You are on page 1of 30

Experimental Design

and Analysis
Philosophy of Science &
Causality seven its true

reject
"

they don't believed


"

human tend to believe what their belief &


what → bias
^

Traditional
by
1 View of Science Sir Francis Bacon
↳ from Aristotelian
opposed approach : use
thinking process
to find of
core of kno .

induction base built kno from observed data


equator -
on
exp .
. ground up
& .

idols false observe Ñwnnis


remove
images
- -

don't On ñnwiniñ uld


-

generalize beyond &

ss
critiques of traditional view
↳ Yal idol
I
assumptions in www.isiawo in
hyp .
→ remove -

image
↳ diff diff Ex Thailand could na vaccination result
person winos
-
.

,
. w .

↳ un
clearly result Ex . 4 results s
supp .
2
hyp .

s dis a
hyp .

↳ fundamental
pros
:
principles e
empiricism , objective ,
rationalism )

2
Assumptions in Science

assumption V5
hypothesis assumption : → assume its true

test to find boo leans value


A
happens .
cause B c
Ésoiisqahuqowni hypothesis →

↳ lawfulness change
time att
change
^

of nature
.

understandable uniform but Taiwan no


: & i
generalize psychology I

uniform
• -

temporal changes e.
g.
times
,
soc . norms

heterogeneity ; sin Ñn
.
universal
↳ finite causation : ai related & unrelated factor
,
causation =/
simplicity
oJuÑÑ =/ IN

3
Principle of
Causality
↳ inference of
replicate Toi
relationship .mn events Ñ a causal cause
,

↳ movement not
observe
we
directly to force
corr .
,
a

↳ causal
relationship E.x.INT by
study operational measurements
we
,
never .
IQ testing scr .

critiques
ss of
empiricist monism quoi ctrl.si •

↳ in
=
>

not stat
STR
depends design
corr .
.
on .

/ scales E=x construct ! ?



single meas .
is
impossibility in soc .
sciences →
vary .
2 IQ test exam → 2

current of causal
ss view
relationship
↳ A B not
why A irinosiñn
→ B

↳ refer to real entities


prob .
&

↳ n.oisoovoswu.at Ñowinoyloi
Modern
Philosophy of Science
4

ss
positivism

, positive minion =

based method of
on
pos .
sciences
, explain beyondpositivism
religion &
metaphysic

=

mathematic logic is : not science


,
but based on real
logical e
sing Jalan minors
;
- naw i

logical positivism logician positivism -

math
+

symbolic logic primary tool - =

verifiability criterion of meaning meaningful when it -

can
empirically verified

critiques law can't verify by observed
:

II. animal instinct observe ALL animal in the → universe

rowñrslsilñ
operation ism

EI . scr .
from diff .
IQ test → diff construct
.

Raven 's Paradox

> Jiyuan .
Dudu wa e if Q
,
then P , ✗ on

ss falsification ism
↳ deductive kno inning win theory www.olailoinoid evidence iwuswoño :iÉo
.
: •


process by falsify risky hyp fail to falsify repeat/ y good theory
theories : .
+ →

↳ Sherlock 's Quote condition research


n'
reasoning an a uuu
, design ,

↳ critiques cannot
theory
: some
falsify
failure etc
unrelated
theory sampling
warn e error
,
.
)

evidence isn't enough


1-

is scientific revolutions
↳ should inte into
science
pret on & donor

Paradigm -

universally accepted ,
set as landmark

make
"

Normal worked
Science -

qñoisiuiñ:O solve landmarks ,


period
found

Anomalies - data Yai fit ñu


paradigm
slowly
emerge µ
Scientific -
crisis to create new
paradigm Ex .
behaviorism →
cog .
revolution

Revolution

critiques : relativism
guarantee the
- not truth
soc . &
psy.in/-epretation reject induction
approach
ss Realism
↳ To nñoÉsñ observer win not observe not accept
'

=
,

↳ sin • is noi cannot them all accurate kno to


recognize
access → use
. .

