causal inferences in path analysis

[lee lime]

Hello,

I have a really basic question about causal inferences in path analysis.

I do have panel data with two waves. My variables of interest are not measured in both two waves.

For example, let's say I am interested in paths like X-> Y -> Z.

To my knowledge, if I want to perform cross-lagged path analysis, X, Y, Z should be measured for each wave.

However, in my data, X is demographics which are assumed not to be changed over time and Y is an item in Wave 1, and Z is an item in Wave 2. But none of X,Y,Z are measured in both wave 1 and wave 2.

I am not sure my explanation is clear enough, but I wonder if I am able to do any causal inferences in this case, or just simple path analysis is fine.

[Keith Markus]

Lime,

Like a number of other inferences, with causal inference you are always free to close the gap between the information in your data and the causal conclusion with assumptions.  Indeed, unless you somehow have direct evidence of a causal connection, there is always a gap to be filled with assumptions.  If you compare your situation to data from a natural experiment, your gap is comparatively larger.  However, if you compare your situation to a fully cross-sectional data set with the same variables, then your gap is comparatively smaller.  The time ordering of your variables allows you to dismiss some rival hypotheses as less plausible (unless you think that the future might cause the past in the system you are modeling).

It is important to distinguish two broad categories of causal inference.  In Holland's phrase: there is inference regarding the causes of a known effect and there is inference regarding the effects of a known cause.  Perhaps a little less opaquely, there are situations in which we do not know what causes what and want to infer what the causes of some variable are.  There are other situations in which we have a pretty confident grasp of what causes what and just want to estimate effect sizes.  The latter type of situation imports more assumptions, and can thus get by with less data.

Without knowing the details of what kind of causal inference you are interested in, my first reaction is that you should be less concerned about missing time points for the variables in your data and more concerned about relevant variables that are not represented in your data at all.

Keith

------------------------
Keith A. Markus
John Jay College of Criminal Justice, CUNY
http://jjcweb.jjay.cuny.edu/kmarkus
Frontiers of Test Validity Theory: Measurement, Causation and Meaning.
http://www.routledge.com/books/details/9781841692203/