An Integrated Model of Condom Use in Sub-Saharan African Youth: A Meta-Analysis

Abstract

Objective:

We tested an integrated social cognitive model derived from multiple theories of the determinants of young peoples’ condom use in Sub-Saharan Africa. The model comprised seven social cognitive antecedents of condom use: Attitudes, norms, control, risk perceptions, barriers, intentions, and previous condom use.

Methods:

We conducted a systematic search of studies including effects between at least one model construct and intended or actual condom use in young people from sub-Saharan African countries. Fifty-five studies comprising 72 independent data sets were included and subjected to random-effects meta-analysis. Demographic and methodological variables were coded as moderators. Hypotheses of the integrated model were tested using meta-analytic structural equation modeling.

Results:

The meta-analysis revealed significant non-trivial sample-weighted correlations among most model constructs. Moderator analyses revealed differences in six correlations for studies that included a formative research component relative to studies that did not. There was little evidence of systematic moderation of relations among model constructs by other candidate moderators. Meta-analytic structural equation models revealed significant direct effects of attitudes, norms, and control on condom use intentions, and of intention, control, and barriers on condom use. Including past condom use increased explained variance in condom use intentions and behavior but did not attenuate model effects. There were also significant indirect effects of attitudes, norms, and control on condom use through intentions.

Conclusions:

Findings provide preliminary evidence to support the integrated condom use model in sub-Saharan African youth. The model provides guidance on potential targets for improving the effectiveness of condom promotion interventions.

Young people living in sub-Saharan Africa have an increased risk of sexual and reproductive health problems, including unwanted pregnancies and sexually transmitted infections (STIs) (WHO, 2016). The sub-Saharan African region has the highest rates of new human immunodeficiency virus (HIV) infections and the highest rates of unintended pregnancies in the world (Hubacher, Mavranezouli, & McGinn; Mayondi et al., 2015; UNAIDS, 2015). Similarly, the sub-Saharan African population is affected by a high prevalence of other sexually-transmitted infections such as syphilis, gonorrhea, bacterial vaginosis, trichomoniasis, and herpes simplex virus type 2 (Chico et al., 2012). Unintended pregnancies and STI incidence are substantially reduced through use of barrier contraceptives, and condom use has been identified as a key behavior in the prevention of both outcomes. However, young people living in sub-Saharan Africa engage in inconsistent or low condom use, despite repeated exposure to messages aimed at promoting condom use (Eggers, Aarø, Bos, Mathews, & de Vries, 2014; Kalolo & Kibusi, 2015; Protogerou, Flisher, & Wild, 2014). Consequently, condom promotion is a priority for public health and infection control in sub-Saharan African nations.

To date, a number of behavioral interventions aimed at promoting condom use in young people have been implemented in sub-Saharan Africa. Reviews of the efficacy of these interventions typically show favorable knowledge and attitude change, but limited or no change in condom use uptake and maintenance (Eaton, Flisher, & Aarø, 2003; Exavery et al., 2012; Protogerou & Johnson, 2014; Scott-Sheldon, Walstrom, Harrison, Kalichman, & Carey, 2013; Wamoyi et al., 2014). The poor success rate demonstrates the need for public health interventions that are based on a fundamental understanding of the determinants of condom use in this population. Interventionists have, therefore, turned to theories from behavioral science and health psychology as basis for understanding condom use in this population, and for informing the development of effective strategies to promote condom use.

Many theories applied to predict and understand condom use are based on the social cognitive tradition (e.g., Bandura, 1986; Fishbein & Ajzen, 2009). These theories adopt an information processing approach and focus on the individual, belief-based factors that affect decisions to engage in health behavior (Conner & Norman, 2015). Prominent social cognitive constructs identified as antecedents of health behavior in these theories include perceptions of severity and susceptibility to disease, and perceived benefits and barriers (e.g., protection motivation theory, Rogers, 1975; the health belief model, Rosenstock, Strecher, & Becker, 1988), attitudes, social norms, and perceptions of control (e.g., the theory of planned behavior, Ajzen, 1985; the I-change model, de Vries, Mesters, van de Steeg, & Honing, 2005; the reasoned action approach, Fishbein & Ajzen, 2009), knowledge, skills, and estimation of costs and benefits (e.g., information-motivation-behavioral skills model, Fisher & Fisher, 1992; health action process approach, Schwarzer, 2008), and self-efficacy (e.g., social cognitive theory, Bandura, 1986). Although some theories propose belief-based constructs as direct antecedents of health behavior (e.g., health belief model, social cognitive theory), many propose motivation or intentions as the most proximal predictor of behavior (Ajzen, 1985; Fishbein & Ajzen, 2009; Rogers, 1975 ). Specifically, intentions are expected to mediate effects of beliefs on health behavior.

Social cognitive theories have shown considerable promise in predicting and explaining variance in condom use and informing the development of safer-sex interventions including condom use. Reviews of research applying these theories to predict condom use, including those focusing on populations in sub-Saharan Africa, have provided cumulative evidence of the psychological antecedents of condom use (e.g., Albarracín, Johnson, Fishbein, & Muellerleile, 2001; Protogerou, Flisher, Aarø, & Mathews, 2012; Sheeran, Abraham, & Orbell, 1999). However, one of the limitations of focusing on specific theories is the potential to neglect particular constructs that may have relevance to the target behavior. In addition, there is considerable redundancy in constructs across theories, such as constructs with similar content labeled differently (Hagger, 2014). Integration of social cognitive theories may provide a solution to these problems through the development of models that are inclusive, by incorporating conceptually distinct constructs, and, parsimonious, by consolidating constructs with like content (Eggers et al., 2014; Hagger, Koch, Chatzisarantis, & Orbell, 2017; Montaño & Kasprzyk, 2008). This goal can be achieved through systematic classification of the social cognitive factors that feature in theories applied to the prediction of health behavior, including condom use, into logical categories. Such classification necessitates close examination of the content of constructs in social cognitive theories and their accompanying measures and collapsing them to arrive at a nuanced, core set of constructs. McMillan and Conner (2007) advocated a ‘core health cognitions’ approach to integration that classifies social cognitive variables from theories based on definition and content to arrive at an optimally comprehensive set of constructs derived from the theories.

