Conference Paper BHPS-2005 Conference: the 2005 British Household Panel Survey Research Conference, 30 June -2 July 2005, Colchester, UK
Assessing hidden bias in the estimation of causal effect in longitudinal data by using a matching estimator with Rosenbaum bounds
Some social, economic, and behavioral outcomes are affected by natural experiments, or treatments that take place not by design but by natural occurrences. Treatment effects in natural experimental settings are difficult to assess because there may be potential selection bias generated by hidden confounders and because counterfactuals are unobservable, regardless of how many control variables the analyst includes in the model. While panel data in general allow the researcher to better gauge causal effects, they are not immune to possible hidden bias. This paper attempts to assess such potential hidden bias of teenage births, a natural “treatment” event, on an outcome variable such as the mother’s postnatal mental wellbeing. When there is no hidden bias, propensity score matching provides an estimate of the treatment effect on an outcome variable. However, matching methods are not robust against hidden bias arising from unobserved variables simultaneously affecting assignment to treatment and the outcome variable. A viable strategy to deal with the problem is the Rosenbaum bounds approach, which provides a “worst-case” scenario. The bounds convey important information about the level of uncertainty contained in matching estimators by showing just how large the influence of a hidden confounder must be to undermine the conclusions of a matching analysis. The first 10 waves of the British Household Panel Survey data are used to illustrate the bounds of the “treatment” effect of teenage motherhood on postnatal mental wellbeing.