↳ brain rat
critiques : in a -
view en matrix Worlds

instrumentalist views -

theory doesn't made to discover the truth


,

it made to
predict &
help us in world

Current View of
ss scientific
Psychology
↳ realist laws behavior
:
,


Popper ian psychological test :

↳ rational data
analysis :

↳ belief
postmodernist realize their
hyp : . /
assumption with bias /

↳ statistics

Experiment
5
and
Casaulity
Conditions for
ss
Causality GGEZ
↳ cause b4 result
↳ cause
changes ,
result
changes
↳ absence of alt .

explanation
" "

Gold Standard
Design
ss
Experimental Design exp & ctrl
group excepts IVs


.
.

randomization
manipulate IVs + diff .
con d.
group design +
complete ctrl .
+
Laboratory Experiment
Experiments
1

↳ characteristics of TRUE
exp . :
manipulate IVs
,
2 or more cond . & random
assignment
↳ explanation >
relationship

pros
of lab .

experiments : infer causal explanation falsify I


theory * main
purpose
*

down causal
narrow
explanation
test
hypothesis
↳ limitations artificial S
conceptual Toi innis construct for proving
:
In
,
prove .

IVs =
ñonnw behavior Yai reflect to real world -

, ,

characteristics limit
personal generalization
works with
manipulated IVs
some ctrl .
unethical
,
difficult to test E=✗ .
safe sex VS unsafe sex

Validity
2

internal
is
validity
↳ sirloin IVs cause DVS
↳ threats to internal
validity
• selection threats - volunteer bias
I

pre
-

existing groups ;
can't random
more
expert boredom . . .
assignment
tailor
, ,

mortality , attrition
; pnowño during process

compare

reactivity -
evaluation
apprehension ;
stimulate I inhibit
performance
effect but not
-

novelty ; wu event # n in Gr . s in Gr .
2 →
strange resp .

Wu
windup .
Ñrsioin
,

experimenter expectancies - bias observation


purposely or
unporposely
influence their Gr
-

influencing p
.
's
response :p .
.

demand characteristics researcher


design that known
expected behavior
cue
by /

p
-

voluntary or
involuntary
external
is
validity
↳ be to other
can
generalized settings
construct

validity
↳ Dvtññ
measurement represent IV &

statistical
is conclusion
validity
↳ hi stat ñiuwi : row
is common
Tai manipulate
misconceptions on statistics and
experimentation
/ var .
2 cont .
var .
Ninon COV

→ corr .

design =/ corr .
coeff .

analysis : corr .
=

requirements for causal relationship


2 of

↳ t test arid n.oqdiwou.nu study design


-
1 ANOVA ± causal
relationship :
3
Planning an
Experiment
"
programmatic research

multiple instantiations of variables IVs diff condition research : same
,
.
series

↳ 98 alt
explanations limitations from previous studies clear problems
.
1 : /
compare more IVs
↳ effect Ex Eastern culture ✓
geivouivnvos .
Western culture ? →

↳ exact
replication conceptual replication
-

"

Steps Setting but


"

in ñnlwoiosnnaos benefit oils depends research


: von
p
I
gain on
-

the Tai make


stage -
un motivation
ship .
into make sense sense →
p .
rooter

→ error in result

% cover
story cÑw last choices
simple
-

> & matched e blocks random

follow
-
control IVs
,
Dvs & extraneous variables ,
post up
V

Step 2 :
Manipulation - if construct
validity laid : treatment eff .
don't
specific
IVs on studied variables

effect but not


placebo : condition
gr
.
better than control
gr .

better than
placebo gr .

p
.

expected result
btw conditions test
compare hyp :
-

-
in observe var .
uisoionss
g) N'Toi
manipulate
Yiu
1UJuw construct

onn
IVs
's construct n'Trinny's
theory Yun
construct test
-

validity :
convergence & discriminant
validity
hi computer " nisi check attention
wnniñwoid e backfires p .

p .
notice the
manipulation
STR to behavior
-

reliability , ,
salience

wisents IVs at stimulus Yoiunnuuuu select random


e
samplings
-
-

how to bias
-

avoiding p
. awareness

1) cue reduction ( btw .

subject , 3) role
play -

2) motivation 4)
multiple study
V

Steps Measuring
: Dvs -
self -

report
-
behavioral measure e direct ✗ indirect ,

physiological measure

Post
Step a :
experimental -
ensure
p.is good •
healthy
follow check into procedures hyp performance
up p
&
p
- -
.
. .

check skill F-¥ wrack mole test reaction time


-

p unique. .

p .
is -
a -

player
check oionsnuninño
p
-
.
4 Ethical Concerns
is
respect of persons

protecting privacy

voluntary participation
force to
overt coercion
join Ex prisoner
• -
.

subtle

coercion -

expect to
join 1 no better
option
Ex join
.
research A to
get extra credit s
pts .

