Newsom 1

USP 656 Multilevel Regression

Winter 2006
xcxcvxvxcvxbfk (ie! 0 mg! 5 mg! "n# $0 mg% &owever! if t'e rese"rc'er w"nts to m"ke inferences
be(on# t'e )"rticul"r v"lues of t'e cg'#f#gfg#fgsfgs#fgs#fg#ffffffffffkkkkk#fg'#g'ble use# in t'e stu#(!
" r"n#om effects mo#el is use# * common ex"m)le woul# be t'e use of )ublic "rt works
re)resenting low! mo#er"te! "n# cvbcbcbcvbcvb
Fixed and Random Coefficients in Multilevel Regression
+'e r"n#om vs fixe# #'g#fg#fg#fg#fg#fg#fg#fk"n# effects is im)ort"nt in multilevel regression ,n
multilevel regression mo#els! bot' level-$ cvcvbcvbvcb to v"r( r"n#oml( "cross grou)s .ec"use of
t'e "ssum)tions "bout t'eir error #istributions! we c"ll t'eir v"ri"nces! oo=var(U0j) "n# 11=var(U1j),
/r"n#om coefficients0
+'is me"ns t'"t we c"n t'ink "bout 0j "n# 1j "s "kin to t'e r"n#om variables , #escribe#
"bove ,nste"# of "ttem)ting to gener"li1e be(on# t'e )"rticul"r v"lues of t'e in#e)en#ent v"ri"ble!
we "re "ttem)ting to gener"li1e be(on# t'e )"rticul"r grou)s in t'e stu#( 2or inst"nce! we m"( '"ve
50 com)"nies! but we wis' to gener"li1e to " l"rger universe of com)"nies w'en we ex"mine t'e
me"ns (interce)ts% or t'e x-( rel"tions'i) (slo)es%
,n &3M! v"ri"nces for t'e interce)ts "n# slo)es "re estim"te# b( #ef"ult +'"t is! w'en (ou
first construct " mo#el! U0j "n# U1j "re estim"te# b( #ef"ult &owever! t'e rese"rc'er '"s t'e c'oice
of setting U0j "n# U1j to be 1ero .( setting t'em to 1ero! we "re testing " mo#el in w'ic' we "ssume
0j "n# 1j #o not v"r( r"n#oml( "cross grou)s ,n f"ct! t'eir v"ri"nce is "ssume# to be 1ero! so t'e(
"re "ssume# to be const"nt or /nonv"r(ing0 "cross grou)s 2or ex"m)le! fixe#! nonv"r(ing interce)ts
woul# im)l( t'e grou) "ver"ge for t'e #e)en#ent v"ri"ble is "ssume# to be e4u"l in e"c' grou)
*lt'oug' we t()ic"ll( refer to t'is constr"int "s /fixing t'e interce)ts0 or /fixing t'e slo)es!0 t'e term is
somew'"t loosel( "))lie#! bec"use we "re re"ll( "ssuming t'e( "re fixe# and nonvarying
Intercept Model and Random Effects ANOVA
,n t'e sim)lest &3M mo#el! t'ere "re no )re#ictors We '"ve one level-$ e4u"tion "n# one level-2
e4u"tion5
0 ij j ij
Y R = +
w'ere Y is t'e #e)en#ent v"ri"ble! 0j is t'e interce)t! "n# Rij is t'e resi#u"l or error 0j c"n be
t'oug't of "s t'e me"n of e"c' grou) +'e level-2 e4u"tion "lso '"s no )re#ictors in it6s sim)lest
form5
0 00 0 j j
U = +
w'ere
0j
is t'e #e)en#ent v"ri"ble!
00
is t'e level-2 interce)t! "n# U
0j
is t'e level-2 error ,n t'is
e4u"tion! 00 re)resents t'e gr"n# me"n or t'e me"n of t'e interce)ts U0j re)resents t'e #evi"tion of
e"c' me"n from t'e gr"n# me"n W'en t'e "ver"ge #evi"tion is l"rge! t'ere "re l"rge grou)
#ifferences
,t c"n be s'own t'"t t'is is re"ll( t'e *N78* mo#el *N78* ex"mines t'e #evi"tions of grou) me"ns
from t'e gr"n# me"n .ec"use we "ssume t'"t t'e grou) me"ns! re)resente# b( 0j "n#! t'us! t'eir
#evi"tions "re r"n#oml( v"r(ing! t'is mo#el is e4uiv"lent to t'e random effects *N78* mo#el We
c"n rewrite our two &3M e4u"tions "s " single e4u"tion! if we )lug in t'e level-2 terms into level-$
e4u"tion
00 0 ij j ij
Y U R = + +
,n t'is e4u"tion! 00 re)resents t'e gr"n# me"n! t'e U0j term re)resents t'e #evi"tions from t'e me"n
(or me"n #ifferences or tre"tment effect! if (ou woul# like%! "n# Rij re)resents t'e wit'in-grou) error or
v"ri"tion ,n more m"t'em"tic"ll( oriente# *N78* textbooks! (ou will see "n *N78* mo#el written
somet'ing like5
Newsom 2
USP 656 Multilevel Regression
Winter 2006
ij j ij
Y e = + +
,f t'e mo#el is "n r"n#om effects *N78*! j is "ssume# to '"ve " r"n#om #istribution ,n multilevel