The present study

Although conceptual and narrative reviews have identified the theory-based factors that relate to condom use in sub-Saharan African youth (Eaton et al., 2003; Protogerou et al., 2012; Protogerou & Hagger, 2017) there is, to date, no quantitative synthesis of single or integrated social cognitive theories focused on condom use in young people from sub-Saharan Africa. There is also very little research on the effects of extraneous factors, such as demographic, environmental, and study parameters, that may interact with psychological constructs from social cognitive theories in predicting young people’s condom use across sub-Saharan African nations. There is, therefore, a need to synthesize evidence from social cognitive theories that have been applied to explain condom use in this population.

The present study addresses this need through a meta-analytic synthesis of research of the social cognitive predictors of condom use in sub-Saharan African youth. We developed an integrated model that includes a core set of social cognitive constructs expected to be associated with condom use model (Figure 1). The model was guided by McMillan and Conner’s (2007) framework, which provides a basis to classify multiple constructs from social cognitive models applied to health behavior into a set of core constructs, along with the processes by which the constructs relate to behavior. McMillan and Conner identified five conceptually-distinct ‘core’ social cognitive constructs attitudes, self-representations, norms, control perceptions, and dispositions to act (intentions) as the core social cognitive correlates of health and risk behavior. The constructs are derived from prominent social cognitive theories applied to health behaviors. The attitudes construct encapsulates risk perceptions and perceived barriers from the health belief model (Rosenstock et al., 1988) and protection motivation theory (Rogers, 1975), and evaluations of the behavior and its outcomes from the theories of reasoned action and planned behavior (Ajzen, 1985; Ajzen & Fishbein, 1980).We chose to retain the distinction between attitudes as beliefs about the behavior, and risk perceptions and perceived barriers in our model considering the potential utility of identifying unique effects for each factor in the development of condom use interventions. The control perceptions construct encompasses beliefs relating to self-efficacy from social cognitive theory (Bandura, 1986), and perceived behavioral control from theories of reasoned action and planned behavior. Norms encompass the injunctive and descriptive norms constructs that feature prominently in social cognitive theory, the theories of reasoned action and planned behavior, the reasoned action approach, and social identity approaches (e.g., White, Terry, & Hogg, 1994). Following McMillan and Conner, we predicted that effects of these core constructs on condom use would be mediated by dispositions to act, captured by the intentions construct, consistent with previously specified models (e.g., Ajzen, 1985; Fishbein & Ajzen, 2009; Rogers, 1975). Alongside this hypothesis, we also proposed direct effects of control perceptions and barriers on condom use. This is consistent with the premise that when these perceptions equate to actual control and real barriers, they serve as ‘proxy’ measures of actual barriers and likely to predict condom use directly, bypassing intentions (Ajzen, 1985; Fishbein & Ajzen, 2009). Overall, McMillan and Conner’s framework enabled the identification of a core set of condom use antecedents, and propose testable hypotheses for the relations among them.

Figure 1.

Path diagram of the integrated condom-use model. Coefficients are standardized parameter estimates (β) with likelihood-based 95% confidence intervals in parentheses. Effects omitted from model for clarity: total effect, control→condom use, β = .246 [.096, .394]; total effect, barriers→condom use, β = −.144 [−.192, −.099].

We aimed to synthesize research on theories that include constructs from the integrated model and test their effects on condom use in sub-Saharan African youth. In addition to examining the averaged sample-weighted correlations for each relation among model constructs, our analysis also permitted testing integrated model predictions. Specifically, we tested the unique prediction of the core components on the model on condom use mediated by intentions in a meta-analytic structural equation model (Cheung, 2015). We also expected model effects to hold when controlling for past condom use, an important pre-requisite in support of the predictive validity of social cognitive models (Ouellette & Wood, 1998), and compared model effects across models that induce and exclude past condom use as a control variable (Figure 2). In addition, we tested the effects of candidate demographic (age, gender, socio-economic status, religion, geographical location) and methodological (time since publication, time lag between psychological and follow-up behavior measures, sample context, study quality, inclusion of formative research) moderators on relations among the social cognitive constructs and condom use intentions and behavior. These moderators have been identified as factors likely to magnify or diminish the relationship between constructs from social cognitive theories and condom use in young sub-Saharan African populations (Eggers et al., 2016; Protogerou & Hagger, 2017). In the event of substantive moderation, we aimed to test differences model effects using separate meta-analytic structural equation models in groups of studies representing each level of the moderator.

Figure 2.

Path diagram of the integrated condom-use model including effects of past condom use. Coefficients are standardized parameter estimates (β) with likelihood-based 95% confidence intervals in parentheses. Effects omitted from model for clarity: total effect, control→condom use, β = .115 [−.061, .280]; total effect, barriers→condom use, β = −.138 [−.252, −.022]; direct effect, past condom use→condom use, β = .433 [.225, .639]; total effect, past condom use→condom use, β = .510 [.306, .714].

Method

Selection criteria

Studies were included if they: (1) sampled young people from sub-Saharan African nations in educational (i.e., elementary, high-school, and higher education students) or non-educational settings (e.g., households, community settings); (2) were cross-sectional, prospective, or intervention-type designs; (3) used a measure of intended or actual condom use as an outcome variable; (4) provided at least one bivariate correlation between a social cognitive construct falling in to seven identified construct categories and condom use; and (5) were full-text peer-reviewed published articles and unpublished theses, written in English. Studies were excluded if they: (1) did not employ condom use as an outcome variable but employed other safer-sex or condom-use related behaviors (e.g., hormonal contraception, abstinence, delaying intercourse, condom use at first intercourse, purchasing, carrying, and negotiating condoms); (2) used composite outcome variables that included some parameter of condom use (e.g., averages of condom use and illegal substance use); (3) were duplicate versions of the original study (e.g., abstract-only report, conference presentation); and (4) were qualitative designs, government reports, and editorial/opinion pieces. There were no publication date restrictions. Following UNESCO’s (2017) conceptualization of youth as “…a period of transition from the dependence of childhood to adulthood’s independence and awareness of our interdependence as members of a community” (para. 1), and the African Youth Charter’s definition of youth as people between 15 and 35 years (African Union Commission, 2006), we included studies that had sampled people up to age 35.