,
not force

but
you
don't
get extra credit s
pts .

inducements treatment E¥

excessive -
over
payoff 1 .
in 2000 B.

↳ inform consent : oinuwuownhdown.mn nos

↳ freedom to withdrawn
confidential data

↳ data
Yair :yñonw
vs confidential
anonymous
-

↳ finish
deception : debrief after
exp .

results Yaiño8Éwoiin's

APA
Policy -

niwoinñu +

don't emotional
-
use when cause
pain e
physical & )

feasible
-

explain early as is +
permit to withdraw

debriefing -
serious matter

always ask
p
.
still consent after debrief

Tai
-

ñnoñp .

& detail anniñwlrl


is benefience
↳ mental
avoid ALL
damage physical e & , : cover human & animal
↳ benefit benefit oils
risk -

analysis : risk ñiñn →


gain GGEZ

confidentially of data

is
justice
↳ fair
compensation worthy payoff accessibility
: 1 in service

↳ share benefits to all member


Power &
Sample Size Determination

1 Null Hypothesis Signi finance Testing e NHS -11


" Fisher Approach
↳ of result fail to Ho true
intepretation non -

sig .
:
reject Ho + is c stats assume
Ho is true ,

is
Neyman - Pearson
Approach added from Fisher

↳ statistical
a
hypotheses Ho : & H
,


type I It agiuvos
& errors →
power analysis &
sample size

↳ decision
emphasis
more
rej on acc .
-
.

cps .
05 &
p .
< .
051

↳ stat effect causation


.
±

is
intepretingp values

prob ñsiwoloio's data
.
ñlñlaioinusion
.


significant effect = n' worn 0 ( non-directional )

↳ Taiwan Taishi indicator eff


p
c. 05
power , good → von CI & .
size oinin


problems :
Ho always likely to be false

enough power ,
any
test can be
sig . Ex .

p
=
.
51
ñsig .

overemphasis on acc .
-

rej .

values confound
p
-

by effect sizes &


sample sizes

have chance for


Type I error ( a )

Type I Type It
2
& Errors "
> a =
.
05 ; prob .

/ go

>
Ho Inna'm reject

False Positive >


Ho cinuoiisi
accept
> lack of small effect sizes
power ,

small N
False
Negative
"
13
↳ power :O
c I -
I

effect ñwdoujoss recommend


prob .
ñ : no c 0.80 I

|
↳ under t.tt treatment has effect but cniou
power : 5 01 )
power
n =
. .

, .

F-¥ test
power
5
; noo
sig a -5
= s
. .

wastes

y
retest
>
unsig 5 e ,
resource

↳ observe
overpower
: n t.tt
,
riou
annoy c
tiny diff .
=

significant ,

↳ not but I
power ,
sig .

Type error

↳ QRP ña data when


:
sample onn
hyp .
or
stop collecting sig .

pi type
I error it

3 Power Analysis
↳ effect size :
large reduce overlap cpb )


sample size ni ñow
pop
ñÉw
: → a. .
→ unis to


alpha criterion aon.tt pb but : → a -
area c
type I , t
↳ statistical
power

GGEZ
One Between
-

Way Subject Design


Between
Subject Design
1 -

↳ cond
1
p .
I .

↳ reduce biases and order effect


pros
:

cons : Toyman individual factor → error t →


collapse area T →
power
t

two t test
group comparisons
" -
e -
,

↳ dfc
to.it depends on N -

ai

N win

IV ÑOU

☒ I 3. XX

f-
-

, 2
=
95% 1.96
95 %
SE 1-
critical
1-
critical

• N T ms t dis .
iii. nÑ 2- ( normal , dis .
Ninths

N t → t doin '
critical

Analysis of Variance ANOVA


2
( ) btw
>
.gr .

↳ Ho ; µ e Omnibus test 1 MS '


=p , =p ,
=
→ s
treatment compare mean
. .

compare
.

, ,

H at least is diff f- =
Ms if Trini ,
pair ñiilaiñ
' '

; I
µ .

error ; s s


,

continuous
IV =

categorical ,
DV = > within
gr
.

↳ rate
per comparison cap
error : =
0.05-1
,

ke k -
1)
#
comparison =

family -
wise error rate -

prob .
of at least a
type I error in
family
> És % siuo :
compare too :

dfw I 1€ F-¥
( I
Xp
= - -
>
,
dfw I ( I 0.51 0.14 ms 14 %
= - -
=
.

