I agree. Models that explicitly formulate sets of mechanisms acting under certain conditions can provide a set of testable predictions. Bill provides one example, from a practical standpoint, why this can be important...what data should be collected (and what data doesn't need to be collected). Also, not explicitly stated, is that such models also inform us how to analyse the data. For instance, a model (not a statistical one) might predict a particular slope for the relationship between two variables (not just that variable x or y is different between populations i and j). Whether or not we find this slope tells us whether the model was valid or not in that particular situation. At this point, we now decide whether to continue using this model, to modify it, or to drop it in favor of some other model.
Now a second point, which will probably be fairly contentious...if data are not collected in light of a scientific model, there is a good chance that they will not be very informative for the progression of science. Though examples do exist, Darwin etc., where haphazardly-collected descriptive data provided the impetus for a highly influential theory, these moments are rare. This is not to say that descriptive studies should be abandoned--often the data collected in a descriptive study are no different than those collected with respect to a model. The difference is, though, that if you have a formulated hypothesis in mind before you collect your data, then your statistical test of the data informs you (with some level of confidence) whether your hypothesis is valid or not (or whether you can lend support to one hypothesis out of a group of several), thus progressing scientific understanding. If you only collect the data (without a model), and find a pattern in it, you can raise the possibility of a hypothesis, but you cannot formally test it (with the same data). And from my experience, most folks stop with the description of the pattern without formally testing the questions that the data have now raised. Thus, I would (in a very strict sense) argue that descriptive studies, on their own, are not necessarily scientific, but can be an important part of a scientific research program, i.e., descriptive studies represent the observational/discovery stage of scientific research, and should be used to formulate and test hypotheses. Lastly, I do think that Ned raises a good point in that, often, the sampling methods used to collect data often do not match the statistical tests that are employed to provide inference. Mike Sears Assistant Professor Department of Zoology Southern Illinois University Carbondale, IL 62901 email: [EMAIL PROTECTED] web: http://equinox.unr.edu/homepage/msears Quoting Bill Silvert <[EMAIL PROTECTED]>: > Ned misses an important point, that statistical models don't give you any > idea what to measure. They simply tell you how to do what you are planning > to do anyway in a way that might give statistically meaningful results. They > > do not have any underlying natural structure. > > When I referred to doing the modelling first I was referring to models that > > actually describe the system and have some scientific basis. > > As an example of what I mean, I was once invited to develop a model of > aquaculture impacts after several years of data had been collected. I began > > the workshop by asking about the nitrogen fluxes, since previous studies had > > shown that these were the most critical variables and would be a key element > > of any model. After a long pause I was informed that nitrogen had not been > measured because no one thought it was important (actually, they didn't have > > the right equipment). If they had built a simple model first we might have > had some useful data to work with. > > And of course I am not ruling out the possibility that the data might > contradict the model. That is fine. That is how science develops. > > So I assure Ned that I am not talking about statistics, but science. > > Bill Silvert > > > ----- Original Message ----- > From: "Ned Dochtermann" <[EMAIL PROTECTED]> > To: "'Bill Silvert'" <[EMAIL PROTECTED]>; <[email protected]> > Sent: Thursday, March 09, 2006 4:23 PM > Subject: RE: "Hamerstrom science" (deliberate non-use of statistical > analysis) > > > > Generally however those concerned, after the fact, about the rigor of > > their > > statistics (or lack thereof) are not reporting naturalistic observations > > but > > attempting to hammer their round data pegs into the square holes of > > already > > established theory. > > > > If your concern is naturalistic observation, you don't have to have too > > much > > concern about whether or not you've properly articulated (or understand) > > the > > underlying statistical model you're testing. > > > > Ned Dochtermann > ------------------------------------------------- This mail sent through https://webmail.unr.edu
