/STI/pregnancy prevention curriculum on the primary outcome unprotected sex in past 3 months at the short-term (6-month) follow up, but no significant effect at the long-term (18-month) follow-up. Few statistically significant effects were seen on the measured psychosocial variables (Coyle et al. 2013).

Methods

All4You2! study design

The intervention was evaluated using a group-randomised controlled trial involving 47 classrooms drawn from 11 district-run alternative high schools in California. The trial procedures were approved by the Institutional Review Board of the primary author’s institution (ETR Associates). Recruitment occurred between 2006 and 2009. Classrooms were randomly assigned using computer-generated random numbers. Students completed three surveys (baseline plus two follow up surveys at 6 and 18 months following baseline). Of the students who were eligible to participate in the study (N=899), roughly 85% completed baseline surveys (N=762 available for analysis). Student retention rates were 80% and 70% at 6 and 18 months, respectively; 86% of the cohort completed at least one follow up survey (83% to 88% by condition).

Intervention

Although students were randomised to one of four conditions (a theory-based HIV/STI/pregnancy prevention curriculum, a service learning unit, the prevention curriculum plus a service learning unit, or an attention control condition), the present study focused only on mediators of the HIV/STI/pregnancy prevention curriculum relative to the control condition. Each intervention consisted of 32 hours of instruction, taught 2–3 times a week. The HIV/STI/pregnancy prevention curriculum addressed essential content and skills related to HIV, other STI, and pregnancy prevention stemming from underlying theoretical models (e.g., social cognitive theory, social influence models,) and key risk behaviours (e.g., condom use, multiple partners, etc.). The attention control condition featured lessons on nutrition and physical education behaviours and was used to control for Hawthorne effects resulting simply from extra group interaction and special attention experienced by students receiving the interventions (Cook and Campbell 1979). For fuller details of the intervention and outcome study, see Coyle et al. (2013).

Measures

One of the primary outcomes of interest in the All4You2! main outcome study was a reduction in unprotected sex within the past 3 months, and this was used as the outcome variable in the mediation analyses. It was coded as 1 for students who reported having sex at least one time without a condom in the past 3 months, and as 0 for students who either reported no instances of sex without a condom or who did not have sex in the past 3 months. This dichotomisation reflects the key messages of the programme – choose not to have sex or use condoms every time if having sex. Thirteen psychosocial variables stemming from the theories underlying the HIV/STI/pregnancy prevention curriculum were analysed as potential mediators. The variables, along with their basic psychometric properties, are shown in Table 1.

Table 1.

Scales and indices with basic psychometric properties.

Construct	Number of Items	Sample Item	Response Format	Alpha a
Behaviorally-related Scales
Exposure to Risky Situations	3	In the past month, how often have you laid down on a couch or bed alone with someone you like?	6- point scale (1 = 0 times, 6 = 16 or more times)	0.74
Volunteering/Helping in Community	1	In the past 6 months, how many times have you volunteered in your community?	4- point scale (1 = 0 times, 4 = 5 or more times)	NA
Psychosocial Indices/Scales
Knowledge Indices
Condom Knowledge	4	When using a condom, a person should always leave some room or space at the end of it.	True/false; count of number correct	NA
HIV Knowledge	2	All STDs, except for HIV/AIDS, can be cured with antibiotics. (reverse coded)	True/false; count of number correct	NA
Perceived Self-efficacy
To get and use condoms	4	If you were going to have sex, could you insist on using a condom if your partner did not want to use one?	4- point scale (1 = I am sure I could not, 4 = I am sure I could)	0.80
To refrain from having sex (“refusal self-efficacy”)	3	Could you keep from having sex if you had strong sexual feelings for someone?	4- point scale (1 = I am sure I could not, 4 = I am sure I could)	0.81
Skills
Refusal Strategies (Skills)	5	In the past month, how many times have you suggested doing something else to keep from having sex or unprotected sex?	4- point scale (1 = 0 times, 4 = 5 or more times)a	0.80
Attitudes and Beliefs
General attitudes toward condoms	15	Asking my partner to use condoms would be like saying “I don’t trust you.”	5- point scale (1 = strongly disagree, 5 = strongly agree)	0.80
General attitudes toward birth control	3	Getting female methods of birth control is a hassle. (reverse coded)	5- point scale (1 = strongly disagree, 5 = strongly agree)	0.74
Beliefs about not having sexual intercourse	5	I believe it is OK for people my age NOT to have sex.	5- point scale (1 = strongly disagree, 5 = strongly agree)	0.63
Normative Beliefs
Perception of peer beliefs about not having sex	4	Most of my friends think it is OK for people my age NOT to have sex.	5- point scale (1 = strongly disagree, 5 = strongly agree)	0.75
Perception of peer beliefs about using condoms	3	Most of my friends think people my age should always use condoms when having sex.	5- point scale (1 = strongly disagree, 5 = strongly agree)	0.46
Perceptions of number of peers who use condoms	1	Most of my friends use condoms if they are having sex.	5- point scale (1 = strongly disagree, 5 = strongly agree)	NA
Communication
Communication Skills	5	I am good at saying what I think.	4- point scale (1 = strongly disagree, 4 = strongly agree)	0.65