"
modeling DV score

↳ Tai :O :) %
cond .
1
; riouiiiuoñup .

population mean
em , Yi =
M

↳ Ii
cord 2
§ group rip joined Yik M +
ok
=
; cak
→ ,
group
mean mean
pop
-
.
.
.

↳ cond 3 orison 'Ñu& % Yik '


Eik '
; p

Mk
in
Mk
-
. .

s error variation

Yik =
µ +
dk + C-
ik

> treatment variation


"
logic of ANOVA
↳ if Ho true all treatment effect F
is : means are
equal → no =

F
0 →

be
= 1

only positive
'

if Ho is false : treatment effects s o → F s 1

↳ dfbtw Kc # ) 1
gr
= .
-

df = Ken -
2) =
N -
k
within
↳ F e s :O
in : treatment iiionñ 0

large gr
.
variance

sample size t.tt

ANOVA
Assumptions
3
> error variance runoutñ


homogeneity of variance
2 2 2

0
ok : 0
= = =
. . .

, ,


normality normal distribution of: residual errors 1


independence observation independents serious
: are
,

4 ANOVA calculation

n = 12

N : 12 ✗ 5

=
60

£
£-1 0.81£
£

Ssfofa , =
( 9. 28 -
0 .
8) ( 1 .
35 -
-1
. . .
+ ( 0.01 -
0 .
8) + C- 0.66 -
0 .
8)

£

SS = 12 ✗ [ I 1.802 -
0 8) £-1 + ( 1.889 -
0 8) I
between
.
. . . .

M
inn

nigjinn


SS =
( 1.28 - 1. 80 2) £-1 (1.35 -
1.802)
£
+ +
( 1.26 -

1. 8021£ -1 ( 1.06 - 1.8021£


within
. . .


Tj Do
# #
Dg
-

£ •

+ (2.01 -
1.889) + ( 0.4 -
1 .
8891£ -1
. . .
+ ( 0.01 -
1.889 1£ + ( -
0.66 -
1. 8891£
5
Measure of Effect
↳ effect size : mean diff .
&
CIA Cohen 's d & f-
, ,

↳ 9:
Yin var .
Ennius ñwiirilunr : Ñuw
'

y s w
'

df
ybiasngsoin

6
APA Write -

Up

7 Balanced & Unbalanced


Design

unequal ; its :ñs
grand n no mean

↳ data ; random
missing sick
forget systematic -

, ,
-

drop out
↳ with unbalanced data
dealing insnais :

data
randomly removing
imputing missing values
by mean / median

8 Violation of Assumptions

normality robust normality
:
,
non - → Q Q -
or
Shapiro -
Wilk

Kruskal Willis
normality → -
or trimmed mean c recommend ,


homogeneity of Levene 's test
'
robust
variance :
, detecting →
,
deal with
unequal s → Welch
↳ independence if violated don't ANOVA
way
: → use one -
Multiple Comparisons
1
Confirmatory vs
Exploratory Data Analysis
is
clustering illusion

erroneously detect pattern
↳ underestimation of
variability : ñnii small
sample Tail ñunonn random
process

is false
positive results

list in ÑWAETN n'718020
'

uowoouiotos false
wuiuiuui
explanation > -

positive
↳ cannot
replication

is
type I error rates
↳ dfw &
Xp ,
c
normally .
05 I

Confirmatory Analysis Multiple Comparisons


2 of
ss
priori comparisons
↳ before collect the data E=x iaononuu from
compare
. 3 o
comparisons
↳ contrast
called
planned comparisons
" "

↳ Fisher 's LSD Test : reduce FW error rates but not recommended

T )

ñnw
planned + ctrl .

dfw

NO
1
>

planned

"
planned contrast
↳ / set of with another / set of
compare gr gr gr gr
. .
. .

↳ contrast coefficient Élsoñ uorilñii :)


:
weight a sum =
o +

compare
o - ons

> ( 0 -
1 0 0 1 )
, , , ,

" " " "


> ( 113 I I / )
-

I z
-

, , z ,
z , z

> ( 9 0 - 1 0 0 )
, , ,
,
↳ linear combination


testing contrasts
of
assuming equal homogeneity

variance -
at var .

better estimation dft → T


power
-

Bonferroni collection Uru Tusi


-
: crit value.

by
' '
FW =
C .
X ( ✗ = ✗ 1 C )

gif
Ok !
value
< -0-5 =

c
p
or -

-
Dunn - Éidcik Test : a' = 1 -
es -
a)
"
c

To Welch adjust p

unequal variance - - Satter


,

↳ CI & Cohen 's d for contrasts

3
Exploratory Analysis of
Multiple Comparisons
hoc
ss
post comparisons
↳ look all
possible pairwise s → error rate T

↳ NOT confirm of
hyp .

