Relationships Between Adolescent Sexual Outcomes and Exposure to Sex in Media: Robustness to Propensity-Based Analysis.

Pregnancies and sexually transmitted infections among U.S. adolescents have proven to be stubborn social and public health problems. Though progress was made between 1990 and 2005 in reducing teen pregnancy in the U.S., the rate (71.5/1,000 teen women annually, Guttmacher, 2010) remains substantially higher than in other nations (Singh & Darroch, 2000). And a recent study of a nationally representative sample of females ages 14 to 19 years found that 38% of those who were sexually active tested positive for sexually transmitted infection (Forhan et al., 2009). The persistence of these problems partly reflects the difficulty of addressing some of the factors that influence adolescent sexual risk, such as poverty and lack of opportunity (Udry & Billy, 1987; Wyatt, 1989). Although other factors are more amenable to influence (e.g., relationships with parents and community, Gavin, Catalano, David-Ferdon, Gloppen, & Markham, 2010, and attitudes, intentions, and perceived norms, Kirby, 2002), the small-group and individual-level interventions that address them reach few youth. In contrast, media depictions regularly reach vast audiences. Youth spend 7.5 hours with media each day – 10 hours and 45 minutes if one accounts for multiple media used simultaneously (Rideout, Foehr, & Roberts, 2010). Media frequently include portrayals of sex, and potential negative consequences of sexual activity and responsible behaviors like use of birth control are seldom depicted (Kunkel, Eyal, Finnerty, Biely, & Donnerstein, 2005; Pardun, L'Engle, & Brown, 2005). From a theoretical standpoint, media depictions stand alongside parents and peers as potential behavioral models for youth (Bandura, 1986). This gives media the potential to put youth at risk and makes the study of media an important area for sexual health research. A growing number of studies link sexual content in media with adolescents' attitudes and sexual activities. Among the strongest are two longitudinal studies that test for relationships between prior exposure to sexual content in the media and subsequent changes in sexual behaviors. In the first of these, Collins and colleagues (2004) surveyed a national sample of 2,002 youths ages 12 to 17 years. Participants reported how often they watched a list of television programs that varied in amount of sexual content, indicated their lifetime experience with a variety of sexual behaviors, and completed more than a dozen measures of background characteristics (e.g., religiosity, parental monitoring). They were surveyed again one year later. Results indicated that baseline virgins who saw more sex on television were more likely to initiate intercourse over the intervening year than those who saw less. Exposure to more sexual content at baseline also predicted progression to more advanced noncoital activities (e.g. from breast touching to genital activities). As social learning theory predicts, the association was specific to content that did not include portrayals of negative outcomes or responsible sexual behavior, and was mediated by self-efficacy, outcome expectancies, and perceived peer norms (Martino, Collins, Kanouse, Elliott, & Berry, 2005). Brown and colleagues (2006) expanded upon this work by linking exposure to sexual content in a broader variety of media to intercourse initiation and advances in noncoital sex. They surveyed 1,017 North Carolina youth at ages 12 to 14 years and again two years later. Exposure to sexual content in television, music, movies, and magazines predicted advancing sexual behavior after other variables were controlled statistically, but only among white youth, who comprised about half of the sample. No relationship was observed among African-American teens, who made up the rest of the sample. Steinberg and Monahan (2010) have questioned the statistical approach used by both studies, regression with covariates. After reanalyzing Brown and colleagues' (2006) data using propensity matching they concluded that the findings from both prior studies are invalid, that “Adolescents' Exposure to Sexy Media Does Not Hasten the Initiation of Sexual Intercourse.” There are several problems with Steinberg and Monahan's analyses and conclusions (Brown, in press; Collins, Martino & Elliott, in press). Briefly, there is no evidence of smaller bias in Steinberg and Monahan's (2010) analysis relative to Brown and colleagues' (2006), and there is evidence of greater variance (larger standard errors/smaller effective sample sizes). Thus, the accuracy of Steinberg and Monahan's estimates is likely less than estimates previously provided by Brown and colleagues and there is no reason to dismiss the prior findings. Nonetheless, Steinberg and Monahan (2010) raise valid concerns about selection and the possible influence of unmeasured confounders that are applicable to both prior studies, as they are to most non-experimental research. All relevant covariates must be included in a regression equation to ensure that the association between a predictor and an outcome is not spurious (Steyer, Gabler, von Davier, & Nachtigall, 2000). In contrast, propensity matching, a statistical technique that allows for a separation of the effects of propensity for exposure (selection) from actual exposure, may render it unnecessary to correct for unobserved factors that influence the outcome (Rubin, 2007), but all factors relevant to the selection of treatment/exposure must be observed and modeled. Use of propensity matching is not necessarily superior to covariate-adjusted regression (Shadish, Clark, & Steiner, 2008; Steiner, Cook, Shadish, & Clark, 2010); it is superior only when there is empirical evidence that propensity scores reduce squared bias to a greater extent than they increase variance, compared to regression approaches. Propensity matching is most useful when 1) a predictor variable cannot be randomly assigned but the factors affecting exposure to the variable can be well-modeled,. and 2) sample sizes are large enough to allow for the reduction in power/increase in variance that can be entailed. Given the difficulty of observing all factors relevant to selection of treatment and the difficulty of observing all factors relevant to an outcome, Robbins and colleagues developed an approach that combines regression with covariates and propensity scoring. Their approach lessens the necessary assumptions of both approaches and provides greater robustness to violations of these assumptions. This “doubly robust” approach, which uses propensity scores to model selection and also adjusts for covariates in the regression models predicting outcomes, assumes only that either the selection model or the regression model is correctly specified, and retains reasonable accuracy even under mild violations of its assumptions (Lunceford & Davidian, 2004; Robins, Hernan, & Brumback, 2000). Here, we use an application of this method to explore whether Collins and colleagues' 2004 findings regarding intercourse initiation and noncoital sex are robust to such potential misspecifications. We also predict pregnancy among youth who are sexually active, an outcome that was linked to television sexual content exposure in a more recent publication using data from the same survey (Chandra et al., 2008).