What Exactly Is Random About Random Effects?

What Exactly Is Random About Random Effects?

c Metho ds of Psychological Research Online 1997, Vol.2, No.2 1998 Pabst Science Publishers Internet: http://www.pabst-publishers.de/mpr/ What exactly is random ab out random e ects? Matthias Siemer Abstract The central question of the pap er is: When should stimuli be treated as random e ects in random- or mixed-e ects ANOVA? The conceptual start- ing p oint is the view of signi cance testing in exp erimental research as fol- lowing the logic of randomization tests. Three designs are discussed: 1 Stimuli nested under treatment conditions, 2 stimuli and treatment condi- tions crossed, and 3 stimuli and treatment conditions counterbalanced. The treatment of stimuli as random e ects in the rst two designs is shown to b e inadequate but is necessary in the third design. Moreover, the analysis has implications for at least two additional topics: 1 The conceptualization of the signi cance test as a randomization test when no randomization takes place, e.g., in quasi-exp eriments and correlational studies, and 2 the validity of aggregating over sub jects and stimuli referred to as aggregation validity . Aggregation validity has to b e distinguished from internal validity and is shown to b e one asp ect of the concept formerly known as external validity, alb eit without app eal to some underlying p opulation. The present pap er starts with the following metho dological question: Under which circumstances is it reasonable to consider stimuli in exp erimental studies to b e random samples drawn from an underlying p opulation of stimuli. By implication these stimuli would have to b e treated as random factors in a mixed- or random- e ects ANOVA-mo del. The answer to this question has to consider the following basic question: What is the exact nature of the supp osed randomness of random e ects? The Discussion fo cuses on the metho dological implications and presupp o- sitions of this assumption of randomness rather than on technical problems of the random-e ects mo del, like the distributions of Quasi-F statistics. 1 Arguments for the treatment of stimuli as ran- dom e ects It is established metho dological standard in cognitive psychology to treat stimuli { e.g. words { as a random factor, just like sub jects. Indeed, cognitive psychologists often carry out two distinct statistical analyses, one with sub jects as random e ects and stimuli treated as xed e ects and the other one with stimuli as random e ects and sub jects treated as xed e ects. The rationale for this approachgoes back to Coleman 1964 and Clark 1973. For instance, Coleman stated that many studies of verbal b ehavior have little scienti c p oint if their con- clusions have to be restricted to the sp eci c language materials that were used in the exp eriment. It has not b een customary, however, to p erform signi cance tests that p ermit generalization beyond the sp eci c materials, and thus there is little evidence that such studies could be Random E ects 140 successfully replicated if a di erent sample of language materials were used 1964, p. 216, italics added . This metho dological requirement has since b een extended to other research areas e.g. Fontenelle, Phillips & Lane, 1985; Hopkins, 1984; Kenny & Smith, 1980; Richter & Seay, 1987; Santa, Miller, & Shaw, 1979; Wickens & Kepp el, 1983 and the statistical treatment of di erent exp erimenters Bonge, Schuldt, & Harp er, 1992; Scheirer & Geller, 1979. The arguments in favor of this metho dological requirement are generally not substantially di erent from those of Coleman 1964. The following paragraph from Fontenelle, Phillips, and Lane 1985 may b e taken as representing this view: In conclusion, generalizing research ndings is a ma jor aim of all exp er- imentation. Although exp erimenters as a rule are very careful to make sure that their results generalize to the p opulation of sub jects, the prob- lem of generalizing to the p opulation of stimuli had b een neglected p. 106, italics added . Before the validity of these arguments can be discussed further, it is necessary to brie y present the widely held conceptualization of inferential statistics in ex- p erimental psychology as following the logic of randomization tests rather than parametric or p opulation based statistics. 2 The signi cance test as randomization test Recentinterpretations of the signi cance test do not fo cus on it as a means to in- fer to an underlying p opulation { at least not in exp erimental contexts. Rather, the signi cance test is seen as following the logic of randomization tests cf. Bre- denkamp, 1980; Edgington, 1995; Erdfelder & Bredenkamp, 1994; Gadenne, 1984; Hager & Westermann, 1983. In randomization tests, the critical random pro cess is not to drawa random sample from an underlying p opulation distribution, but rather to randomly assign sub jects or stimuli to exp erimental conditions. This random assignment is required to secure sto chastic indep endence of treatment and potential ly confounding variables PCVs. PCVs designate all other factors that have an e ect on the dep endentvariable/s in the sense of b eing parts of sucient conditions