In that case I'd argue that your use of path analysis was essentially exploratory in nature and thus would carry with it the same sort of caveats as any exploratory analysis whether conducted using regression (e.g. step-wise regression) or any other exploratory analysis.
Ned Dochtermann -- Ned Dochtermann Department of Biology University of Nevada, Reno 775-784-6781 [email protected] www.unr.nevada.edu/~dochterm/ -- Quoting William Silvert <[email protected]>: > It is a long time since I used any kind of standard statistical analysis, > including Path Analysis, so I may be a bit rusty on this, but I do not think > that the inference of causality is necessarily based on a priori > assumptions. In one case (Silvert, William. 1982. Top-down modeling of > multispecies fisheries. In "Multispecies Approaches to Fisheries Management > Advice," M. C. Mercer, ed. Can. Spec. Publ. Fish. Aquat. Sci. 59:24-27.) > where I was investigating the role of environmental factors in fisheries I > tested a model based on the hypothesis that both effort and temperature > affected fish catches. But something seemed wrong, and when I tackled the > problem with Path Analysis I came up with the surprising result that catch > determined effort, and not the other way around. This seemed like nonsense, > but I happened to meet the authors of the data report I was using, and they > were greatly upset that I was analysing their data. Finally they confessed > that they had been told to collect effort data, but since they were unable > to do this they made up the numbers by assuming that whatever fish a boat > caught most of must have been what they were fishing for. Thus Path Analysis > led me kicking and screaming to the right result - catch really did > determine effort! > > Bill Silvert > > ----- Original Message ----- > From: "Ned Dochtermann" <[email protected]> > To: "William Silvert" <[email protected]> > Cc: <[email protected]> > Sent: sábado, 5 de Dezembro de 2009 20:04 > Subject: Re: [ECOLOG-L] an example of a false correlation > > > > It is interesting that so many researchers bring up the issue of > > "correlation > > doesn't demonstrate causality" and yet are more than happy to argue > > causation > > based on regression, which usually relies on the same sort of data as used > > for > > standard correlational analyses. > > > > Regression analyses are not typically used for experimental data and thus > > are > > asserting causation based on correlation. The practical difference between > > regression and correlation is the a priori hypothesizing of causation (how > > many > > regressions are really conducted post hoc after examining scatter plots is > > a > > whole other can of worms). The exact same statistical results in terms of > > "F's", "P's" and R^2 will be obtained if you switch which variables are > > dependent and independent in a linear analysis (of course intercepts and > > coefficients would differ). > > > > I mention this in relation to the below post because when discussing path > > analysis with a colleague awhile back the comment was made that > > "correlation=/=causation" to which my reply was "that's what you do with > > regression". Although path analysis does use correlation matrices (or > > covariance matrices) in its analysis I consider it to be more akin to > > multiple > > regression as you're testing a priori causal hypotheses. > > > > Next time someone asserts correlation doesn't demonstrate causation ask > > them > > when the last time was that they used regression. The ensuing argument can > > be > > quite amusing. > > > > Ned Dochtermann > > > > -- > > Ned Dochtermann > > Department of Biology > > University of Nevada, Reno > > > > 775-784-6781 > > [email protected] > > www.unr.nevada.edu/~dochterm/ > > -- > > > > > > > > Quoting William Silvert <[email protected]>: > > > >> Since I think that Mao Tse Tung and Pol Pot were heavy smokers, to say > >> nothing of Fidel Castro, I guess that where Malcolm is going is to show > >> that > >> by selecting your data you can obtain any correlation you want. > >> > >> The point that correlation alone does not prove causality is one that > >> statisticians are always making, but mention should be made of Path > >> Analysis, a technique based on multiple correlation, that is effective in > >> establishing causal pathways. It is used extensively in the social > >> sciences, > >> but only rarely in ecology. I think that it deserves more attention, I > >> have > >> certainly found it of value. > >> > >> However a benthic ecologist I know who used Path Analysis died young, so > >> although that is only one data point, perhaps Path Analysis should be > >> avoided! > >> > >> Bill Silvert > >> > >> ----- Original Message ----- > >> From: "malcolm McCallum" <[email protected]> > >> To: <[email protected]> > >> Sent: sexta-feira, 4 de Dezembro de 2009 18:13 > >> Subject: [ECOLOG-L] an example of a false correlation > >> > >> > >> > Anyone who teaches stats might be interested in this false correlation > >> > I > >> > just stumbled on. > >> > > >> > 1) Stalin did not smoke > >> > 2) Hitler did not smoke > >> > 3) Mussolini (sp?) did not smoke. > >> > 4) Roosevelt smoked > >> > 5) Churchill smoked > >> > > >> > I guess you can see where I'm going with that! > >> > :) > >> > > >> > -- > >> > Malcolm L. McCallum > >> >