↳ DON'T HAR
King
↳ equal variances Tukey 's : HSD
adjust value
p
- -

Howell GGEZ
unequal Game
variances : -
Between Subject Factorial
-

Design
1
Describing Factorial
Design
↳ use when : 2 or more IV
↳ Ex
( 2
Ways )
bystander
.
3 ✗ 2
urgent design → 2 var .

, bystander 3 Iv . &
urgent a Iv .

.
6 cond .

s mean from no -

bystander cond .

ignore urgent
+ cond .

nn role + col .

ignore bystander
from cond cond
s mean non -

urgent .
+ .

Factorial
Why Design
2

is examine interaction effect


↳ interaction does A ?
depends: 13 on


improve
construct
validity test : more
complex &
predictions ,
riskier
hyp higher .

.
credit
↳ null effect
wosun test nñolaiino : diff .

single .
dissociation -
✗ → Y but ✗ * Z e discriminant
validity ,

low unreliable
oink - weak
manipulation ,
power
entities
,

• double dissociation -
✗ → y but ✗ * z + w → z but w * y
2- ✗ X Z WH y
✗ → Y W →

meaningful intepretation
,

discriminant
convergent
,

& →

"

tibial ; Tailor
von
explanation
recruited
is
reducing variance error +
exp .

group
is

↳ Person Experimental Design categorize nwÑwn§w


✗ :
30
High '
- ctrl .

group
is

>
exp group 15


.

Med

blocking
30
extra Ñw IV in IV
.

: in
use var c → → I '
-

ctrl is
group
. .
.

↳ randomized block >


exp group 18

design random man block


.

: so low
=
ctrl
↳ group
. 15

t I F.tt T irvin
mse Sse sig
error → → → →
.
power .

of effects
"
general ivity
-

&
boundary
↳ check confound
↳ oioslaial interaction at nduswo
,
power

intepret CI & Effect size
Effects Factorial
Design
3
in

is interaction effects

↳ if interaction sin effects


is
sig .

.
ignore main

Ls
Design Complexity GGEZ
limit
'

sinai ñ a
way
or
way
-
3-

↳ twin I IV size in'wÑwÑ7 2X


Samp
.

↳ iizlrwauinorr 2D ( 3-
way I
Factorial ANOVA

1 Structural Model
C ( d 1)
Mjk Mt 13
-
+

i in factor A IV. factor B IV. k


p . #
gr .

j &
> interaction effect
-1

>

1.
grand
v v v
mean
A effect 13 effect error

(
Maj M ' )
M
MBK
'
-
-

is effect size

↳ factor Tai , Éuoiosiuo : Ii


partial
' ' '
reflect
partial
'

morning
use ai
y
s
y :
y →
y

"N factor
total
rilsiowñou

( bias
by pop .
'

2 Factorial ANOVA ii.


data

is SS
total
v

Grand mean n =
10

= (9-11.61) £-1 . . .
+
( g, ,
y , .gg ,
,

# cond.var.EU marginal mean c


ignore )
g

dftota , N
r
• = -

"
ssmaine main effect ,

2
condition ; =
10 ✗ 2 [ ( 6.75 -
11.611 1- . . .
+ ( 15.65 -
21.611£
10 5 [ ( 10.06 11.611£ +
( 13.16 11.611£
age ;
= ✗ - -

• df = C -
1

is SS
interaction

+
mean each cell block

=
10 [ ( 7.0 - 11.611£ -1
. . .
+ ( 19.3 - 11.611£
o
df =
( C -
1) ( A- 1)

→ SS
error

is
simple effects 1 conditional effects
↳ effect IV ri Iv factor
no , s .
n' o
Éwq
↳ airwoman
'
interests effect into ctrl .

dfw
ts One ANOVA uunnoin :ñw
way
~ -


SS =
10 ✗
[( 12.0 -
15.65 1£ + ( 19.3 -
15.65121
intention
• df = A -
I
3
Type of Sum of
squares
↳ I
SS
type sequential coiioiun .la Wion order diff result
'
: var .
aiwo , → .

not
E.x.la A n' B
'

on → Arnon B Tau ctrl A .

