I am running a series of path analyses using AMOS and could desperately use some help around how to control for confounding demographic variables;
Q1. I have read that while some demographic variables may correlate with the outcome and predictor variables, if this effect size is small it may not be worth controlling for these effects. Is this true? and what effect size would be the cut off?
Q2. What are the specific steps to adding confounding variables into my model? My understanding is that you add the variable as another exogenous variable and add single-direction arrows point at the endogenous variables, and then bi-directional arrows pointing at all other exogenous variables. Is this correct? (I have also seen some mention of controlling for confounds by adding them as moderators instead?)
Q3. What happens when the confounding variable is nominal? E.g. Employment status (1 = full time, 2 = part time, 3 = unemployed etc..). What are the steps in converting this to a variable that can be added into a model?
I have tried to find resources to address these questions to no avail and would very much appreciate any assistance or recommended reading.
Hi Lisa,
beyond mere predictive modeling (like regression), there is no "theory-less" procedure to incorporate controls in the causal model. Control variables are those variables that--based on your / the literature's assumptions--play a causal role in the modelled system such that omission would lead to a bias of the focal effects. This will happen if the respective control variable is embedded in a "backdoor path" that creates a non-causal, spurious covariance between IV and DV. The essential characteristic of such a path is that it starts with an arrow pointing at the IV and it ends with an arrow pointing to the DV. In the most straightforward way, this is the case for simple confounders that affect the IV as well as the DV.
This definition also saves you from the most fundamental mistakes, namely accidently controlling for mediators (for which the first characteristic fails) and colliders (for which both characteristics fail).
Long story short, you have to create a path diagram that expresses not only your beliefs about the causal interplay but also with regard to the role of the control variable. The causal estimates of the model represent this essential conditional nature (if the model with its structure of relevant variables is correct THEN the estimates represent the respective causal impact).
There is a nice browser software called dagitty.net with which you can draw path diagrams and decide which variables have to be controlled. Often, control variables themselves have some causal interconnection which implies that controlling only some of them (or even one) is sufficient.
See
Vahratian, A., Siega-Riz, A. M., Savitz, D. A., & Zhang, J. (2005). Maternal pre-pregnancy overweight and obesity and the risk of cesarean delivery in nulliparous women. Annals of Epidemiology, 15(7), 467-474. doi:10.1016/j.annepidem.2005.02.005
Vanderweele, T. J. (2019). Principles of confounder selection. European Journal of Epidemiology, 3, 211-219. doi:10.1007/s10654-019-00494-6
Elwert, F. (2013). Graphical causal models. In S. L. Morgan (Ed.), Handbook of causal analysis for social research. (pp. 245-273). Dordrecht Heidelberg New York London: Springer.
Best,
Holger
Mostly the socio-demographic variables are used as moderating variables of the model.
Hi again,
I just saw that I did not respond to your specific questions
Q1: Again, if your theory says that a variable should be controlled, do it. Don't judge based on any empirical stats (correlations, effects) whether to control it. Sample estimates are just error-prone estimates and--especially in low-N situations, power may be reduced. The essence of setting up the model (with controls) deals with issues on the population level and on the sample level *on average across all sample*. This is all you can achieve as a scientist: Being correct on average.
Q2: There are no steps. Just set up the structure as you think it is. Normally this will mean that you incorporate controls with effects on all variables. This has the positive result that confounding effects are reduced but at the same time, there is nothing testable when doing so. Hence, be sure that your overall model has still some "holes in the chees"
Q3: When your control variables are categorical, proceed as in any other dummy situation, that is, create k-1 dummy variables and use these.
Best,
Holger
Holger Steinmetz Yes this makes sense. Thank you very much for your help
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling