2xk 2-factor mixed gro ups anov a · 2015. 3. 13. · mixed gro se of this stu emester. pra am...
TRANSCRIPT
2xk
givendifficexamdepe ProcThereeffec Initia
2-wa
Main
Initia Get glm T B /wsf /me /prin /plo /wsd /des
2-Factor
The purpon during the sulty as the ex
m problems (=endent variabl
cess: e are a lot of
cts will require
al Analysis Get descr Determine Consider
ay Interactio Get 2-way
n Effects Get estim Why are t
al Analysis
descriptive
TestPerf1 TesBY PractDif factor=AllTes
ethod=sstype(nt=descriptive
ot=profile(AllTdesign=AllTesign=PractDif
Mixed Gro
ose of this stusemester. Praxam problems=3). Different le was perform
steps to a coe different sub
riptive meanse what effectswhat main ef
ons y cell means &
ated marginathe “Descriptiv
e means, plo
stPerf2 TestP
sts 4 (3) e ests*PractDifsts f.
oups ANOV
udy was to exactice Difficults (=1), they w
sections of thmance on eac
mplete analysbsets of these
, plots & F-tess are significaffects are likel
& follow-up a
al means & fove” and “Estim
ots & F-test
Perf3 TestPerf
f)
VA
xamine the Prty was a 3-co
were easier thhe course wech of the four
sis of a 2-waye. Here’s a pre
sts ant ly to be intere
nalyses to de
llow-up analymated” margi
ts
f4 list “by giv co ge ge ide ide
ractice Difficuondition variabhan the exam ere randomly ar examination
y design. Diffeview…
esting – based
escribe the 2-w
yses to descrinal means dif
ts DV -- list eay” IV ve a name to rrects each et descriptive ct plot of cell m
entifies WG IVentifies BG IV
The “Des The margthe cell m For exam( (66.746 Notice thagiven (mo
lty on exam pble - practiceproblems (=2assigned to res given durin
ferent pattern
d on the aggr
way interactio
be each mainfferent ?
ach variable t
the WG IV (ceffect for all otcell and marg
means (x-axisV V (interactions
scriptive Statis
ginal means ameans being a
mple, the mar1*20) + (77.5
at the marginore below!).
performance fe problems we2), or they weeceive the thrg the semest
ns of significa
regations invo
on
n effect
that is DV for
can’t match anther effects ginal means s * “separate
s are automat
stics” are the
are weighted baggregated.
rginal mean fo5523*24) + (7
al means for
for each of there either aboere more difficree difficulty lter.
nt and non-si
olved
one of the IV
ny variable na
lines” )
tically genera
raw or “unco
by the differe
or the TestPe77.5091*25)) /
the BG main
e four exam out the same cult than the evels. The
gnificant
V conditions
ame)
rated)
rrected” mean
ntial sizes of
erf1 is / 69 = 74.404
effect are no
ns.
44
t
The betweedifficulty is a
From the mlike relative difficulty grofour exams.difficulty praeasier on ExoutperformeoutperformeExams 3 ansame and e
The ANtwo sum The WGinteracttable, w The “Sptraditionare varip-value assump Both theeffect of
n groups maialso significan
eans and theperformance
oups changed. Harder and sactice both ouxam 1. For Exed same, whiced easier. Whnd 4, harder oeasier.
OVA results mmary tables.
G main effect ion are shown
with multiple F
phericity Assunal approach.ous attemptsfor departure
ptions of the m
e inteaction af test-retest a
n effect of prant.
e plots, it lookse of the exam d across the same
utperformed xam 2 harderch hile for both
outperformed
are given in .
and the n in one
F-tests.
umed” is the The others to correct the
es from the model.
and the main are significant
actice
s
r
e
.
Consider what lower-order effects we will need to check for descriptive/misleading patterns Because of the significant 2-way, the means patterns of each main effect will have to be carefully checked against the corresponding simple effects to determine if they are descriptive or misleading. Remember, this will have to be done whether the main effect is significant or not – main effect nulls can be misleading! Consider what lower-order effects are likely to be interesting – based on the aggregations involved PractDif
These conditions are really pretty arbitrary. The an average of the four tests seems a reasonable aggregate to represent the semester’s performance, so this
main effect might be interesting, especially if the main effect pattern is descriptive for a majority of the exams. Four Exams
The marginal means are of dubious value, because it is unlikely that a group of students will practice for an exam with a variety of differently difficult practice problems. And so, it is not clear what population would be represented by the aggregate of the easier, harder, and similar difficulty performances
So, this main effect is only likely to be interesting if that main effect is descriptive, and so, it describes the behavior of those who practiced with similarly difficult, harder, and easier materials.