Search Strategy

We conducted a search of electronic databases including Web of Science, PubMed, PsycINFO, and Google Scholar up to August 2016. In addition, we hand-searched reference lists of studies included from the database search, as well as relevant systematic reviews and meta-analyses, for additional eligible studies. Each sub-Saharan African country was a key word in the literature search, in addition to ‘Sub-Saharan Africa’, ‘West Africa’, ‘East Africa’, ‘Southern Africa’, and ‘South Africa’. We combined these words with key terms describing sexual risk-taking (‘sex’, ‘condom’, ‘HIV’, ‘AIDS’, ‘sexually transmitted disease’, ‘STD’, ‘sexually transmitted infection’, and ‘STI’). Moreover, we combined the above terms with the names of separate social cognitive theories and variations of the overarching constructs drawn from McMillan and Conner’s classification system. Study selection and reasons for exclusion are presented in a flow chart (Figure 3) based on PRISMA guidelines (Moher, Liberati, Tetzlaff, Altman, & The PRISMA Group, 2009). Two co-authors independently screened the abstracts for eligibility; then the full copies of eligible titles were screened using a priori inclusion-exclusion criteria, which resulted to the final list of included studies.

Figure 3.

Flow of studies through meta-analysis based on PRISMA guideline

Classification of constructs

Guided by our proposed model based on McMillan and Conner’s (2007) ‘core health cognitions’ framework, we identified constructs from eight social cognitive theories applied to condom use in the included studies: the theories of reasoned action and planned behavior (Ajzen, 1985; Ajzen & Fishbein, 1980); the health belief model (Rosenstock et al., 1988); social cognitive theory (Bandura, 1986); information-motivation-behavioral skills model (Fisher & Fisher, 1992); the health action process approach (Schwarzer, 1992); the AIDS risk-reduction model (Catania, Kegeles, & Coates, 1990); protection motivation theory (Rogers, 1975); and the I-change model (de Vries et al., 2005). Constructs from these theories were matched with one of the seven constructs in our model: attitudes, norms, control, risk perceptions, barriers to condom use, intentions, and previous condom use. The process required a content analysis of the measure tapping the study construct and matching it with the appropriate model construct. We allocated ‘content to category’ matching independently and compared our classifications. The classifications were the identical, which was expected given that most studies used inventories developed by the social cognitive theorists themselves. Protogerou and Hagger (2017) describe the detailed process of construct classification.

Data Extraction

As all studies included in the current analyses were correlational in design, the zero-order Pearson correlation coefficient (r) was selected as the effect size metric. Effect sizes and associated sample sizes for relations among the model constructs were extracted from the source studies. In cases where an effect of interest was tested but insufficient data were reported to compute an effect size, we requested the data from the study authors. We also approached authors if information relevant to candidate moderators was missing. In the event that zero-order correlation coefficients among variables were not available but other effect size statistics were (e.g., t-tests, χ², odds ratios), we used these to produce zero-order correlation coefficients using appropriate transformations. In cases where studies included multiple measures of the dependent variable, condom use, we used the measure that was most closely matched to the target behavior (frequency of condom use). For studies that included relationships between condom use and more than one measure of the seven core social cognitive constructs identified a priori (e.g., self-efficacy and perceived behavioral control), we produced an aggregated effect size by taking the average of the effects (Marín-Martínez & Sánchez-Meca, 1999). For studies that had adopted a prospective design and included follow-up measures of condom use behavior on multiple occasions, we used the dependent measure taken at the most distal time point to compute the effect size given that long-term follow up behavioral data are relatively rare in this literature. Of the 12 eligible prospective studies included in the analysis, only one had multiple follow-up measures of condom use.

We coded studies into groups on the following moderator variables: participant age (younger or older); gender (predominately male, predominately female, balanced, male-only, or female-only); socio-economic status (high, low, or not stated); predominant religion (Christian, other, or not stated); geographical location (peri-urban, rural, or both); sample context (school, higher education, or community); study quality (acceptable or questionable); and the inclusion of a preliminary formative research component (included formative component or did not include formative component). In addition, time since publication (defined as number of years since publication to the end of the current search period), time lag between measures of social cognitive constructs and follow-up measures of behavior (measured in weeks), and study quality were treated as continuous moderators. Testing for age as a moderator presented some challenges as there was substantive within-study variability in the age of participants. We therefore performed moderator analyses with age as a categorical variable, defined as above or below the sample median of 18.5, and as a continuous variable using the sample average. This value corresponds with the end of adolescence, when young people are expected to assume many more responsibilities (UNESCO, 2017). However, we recognize the inherent limitations of dichotomizing age as an ostensibly continuous variable, and of using the sample average as a continuous variable, which does not account for within-study variability. We also conducted moderator analysis with study quality as a categorical and continuous variable given that methodological tools have, by convention, been dichotomized into ‘acceptable’ and ‘questionable’ categories according to guideline cutoff scores, but can also be treated as a continuous variable. The extant literature has suggested that these variables have the potential to influence the relationship between social cognitive constructs and condom use in sub-Saharan African settings (Eggers et al., 2016; Protogerou & Hagger, 2017; Protogerou & Johnson, 2014). Nonetheless, sub-Saharan African studies have also found that these variables can either increase, or decrease, or leave the relationship between social cognitive constructs and condom use unaffected (e.g., Bryan, Kagee, & Broaddus, 2006; Heeren, Jemmott, Mandeya, & Tyler, 2008; Protogerou, Flisher, Wild, & Aarø, 2013). Given these inconsistencies, we made no a priori predictions about the direction of moderator effects and viewed all moderator analyses as exploratory.

Data Analysis

Random-effects meta-analysis was used to compute averaged sample-weighted correlations (r₊) among model constructs using Hedges and Vevea’s (1998) methods and SPSS macros developed by Field and Gillett (2010). The averaged sample-weighted correlations were considered non-trivial in value if they equaled or exceeded a small effect size (< .10) in Cohen’s (1992) taxonomy of effect sizes. We also conducted heterogeneity tests of the averaged correlations using Cochran’s (1952) Q and the I² statistics (Higgins & Thompson, 2002). Statistically significant Q values and I² values exceeding 25% are indicative of substantial heterogeneity in the correlations (Higgins & Thompson, 2002). We evaluated the presence of small-study bias in the averaged sample-weighted correlations by computing statistics based on plots of the correlations from each study against study precision (usually the reciprocal of the study sample size). Asymmetry in the predicted ‘funnel’ shape of the plots is considered evidence of small study bias, that is, the tendency for studies included in the analysis to exhibit large effects relative to their size. This is often taken as a potential indicator of publication bias. We used Begg and Mazumdar’s (1994) rank-order correlation to test for the interdependence of variance and effect size, with a significant correlation indicative of the presence of publication bias.

