From: "Wendy K. Olsen" <W.K.Olsen-AT-Bradford.ac.uk> Subject: Re: BHA: Regression paper; problem with example Date: Thu, 24 Sep 1998 14:11:25 +0100 (GMT Daylight Time) Dear Everyone, This is my first message to the list although I've been reading the list selectively for months. I couldn't attend the conference - too bad for me! Doug Porpora, you said that the regression equation would tell you whether race was appearing to have an effect in addition to (or controlling for) economic factors. I enclose at the bottom your statement as QUOTE. But you ignore the crucial problem that the simple regression cannot discern between the concurrent effect of 'race' and an 'economic' factor when and if they rise and fall together. (In other words, if many persons are both of one race and have low values for the measured economic variable, and many persons of another 'race' have high values, regression will say 'race adds nothing to the explanation by economic factors'.) This is the problem highlighted in Lawson's book. There are two constructive responses to this problem. One is to get some new data that hits more directly the possible racism of bank employees -- opinion data, person-wise loan approval data, or something. Another is to use the data in a way that have multiple dependent variables. Although imperfect, path analysis (using two, three, or more regressions) may help utilise evidence in a more complex way than single regressions ever can do. (I am assuming that we have cross-sectional, not time-series data, since I too have just about given up hope on econometrics -- which is almost defined as the statistical analysis of time series data.) Path analysis is described in a working paper I've been developing; it has its defects. But one thing it allows you to do is distinguish operationalised causal mechanisms working at three or four units of analysis, as they affect a final outcome which is likely to be noticeably influenced by some of the underlying factors. E.g. path analysis could have variables at several levels, like: INDIVIDUals a)in bank BANK e.g. policies b) in household e.g. gender, age, earnings HOUSEHOLDS e.g. Hh Income REGION e.g. neighbourhood facilities atmosphere no. of bankers' visiting etc. -----> LOAN REQUEST e.g. application details proposal risk analysis etc. ---> LOAN DECISION. I've shown these objects moving to the right on my screen, as we move toward the final outcome; intermediate factors may be very important to the final outcome. A major one in your example is that 'race', or perceptions of racism, may affect the odds of a person putting in a loan application. Self-selection then gives the bankers a biased sample. This affects the likelihood (and we need to study the bias as an empirical question) of a particular 'race'(!)'s loan application getting approved. Path analysis is promising, and if anyone on the list wants to explore the use of CR ideas to hone path analysis for heuristic scientific research, get in touch. QUOTE FROM DP: In my talk, I offered some examples where multiple regression analyses applied to current history have politically important implications. One example concerned whether American banks were red-lining -- refusing loans to neighborhoods on the basis of race. The banks claimed they heeded only economic variables . . . We can compare the importance of neighborhood racial characteristics with neighborhood economic characteristics, each controlling for the other. If we see that in the recent past, race made an important difference even controlling for economic factors, that undermines the banks' claims that they were not redlining. Does the regression equation applied to this situation represent a universal law? No. Does it tell us anything about what will be in the future? Hopefully not. If the regression equation had its intended political effect, then a subsequent regression equation ought to yield very different results. On Thu, 24 Sep 1998 00:17:19 -0400 Doug Porpora <porporad-AT-duvm.ocs.drexel.edu> wrote: > Hi Everyone, > > Tony, I've given some thought to the question you posed to me. First, you > asked me to clarify which of the following positions i am defending: > > >1) regression analysis does not require a closure for the "parameter > >estimates", etc., to be meaningful/interpretable. > > > >2) regression analysis only requires a closure within the sample > >period or domain. > > You further ask, > > >If 1) I would be interested in how this conclusion is reached; if > >2) I would be interested in wether you think this will get us far in > >social world. > > These questions are wonderful because they are far more difficult than they > appear. I probably need more time to think them through, but since I only > asked for a day, here is what I have come up with. I think in a way I want > to defend both (1) and (2). I think that in some important disputes, we > have the closure you describe in (2) and that this is all we need. But I > also think that closure is not even necessary for regression