^aCronbach’s Alpha based on students’ baseline responses.

Data analysis

The All4You2! study design consisted of repeated measurements nested within students nested within classrooms nested within schools. Thus, in keeping with the main outcome study, 4-level repeated measures multilevel regression models were used to estimate the a, b, and c′ effects in order to adjust standard errors for the presence of intra-class correlation among measurements on the same student, and students among the same classrooms and schools. Models included a continuous time variable indicating the exact follow-up time point for a given student, interaction terms between the time variable and the intervention indicators (for the a path estimate) or between the time variable and the mediator variable (for the b path estimate), and were adjusted for the following common set of covariates for which there was some imbalance across study arms at baseline: age, language spoken in the home (measured as the amount of English spoken), suspension in the past year and probation status. Although models contained all three intervention indicators for completeness, given the focus of the present paper only the effect of the HIV/STI/pregnancy prevention curriculum relative to the control condition was examined in the mediation models. For the models in which the dichotomous outcome, unprotected sex in the past 3 months, was the dependent variable (the b and c′ path models), multilevel logistic regression models were used. Therefore, all estimates reported were standardised in order to carry out significance testing (MacKinnon 2008; MacKinnon and Dwyer 1993). Models in which the mediator was the dependent variable (the a path models) used linear multilevel regression models. All analyses were carried out using MLwiN software v. 2.23 (Rasbash et al. 2011). It should be noted that the estimates of the magnitude of the intervention effects reported here differ slightly from those reported in the main outcome study because they are standardised and because a common set of covariate controls was used for all mediation models.

The mediation process was assessed by examining the significance level, size and direction of the a and the b effects using Wald tests and by examining the significance level, size and direction of the ab effect estimate using the Sobel test (MacKinnon 2008; Hayes 2009; Sobel 1982). Following tradition, effects are referred to as statistically significant if their significance level is p < 0.05; however given this is an exploratory study, p-values less than 0.10 were considered potentially important as well. The magnitude of each indirect effect was measured using the proportion of the absolute total effect that was mediated for each psychosocial variable: |ab|/(|c′|+|ab|) (MacKinnon 2008). In keeping with the main outcome study, the present study assessed the mediating effects of each psychosocial variable separately at 6 month and 18 month follow-ups, using appropriate linear combinations of regression coefficients to obtain separate 6-month and 18-month estimates. In other words, the study did not consider lagged mediation models in which a psychosocial variable measured at 6 months would be considered as a mediator for an outcome measured at 18 months, given there are no theory-based standards that dictate the time lags between the mediators and behaviour changes.

Results

Tables 2 and 3 provide estimates of the standardised a, b, c′ and ab effects and corresponding standard errors from the multilevel regression models.

Table 2.

Mediation of intervention effects at six months.

