Newsom 1

USP 656 Multilevel Regression

Winter 2006
bjkjjhgjfjfjfjfjhfjffsdfk i!e!" 0 mg" 5 mg" #nd $0 mg%! &owever" if the rese#r'her w#nts to m#ke
inferen'es be(ond the )#rti'ul#r v#lues of the 'ghdfdgfgdfgsfgsdfgsdfgdffffffffffkkkkkdfghdghble used
in the stud(" # r#ndom effe'ts model is used! * 'ommon e+#m)le would be the use of )ubli' #rt
works re)resenting low" moder#te" #nd 'vb'b'b'vb'vb
Fixed and Random Coefficients in Multilevel Regression
,he r#ndom vs! fi+ed dhgdfgdfgdfgdfgdfgdfgdfk#nd effe'ts is im)ort#nt in multilevel regression! -n
multilevel regression models" both level.$ 'v'vb'vbv'b to v#r( r#ndoml( #'ross grou)s! /e'#use of
the #ssum)tions #bout their error distributions" we '#ll their v#ri#n'es" oo=var(U0j) #nd 11=var(U1j),
0r#ndom 'oeffi'ients!1
,his me#ns th#t we '#n think #bout 0j #nd 1j #s #kin to the r#ndom variables - des'ribed
#bove! -nste#d of #ttem)ting to gener#li2e be(ond the )#rti'ul#r v#lues of the inde)endent v#ri#ble"
we #re #ttem)ting to gener#li2e be(ond the )#rti'ul#r grou)s in the stud(! 3or inst#n'e" we m#( h#ve
50 'om)#nies" but we wish to gener#li2e to # l#rger universe of 'om)#nies when we e+#mine the
me#ns inter'e)ts% or the +.( rel#tionshi) slo)es%!
-n &4M" v#ri#n'es for the inter'e)ts #nd slo)es #re estim#ted b( def#ult! ,h#t is" when (ou
first 'onstru't # model" U0j #nd U1j #re estim#ted b( def#ult! &owever" the rese#r'her h#s the 'hoi'e
of setting U0j #nd U1j to be 2ero! /( setting them to 2ero" we #re testing # model in whi'h we #ssume
0j #nd 1j do not v#r( r#ndoml( #'ross grou)s! -n f#'t" their v#ri#n'e is #ssumed to be 2ero" so the(
#re #ssumed to be 'onst#nt or 0nonv#r(ing1 #'ross grou)s! 3or e+#m)le" fi+ed" nonv#r(ing inter'e)ts
would im)l( the grou) #ver#ge for the de)endent v#ri#ble is #ssumed to be e5u#l in e#'h grou)!
*lthough we t()i'#ll( refer to this 'onstr#int #s 0fi+ing the inter'e)ts1 or 0fi+ing the slo)es"1 the term is
somewh#t loosel( #))lied" be'#use we #re re#ll( #ssuming the( #re fi+ed and nonvarying!
Intercept Model and Random Effects ANOVA
-n the sim)lest &4M model" there #re no )redi'tors! We h#ve one level.$ e5u#tion #nd one level.2
e5u#tion6
0 ij j ij
Y R = +
where Y is the de)endent v#ri#ble" 0j is the inter'e)t" #nd Rij is the residu#l or error! 0j '#n be
thought of #s the me#n of e#'h grou)! ,he level.2 e5u#tion #lso h#s no )redi'tors in it7s sim)lest
form6
0 00 0 j j
U = +
where
0j
is the de)endent v#ri#ble"
00
is the level.2 inter'e)t" #nd U
0j
is the level.2 error! -n this
e5u#tion" 00 re)resents the gr#nd me#n or the me#n of the inter'e)ts! U0j re)resents the devi#tion of
e#'h me#n from the gr#nd me#n! When the #ver#ge devi#tion is l#rge" there #re l#rge grou)
differen'es!
-t '#n be shown th#t this is re#ll( the *N89* model! *N89* e+#mines the devi#tions of grou) me#ns
from the gr#nd me#n! /e'#use we #ssume th#t the grou) me#ns" re)resented b( 0j #nd" thus" their
devi#tions #re r#ndoml( v#r(ing" this model is e5uiv#lent to the random effects *N89* model! We
'#n rewrite our two &4M e5u#tions #s # single e5u#tion" if we )lug in the level.2 terms into level.$
e5u#tion!
00 0 ij j ij
Y U R = + +
-n this e5u#tion" 00 re)resents the gr#nd me#n" the U0j term re)resents the devi#tions from the me#n
or me#n differen'es or tre#tment effe't" if (ou would like%" #nd Rij re)resents the within.grou) error or
v#ri#tion! -n more m#them#ti'#ll( oriented *N89* te+tbooks" (ou will see #n *N89* model written
something like6
Newsom 2
USP 656 Multilevel Regression
Winter 2006
ij j ij
Y e = + +
-f the model is #n r#ndom effe'ts *N89*" j is #ssumed to h#ve # r#ndom distribution! -n multilevel
regression" be'#use we #ssume U0j to v#r( r#ndoml(" the sim)le &4M model with no level.$ or level.2
)redi'tors is e5uiv#lent to the r#ndom effe'ts *N89* model!
