“The Future of Longitudinal Studies:
What we know; What we don’t know; What we need to know”

Inferring Causality from Longitudinal Studies
Chaired By: Elizabeth Owens, University Of California, Berkeley
Friday, March 21, 2003

Diana Baumrind
"When are Causal Inferences Justified in the Debate about Physical Discipline “Effects”?"
University of California, Berkeley

Causal inferences referring to “effects” of physical discipline are not justified from the primary studies synthesized in Gershoff’s (2002) meta-analytic review because the great majority, even when longitudinal, were correlational and failed to meet the elementary requisite conditions for making causal inferences from correlational data. In the opinion of Overton, Editor of the SRCD Monographs, data from most social science research, including experimental research, do not justify causal terms such as “influence”, “affect”, or “the risk factor caused the behavior.” However, public policy application of such research too often presumes it can support causal inferences. For example, if divorce or physical punishment is not at least a potential causal risk factor for the detrimental child attributes with which it is correlated, then social policy recommendations based on such findings, even from longitudinal research, would be rash.

Although correlational evidence can never establish cause, such evidence can support a principled argument that one characteristic such as “spanking” is a plausible causal component in producing another characteristic such as aggression, with which it is correlated, provided that at least four conditions are all met: 1) the phenomenon corresponding to “spanking” must be bracketed to exclude very severe levels of physical punishment; otherwise the correlation will merely reflect abusive levels, 2) temporality must be established by controlling for the baseline “outcome” variable in a prospective longitudinal design, 3) plausible rival methodological and substantive hypotheses must be rigorously tested by statistically controlling for plausible “third” variables, 4) reliable and valid measures of putative dependent, independent and third variables must be employed. In addition, ancillary desirable conditions include an effect size sufficient to rule out the alternative hypothesis that the association is due to a weak unmeasured confounder, and consistency in most contexts and populations.

Unfortunately most correlational studies, even when longitudinal, do not meet these criteria. For example, Gershoff’s meta-analysis which inferred a causal relation between physical discipline and various negative outcomes, did not meet these criteria well; yet social policy implications were quickly and widely drawn.

In order to more adequately test whether there exists evidence to support a principled argument that physical discipline causes the child attributes with which it is linked, Baumrind and Owens, using archival data from Baumrind’s Family Socialization and Developmental Competence longitudinal program of research conducted a study of children ages 4, 9 and 14 years, and their parents (based on 50 hours of observation and interviews) which met all the requisite and desirable criteria previously summarized. For example, parents who used physical punishment abusively (“red zone” families) were separated from those whose use was normative in frequency and intensity; a measure of initial child misbehavior was partialed out; a plausible third variable, namely a reliable measures of consistent, responsive discipline was covaried out; and reliable and valid measures of parent and child measures obtained from independent sources were employed. Before the “red zone” families were removed the results looked similar to those of Gershoff. However, once these very high risk families were removed and plausible third variables covaried out correlations were close to zero. Furthermore, the link between a measure of verbal discipline was at least as large as that of physical discipline and in addition physical discipline added no significant variance to that of the parent typology.

In sum, there was no evidence to support inferences from the correlational studies in Gershoff’s meta-analysis to suggest that mild to moderate spanking is associated with negative outcomes. Our research illustrates the importance of testing plausible rival hypotheses and keeping a very good longitudinal data set such that potentially relevant third variables pertaining to alternative hypotheses can be entered into statistical models first, and intervention selection biases can be ruled out (e.g., spanking may be a “marker” for other negative relational factors—not the “cause” of negative outcomes). Our research also highlights the importance of having reliable and valid measures of dependent, independent, and third variables—so that it may be possible to statistically control for these in prospective designs.