to be > applicable -- the claim you label (1). > > Let me deal with (2) first. I think regression is often one piece of > evidence that can be adduced to resolve debate about a given > sociohistorical context. I stopped myself from saying something in my talk > that I am not sure I am ready to stand by, but let me float it here: > > In one sense, the past is a closed system: Those mechanisms that operated > operated and no others did or ever will in the past. (I am assuming that a > closure can encompass more than one mechanism as long as the operative > mechanisms are isolated from all others.) > > If the past is a closed system, then if we want to know which mechanisms > actually operated in a given sociohistorical context and if we want to know > their relative importance, then regression -- and particularly multiple > regression -- offers one valid piece of evidence. > > In my talk, I offered some examples where multiple regression analyses > applied to current history have politically important implications. One > example concerned whether American banks were red-lining -- refusing loans > to neighborhoods on the basis of race. The banks claimed they heeded only > economic variables such as a neighborhood's median income or housing value. > Well, if we look at the past, what's done has been done, and a multiple > regression equation can tell us the relative importance of the different > factors involved. > > We can compare the importance of neighborhood racial characteristics with > neighborhood economic characteristics, each controlling for the other. If > we see that in the recent past, race made an important difference even > controlling for economic factors, that undermines the banks' claims that > they were not redlining. > > Does the regression equation applied to this situation represent a > universal law? No. Does it tell us anything about what will be in the > future? Hopefully not. If the regression equation had its intended > political effect, then a subsequent regression equation ought to yield very > different results. > > Another example I cited was from the work of Christopher Chase-Dunn and > Denny Braun, who both use multiple regression on the recent past to show > that multinational penetration of third world countries has deleterious > effects on development and inequality, controlling for rival explanations. > Again, the effort here is to determine what has been the case so far in a > very specific historical context. The regression equations are not > themselves the explanation for what happened but only evidence for an > explanation. They are certainly not meant to be taken as universal laws, > applicable anywhere, anytime. > > Arguably, if the past can be treated as a closed system, then regression > equations applied to the past are premised on closed systems. I think that > the use of regression in cases like this does address some politically > important issues and in that sense does get us as far as anything will. > > Do systems have to be completely closed, however, for regression analysis > to be valid? No. As I understand regression, the error term can accomodate > a porous system where the extraneous factors at least approximate > randomness in their effect. So I do not think that regression requires a > perfectly closed system. Of course, more closure is necessary to make more > of regression than an evidentiary tool, to make it into the basis for > nomothetic explanation. I don't think that obtains. > > The above is my general line at the moment, but I confess I am not entirely > satisfied with what I have said. I confess Tony's questions require more > than the day I requested to think this through. Let me hasten to add that > I was an early critic of the use to which sociologists typically put > regression analysis -- to search for timeless, contextless laws for > nomothetic explanation. I don't support that. I also don't think it makes > much sense to attempt to fix with any precision the exact beta weights > associated with economic variables as if they will remain stable. And much > of what goes on in econometrics sounds even worse than what has tended to > go on in sociology. But as with postmodernism, I think there is something > of a baby there that should not be tossed out with the volume of bathwater. > > Well, it's 11:30 p.m. and my mind has just suddenly stopped. I welcome any > and all feedback from anyone. Thanks, Tony, for asking some questions that > are much more difficult than they seem. I will keep thinking about it. > > doug > > p.s. Tony, the private copy I just sent you was inadvertant. I meant to > send to the list. > > doug porpora > dept of psych and sociology > drexel university > phila pa 19104 > USA > > porporad-AT-duvm.ocs.drexel.edu > > > > > --- from list bhaskar-AT-lists.village.virginia.edu --- ---------------------- Wendy K. Olsen w.k.olsen-AT-bradford.ac.uk --- from list bhaskar-AT-lists.village.virginia.edu ---
Display software: ArchTracker © Malgosia Askanas, 2000-2005