Hi! Path Analysis or Structural Equation Modeling (SEM) is used to test the fit of a hypothetical model with your empirical data. It is used not only to assess relationship between two or more variables, but it enables you to build complex models built up from your research variables, and to test whether this is a valid (good fitting) model. E.g. you can look for mediating variables and different pathways through which variables effect each other. AMOS is a software running under SPSS, and it is a very useful graphic interface to build models. I have found this website of David A. Kenny (http://davidakenny.net/cm/causalm.htm) very useful for directions.
Thanks Andras! I am going to find further directions form the site you posted above but you have already provided me with the basic description of that method.
Hello Rajendra, the model construction is not an authomatic process. It must be guided by the researcher. The estimation algorithm to be applied depends on type of data. For example, maximum likelihood could be optimal if data are normally distributed. A set of goodness-of-fit indices are also available in the most used SEM programs (AMOS, LISREL, EQS, M-PLUS, R, etc.). You could read the paper of Hu and Bentler (1999) entitled: Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. It is published in the journal "Structural Equation Modeling", volume 6, issue 1, pages 1-55.
No one yet has introduced the fact that Path Analysis requires both exploratory and confirmatory factor analyses. Because such analyses are completely open to the beliefs and biases of the researcher, even with the addition of SEM, Path Analyses are primarily interpretative rather than definitive analyses. The path diagrams that are the main "result" in PA are notoriously hard to interpret, much less to understand. The American Psych Assoc has two books on reading and understanding mulivarate stats that are written very simply, and PA in one of them. That all said, I am sure that there is a less onerous and difficult to understand analysis than PA for your data. I never gotten excited about a Path Analsis, and I'm sure I never will.
Path analysis is one of the "families" of statistical techniques that fall under the umbrella of Structural Equation Modelling (SEM) framework. Put it simple, it is as if you run simultaneous multiple regressions among and across observed variables that you postulate have a certain causal relationship and you want to fit the data to the model.
The direction (postive or negative), the magnitude of the relationships (size of the effects, path coefficients/regression coefficients), and the interpretation of the results depend on various factors, starting from the type of variables you have at hand [i.e. on the level of measurement (binary, categorical, interval or ratio scale, etc.), observed or latent], the sample you have, the assumptions you make, the type of algorithm you use, etc.
Of course, as always, everything should start from a research question/problem. Then comes the choice of the appropriate test/research methodology to answer your questions.
In which research issues and for what research purposes path analysis could prove to be more useful? =>
Again, it depends on the research questions you have/type of problem you have to solve and on what alternative technique you have in mind.
I would build upon the second part of Marcus’s answer (above) and reinforce the importance of not only having a clear research question, but also ensuring that the research model (i.e., the overall path diagram that enjoins the various exogenous and endogenous variables) being tested is grounded in (strong) theory. It is all too easy to go on ‘fishing expeditions’ when doing this kind of modelling; that is, adding new pathways here and there to improve the overall ‘fit’ of the model (see Albert’s post about goodness-of-fit indices). At risk of making a too-obvious comment, investing serious time developing a robust conceptual and theoretical understanding of the phenomena under investigation before performing the path analyses will be time very well spent.
It can be done the other way too. Once you find strong paths, you can test them on various known theories or relationships that are found in literature.Of course one has to start with some theoretical background and constructs and not fire in drak.But the model certainly can be used for exploration purposes also and not only for confrimation of a theoretical framework.
Path analysis is a type of Structural Equation Modeling (SEM). The SEM includes several things but two of its applications are very common - the Confirmatory Factor Analysis and Path Analysis. In SEM terminology the former is called testing of "measurement model" and the later "structural relationship model". The path analysis is also sometimes referred as Causal Modeling as we test a specific pattern of relationship among variables in which some are assumed to be the cause of the other(s). However, this label is a misnomer as we can not establish cause - effect relationship in true sense using path analysis.
In simple terms, the Path analysis involves testing a theoretically or empirically determined specific pattern of relationship among a set of variables. Say, we are working with four variables A, B, C, and D. Then using path analysis we can test the relationship among these variables as defined by some theory or hypothesis based on prior experience or empirical evidences. For instance, if any theory suggest that A is a variable that may cause changes in both B and C, and B and C are the variables responsible for change in D. It is also assumed that A has some direct effect on D also. Then we can specify the relationship among the said variables as stated above and test it using some software based on structural equation modeling or even regression analysis based software or macros. The said model may be displayed visually also in which we'll link A to B and C with an arrow pointing to B and C (the arrows represent the path from A to B and C and suggest that A is leading to B and C). After this, we'll link B and C to D but now the arrows will originate from B and C and will point towards D. Finally, we;ll add A to D with arrow pointing to D to represent the direct relationship of A and D. This pattern of relationship among the said variables represent the nature of relationship as suggested by the earlier theory. Now we'll test this model of relationship using specific software.
The results of this testing has two important components. The first is the overall fit of the model. This is assessed using the available fit indices such as Chi Square test, chi square and df ratio, the GFI, CFI, TLI, RMSEA etc. For a model to be called good fit, the Chi square should be non significant (but for large samples it may be significant and may not a pose problem if other indices suggest good fit), the chi square to df ratio should be less than 5, the GFI, CFI, TLI, should be .95 or higher (but a value above .9 also suggest an adequate fit) and the RMSEA value should be less than .05.
The second important component of the results involve testing the significance of the paths and this should be done when there is an evidence that the overall model is fit. This includes testing of significance of direct effects (paths) and indirect effects (if it is in the model). In the above example, there are five direct paths ( A to B, A to C, B to D, C to D and A to D) and two indirect paths (A to D through B and A to D through C). If indirect paths are found to be significant then one can infer that the relationship between A and D is mediated by B and C. If only one indirect path (say A to D through B) is found significant and the other is non-significant then well infer that the relationship between A and D is mediated by B but not by C. suppose that the direct path from A to D is also significant then one can infer that A has some significant direct effect on D but also has some indirect effect on D through B and C, In other words, the relationship between A and D is partially mediated by B and C. similarly, the other paths also need to be interpreted.
To have more confidence in the tested model, one should also test the alternative relationship among the variables and if these alternative models do not fit or have a poor fit than the one that was hypothesized then one may say with more confidence that hypothesized causal relationship is more likely to exist.
I am sorry but SEM does not consider causal effect.It cant.that researcher has to decide form collateral KNOWLEDGE And theory .Please correct me if I am wrong
In addition to SEM being able to establish causal relations from associations, it is essential to note that SEM is not equipped to handle nonlinear causal relationships