	Program’s effect on the mediator			Mediator’s effect on the outcome			Program’s effect on the outcome adjusting for the mediator			(Sobel test) estimate of mediated effect			Proportion of effect that is mediated
	a	(SE)	p-value	b	(SE)	p-value	c’	(SE)	p-value	ab	(SE)	p-value
Knowledge
Condom knowledge	0.003	0.006	0.617	0.057	0.051	0.264	−0.195	0.062	0.002 **	0.000	0.000	0.648	0.089%
HIV knowledge	0.009	0.010	0.379	−0.054	0.055	0.322	−0.111	0.077	0.148	0.000	0.001	0.511	0.441%
Perceived Self-efficacy
Perceived self-efficacy to get and use condoms	−0.016	0.013	0.241	−0.140	0.044	0.002**	−0.151	0.078	0.053	0.002	0.002	0.271	1.438%
Perceived self-efficacy to refrain from having sex	0.035	0.020	0.074	−0.096	0.042	0.023*	−0.139	0.078	0.076	−0.003	0.002	0.160	2.385%
Skills
Refusal strategies	0.004	0.018	0.806	−0.052	0.047	0.274	−0.158	0.067	0.019*	0.000	0.001	0.810	0.144%
Attitudes and Beliefs
General attitudes toward condoms	−0.009	0.014	0.527	−0.193	0.044	0.000*	−0.138	0.077	0.075	0.002	0.003	0.531	1.225%
General attitudes toward birth control	−0.006	0.017	0.715	−0.021	0.041	0.601	−0.141	0.079	0.075	0.000	0.000	0.765	0.096%
Beliefs about not having sexual intercourse	0.025	0.018	0.157	−0.094	0.045	0.038*	−0.135	0.077	0.083	−0.002	0.002	0.242	1.712%
Perceptions of peer beliefs about not having sexual intercourse	0.021	0.021	0.321	−0.092	0.043	0.033*	−0.135	0.076	0.077	−0.002	0.002	0.368	1.432%
Perceptions of peer beliefs about using condoms	0.003	0.017	0.874	−0.192	0.042	0.000**	−0.139	0.077	0.072	−0.001	0.003	0.874	0.378%
Perceptions of number of peers who use condoms	−0.016	0.018	0.361	−0.096	0.040	0.016*	−0.140	0.075	0.064	0.002	0.002	0.392	1.111%
Risky Situations
Exposure to risky situations	−0.046	0.032	0.156	0.447	0.041	0.000**	−0.102	0.066	0.123	−0.021	0.015	0.160	16.829%
Communication
Communication Skills	−0.011	0.010	0.281	0.021	0.044	0.634	−0.139	0.077	0.073	0.000	0.001	0.663	0.170%

Notes: s Regression path coefficients have been standardized.

^*Significantly mediated intervention effect p < 0.05;

^**Significantly mediated intervention effect p < 0.01.

Table 3.

Mediation of intervention effects at 18 months.

