| Summary: | The claim that early childhood television exposure causes later attention problems (Christakis et al., 2004) remains strongly held by both the popular media and many researchers in the field, despite the fact that re-analyses and meta-analyses have directly challenged this finding (Foster & Watkins, 2010; Kostyrka-Allchorne et al., 2017; Nikkelen et al., 2014). To further examine the validity of the original claim, we subjected the same dataset (the National Longitudinal Survey of Youth) to a “multiverse analysis” (Steegen et al., 2016). Because research requires a series of analytic decisions, some of which are arbitrary, any individual finding may be more or less dependent on the analysis strategy used. Thus, we employed more than 100 analytic models to see how robust the purported effect might be across a variety of analytic decisions.
As in Christakis et al. (2004), data were obtained from the National Longitudinal Survey of Youth - 1979. Our variable selection process was based on the one reported in the original study with some additions. All downloaded data, analysis codes, and supplemental figures are available anonymously at goo.gl/93uWt4. One added covariate of particular interest was children’s temperament, because difficult temperament (i.e., short attention span, negative reactivity) has been associated with both attentional flexibility (Smith et al., 1997) and concurrent TV watching (Thompson et al., 2013).
We first conducted 4 linear regressions and 36 propensity score analyses, varying analytic parameters including: the cut point for distinguishing “high TV” from “low TV” exposure; age at TV exposure [approximately 1.5 years vs. 3 years]; whether a doubly-robust analysis was used; whether the attention outcome was standardized within sex; and whether estimating the average treatment effect (ATE) or the average treatment effect for the treated (ATT). Only 4 of the 40 analyses was consistent with a causal effect of TV on attention (all four were variants of PSA using the unstandardized attention measure). Figure 1 shows a summary of the analyses using the unstandardized attention measure. Note that effect sizes cluster around zero. In an attempt to replicate the original analytic strategy, we also conducted 42 logistic regressions. The logistic regressions required dichotomizing the attention measure to discriminate problematic from non-problematic behavior. Because there were no a priori reasons to choose a particular cut point, we systematically varied the cut point from 110 to 130. As shown in Figure 2, we found that significant effects only emerge at cut points of 123 and 124, similar to the cut point of 120 chosen by Christakis et al. (2004). Given that only four analytic paths out of 82 showed a significant effect, Christakis et al.’s findings appear extremely model dependent, leading us to conclude that early exposure to TV has no real effect on later attention. Finally, we found that difficult temperament was in fact predictive of hours of TV watching in early childhood, at about the same magnitude as BMI and parent education. This finding is in line with recent investigations of temperament and screen media use (e.g., Thompson et al., 2013).
|
|---|