Remember – – non-significant lower-order effects that are involved in a significant higher order effect must be compared to the corresponding simple effects, to determine whether they are descriptive or misleading!!! 2-way Interaction Pairwise Comparisons You will usually want both sets of simple effects. One of those sets will be used to describe the pattern of the significant interaction. Each set will be used to determine if the corresponding main effect pattern is descriptive or misleading. Select the set of simple effects that most directly addresses the research question or research hypothesis The statement that, “The purpose of this study was to examine the Practice Difficulty on exam performance for each of the four exam given during the semester. .” makes the selection of the simple effects to use to describe the interaction straightforward. From this, we’ll want to focus on the simple effect of practice difficulty (easier, harder, similar) and then examine how this simple effect is different for each of the four exams.
Obta /emm
The p This
aining and d
means=tables
pairwise effec
interaction pa
describing
s(AllTests*Pra
cts describing
Test1
Test2
Test3
Test4
attern allows
the pairwis
actDif) compa
g the pattern o
us to anticipa
e simple ef
are(PractDif)
of the interact
ate that the ma
ffects of Pra
tion are:
Easier v Sa
<
<
=
=
ain effect patt
actice Diffic
this asksfor the 2
the orde“table” cmeans
the variacommantest
These are thDescriptivesthe tables co Remember tin this MG fapairwise comerror term. simple effecPractice Diffomnibus AN .
The F-tessimple effour exam With 3 Prfollow-upthese sim The pairweffects ar
me Easi
tern of Practic
culty for eac
s for the an a2-way interacter of the variacommand con
able specifiednd tells which
he same cell s table aboveommand.
that this is theactorial. So, tmparisons areNotice that th
ct F-tests matcficulty BG ma
NOVA above.
sts tell us thatffect of Practicms.
ractice Difficup analyses to emple effects
wise comparisre shown on t
ier v Harder
<
<
<
<
ce Difficulty w
ch Exam
analysis of thetion
ables in parenntrols the disp
d in the “comph set of simple
means as in t, but rearrang
e set of BG sthe simple F-e computed uhe dferror (66ch those from
ain effect test
t there is a sigce Difficulty fo
ulty conditionsexplicate the
sons for thesethe next page
Same v
=
<
<
<
will be mislea
e cell means
nthesis of the play of the
pare” e effects to
the ged to match
imple effects tests and
using a BG ) for the
m the test of in the
gnificant or each of the
s, we will needpattern of
e simple e.
Harder
=
<
<
<
ading
e
d
If yopair t = andma
ou are askedrwise compar
Mean Differe
d the df = the in effect in the
the t-value forison is…
ence / Std. Er
dferror for tese overall ANO
or each
rror
sting the BG OVA (66).
Obta /emm
The Fapprowe w If you t = df = The p The p Easie Same Hard
aining and d
means=tables
F-tests SPSSoach to comp
will focus on th
u are asked fo
= Mean Diffe= dferror from
pairwise com
pattern of the
er
e Difficulty
er
This
describing
s(PractDif*AllT
S provides for paring the effehose.
or the t-value
erence / Std. m the WG ma
parisons for t
e interaction is
Test1 v T
=
=
=
s interaction p
the pairwis
Tests) comp
these within-ect means. S
correspondin
Error ain effect and
these simple e
s:
Test2 Test1
pattern allows
e simple ef
pare(AllTests
-subjects simpSince the pairw
ng to a p-valu
d Interaction F
effects are sh
v Test3 Te
<
=
=
us to anticipa
ffects of Exa
)
ple effects arewise comparis
e for any pair
F-tests
hown on the n
est2 v Test4
<
=
<
ate that the m
am for each
this askfor the
the ord“table” means
the varcommatest
The cell me“Descriptive.
e based on a sons provide
rwise compar
next page.