	Program’s effect on the mediator			Mediator’s effect on the outcome			Program’s effect on the outcome adjusting for the mediator			(Sobel test) estimate of mediated effect			Proportion of effect that is mediated
	a	(SE)	p-value	b	(SE)	p-value	c’	(SE)	p-value	ab	(SE)	p-value
Knowledge
Condom knowledge	0.004	0.008	0.624	0.107	0.071	0.133	−0.489	0.228	0.033 *	0.000	0.001	0.641	0.31%
HIV knowledge	−0.010	0.013	0.444	0.104	0.078	0.185	−0.121	0.373	0.746	−0.001	0.002	0.507	3.02%
Perceived Self-efficacy
Perceived self-efficacy to get and use condoms	−0.001	0.017	0.974	−0.106	0.062	0.088	−0.137	0.350	0.695	0.000	0.002	0.974	0.16%
Perceived self-efficacy to refrain from having sex	0.009	0.024	0.716	−0.128	0.063	0.045*	−0.141	0.349	0.687	−0.001	0.003	0.721	2.85%
Skills
Refusal strategies	−0.003	0.022	0.882	0.057	0.067	0.397	−0.272	0.334	0.416	0.000	0.001	0.884	0.25%
Attitudes and Beliefs
General attitudes toward condoms	0.003	0.017	0.868	−0.214	0.063	0.001**	−0.160	0.350	0.648	−0.001	0.004	0.868	1.35%
General attitudes toward birth control	−0.006	0.023	0.814	−0.088	0.061	0.148	−0.182	0.350	0.603	0.000	0.002	0.816	0.97%
Beliefs about not having sexual intercourse	0.006	0.021	0.762	−0.133	0.064	0.039*	−0.174	0.347	0.616	−0.001	0.003	0.765	1.74%
Perceptions of peer beliefs about not having sexual intercourse	0.010	0.026	0.696	−0.079	0.065	0.222	−0.157	0.342	0.646	−0.001	0.002	0.709	1.84%
Perceptions of peer beliefs about using condoms	0.003	0.022	0.901	−0.200	0.063	0.002*	−0.161	0.350	0.646	−0.001	0.004	0.901	1.25%
Perceptions of number of peers who use condoms	−0.009	0.025	0.719	−0.035	0.061	0.565	𢄠.138	0.339	0.684	0.000	0.001	0.760	0.84%
Risky Situations
Exposure to risky situations	−0.086	0.040	0.031*	0.495	0.061	0.000**	0.032	0.339	0.925	−0.043	0.020	0.038*	84.74%
Communication
Communication Skills	−0.007	0.013	0.584	0.017	0.063	0.794	−0.165	0.344	0.632	0.000	0.001	0.814	0.26%

Notes: s Regression path coefficients have been standardised

^*Significantly mediated intervention effect p < 0.05;

^**Significantly mediated intervention effect p < 0.01.

6-month follow-up

As seen in the main outcome study (Coyle et al. 2013), at the six-month follow-up unprotected sex in the past three months was significantly impacted upon by the intervention (standardised c = –0.15; p = 0.04); none of the psychosocial variables measured were significantly impacted upon by the intervention at the six-month follow-up, although perceived self-efficacy to refrain from having sex came close at p = 0.07 (standardised a = 0.035). On the other hand, there were multiple b estimates that were statistically significant at the traditional p < 0.05 level representing relationships, in the expected direction, between the following psychosocial variables and unprotected sex in the past three months, after adjusting for the intervention effect: self-efficacy to get and use condoms, self-efficacy to refrain from having sex, attitudes towards condoms, beliefs about not having sex, perceptions of peer beliefs (peer norms) about not having sex, perceptions of peer beliefs (peer norms) about using condoms, perceptions of number of peers who use condoms, and exposure to risky situations.

Some degree of partial mediation was indicated by the c′ estimate for all psychosocial variables except condom knowledge, self-efficacy to get and use condoms, and refusal strategies. However, the results of the Sobel test for significance of the estimate of the mediated effect, ab, showed no statistically significant mediators of the intervention at the 6-month follow-up. However, the indirect effect of exposure to risky situations accounted for 16.8% of the total intervention effect (p = 0.16). The next highest proportion of intervention effect was accounted for by the mediator self-efficacy to refuse sex at 2.4% of the total effect (p=0.16).

18-month follow-up

As in the main outcome study (Coyle et al. 2013), at the 18-month follow up unprotected sex in the past 3 months was not significantly impacted upon by the intervention (standardised c = −0.05; p = 0.61); of the psychosocial variables measured, only exposure to risky situations was significantly impacted by the intervention (standardised a = −0.086; p = 0.03). There were multiple statistically significant (p < 0.05) b effects in the expected direction: self-efficacy to refrain from having sex, attitudes towards condoms, beliefs about not having sex, perceptions of number of peers who use condoms, and exposure to risky situations.