* *
recommend
98 B n' on A → "
A Tau ctrl 13 .

↳ II but not ctrl


SS
type : each nouqwñwios ca qw
.
var .
B & B
qw A 1 .
interaction
inlaid interaction → do I ñni, I ñnii
power
-

↳ Ss
type 111 : each var .

nouqwrisuwa + interaction
,
recommended * *

A z
A
,

B ±, B

int .
* int .

Wario un 'm inn : 7h03 unuoiou interaction

↳ if balanced data I I II
summary
: → = =

if unbalanced data →
Type II

non -

sig .
interaction →
Type I e
power
t

mostly use →
Type I
Within -

Subjects Design
Repeated Design
90
1 -
Measures 2 Iv
Iv
.

98
pair sample
t test Toi
Toi
> 2
repeated measure ANOVA


.

p .

woman
cond .

, ñaÑÑViÉouq
↳ Ex Reading
long - cased
study : trio
survey or exp .
iiuuñsioni .
Hrs .
;
Wk .
1 -
a -
3

Ls matched ñsiattr townie E± Twins


pair :
compare p .
. .

exp .

is
advantages
↳ their
p .
are own
ctrl.gr . :
personal characteristics are match → error matched →

t F T irvin Power is
error I →
,

sig . c


Tip .
siouas e
compare
with btw
design @ same
power
)

2
Disadvantages of
Repeated -
Measures
Design
→ order effect
↳ not effect
DV
changed by IV order ,
l Brian .
Hongi )

task

practice effects -

perform ring →
improve DV
performance
- habituation É our
,

detect time effects


-
can
by adding uñsuun

fatigue effect -
if tired ; performance It over time
if bored ; but
-

response
time it
performance It

effect effect from 1 to another cond


carryover carryover
a -
.

Order donate
a
; +0$ → iuño
empathy ?

sensitization effects / contrast 1 context effect


compare

-

from
-

reactivity ; exp .
1 cond .

changes resp .

in other cond .

e
std .

changes

- harder to ctrl .

;
Ii btw nannw
Controlling Order Effect
3

is
counterbalancing
↳ assume iniiñ order effect luwian k
§
niouq

complete counterbalancing - all


possible order c
permutation , = k!

-
order effects sina.rsriwldi.es
- N =
no k too much !
g
from
• random
counterbalancing -

II. choose 12 60

- Intron trial IV. ,


more than cond.lv .

latin
"

square design
↳ # order =
# condition


A noun
.n
order C A @ I →
4)

a nñ
practice


A noun
.n
order C A @ I →
4)


A onwñouoioiowownou C A → 131 C l D)

• uñ
practice carryover +

partially order ctrl .

>s washout period


↳ iiñ
fatigue +
carryover

GGEZ
Repeated -
Measures ANOVA

1
Logic Assumptions
treatment
of
Repeated Measures ANOVA
oio oinsñrs
resp .

& -
.

↳ / 1
> One
Way
-

I
variances : btw subject eind diff
-
treatment . .
) + ccond . ) + errors
ep .
error ,
can't handle

90
partial out btw
subject effects Way nnwtñ
One

-

loix-ridiff.in
-

each cond
I
.

set treatment
'


mean
centering ij =p x-p -
✗ . score -

x-p to 0 → iuñonoi

e not recommend ,

^ , btw effect a o

j7UMc7 . . .

↳ luminous £

separate
'

oonntwéowq
~ a

variance -
nun s →
-
2

~ I
ng

↳ v1

assumption sphericity Sta : -


'

! s
o
'
dis c


- - -

DnYaiÑwos ctwiiño :iñn type II Toi


sphericity , I / error ,

Mauchly 's Test - not robust to non -

normality .
power
I

Huynh Feldt -

Greenhouse Geisser
• -
-
iou
power
T uoiatosñw Type I error of c * recommend * I

MAN OVA advanced methods Taio


e ,
sphericity

-
now

2 Calculation
Repeated -
Measures ANOVA :

→ SS Sstota ,
=
[ ( X -
Ii
total
&
= [ ( 21 -
13.2441£ -1
. . .
+ ( 9 -
13.244 )

kn

dftofa , a -
I a (5×9) -
1 a 44

" SS
subject
sssubj.ec/-=wEcx-s-x-#
&
=
5 •
[ ( 12.6 -
13 ,
247£ -1
. . .
1- ( 16,2 -
13.244 )

dfsubject =
n -
I - 9- 1 i 8

'