for causal in uence on the dep endent variable/s. The latter re ects the fact that psychological hyp otheses and explanations are essentially incomplete: They entail only one or few causes as parts of a complete causal network. Moreover, they are neither necessary nor sucient in themselves but have to b e supplied by other background- conditions to b e e ective see Siemer, 1993 for a more detailed discussion of this prop erty of psychological explanations. PCVs b ecome actual ly confounding variables ACVs inasmuch as the sto chastic indep endence of PCV and treatment do es not hold. The randomization pro cedure is therefore required to secure the internal or ceteris-paribus "other things equal" validity of an ex- p eriment Hager & Westermann, 1983. If the treatment pro cedure itself induces confounding variables, randomization do es not ensure internal validity, since the confounding takes place after randomization e.g., Co ok & Campb ell, 1986. According to this rationale, a randomization test informs ab out the probability of the empirical data in terms of di erences in central tendencies, under the as- sumption that they are exclusively the result of the random assignment pro cedure, that is, chance. Most imp ortantly, the distribution of the test statistic is generated solely on the basis of the sample data. To summarize, randomization is necessary to secure the internal validityofan exp eriment that tests causal hyp otheses. Therefore, the randomization test answers c MPR{online 1997, Vol.2, No.2 1998 Pabst Science Publishers Random E ects 141 the question of how likely the results are, under the assumption that the treatment has no e ect H and all e ects are exclusively a consequence of the randomization 0 pro cedure. Consequently, the randomization test with appropriate -level protects from falsely accepting a causal hyp othesis, and consequently increases the internal validity of an exp eriment. In particular, inferential statistics do not provide the inductive generalization from a sample to a p opulation, that is, the "external validity" or exp ected probability to replicate the results of an exp eriment. Therefore, the question of interest can be rephrased as: Under which circum- stances do es the treatment of stimuli as random e ects raise the internal validityof a study? Note that in the present context ANOVA is treated as an approximation to randomization tests. This is p ossible b ecause Monte-Carlo studies have shown that randomization tests and their parametric equivalents in most cases lead to very similar results summarized in, e.g., Bredenkamp, 1980. The main reason for this approach is that ANOVA is a much more common metho d for the evaluation of exp erimental designs than randomization tests, at least for the designs considered. Therefore, from a statistical p ointof view, I will discuss random e ects or mixed ANOVA mo dels, whereas from a conceptual or metho dological p oint of view, these ANOVAs will b e treated as approximations to randomization tests. 3 Sub jects as random factors To illustrate this approach I will discuss a design that requires the treatment of subjects as random e ects { b oth at a statistical and a conceptual level: A single factor rep eated measurements design with two treatment levels. It is assumed, that sequence of the factor levels and sub jects are randomly combined. The resp ective randomization test informs ab out the probability of empirical mean di erences, given the H that these di erences are exclusively the result of the random combi- 0 nation of the sequence of the factor levels and the sub jects. That is, the probability that the e ect is the result of chance uctuations in the reactions of the sub jects across the two levels. The notion of "chance uctuation" subsumes the variation of the in uences of all PCVs at the two p oints of measurement. Consequently, the distribution of the test statistic is generated by the p ermutation of the combinations of reactions and treatment levels within sub jects e.g., Edgington, 1995. In the ANOVA evaluation of this design, sub jects are treated as random ef- fects in a mixed-mo del approach. This is re ected by the fact that the F -ratio for the exp ected mean squares of treatment variability is compared to the exp ected mean squares of the interaction of treatmentby sub jects, consisting of the variance 2 2 see Table 1.

View Full Text

Details

  • File Type
    pdf
  • Upload Time
    -
  • Content Languages
    English
  • Upload User
    Anonymous/Not logged-in
  • File Pages
    13 Page
  • File Size
    -

Download

Channel Download Status
Express Download Enable

Copyright

We respect the copyrights and intellectual property rights of all users. All uploaded documents are either original works of the uploader or authorized works of the rightful owners.

  • Not to be reproduced or distributed without explicit permission.
  • Not used for commercial purposes outside of approved use cases.
  • Not used to infringe on the rights of the original creators.
  • If you believe any content infringes your copyright, please contact us immediately.

Support

For help with questions, suggestions, or problems, please contact us