Читать книгу Applied Univariate, Bivariate, and Multivariate Statistics - Daniel J. Denis - Страница 16

One of the most prominent advances in social statistics is that of structural equation modeling. With SEM, as we will survey in Chapter 15, one can model complex networks of variables, both measurable and unmeasurable. Structural equation modeling is indeed one of the most complex of statistical methods in the toolkit of the social scientist. However, it is a perfectly fair and reasonable question to ask whether structural equation modeling has helped advance the cause of social science. Has it increased our knowledge of social phenomena? Advanced as the tool may be statistically, has the tool helped social science build a bigger and better house for itself?

Such a question is open to debate, one that we will not have here. What needs to be acknowledged from the outset, however, is that statistical complexity has little, if anything, to do with scientific complexity or the guarantee of scientific advance. Indeed, the two may even rarely correlate. A classic scenario is that of the graduate student running an independent‐samples t‐test on well operationally defined experimental variables, yet feeling somewhat “embarrassed” that he used such a “simple” statistical technique. In the lab next door, another graduate student is using a complex structural equation model, struggling to make the model identifiable through fixing and freeing parameters at will, yet feeling as though she is more “sophisticated” scientifically as a result of her use of a complex statistical methodology. Not the case. True, the SEM user may be more sophisticated statistically (i.e., SEM is harder to understand and implement than t‐tests), but whether her empirical project is advancing our state of knowledge more than the experimental design of the student using a t‐test cannot even begin to be evaluated based on the statistical methodology used. It must instead be based on scientific merit and the overall strength of the scientific claim. Which scientific contribution is more noteworthy? That is the essential question, not the statistical technique used. The statistics used rarely have anything to do with whether good science versus bad science was performed. Good science is good science, which at times may require statistical analysis as a tool for communicating its findings.

In fact, much of the most rigorous science often requires the most simple and elementary of statistical tools. Students of research can often become dismayed and temporarily disillusioned when they learn that complex statistical methodology, aesthetic and pleasurable on its own that it may be (i.e., SEM models can be fun to work with), still does not solve their problems. Research wise, their problems are usually those of design, controls, and coming up with good experiments, arguments, and ingenious studies. Their problems are usually not statistical at all, and in this sense, an overemphasis on statistical complexity could actually delay their progress to conjuring up innovative, ground‐breaking scientific ideas.

The cold hard facts then are that if you have poor design, weak research ideas, and messy measurement of questionable phenomena, your statistical model will provide you with anticlimactic findings, and will be nothing more than an exercise in the old adage garbage in, garbage out. Quantitative modeling, sophisticated as it has become, has not replaced the need for strict, rigorous experimental controls and good experimental design. Quantitative modeling has not made correlational research somehow more “on par” with the gold standard of experimental studies. Even with the advent of latent variable modeling strategies and methodologies such as confirmatory factor analysis and structural equation modeling, statistics does not purport to “discover,” for real, hidden variables. Modeling is simply concerned with the partitioning of variability and the estimation of parameters. Beyond that, the remainder of the job of the scientist is to know his or her craft and to design experiments and studies that enlighten and advance our knowledge of a given field. When applied to sound design and thoughtful investigatory practices, statistical modeling does partake in this enlightenment, but it does nothing to save the scientist from his or her poorly planned or executed research design. Statistical modeling, complex and enjoyable as it may be on its own, guarantees nothing.