At 18-months, the c′ effect estimates indicated a high degree of partial mediation for the psychosocial variable exposure to risky situations only. Indeed the mediated effect of this variable was statistically significant (p = 0.04) and accounted for 84% of the total intervention effect at this follow-up. The mediated effects of HIV knowledge and self-efficacy to refrain from having sex accounted for roughly 3% of the total intervention effect each, but these indirect effects were not close to statistically significant.

Discussion

This study found that many of the psychosocial variables targeted by the All4You2! intervention were indeed related to one of the main behavioural outcomes (unprotected sex in the past three months), providing evidence that the conceptual theories from which those variables stem are appropriate. In particular, after controlling for the effect of the intervention, the relationships between the behavioural outcome and the psychosocial variables self-efficacy to refuse sex, attitudes towards condoms, beliefs about refraining from sex, exposure to risky situations, and peer norms about condoms were all statistically significant and in the expected direction at both 6-month and 18-month follow ups. Additionally, at the six-month follow up, self-efficacy to get condoms, peer norms about refraining from sex, and perceptions of peer use of condoms were significantly associated with the behavioural outcome. Baumler et al. (2012) found many of these same variables were related to their behavioural outcome (sexual initiation) in mediation analyses they conducted of a similar theory-based HIV/STI/pregnancy prevention intervention for middle school youth [It’s Your Game (IYG)], further supporting the use of these theories for these interventions.

The a estimates of the intervention’s effect on the psychosocial mediators were statistically significant only for one of the 13 psychosocial variables measured at 18-month follow-up: exposure to risky situations. The impact of the intervention on self-efficacy to refuse sex was close to significant at the 6-month follow-up (p = 0.07). Baumler et al. (2012) also found that few psychosocial variables (3 of 13 – self-efficacy to refuse sex, STI signs and symptoms and peer norms about condoms) were significantly impacted upon by their intervention. The lack of impact on the psychosocial variables may be the result of a variety of factors, such as: insufficient exposure to the intervention; a need to narrow and deepen the focus of the intervention on fewer mediating factors; and/or a lack of sensitivity of the measures. The participants in All4You2! received an average of 70% of the lessons, a reflection of the relatively short amount of time young people tend to stay in alternative school settings; this reduced exposure may have reduced the intervention impact.

Exposure to risky situations emerged as the most important mediator of the intervention effect at both follow-up time points; it accounted for 16% of the intervention effect at 6 months, and it was a statistically significant mediator at the 18-month follow-up where it accounted for 84% of the intervention effect. The variable exposure to risky situations measured how often a student had been at a party or friend’s house or with a boyfriend or girlfriend without a parent/guardian present, and how often s/he had laid down on a couch alone with someone s/he liked. The emergence of this construct as an important mediator at both follow-up time points in the present study suggests one area for deepening intervention work with alternative school youth, particularly given data suggesting the majority of these youth have had sexual intercourse. Of interest, Baumler et al. (2012) also found that exposure to risky situations, though not a statistically significant mediator, accounted for the largest proportion of the effect of their intervention on middle school students. Future interventions may want to focus more on this concept, ensuring students can identify risky situations and, more importantly, develop realistic strategies for avoiding or getting out of these situations. Self-efficacy to refuse sex was close to being a significant mediator at the six-month follow-up, and also may remain an important focal point of these programs, consistent with the underlying social cognitive theory. Further support for the importance of focusing on self-efficacy to refuse sex in HIV/STI/pregnancy prevention interventions was found in Baumler et al. (2012) where the variable was a significant mediator of the intervention effect on their main outcome of interest, sexual initiation.