Effects of categorical moderator variables on the correlations among model constructs were tested by conducting separate meta-analyses for correlations among constructs in groups of studies defined by each level of the moderator (e.g., younger and older participants). Comparisons were made using 95% confidence intervals about the averaged sample-weighted correlations in each moderator group with a formal test provided by Welch’s t-test. Categorical moderator analyses were conducted in cases where there were at least two studies at each level of the moderator. We also conducted a series of meta-regression analyses to examine to examine effects of moderators that were continuous in format (e.g., study quality, average sample age, time lag between measures of psychological constructs and follow-up measure of condom use, time since publication) using Wilson’s (2001) Metareg macros for SPSS. As moderator variables may be related, the meta-regressions also allowed us to examine unique effects of selected categorical and continuous moderator variables on the correlations among model variables. As the numbers of studies that included a follow-up measure of condom used numbered very few (k = 12), we opted to conduct single-variable meta-regressions for the time lag moderator to maximize statistical power, but acknowledge the limitation of this analysis as it does not test the unique effect of this moderator when controlling for other moderators. In this analysis, we did not predict that time lag would moderate correlations among concurrently-measured variables. However, recognizing that effects in prospective models tend to decline over time (Gollob & Reichardt, 1987), we expected effects between psychological variables and prospectively-measured condom use would be smaller with increasing time lag. In instances where a moderator indicated a significant prediction of the effect size, we used the ‘moving constant’ technique to identify the effects size at meaningful levels of the moderator (Johnson & Huedo-Medina, 2011).

Hypothesized relations among constructs in the integrated model were tested using meta-analytic structural equation modeling using the MASEM (Cheung, 2015; Cheung & Hong, 2017) package on R (R Development Core Team, 2017). Traditional approaches to analyzing multiple relations among social cognitive variables have adopted a univariate approach, which involves subjecting matrices of averaged sample-weighted correlations among variables in the model derived from meta-analysis to a path analysis (Hagger, Chan, Protogerou, & Chatzisarantis, 2016; Hagger & Chatzisarantis, 2016). Such approaches have inherent limitations such as using the same sample size to estimate the models and treating the averaged correlation matrix as a covariance matrix, which may lead to inaccuracies in the estimated standard errors, confidence intervals, and chi-square values of the resulting models (Cheung, 2015). The meta-analytic structural equation modeling approach is a two-stage alternative that overcomes the limitations of the univariate approach. In the first stage, transformations are applied to correlation matrices from individual studies to account for study-specific random effects so that they can be analyzed as covariance matrices a structural equation model. Specifically, the analysis yields a pooled correlation matrix, which represents the estimated average correlation matrices of the population, and the associated asymptotic sampling covariance matrix, representing the precision of the estimated average correlation matrix. In the second stage, the a priori model is fitted to the covariance matrix from the first stage. Missing data are handled by use of full information maximum likelihood estimation¹.

We estimated two models in our meta-analytic structural equation modeling analysis: A model testing the hypothesized effects among study constructs as stipulated in the proposed integrated model (Figure 1), and a modified model that included past condom use as a predictor of all other constructs in the model (Figure 2). Fit of the proposed model with data from the meta-analysis was evaluated using multiple goodness-of-fit indices: the model goodness-of-fit chi-square, the comparative fit index (CFI), the Tucker-Lewis index (TLI), the standardized root mean square of the residuals, and the root mean error of approximation (RMSEA). A non-significant chi-square, CFI and TLI values that approach or exceed .95, a SRMSR value of less than .008, and a RMSEA value of .005 or less indicate good fit of the model with the data (Hu & Bentler, 1999). Variability about parameter estimates including direct and indirect effects was estimated using likelihood-based confidence intervals. Likelihood based confidence intervals have an advantage of Wald confidence intervals based on the standard errors as they capture asymmetry in the parameter distributions (Cheung, 2007, 2009). Based on guidelines proposed by Seaton, Marsh, and Craven (2010) we adjudged a .10 value for parameter estimates to be the minimum considered for the effect to be non-trivial and have “meaningful value”, with smaller values (< .075) regarded as unsubstantial, even if they achieve statistical significance. We evaluated whether inclusion of past condom use as a predictor in the model would result in an attenuation of effects of study constructs (attitudes, norms, control, risk perceptions, and barriers) on intentions, and of intentions on behavior, by comparing the confidence intervals of the parameter estimates from the model excluding past behavior with those from the model including past behavior. In cases where our moderator analyses indicated systematic effects of a moderator on relations among model constructs, we estimated the meta-analytic structural equation model in sets of studies at each level of the moderator. Heterogeneity statistics from the first stage of the model were used to evaluate whether the moderator had resolved the heterogeneity in correlations among variables. We used the standard method to test for differences in parameter estimates at each level of the moderator (Schenker & Gentleman, 2001). Specifically, we computed 95% confidence intervals about the difference in the parameter estimates of interest. Confirmation of a statistically significant difference in a parameter estimate is supported if the confidence interval does not contain zero.

Appraisal of Study Quality

Quality of included studies was appraised using the evidence-based librarianship (EBL) critical appraisal checklist (Glynn, 2006) and was included as a moderator of model effects. Study quality was evaluated on four domains: population (e.g., representativeness, appropriateness of selection criteria, response rate); data collection (e.g., clarity and validation of instruments, inclusion of instruments in report); study design (e.g., appropriateness of methodology, replicability, ethics approval); and results (e.g., clarity and accuracy of results, recommendations for future research, external validity). Each research domain is critically appraised by checking a “yes” (Y), “no” (N), “unclear” (U), or “not applicable” (N/A) next to each criterion. A total score (T) was computed by assigning a score of 1 to positive (Y) and 0 to negative (N, U, or NA) responses, expressed as a percentage, as well as separate scores for each domain. In line with the tool’s guidelines, we created a dichotomous study quality variable with studies receiving a T-score of less than 75% considered of questionable quality, while studies with a score of 75% or above were considered of acceptable quality. All included studies were scored by one co-author, and a second co-author assessed the quality of a randomly selected sub-sample of studies (k = 15).