Test2 v Te
=
=
=
main effect of
h level of Pr
ks for the an a2-way interac
der of the variacommand co riable specifieand tells whic
eans will be the Statistics” a
somewhat dthe importan
rison…
est3 Test2 v
<
=
Exam will be
ractice Diffi
analysis of thction ables in pare
ontrols the dis
ed in the “comch set of simp
he same as gabove.
ifferent “multint portion of th
v Test4 Te
<
=
<
misleading.
culty
he cell means
nthesis of thesplay of the
mpare” le effects to
given in the
variate” he analysis,
est3 v Test4
=
<
=
e
Pleasthe S
se note that Std Error use
the Std Erroed for these
ors used for tBG pairwise
the WG pairw comparison
wise comparns. See the d
risons up abodiscussion in
ove are subsn the next se
stantially smction!
maller than
Alte This baseError Why showlargemean Anota mixa mo SPSSStd Evs. Rconscomppowecompand o An athat imadewithinpopuThe eerror
Why
1. ItsSu
2. Ttoc
rnative Ana
is a MG moded on another rs of the follow
care? Becauwn above) thaer or smaller sn difference c
her issue withxed factorial, ore powerful te
S uses a BG Errors = 2.497Retest, Std Er
equently morparisons prover! This has lparisons. Whour intent was
lternative is tos then used te when we con-group varia
ulations that aextension in tr term to comp
do people wh
t is based on standard errorSo, it providesusing this appThis approacho get SPSS tcompare thes
alysis of Ce
el. The WG merror term. H
w-ups vary wi
use the followat uses the stasample sizes,can be signific
h mixed groupthe interactioest than woul
error term to 7 & 2.535) anrors = .734 & re powerful, thvides evidenceed some to re
hile solid statiss to describe
o use this WGto compare anompute the ANability from theall have the sathe WG error pare the asso
ho like this ap
the same estr than the pais a more powproach (though allows the co provide thise means.
ll Means
main effect anHowever, the fith sample siz
w-up analysesandard error i the standard
cant for one c
ps designs invn is tested asd a correspon
compare the d it uses a W.670). The W
han for the BGe of an interaecommend alstical advice, how this effec
G error term thny two cell meNOVA error tee different desame variabilityterm approac
ociated cell m
pproach like it
timate of variarwise error te
werful significah the reverse
comparison ofs significance
nd interactionfollow-up anaze.
s are based on the denomi
d error can becomparison an
volves the chs a within-grounding compar
BG simple efWG error term WG error termG simple effecction pattern,lways examinwhat are we ct is moderate
hat was usedeans. This is erm for BG msign cells, becy, so the aggrch is that sinceans to explic
t?
ability, but larerm approachance test, ande can happen f nonadjacenttest, but the
n F-tests are balyses are eac
n a t-test (thanator. So, de
e larger (smallnd not signific
oice of the erups effect, usrison using a
ffect within thto compare th
ms are smallect. One poss while the BG
ning significanto do when thed by the WG
to test the inan extension
models. That cause they earegate of themce the proper cate the patte
rger sample s use by SPSS
d more pairwison occasion)
t cells meansComputators
based on onech based on a
at isn’t shown epending on wler ns) or smacant for anoth
rror term to ussing the withinbetween gro
he MG interache WG simplr and the WG
sible consequG comparisonnt MG interache BG IV is thG IV?
nteraction as tn of the “homoassumption isach represenm all is the beerror term to
ern of the inte
size and the WS, especially se cell mean ). . Sometimes
s will give us a
e error term aa specific erro
in the outputwhether the cealler (larger nsher.
se to test the n-groups erroups model an
ction e.g., Ease effect with t
G pairwise comuence that thens do not, simtions using thhe “primary” v
the basis for cogeneity of vas that it maket a sample ta
est estimate otest the inter
eraction.
WG error termwhen compacomparisons
s, with larger dan LSDmmd t
nd the BG maor term, and t
t, but how to cells being coms), and the sa
pairwise simr term, gener
nd error term.
sier vs. Samethe interactionmparisons aree examinationply because o
he WG pairwisvariable in the
computing anariance” assues sense to coken from diffe
of the variabiliaction, it is al
m, and, so, usring the BG s
s are significa
designs, therethat we can u
ain effect is the Standard
compute it is mpared have ame cell
ple effects. Inrally leading to
e Difficulty, n (e.g., Test e
n of the WG of differential se e factorial,
n LSD value mption that isombine the erent ity of each. so the proper
es a smaller simple effectsntly different
e is no easy use to
n o
s
r
.