regression! bec"use we "ssume U0j to v"r( r"n#oml(! t'e sim)le &3M mo#el wit' no level-$ or level-2
)re#ictors is e4uiv"lent to t'e r"n#om effects *N78* mo#el
W'en we "## )re#ictors to t'e level-$ e4u"tion! t'e( "re cov"ri"tes "n# t'e mo#el becomes "
r"n#om effects *N978* in w'ic' t'e me"ns "re "#:uste# for t'e cov"ri"te
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 "ssume# to be me"sure# wit'
me"surement error +'e scores "re " function of " true
score "n# r"n#om error; (2% t'e v"lues come from "n#
"re inten#e# to gener"li1e to " muc' l"rger )o)ul"tion
of )ossible v"lues wit' " cert"in )rob"bilit( #istribution
(eg! norm"l #istribution%; (<% t'e number of v"lues in
t'e stu#( is sm"ll rel"tive to t'e v"lues of t'e v"ri"ble "s
it "))e"rs in t'e )o)ul"tion it is #r"wn from Fixed
variable5 ($% "ssume# to be me"sure# wit'out
me"surement error; (2% #esire# gener"li1"tion to
)o)ul"tion or ot'er stu#ies is to t'e s"me v"lues; (<%
t'e v"ri"ble use# in t'e stu#( cont"ins "ll or most of t'e
v"ri"ble6s v"lues in t'e )o)ul"tion
,t is im)ort"nt to #istinguis' between " v"ri"ble t'"t is
varying "n# " v"ri"ble t'"t is random * fixe# v"ri"ble
c"n '"ve #ifferent v"lues! it is not necess"ril( inv"ri"nt
(e4u"l% "cross grou)s
Random variable:
)'otogr")'s
re)resenting in#ivi#u"ls
wit' #iffering levels of
"ttr"ctiveness
m"ni)ul"te# in "n
ex)eriment! census
tr"cks
Fixed variable:
gen#er! r"ce! or
intervention vs control
grou)
Pre#ictor
v"ri"bles in M3R
gener"ll(
"ssume# to be
fixe#
Effects Random effect5 ($% #ifferent st"tistic"l mo#el of
regression or *N78* mo#el w'ic' "ssumes t'"t "n
in#e)en#ent v"ri"ble is r"n#om; (2% gener"ll( use# if
t'e levels of t'e in#e)en#ent v"ri"ble "re t'oug't to be
" sm"ll subset of t'e )ossible v"lues w'ic' one wis'es
to gener"li1e to; (<% will )rob"bl( )ro#uce l"rger
st"n#"r# errors (less )owerful% Fixed effect5 ($%
st"tistic"l mo#el t()ic"ll( use# in regression "n# *N78*
"ssuming in#e)en#ent v"ri"ble is fixe#; (2%
gener"li1"tion of t'e results "))l( to simil"r v"lues of
in#e)en#ent v"ri"ble in t'e )o)ul"tion or in ot'er
stu#ies; (<% will )rob"bl( )ro#uce sm"ller st"n#"r#
errors (more )owerful%
Random effect5
r"n#om effects *N78*!
r"n#om effects
regression
Fixed effect5 fixe#
effects *N78*! fixe#
effects regression
,nterce)t onl(
mo#els in M3R
"re e4uiv"lent to
r"n#om effects
*N78* or
*N978*
Coefficients Random coefficient: term "))lies onl( to M3R
"n"l(ses in w'ic' interce)ts! slo)es! "n# v"ri"nces c"n
be "ssume# to be r"n#om M3R "n"l(ses most
t()ic"ll( "ssume r"n#om coefficients 7ne c"n
conce)tu"li1e t'e coefficients obt"ine# from t'e level-$
regressions "s " t()e of r"n#om v"ri"ble w'ic' comes
from "n# gener"li1es to " #istribution of )ossible v"lues
=rou)s "re conceive# of "s " subset of t'e )ossible
grou)s
Fixed coefficient: " coefficient c"n be fixe# to be non-
v"r(ing (inv"ri"nt% "cross grou)s b( setting its between
grou) v"ri"nce to 1ero
R"n#om coefficients must be v"ri"ble "cross grou)s
9once)tu"ll(! fixe# coefficients m"( be inv"ri"nt or
v"r(ing "cross grou)s
Random coefficient5
t'e level-2 )re#ictor!
"ver"ge income! is
use# to )re#ict sc'ool
)erform"nce in e"c'
sc'ool ,nterce)t
v"lues for sc'ool
)erform"nce "re
"ssume# to be "
s"m)le of t'e interce)ts
from " l"rger )o)ul"tion
of sc'ools
Fixed coefficient5
slo)es or interce)ts
constr"ine# to be e4u"l
over #ifferent sc'ools
.ot' use# in
M3R Slo)es "n#
interce)t v"lues
c"n be
consi#ere# to be
fixe# or r"n#om!
#e)en#ing on
rese"rc'ers
"ssum)tions "n#
'ow t'e mo#el is
s)ecifie# +'e
v"ri"nce of t'e
slo)es or
interce)ts "re
consi#ere#
r"n#om
coefficients