The lack of significant mediating effects of the psychosocial variables examined at the 6-month follow-up, when there was a significant intervention effect on the behavioural outcome, and the inability of the strong meditational influence of exposure to risky situations to transmit, by itself, a significant effect on the behaviour at 18 months, may be due to several factors. First, the All4You2! intervention’s impact on unprotected sex in the past 3 months may be influenced by additional mediators not measured in this study. Second, the psychosocial variables may have worked together in more complex, potentially cumulative processes to influence the outcome. Flay, Petraitis, and Hu (1999) argued that the complexity of the processes underlying health risk behaviours involving substance use calls for a broadening of the theories upon which many prevention intervention curricula are based, as well as a need for more specificity with respect to what parts of theories are most applicable; this argument is likely very relevant to HIV/STI/pregnancy prevention programmes and the underlying psychosocial variables they target to reduce risk behaviours. Third, the intervention component that focused on what to do in risky situations may not have had sufficient emphasis in relation to other intervention components. On the other hand, there was a significant reduction in exposure to risky situations at 18 months, and this actually could account in part for the lack of an effect on the behavioural outcome (unprotected sex) at the 18-month follow-up: with exposure to fewer risky situations by 18 months, students had fewer opportunities to have unprotected sex.

The mediation analyses conducted for this study are a first step and address only one aspect of the causal pathway underlying a theory-based prevention intervention such as All4You2!. Conducting growth process mediation analyses and multilevel structural equation modeling (MSEM) analyses that allow for decomposition of mediating effects into individual student and group (school or classroom) levels also would provide additional information about the nature of these causal pathways. Furthermore, mediation analyses need to be carried out for multiple HIV/STI/pregnancy prevention interventions across various populations to build a better understanding of where and among what populations particular aspects of interventions should be tailored to ultimately improve the magnitude of risk reduction (Kirby, Laris, and Rolleri 2007; Bell et al. 2007; DiClemente et al. 2008; Hardnett et al. 2009).

Limitations

There are a number of limitations to the study which are worth commenting on. Measures in this study were self-report which carry with them a variety of potential inaccuracies (Upchurch et al. 2002; Brener et al. 2003). It is unknown the degree to which results may generalise beyond the study population; in particular, they may not generalise beyond young people attending alternative schools in the USA, who tend to be older (mean age at baseline for All4You2! students was 16.2) and have higher risk for educational failure. Further, statistical power in school-based studies may be limited given study resources and the necessity of using group-randomised trial designs.

The Sobel test was used in this study for testing for mediated effects; potentially more accurate methods for testing mediation and building confidence intervals (CIs) can sometimes be obtained using the actual distribution of the product or bootstrap methods (MacKinnon 2008; Hayes 2009; MacKinnon, Lockwood, and Williams 2004) because these methods do not assume that the mediated effect has a normal distribution. However, there were problems obtaining estimates using these methods through the available PRODCLIN programme¹ (MacKinnon, et al. 2007).

Finally, our analyses did not include moderators or subgroup analysis, and these should be investigated. Mediation analyses conducted across different study populations may be used to tailor interventions to target only mediators most salient to specific adolescent subgroups (e.g, defined by gender, age and race/ethnicity) for whom they are being developed (Kirby, Laris, and Rolleri 2007; Bell et al. 2007; DiClemente et al. 2008; Hardnett et al. 2009).

Conclusions

Results of this study suggest specific areas for deepening the focus of HIV/STI/pregnancy prevention interventions for alternative school youth: helping them identify and avoid exposure to risky situations and improving their self-efficacy to refuse sex. The other psychosocial variables targeted by the intervention were indeed strongly related to the targeted behavioural outcome yet proved difficult to influence themselves, which likely resulted in their not being significant mediators. The lack of findings of a broad set of psychosocial variables that acted as mediators for the intervention suggests the need for adjusting and possibly augmenting existing theories, focusing the intervention on fewer factors most related to the outcome behaviour of interest to allow deeper, more efficient instruction, and/or continuing to refine how these factors are measured.

Different causal mechanisms are beginning to emerge for similar theory-based interventions when implemented in different populations (e.g., Baumler et al. 2013). Additional mediation analyses are needed that examine different aspects of the underlying mechanisms of these types of interventions for different populations. This is critical work that may help interventionists better understand the complex processes that lead to behaviour changes and how to tailor their interventions to maximise resources and impact (Kirby, Laris, and Rolleri 2007; Bell et al. 2007; DiClemente et al. 2008; Collins, MacKinnon and Reeves 2013).