Results

Study Characteristics

Fifty-five studies, comprising 72 independent data sets (N = 55,069), were included in the meta-analysis. A list of included studies is provided in Appendix A (supplemental materials) and study characteristics are summarized in Appendix B (supplemental materials). Studies appeared between 1992 and 2016 as published journal articles (k = 50, 91%) or postgraduate dissertations (k = 5, 9%). Most studies employed a cross-sectional design (k = 43, 75%) and reported a preliminary formative research phase (k = 34, 64%). Average sample age ranged between 12.1 and 26.5 (M = 19, SD = 3.43). Male participants were over-represented, with 23 studies (42%) including predominately-male samples (male composition > 50%), and 13 studies (24 %) including male-only samples. Thirteen sub-Saharan African nations were represented (Botswana, Cameroon, Ethiopia, Ghana, Guinea, Kenya, Namibia, Nigeria, Rwanda, South Africa, Tanzania, Uganda, and Zimbabwe), with a large number of studies being conducted in South Africa (k = 16, 29%). About half studies were conducted in (peri)urban (k = 30, 54%) and school settings (k = 27, 49%).

Appraisal of Study Quality²

The majority of studies (k = 45, 88%) received a T-score of < 75% indicating questionable overall quality. In terms of the separate domains of the EBL checklist, studies exhibited highest quality scores in the results validity domain (M = 65.45, SD = 25.37), with most studies reporting results clearly (k = 47, 85%), accurately (k = 42, 76%), and completely (k = 48, 87%). Studies received the lowest quality scores in the sample validity domain (M = 42.75, SD = 22.59), revealing potential selection/sampling (k = 38, 69%) and representativeness (k = 25, 45.5%) biases. Inspection of individual checklist items revealed that the strongest study domain was the reporting of results, with 48 studies reporting full disclosure of findings (87.3%). The weakest study domain was replicability with only ten studies (18%) providing sufficient methodological detail to allow replication. Inter-rater reliability analysis indicated substantial agreement in study quality ratings across all items (mean Cohen’s κ = .73, 95% CI = .67, .80) and total study quality scores (intraclass correlation = .949, p < .001) for the randomly selected sub-sample of studies.

Sample-Weighted Correlations

Averaged sample-weighted correlations among model constructs in the meta-analysis are presented in Table 1, along with confidence intervals, heterogeneity tests, and publication bias statistics³. The averaged correlations among all but seven of the constructs were statistically significant and of sufficient size to be considered non-trivial⁴. However, correlations between barriers and all other constructs, between attitudes and risk perceptions, and between control and condom use, were not statistically significant, with confidence intervals that included zero. Levels of heterogeneity ranged from small to moderate in the current sample according to Q and I² statistics.

Table 1.

Averaged Sample-Weighted Correlation Coefficients (r₊) and Heterogeneity Statistics for Effects Among Constructs from the Integrated Condom Use Model

Relationshipa	k	N	r+	CI95		τ	Q	df	I2	τk
				LL	UL
Attitudes-Norms	29	26935	.35	.28	.42	.04	53.82**	28	47.97	−0.19
Attitudes-Control	27	28135	.31	.20	.42	.10	33.00	26	21.19	0.06
Attitudes-Risk perceptions	16	10862	.07	−.06	.20	.07	16.35	15	8.26	−0.19
Attitudes-Barriers	7	3015	.04	−.11	.19	.04	12.54	6	52.16	0.15
Attitudes-Intentions	39	30408	.38	.31	.45	.06	63.98**	38	40.60	0.03
Attitudes-Condom use	8	3668	.20	.13	.26	.01	7.78	7	10.04	0.05
Attitudes-Past condom use	30	32101	.20	.15	.25	.17	36.74	29	21.08	0.13
Norms-Control	28	27035	.36	.26	.45	.08	38.60	27	30.05	−0.19
Norms-Risk perceptions	7	4040	.28	.18	.38	.01	7.59	6	20.95	−0.71*
Norms-Barriers	2	302	.17	−.06	.38	.02	1.00	1	0.00	−1.00**
Norms-Intentions	39	30730	.40	.32	.47	.07	57.84*	38	34.29	−0.05
Norms-Condom use	8	3668	.21	.14	.28	.01	30.84***	7	8.19	0.28
Norms-Past condom use	27	29283	.22	.13	.31	.06	20.18	26	0.00	0.10
Control-Risk perceptions	9	6401	.30	.13	.45	.07	9.95	8	19.58	−0.05
Control-Barriers	2	302	.07	−.04	.18	.00	0.72	1	0.00	−1.00
Control-Intentions	37	30539	.39	.31	.47	.09	59.71**	36	39.71	−0.25*
Control-Condom use	8	3668	.17	−.04	.37	.09	6.09	7	0.00	0.57*
Control-Past condom use	34	40046	.27	.19	.34	.05	32.49	34	0.00	0.09
Risk perceptions-Intentions	17	7517	.24	.12	.36	.07	17.77	16	9.99	−0.19
Risk perceptions-Condom useb	1	1006	.08	‒	‒	‒	‒	‒	‒	‒
Risk perceptions-Past condom use	15	12472	.05	.00	.10	.01	27.88*	14	49.79	0.07
Barriers-Intentions	5	1305	.05	−.23	.32	.01	3.27	4	0.00	0.00
Barriers-Past condom use	6	3041	.03	−.18	.24	.07	8.18	5	38.89	−0.07
Intentions-Condom use	9	3988	.28	.20	.35	.01	16.83*	8	0.00	0.28
Intentions-Past condom use	15	23363	.31	.18	.43	.07	20.03	14	30.11	0.10
Condom use-Past condom use	5	2050	.49	.21	.70	.13	3.84	4	0.00	−0.20

Note. r₊ = Sample-weighted average correlations; k = Number of studies; N = total sample size; CI₉₅ = 95% confidence intervals for sample-weighted averaged correlation; LL = Lower limit of 95% confidence interval; UL = Upper limit of 95% confidence interval; τ = Estimated variance in population (Fisher-Transformed correlation); Q = Cochran’s (1952) Q homogeneity statistic; df = Degrees of freedom for the Q statistic; I² = Higgins and Thompson’s (2002) I² statistic; τ_k = Begg and Mazumdar’s (1994) ranked correlation statistic based on Kendall’s τ.

^aEffect sizes for the relationships between risk perceptions and barriers and between barriers and condom use are omitted as no studies in the current sample tested these effects;

^bOnly one study in the current sample provided a test of the relationship between risk perceptions and condom use which precluded a meta-analytic synthesis, the reported effect size is raw effect size reported in the study.

