Winter 2006 dsfsdfsdfsdfsdfdfsdffgfd (i.e., 0 mg, 5 mg, and 0 mg!. "owever, if t#e resear$#er wants to ma%e inferen$es &e'ond t#e (arti$ular values of t#e $g#dfdgfgdfgsfgsdfgsdfgdffffffffff%%%%%dfg#dg#&le used in t#e stud', a random effe$ts model is used. ) $ommon e*am(le would &e t#e use of (u&li$ art wor%s re(resenting low, moderate, and $v&$&$&$v&$v& Fixed and Random Coefficients in Multilevel Regression +#e random vs. fi*ed d#gdfgdfgdfgdfgdfgdfgdf%and effe$ts is im(ortant in multilevel regression. ,n multilevel regression models, &ot# level- $v$v&$v&v$& to var' randoml' a$ross grou(s. .e$ause of t#e assum(tions a&out t#eir error distri&utions, we $all t#eir varian$es, oo=var(U0j) and 11=var(U1j), /random $oeffi$ients.0 +#is means t#at we $an t#in% a&out 0j and 1j as a%in to t#e random variables , des$ri&ed a&ove. ,nstead of attem(ting to generali1e &e'ond t#e (arti$ular values of t#e inde(endent varia&le, we are attem(ting to generali1e &e'ond t#e (arti$ular grou(s in t#e stud'. 2or instan$e, we ma' #ave 50 $om(anies, &ut we wis# to generali1e to a larger universe of $om(anies w#en we e*amine t#e means (inter$e(ts! or t#e *-' relations#i( (slo(es!. ,n "3M, varian$es for t#e inter$e(ts and slo(es are estimated &' default. +#at is, w#en 'ou first $onstru$t a model, U0j and U1j are estimated &' default. "owever, t#e resear$#er #as t#e $#oi$e of setting U0j and U1j to &e 1ero. .' setting t#em to 1ero, we are testing a model in w#i$# we assume 0j and 1j do not var' randoml' a$ross grou(s. ,n fa$t, t#eir varian$e is assumed to &e 1ero, so t#e' are assumed to &e $onstant or /nonvar'ing0 a$ross grou(s. 2or e*am(le, fi*ed, nonvar'ing inter$e(ts would im(l' t#e grou( average for t#e de(endent varia&le is assumed to &e e4ual in ea$# grou(. )lt#oug# we t'(i$all' refer to t#is $onstraint as /fi*ing t#e inter$e(ts0 or /fi*ing t#e slo(es,0 t#e term is somew#at loosel' a((lied, &e$ause we are reall' assuming t#e' are fi*ed and nonvarying. Intercept Model and Random Effects ANOVA ,n t#e sim(lest "3M model, t#ere are no (redi$tors. We #ave one level- e4uation and one level-2 e4uation5 0 ij j ij Y R = + w#ere Y is t#e de(endent varia&le, 0j is t#e inter$e(t, and Rij is t#e residual or error. 0j $an &e t#oug#t of as t#e mean of ea$# grou(. +#e level-2 e4uation also #as no (redi$tors in it6s sim(lest form5 0 00 0 j j U = + w#ere 0j is t#e de(endent varia&le, 00 is t#e level-2 inter$e(t, and U 0j is t#e level-2 error. ,n t#is e4uation, 00 re(resents t#e grand mean or t#e mean of t#e inter$e(ts. U0j re(resents t#e deviation of ea$# mean from t#e grand mean. W#en t#e average deviation is large, t#ere are large grou( differen$es. ,t $an &e s#own t#at t#is is reall' t#e )N78) model. )N78) e*amines t#e deviations of grou( means from t#e grand mean. .e$ause we assume t#at t#e grou( means, re(resented &' 0j and, t#us, t#eir deviations are randoml' var'ing, t#is model is e4uivalent to t#e random effects )N78) model. We $an rewrite our two "3M e4uations as a single e4uation, if we (lug in t#e level-2 terms into level- e4uation. 00 0 ij j ij Y U R = + + ,n t#is e4uation, 00 re(resents t#e grand mean, t#e U0j term re(resents t#e deviations from t#e mean (or mean differen$es or treatment effe$t, if 'ou would li%e!, and Rij re(resents t#e wit#in-grou( error or variation. ,n more mat#emati$all' oriented )N78) te*t&oo%s, 'ou will see an )N78) model written somet#ing li%e5 Newsom 2 USP 656 Multilevel Regression Winter 2006 ij j ij Y e = + + ,f t#e model is an random effe$ts )N78), j is assumed to #ave a random distri&ution. ,n multilevel regression, &e$ause we assume U0j to var' randoml', t#e sim(le "3M model