Desc /emm
Alte You wthe pthe n The “For t Pleasappro
Whicand t
cribing the
means=tables
rnative Ana
will sometimepattern of the n for the marg
“n” for the LShis design, w
se note: Becoach taken us The emm
the effect estimated
This approgreater th
ch one to use?to be very cle
BG Main Ef
s(PractDif) com
alyses of Ma
es see folks omain effect. inal means is
D computatioith 4 WG con
cause this desing the emmeans analysisof review atte
d from the mooach compare non-orthog
? As you migear which one
ffect of Prac
mpare(PractD
arginal Mea
obtain an LSDThat LSDmm
s different from
on is the numbditions, and 6
sign is non-omeans analysis tested and dendance and del. es the raw monality (uneq
ght expect, op you are pres
ctice Difficu
Dif)
ans of Pract
Dmmd value amd value will dm the n of the
ber of data po69 cases spre
rthogonal (has above! described thethe interactio
arginal meanual-n) of the d
pinions differ, senting.
ulty
tice Difficult
and use it to cdiffer from thee cell means.
oints, not the ead across 3
as unequal n),
e effect of pracon. That is wh
s (without cordesign, the m
and the impo
You shoulthis main e“Descriptivanalysis o The F-testabove, beunique comodel. Samean diffemeans of other effec The pairwthe main e
However, interactionThis main because it
ty
compare the me value used t
number of caBG condition
, this analysis
ctice difficultyhy it compare
rrection for thmore these two
ortant things a
ld notice that effect were noves” table at t
output!
t matches whecause both antribution of taid differentlyerence amongthe groups, acts in the mod
wise comparisoeffect of Pract
Easier < H
we know fromn that this is oeffect must b
t is potentially
marginal meato compare ce
ases, each mas, this would
s is important
y after corrected the estimat
he other effeco analyses ar
are to know w
the marginal ot given in thethe beginning
hat’s in the ANare for the corthis main effey, both are tesg the estimateafter correctindel.
ons show thetice Difficulty
arder < Sa
m the pattern only descriptivbe communicay misleading.
ans, to test anell means abo
arginal mean be 4 * 23 = 9
ly different fro
ting practice dted marginal
ts in the modre likely to diff
what “your kin
means for e g of the
NOVA table rrected or ct to the sting the ed marginal g for the
e pattern of to be:
ame
of the ve for Test2 ated carefully
nd describe ove, because
is based on. 93.
om the
difficulty for means –
el). The ffer!
d” expects
y,
Desc /em
The Feffecapprocompwill fo If youany p t = df =
The p Howepoten
cribing the
mmeans=table
F-tests SPSScts are based oach to compparisons provocus on those
u are asked fopairwise comp
= Mean Diffe= dferror from
pattern of the
ever, we knowntially mislead
WG Main E
es(AllTests)
S provides for on a somewh
paring the effevide the impore.
or the t-value parison…
erence / Std. m the interact
e Main Effect i
Test1 v T
=
w from the pading. There is
Effect of Tes
compare(A
these within-hat different “ect means. Srtant portion o
correspondin
Error tion F-test
is:
Test2 Test1
attern of the ins no practice
sts
llTests) Yothe74
-subjects simpmultivariate”
Since the pairwof the analysis
ng to a p-valu
v Test3 Te
<
nteraction thadifficulty cond
ou should notie marginal me.40, Test2 = 7
ple
wise s, we
e for
Alsoand matc“Des Becadesiare alike amodmod So, beinmeaquannot tvaria
est2 v Test4
<
at this main efdition for whic
ice that the means from the76.22, Test3
o, the F-test fothe pairwise
ch) are not coscriptive Stati
ause there argn conditionsall collinear (nall other mult
del tests the udel, controlling
in a factorial ung tested are dans, the samentitative variathe same as table and the c
Test2 v Te
=
ffect must be ch there is thi
means shown e “Descriptive= 77.97 & Te
or “AllTests” icomparison s
omparing the stics” above.
re unequal sas, the main efnonorthogonativariate analyunique contribg for the othe
using Type IIdifferent than
e as a multipleables will test the bivariate ccriterion!
est3 Test2 v
<
communicateis pattern of t
here are not e Statistics” abest4 = 81.57)
n the ANOVAshown below data means s
ample sizes affects and theal, or correlateyses using Tybution of eachr effects in th
I SS, the main the raw datae regression ia regression correlation be
v Test4 Te
<
ed carefully, best effects.
the same as bove (Test1 =
A table above (which shown in the
mong the interaction ed). Thus,
ype III SS, theh effect to the e model.
n effects a marginal including weight that is
etween a
est3 v Test4
<
because it is
=
e
s