^*p < .05

^**p < .01

^***p < .001

Moderator Analyses

Results of moderator analyses are presented in Appendixes C, D, and E (supplemental materials). We found relatively few effects of the moderators on model relations for moderator analysis with categorical moderator variables (Appendix C). The formative research moderator had the most pervasive effect with six of the 25 relations⁵ among the integrated model variables demonstrating significant differences. Specifically, the effects of attitudes on risk perceptions, intentions, and past condom use, the effects of norms and control on risk perceptions, and the effects of intentions on condom use, were stronger for studies that had included a formative research component, compared with studies that had not. In contrast, we found very few moderator effects for the participant age, gender, socio-economic status, predominant religion, geographical location, sample context, and study quality moderator groups.

Focusing on the meta-regression analyses (Appendixes D and E), the single-moderator analyses revealed significant effects of time lag on study effect size for the relations between attitudes and norms (β = .758, p < .001) and intentions (β = .541, p = .020), and between intentions and norms (β = .626, p = .006) (Appendix D). Examining the correlations for different values of time lag using the moving constant technique indicated stronger, statistically significant correlations for each relationship when the time lag was shorter (Appendix E). In the multiple-moderator meta-regression analyses, in which study effect sizes were regressed on continuous (age, study quality score, time since publication) and dichotomous categorical (formative research) moderators, the effects of formative research was most pervasive (Appendix D). Specifically, the effects of attitudes on norms (β = −.472, p = .014), risk perceptions (β = −.661, p = .004), intentions (β = −.399, p = .016), and past condom use (β = −.411, p = .035) were all stronger in studies that included formative research, corroborating the categorical moderator analyses for this variable. In addition, we found significant effects of age for the attitude-norms (β = −.573, p = .001), attitudes-risk perceptions (β = −.642, p < .001), norms control (β = −.483, p = .012), and intentions-past condom use (β = −.911, p < .001) relationships. There was also a statistically significant effect of study quality on the attitude-risk perceptions relationship (β = −.937, p = .025). Moving constant analyses revealed that effects were generally stronger at higher levels of the age moderator, with the exception of the intentions-past condom use relationship, which was smaller with increased age. The attitude-risk perceptions relationship was larger and more negative in higher quality studies (Appendix E).

Meta-Analytic Structural Equation Models⁶

Standardized parameter estimates and likelihood-based confidence intervals for the meta-analytic structural equation model of the integrated condom use model excluding past behavior, and the model including past condom use are presented in Figures 1 and 2, respectively. In addition, indirect effects, confidence intervals of parameter estimates, and tests of difference in path coefficients across the models including and excluding past condom are provided in Appendix F (supplemental materials). The model excluding past condom use exhibited good fit with the data (χ² = 8.212, df = 3, p = .042; CFI = .991; TLI = .934; SRMSR = .036; RMSEA = .005). Model parameter estimates revealed statistically significant, non-trivial direct and positive effects of attitudes, norms, and control on condom use intentions, while effects for risk perceptions and barriers were not significant and trivial in size. We also found statistically significant, non-trivial direct and positive effects of intention and control on condom use, and negative effects of barriers on condom use⁷. In addition, the analysis yielded statistically significant indirect effects of attitudes, norms, and control on condom use mediated by intentions⁸. Indirect effects of risk perceptions and barriers on condom use were not statistically significant. The model accounted for statistically significant non-zero proportions of the variance in intentions (24.88%) and condom use (14.73%). Overall, results supported the hypothesized pattern of effects in the proposed model. The model including past condom use as a predictor also exhibited good fit with the data (χ² = 3.728, df = 3, p = .292; CFI = .999; TLI = .991; SRMSR = .022; RMSEA = .002). The inclusion of past condom use resulted in some observed reductions in the magnitude of the effects of intention and control on condom use. However, these reductions were not statistically significant according to the standard method (Schenker & Gentleman, 2001; Appendix F). The inclusion of past behavior also resulted in no significant differences in the magnitude of the indirect effects of attitudes, norms, and control on behavior mediated by intention. There were non-trivial, statistically significant effects of past condom use on all model variables and statistically significant total indirect effect of past condom use on condom use mediated by the social cognitive variables. Inclusion of past condom use as a predictor in the model resulted in a minor increase in the proportion of variance explained in condom use intentions (26.39%), and a much larger increase in variance explained in behavior (29.71%), reflecting the substantive direct effect of past condom use on future condom use.

Given the pervasive effect of the formative research moderator, we examined whether the pattern of relations in the integrated condom use model varied in groups of studies that included formative research and those that did not. We therefore aimed to conduct separate meta-analytic structural equation models in groups of studies determined by the formative research moderator and compare the magnitude of the effects⁹. It is important to note that several of the model relations were not tested in the group of studies without a formative research component, which meant that several of the cells in the pooled correlation matrix were empty, precluding estimation of the model in this group. However, there were sufficient data to estimate the model in the group of studies including formative research, so we conducted a sensitivity analysis to evaluate whether our conclusions regarding the pattern of effects in our model were affected by excluding studies that did not include a formative research component. The model for the sample of studies excluding studies without a formative research component exhibited satisfactory fit with the data (χ² = 3.940, df = 3, p = .268; CFI = .999; TLI = .989; SRMSR = .031; RMSEA = .003)¹⁰. Full model results appear in Appendix G (supplemental materials)¹¹. There were no significant differences in the pattern of effects in the model estimated in the full sample of studies and the model estimated in the sample of studies excluding studies without a formative research component according to the standard method (Schenker & Gentleman, 2001; Appendix G). As we found little evidence of systematic moderation of relations among study variables for any of the other moderator variables, we did not conduct additional moderator or sensitivity analyses of model effects for any of the other moderators.

Publication Bias

Begg and Mazumdar’s (1994) rank correlation test suggested little evidence of systematic publication bias across effects among model constructs. The test revealed potential bias in three of the 25 averaged sample-weighted correlations: norms–risk perceptions (τ (k = 7) = −.714, p = .024), control–intentions (τ (k = 37) = −.251, p = .029), and control–condom use (τ (k = 8) = .571, p = .048).

