How measures of perception from survey data lead to inconsistent regression results: evidence from adolescent and peer substance use

Abstract

In studies of peer group behavior, the direct measure of peer group behavior is often not available, and so is replaced by perceptions from survey respondents. This study shows that regression estimators are inconsistent when the correctly measured independent variable of group behavior is replaced with perceived measures from survey respondents. The inconsistency is due to three sources: projection of own behavior onto the group, rescaling the marginal effect of the group, and simple random measurement error. We discuss why each effect may cause inconsistency, derive formulas for the probability limit to quantify their effects, and illustrate with three examples of adolescent smoking and drinking.

Peer Effects in UK Adolescent Substance Use: Never Mind the Classmates?

“…Krauth (2005), for example, suggests that selection accounts for around half of the overall association between own smoking and reports of friends' smoking, and if this carried over to our own estimates discussed in the following section then we would still see a large and positive association between own and friends' smoking. Norton et al (2003) suggests the association between own smoking and perceived friends' smoking is around 1.5 times larger than that between own smoking and actual friends' smoking, and again if this carried over to our own case, and we were interested in actual peer smoking rather than perceived peer smoking, we would still be left with a large and positive peer effect in smoking between friends.…”

“…that only friends influence own behaviour, using simple dichotomous measures of perceived prevalence of substance use amongst friends. Results are presented in Table 3 rows 3 and 5, and essentially provide UK estimates of peer effects along the lines of Norton et al (2003) and (the naive estimators of) Krauth (2005Krauth ( , 2007. Results where school dummies are included to capture common unobserved influences on substance use are presented in Table 3 rows 4 and 6.…”

“…Clark and Lohéac, 2007) may be perfectly consistent with large peer effects between friends (e.g. Norton et al, 2003). And studies that estimate large and statistically significant peer effects between classmates may be mostly picking up peer effects between friends, some of whom happen to be in the same school class.…”

Estimating the impacts of household behavior on youth smoking: evidence from Ontario, Canada

“…2,3 For instance, Norton et al (1998) obtain a positive and statistically significant relation between individual and peer smoking while controlling for parental smoking but are unable to control for corresponding behavior by siblings. Similarly, Gaviria and Raphael (2001), Norton et al (2003), and Krauth (2005) find a significant correlation between youth and peer smoking but do not control for family 2 Research by social scientists on the impacts of peer behavior on youth smoking participation is voluminous. For an excellent survey please refer to Kobus (2004).…”

“…8 4 Gaviria and Raphael (2001) employ a sample of tenth graders from the 1990 National Education Longitudinal Study (NELS). Norton et al (2003) use data on grade 9 students in a North Carolina study, and grade 5 and 6 students in an Illinois study. Krauth (2004) uses data from Health Canada's 1994 Youth Smoking Survey.…”