Poor Instruments Lead to Poor Inferences: Comment on Roberts, Glymour, and Koenen (2013)

Abstract

Roberts, Glymour, and Koenen (2013) reported results of an empirical study concerning the nature of the association between childhood sexual abuse and adult homosexuality. Sexual contact with adults (including childhood sexual abuse) is especially common among young people who will become homosexual adults; the association is robust and worthy ofempirical attention. A number of hypotheses, some of which posit that childhood sexual abuse causes adult homosexuality and some of which do not, are consistent with the correlation. The study by Roberts et al. attempted to test the hypothesis that childhood sexual abuse is a cause of adult homosexuality. The study had an admirably large and representative sample, and it employed the sophisticated statistical methodology of instrumental variables regression, which can under some circumstances provide good evidence for causation using correlational data. Indeed, Robertsetal. concludedfromtheiranalyses that‘‘childhoodmaltreatment,’’including sexual abuse, causes adult homosexuality, with the effect of sexual abuse being especially strong for males. Unfortunately, Roberts et al. haveapplied instrumental variables regression inappropriately, and their analysis cannot plausiblyprovidegoodevidence thatchildhoodsexualabusecauses adult homosexuality. The primary defect of their analysis concerns their choice of instruments: presence of a stepparent, poverty, parental alcohol abuse, and parental mental illness. Roberts et al. claimed that these instruments are equivalent to ‘‘natural experiments’’and assert that they may be used in instrumental variables analyses because they are‘‘not known to be influenced by or to directly influence nascent sexual orientation’’(p.162). Belief in the reasonablenessof thisequation—of presence of a stepparent, poverty, parental alcohol abuse, and parental mental illness with‘‘natural experiments’’—is certainly a prerequisite for having any faith in the analyses provided by Roberts et al. But it is not a reasonable equation. The gist of the critique that we elaborate herein is that thinking of variables, suchaspresenceofastepparent,poverty,parentalalcoholabuse, and parental mental illness as comprising natural experiments violates the common understanding of‘‘natural experiment,’’is inconsistent with some facts already known about sexual orientation, and contrasts sharply and unfortunatelywith the most admired and widely accepted applications of the instrumental variables approach. Not only do Roberts et al.’s results fail to provide support for the idea that childhood maltreatment causes adult homosexuality, the pattern of differences between males and females is opposite what should be expected based on better evidence, including evidence that relies on what are much more plausibly considered‘‘natural experiments.’’