Discussion

The purpose of the current research was to develop an integrated model of the determinants of young peoples’ condom use in sub-Saharan Africa, meta-analyze studies adopting model constructs to predict condom use in this population, and test the model using meta-analytic structural equation modeling. We also aimed to assess the impact of candidate moderators on the obtained relationships among theoretical constructs and between theoretical constructs and condom use. The model was developed from research applying social cognitive theories to condom use with model constructs classified along the lines of McMillan and Conner’s (2007) ‘core health cognitions’ approach. Attitudes, norms, control, risk perceptions, barriers to condom use, intentions, and previous condom use, were included in the proposed model, with intentions serving to mediate the effects of the core constructs on condom use. We identified studies including correlations among these constructs in a comprehensive database search and subjected them to random-effects meta-analysis. Studies were also coded into groups on the following moderator variables: age, gender, socio-economic status, religion, geographical location, time from publication, sample context, study quality, follow-up time lag, and inclusion of formative research. We found statistically significant averaged sample-weighted correlations among the majority of the model constructs (i.e., all but seven). Testing the proposed model using meta-analytic structural equation modeling revealed significant, direct, non-trivial, positive effects of attitudes, norms, and control on condom use intentions. We also found significant, positive, non-trivial effects of intention and control on condom use, and a significant negative effect of barriers on condom use. In addition, we found significant indirect effects of attitudes, norms, and control on condom use, mediated by intentions. Inclusion of past behavior in the model resulted in some observed attenuation of effects in the model, particularly the effect of control on condom use, but there were no significant differences. Past behavior also accounted for a substantive proportion of the variance in condom use. Moderator analyses revealed few differences on study variables across moderator groups. Of the moderators, the inclusion of formative research had the most pervasive effect on model relations. A sensitivity analysis using meta-analytic structural equation model was conducted to test whether exclusion of studies that did not include a formative research component altered our conclusions with respect to model effects. The analysis revealed that proposed model effects hold with no substantive variation in effect sizes as a result of excluding studies without formative research.

Implications for the Integrated Model

The integrated condom-use model accounted for substantial variance in condom use intentions and actual behavior in sub-Saharan African youth. Attitudes, norms, control, and barriers to condom use were prominent predictors, while risk perceptions had weak effects. Results are broadly consistent with research that has applied social cognitive theories such as protection motivation theory and the theory of planned behavior to condom use in other populations (e.g., Albarracín et al., 2001; Bengel, Belz-merk, & Farin, 1996; Sheeran & Orbell, 1998), although the sizes of the effects, particularly the intention-condom use relationship is smaller in the current study. It is important to note that effect sizes of attitudes, norms, and control factors on behavior through intentions were similar in magnitude. Norms may have an important role in this context, and similar findings have been reported in research applying social cognitive theories such as the theory of planned behavior and the reasoned action approach to condom use and other risk behaviors (McEachan, Conner, Taylor, & Lawton, 2012; McEachan et al., 2016; Sheeran et al., 1999). Given the equal, additive effects of the three components, interventions that target all three sets of beliefs may be optimally effective in promoting condom use. Therefore, messages that promote the advantages of condom use and downplay drawbacks, promote obligation to sexual partners and significant others, and enhance confidence in using condoms may be the most effective strategies to enhance condom use. Consistent with the premise of the theory of planned behavior (Ajzen, 1985) that control-related perceptions and barriers will directly predict behavior to the extent that they align with actual constraints on behavior, the direct negative effect of barriers on condom use in the present analysis, and the lack of an indirect effect through intentions, suggests that perceived barriers to condom use (e.g., perceived stigma, beliefs in reduced pleasure or effectiveness, religious beliefs) may be less influential than actual barriers (e.g., access to condoms and refusal of partners to use condoms). These findings also support a basic premise of the health belief model (Rosenstock et al., 1988), which specifies direct effects of beliefs about barriers on health behavior. Reducing these barriers may also be an appropriate avenue for intervention. The non-significant, trivial effects of risk perceptions is consistent with extant integrated models that have accommodated this factor alongside beliefs about behavior (e.g., Barg et al., 2012; Hattar, Pal, & Hagger, 2016; Maher & Conroy, 2016). It seems that beliefs about benefits and costs, significant others’ influence, and personal capacity are more salient than beliefs relating to risk. In the context to condom use this is extremely pertinent as many campaigns and educational programs aimed at condom use have focused on raising individuals’ beliefs about risk including highlighting perceptions of vulnerability to STIs and the severity of these conditions (e.g.,Harvey, Stuart, & Swan, 2000). Such approaches may be less effective compared to targeting personal beliefs on condom use.

The inclusion of past condom use in the model led to observed attenuation of some of the effects in the model and a strong direct effect of past condom use on subsequent condom use. Reduction of the effect of intention on behavior by past behavior is consistent with other tests of social cognitive models in health behaviors in primary studies and meta-analyses (Hagger & Chatzisarantis, 2016), including research examining condom use (Albarracín et al., 2001). However, the reduction in the size of the effects in the current analysis with the inclusion of past condom use was relatively trivial and not statistically significant. Inclusion of past condom use in the model, however, accounted for a substantive proportion of additional variance in condom use. Importantly, there were statistically significant direct and indirect effects of past condom use on future condom use through the model constructs. According to Ouellette and Wood (1998), effects of past behavior may serve two functions. Direct effects of past behavior on subsequent behavior may model habitual effects, effectively behavioral stability, and represent the non-conscious, automatic processes by which behavior is enacted. Indirect effects through intentions and other social cognitive variables may reflect deliberative, rational decision-making processes. Given that the direct effect was much larger than the indirect effect of past condom use on condom use, current findings suggest that condom use in sub-Saharan African youth has a strong habitual component. One interpretation of this pattern of effects is that the decision to use condoms, or not, is predominantly determined by previous experience and past habits. It also implies that intentional processes might play a less important role. Of course, the significant, non-trivial effects of intentions on condom use mean that effects of intentions are not negligible, and intervening to change intentions remain a viable prospect, however, their impact may have less effect in the face of strong habits. In such cases, it may be necessary to use intervention strategies aimed at promoting habitual use of condoms, or means to break habitual non-use. Such strategies include cue-identification and management and self-monitoring. Getting individuals to recognize the potential cues or prompts to unwanted behaviors and manage them is one strategy to circumvent habitual action. Similarly, flagging cues to carry or ensure availability of condoms (e.g., going to a social gathering, visiting a potential partner), and also situations where one is likely to negotiate their use with a prospective partner, and dealing with potential negative responses, may assist in preventing habitual non-use. It is also possible that the direct effect of past behavior on condom use is due to shared method variance between the two behavioral assessments. This may have been a possibility given that condom use was measured exclusively by self-report. However, the time lag between assessments may have mitigated such effects.

