Selection of covariates used for identification when conducting control plans : A simulation study

Abstract: In the case where a causal structure between variables is known and can be represented by a directed acyclic graph, this thesis examines the procedure of conducting a control plan. A control plan is an operation where a treatment variable is set according to some function of other variables, with the goal of bringing the response variable close to a specic value, as well as to reduce its variance. Evaluating such control plans through trial and error however, can be a costly and time consuming task. Fortunately, it is possible to estimate the effects of conducting such a control plan prior to actually implementing it, through the use of observational data. Besides the variables used in the control plan, it is often necessary to account for other variables as well in order to achieve unbiased estimations. Focusing mainly on these variables, simulations will show how different sets of variables will affect the variance of the estimations, in terms of the effect of the control plan, both on the mean and on the variance of the response. Besides using normally distributed variables, an attempt will also be made to evaluate binary treatment variables. However, lacking theoretical results, no firm conclusions can be drawn.