wit# no level- or level-2 (redi$tors is e4uivalent to t#e random effe$ts )N78) model. W#en we add (redi$tors to t#e level- e4uation, t#e' are $ovariates and t#e model &e$omes a random effe$ts )N978) in w#i$# t#e means are ad:usted for t#e $ovariate. Newsom 3 USP 656 Multilevel Regression Winter 2006 Summary a!le Random vs" Fixed #efinition Example $se in Multilevel Regression Varia!les Random variable5 (! is assumed to &e measured wit# measurement error. +#e s$ores are a fun$tion of a true s$ore and random error; (2! t#e values $ome from and are intended to generali1e to a mu$# larger (o(ulation of (ossi&le values wit# a $ertain (ro&a&ilit' distri&ution (e.g., normal distri&ution!; (<! t#e num&er of values in t#e stud' is small relative to t#e values of t#e varia&le as it a((ears in t#e (o(ulation it is drawn from. Fixed variable5 (! assumed to &e measured wit#out measurement error; (2! desired generali1ation to (o(ulation or ot#er studies is to t#e same values; (<! t#e varia&le used in t#e stud' $ontains all or most of t#e varia&le6s values in t#e (o(ulation. ,t is im(ortant to distinguis# &etween a varia&le t#at is varying and a varia&le t#at is random. ) fi*ed varia&le $an #ave different values, it is not ne$essaril' invariant (e4ual! a$ross grou(s. Random variable: (#otogra(#s re(resenting individuals wit# differing levels of attra$tiveness mani(ulated in an e*(eriment, $ensus tra$%s Fixed variable: gender, ra$e, or intervention vs. $ontrol grou(. Predi$tor varia&les in M3R generall' assumed to &e fi*ed Effects Random effect5 (! different statisti$al model of regression or )N78) model w#i$# assumes t#at an inde(endent varia&le is random; (2! generall' used if t#e levels of t#e inde(endent varia&le are t#oug#t to &e a small su&set of t#e (ossi&le values w#i$# one wis#es to generali1e to; (<! will (ro&a&l' (rodu$e larger standard errors (less (owerful!. Fixed effect5 (! statisti$al model t'(i$all' used in regression and )N78) assuming inde(endent varia&le is fi*ed; (2! generali1ation of t#e results a((l' to similar values of inde(endent varia&le in t#e (o(ulation or in ot#er studies; (<! will (ro&a&l' (rodu$e smaller standard errors (more (owerful!. Random effect5 random effe$ts )N78), random effe$ts regression Fixed effect5 fi*ed effe$ts )N78), fi*ed effe$ts regression ,nter$e(t onl' models in M3R are e4uivalent to random effe$ts )N78) or )N978). Coefficients Random coefficient: term a((lies onl' to M3R anal'ses in w#i$# inter$e(ts, slo(es, and varian$es $an &e assumed to &e random. M3R anal'ses most t'(i$all' assume random $oeffi$ients. 7ne $an $on$e(tuali1e t#e $oeffi$ients o&tained from t#e level- regressions as a t'(e of random varia&le w#i$# $omes from and generali1es to a distri&ution of (ossi&le values. =rou(s are $on$eived of as a su&set of t#e (ossi&le grou(s. Fixed coefficient: a $oeffi$ient $an &e fi*ed to &e non- var'ing (invariant! a$ross grou(s &' setting its &etween grou( varian$e to 1ero. Random $oeffi$ients must &e varia&le a$ross grou(s. 9on$e(tuall', fi*ed $oeffi$ients ma' &e invariant or var'ing a$ross grou(s. Random coefficient5 t#e level-2 (redi$tor, average in$ome, is used to (redi$t s$#ool (erforman$e in ea$# s$#ool. ,nter$e(t values for s$#ool (erforman$e are assumed to &e a sam(le of t#e inter$e(ts from a larger (o(ulation of s$#ools. Fixed coefficient5 slo(es or inter$e(ts $onstrained to &e e4ual over different s$#ools. .ot# used in M3R. Slo(es and inter$e(t values $an &e $onsidered to &e fi*ed or random, de(ending on resear$#ers assum(tions and #ow t#e model is s(e$ified. +#e varian$e of t#e slo(es or inter$e(ts are $onsidered random $oeffi$ients.