When we #dd )redi'tors to the level.$ e5u#tion" the( #re 'ov#ri#tes #nd the model be'omes #
r#ndom effe'ts *N:89* in whi'h the me#ns #re #djusted for the 'ov#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 variable6 $% is #ssumed to be me#sured with
me#surement error! ,he s'ores #re # fun'tion of # true
s'ore #nd r#ndom error; 2% the v#lues 'ome from #nd
#re intended to gener#li2e to # mu'h l#rger )o)ul#tion
of )ossible v#lues with # 'ert#in )rob#bilit( distribution
e!g!" norm#l distribution%; <% the number of v#lues in
the stud( is sm#ll rel#tive to the v#lues of the v#ri#ble #s
it #))e#rs in the )o)ul#tion it is dr#wn from! Fixed
variable6 $% #ssumed to be me#sured without
me#surement error; 2% desired gener#li2#tion to
)o)ul#tion or other studies is to the s#me v#lues; <%
the v#ri#ble used in the stud( 'ont#ins #ll or most of the
v#ri#ble7s v#lues in the )o)ul#tion!
-t is im)ort#nt to distinguish between # v#ri#ble th#t is
varying #nd # v#ri#ble th#t is random! * fi+ed v#ri#ble
'#n h#ve different v#lues" it is not ne'ess#ril( inv#ri#nt
e5u#l% #'ross grou)s!
Random variable:
)hotogr#)hs
re)resenting individu#ls
with differing levels of
#ttr#'tiveness
m#ni)ul#ted in #n
e+)eriment" 'ensus
tr#'ks
Fixed variable:
gender" r#'e" or
intervention vs! 'ontrol
grou)!
Predi'tor
v#ri#bles in M4R
gener#ll(
#ssumed to be
fi+ed
Effects Random effect6 $% different st#tisti'#l model of
regression or *N89* model whi'h #ssumes th#t #n
inde)endent v#ri#ble is r#ndom; 2% gener#ll( used if
the levels of the inde)endent v#ri#ble #re thought to be
# sm#ll subset of the )ossible v#lues whi'h one wishes
to gener#li2e to; <% will )rob#bl( )rodu'e l#rger
st#nd#rd errors less )owerful%! Fixed effect6 $%
st#tisti'#l model t()i'#ll( used in regression #nd *N89*
#ssuming inde)endent v#ri#ble is fi+ed; 2%
gener#li2#tion of the results #))l( to simil#r v#lues of
inde)endent v#ri#ble in the )o)ul#tion or in other
studies; <% will )rob#bl( )rodu'e sm#ller st#nd#rd
errors more )owerful%!
Random effect6
r#ndom effe'ts *N89*"
r#ndom effe'ts
regression
Fixed effect6 fi+ed
effe'ts *N89*" fi+ed
effe'ts regression
-nter'e)t onl(
models in M4R
#re e5uiv#lent to
r#ndom effe'ts
*N89* or
*N:89*!
Coefficients Random coefficient: term #))lies onl( to M4R
#n#l(ses in whi'h inter'e)ts" slo)es" #nd v#ri#n'es '#n
be #ssumed to be r#ndom! M4R #n#l(ses most
t()i'#ll( #ssume r#ndom 'oeffi'ients! 8ne '#n
'on'e)tu#li2e the 'oeffi'ients obt#ined from the level.$
regressions #s # t()e of r#ndom v#ri#ble whi'h 'omes
from #nd gener#li2es to # distribution of )ossible v#lues!
=rou)s #re 'on'eived of #s # subset of the )ossible
grou)s!
Fixed coefficient: # 'oeffi'ient '#n be fi+ed to be non.
v#r(ing inv#ri#nt% #'ross grou)s b( setting its between
grou) v#ri#n'e to 2ero!
R#ndom 'oeffi'ients must be v#ri#ble #'ross grou)s!
:on'e)tu#ll(" fi+ed 'oeffi'ients m#( be inv#ri#nt or
v#r(ing #'ross grou)s!
Random coefficient6
the level.2 )redi'tor"
#ver#ge in'ome" is
used to )redi't s'hool
)erform#n'e in e#'h
s'hool! -nter'e)t
v#lues for s'hool
)erform#n'e #re
#ssumed to be #
s#m)le of the inter'e)ts
from # l#rger )o)ul#tion
of s'hools!
Fixed coefficient6
slo)es or inter'e)ts
'onstr#ined to be e5u#l
over different s'hools!
/oth used in
M4R! Slo)es #nd
inter'e)t v#lues
'#n be
'onsidered to be
fi+ed or r#ndom"
de)ending on
rese#r'hers
#ssum)tions #nd
how the model is
s)e'ified! ,he
v#ri#n'e of the
slo)es or
inter'e)ts #re
'onsidered
r#ndom
'oeffi'ients!