Effects of Moderators

Overall, few of the averaged sample-weighted correlations among model constructs were affected by moderator variables. Results provide little evidence to indicative of a systematic pattern of moderation. Heterogeneity in effect sizes was low to moderate in most cases, which may explain why candidate moderators did not substantially impact construct relationships. The only exception to this pattern was the formative research moderator. Six of the effects were stronger in studies that included a formative research component, in which study measures were developed and piloted in advance, relation to studies that did not include formative research. Studies that included formative research in the development of their study measures, particularly ensuring the content of items were relevant to the sample (e.g., by identifying salient beliefs) and improving the correspondence between the measures and the target behavior, are likely to have improved the precision of the study measures and, as a consequence, reduced method variance. These findings suggest that the inclusion of formative research is an important methodological step in developing measures when testing social cognitive models in condom use research in sub-Saharan Africa youth.

It is important to note that these differences did not translate to variations in effect sizes when testing the full model. Our sensitivity analyses did not indicate any variation in model effects when studies without a formative research component were excluded. Given that the moderator analysis affected relatively few relations the integrated model, it is likely that the moderator effects were insufficiently pervasive to have a substantive effect on the full network of constructs in the proposed model. Overall, these data provide preliminary evidence that exclusion a formative research component is unlikely to have meaningful effect on model tests and is unlikely to affect conclusions drawn. However, we were unable to test the model in a set of studies that did not include formative research due to insufficient studies testing some of the model relations. This precluded a comparison of models for mutually-exclusive sets of studies that included and did not include a formative research component.

Strengths, Limitations, and Avenues for Future Research

The current study is the first to develop and test an integrated condom use model in sub-Saharan African youth based on a meta-analytic synthesis of studies testing social cognitive models in this context. Our model and research synthesis makes a number of important contributions. Consistent with the advocacy of integrated theoretical approaches to provide efficient and comprehensive means to explain behavior (Hagger et al., 2017; Hamilton, Kirkpatrick, Rebar, & Hagger, 2017; Montaño & Kasprzyk, 2008), we developed our model by integrated constructs from multiple social cognitive models applied to condom use following McMillan and Conner’s (2007) ‘core cognitions’ framework. This endeavor reduced redundancy across multiple constructs, increased parsimony in predictors, and identified the social cognitive predictors that may be optimally effective in predicting condom use and associated processes. Our analysis also demonstrated the unique effects of the integrated model constructs on condom use intentions and behavior across studies included in the analysis using meta-analytic structural equation modeling. Finally, we tested the effects of a number of candidate moderators of effects among the integrated model constructs.

We should also acknowledge the limitations of the current analysis, many of which are related to shortcomings in the included studies. For example, the majority of studies were classified as of questionable quality based on scores on our study quality toolkit. This notwithstanding, study quality did not systematically influence any of the averaged sample-weighted correlations among model variables. Nevertheless, a separate methodological artifact, the inclusion of formative research, did have an effect on some of the study relations, indicating the importance of developing appropriate measures that are likely to capture constructs of interest with greater precision. A related issue is the relatively small number of studies in some levels of the moderator variables. In some cases, moderator groups sample sizes included fewer than five studies, which likely reduces the precision of the estimate and increases the sensitivity of the effect size to errant effects. A related problem was insufficient data to test the effect of the formative research moderator on relations among constructs the integrated model using meta-analytic structural equation modeling. In addition, in order to estimate the structural equation models, we had to complete two empty cells in the set of correlation matrices with data from primary research on condom use in other contexts. While we took care in filling these cells with data from research on condom use in closely-related populations, current findings should be interpreted with this caveat in mind. As the number of studies in the field increases, future studies may be able to test the model on complete sets of correlation matrices and test effects of moderators with greater accuracy. Although a number of the current studies were prospective in design and reported including a follow-up measure of condom use, a further limitation of the current studies was the preponderance of cross-sectional correlational data. Such data provide no basis on which to infer causal relations among model constructs, and the direction of effects is inferred from theory alone. This means that other statistically plausible models that fit the current data could be found, even if the pattern of effect may be theoretically contraindicated. We therefore advocate future research that aims to manipulate some of the key predictors of condom use intentions and behavior in sub-Saharan African youth and examine the effects on behavioral outcomes. A further issue is that variation in time lag between measures of the psychological variables and prospectively-measured condom use across studies may also have added additional method variance to the sets of relations. Although our time lag moderator analysis did not indicate that such variation was linked to variation in effect sizes, it remains a potential source of variance. As the literature expands, the probability of more studies with a greater variation in lag times increases, so examining time lag as a moderator should be a consideration in future analyses. Finally, our model is limited in that it focuses exclusively on social cognitive predictors and an individualist approach. The model is, therefore, silent on other influences such as implicit, relational, societal, and structural factors (Johnson, Scott-Sheldon, & Carey, 2010), which have been shown to impact condom use (Protogerou & Hagger, 2017). Related to this, although we found non-trivial, statistically significant effects among key model constructs, the effects were relatively small. Finally, a substantive proportion of the variance in condom use intentions and behavior remained unexplained. Future theoretical and empirical work should seek to incorporate additional variables into the model and evaluate the extent to which they add to the prediction of condom use in this population.

Conclusion

Our integrated model provides cumulative evidence of a core set of social cognitive determinants of condom use intentions and behavior among sub-Saharan African youth derived from multiple models and theories. The model is the first to demonstrate the key predictors of condom use based on a synthesis of research in this population and context, and also identified salient predictors through a systematic synthesis of constructs and measures across social cognitive models that have been applied in this context. We anticipate that our model may provide a basis for future research examining the predictors, and further corroboration of the model predictions in primary research is needed. The model may also assist in the development of interventions, particularly those that may assist in breaking habits. We also expect our findings to inform future research, particularly the need to conduct formative research in developing measures, and the need for more experimental and intervention research aimed at manipulating key constructs in the model. We also view the model as flexible and modifiable, and we look to future high quality tests of its premises to provide further data on how the model may be modified to improve its predictive validity.