" SS SS treatment ' n E e It -


It
treatment
&
a 9 .
I ( 22 .
333 -
13 .
244 1£ +
. . .
+ ( 6 . 778 -
13.244 )


df treatment - k -
1 i 5- 1 a 4

SS SS .
SS SS SS
treatment
- -

is
error
error total
subject

dferror =
( n -
1) ( k -
1)
is F Table
Tinnon inn : ssb.tw uonn.no :owño s
'
nos ind diff
. .
→ error HUUuw%iwn=isÑN§ ooiiugvos ind .
diff qiiioionirin diff
.
.
nnooouriuonlioii-ruoini.nu/lMnoinsrTW
, ,
-1

Hunan 'lw error sized


µ

treat
SS
/ df / error
1-

contrasts
'
. jnoinrioniw F - t

'

I :/
'
'
+
(1) +
"
+
1- ;)

GGEZ
Factorial
Designs with Within -

Subjects Factor

Threats to Internal
Validity
1


history
↳ outside events
changes p
.

responses
is maturation
↳ nat .

changes
testing Yaioioualsnsñn
requires N at least 2×1


e

affect test
pretest post
^
,
scr .
scr .

↳ no
pretest or Solomon 4-
Group Design
ss instrumentation changes
↳ iñusvog insensate F-§ low
n .
I .

battery in
pattern Toi
↳ observer drift : nurse nor become less reliable
fatiuge habituation
e.
g. ,

Sol sample order F-¥


'
random .
coder # I -

1,2 4 5,3
, ,

coder # 2 -

3,4 ,
1
, 2,5

iron observed behavior


change in cat criteria
.
-

set classification
system E±
'
Click !
'

.
narisoiñuñ →

statistical
regression
is
treatment ñu effect Toi


inn error

p
extreme to their solved ctrl
scr .

regress gr
.
mean
,
by .

group

2
Between Subject and Within -

Subject Factorial
Design : Mixed Factorial
Design
is
typical cases


pre
-

post exp .
-
ctrl .

group
: random
assignment not
always perfect
pretest scr . can reduce error

↳ establish
boundary
↳ both btw -

subject & within -

subject factors : btw sensitive to demand characteristics


within increases
power

detecting order effects

ss effects s diff cond .


.

> between
btw error
>
gr
-
.

total s diff time .

variations within s interaction

> within -

gr .
error
I 6

Mixed Factorial
Design Calculation
2

3
:
:
3

8 4
n z

→ SS SS total =
ICY -
51£ 5

total 7-

=
(150 -
169.021) £-1 ☐ o .
-1 ( 122 - 169.0211£
o df z en ki ) -
1 a ( 8. 3 . b) -
I >
143

ss SS SS btw
2
I [ ( is -

F)
btw
"
a
6 •
[ ( 88.333 -
169.27 )
£-1 . . .
-1 (172.33 -

169.021)
k a 3

• df a ( nk ) -
2 a C 8 ✗ 3) -

l a
23

"
ssgroupssgroup-ni-cy-o.IT [ 169.0211£
&
I
=
8 ✗ 6 ( 130.875 -
+ . . .
+ ( 144 . 667 -
169.027 )


df a k -
1 a 2

" SS SS =

SSB -
SS
group
er ror
error ,
,

2
SS = nk [ ( I, 5)
SS
-

" interval
interval
2
8 3 [ 286.208 169.021 )
&
-1 -1 ( 149.125 169.0211£ ]
GGEZ
-
-
.
. . .

• df z I -
I z 5

I 12
SS.ee/1SSce11s-nEcY-oz
-

>s

&
a 8 .
[ ( 213 . 875 -
769.021 1£ -1
. . .
+ ( 138 .
625 -
169.027 ) I

• df , c ki ) -
1 2 18 -
I >
17

" SS
interaction SSIxo-sscell-ssgroup-ssinte.ua#
SS SS within SS total SS
Sserrorw Ssw Ss interval 552×6
a -
• • - - -

is btw
within

effects
"
simple
F-§ 286.2081£ 2081£ I
.

SSG .
at ynf . y
a
8 . [ ( 213.875 -
-1
. . .
-1 ( 290.125 -
286 .
Field
Experiment
Field
1
Experiment
↳ realistic artificial task
vs
settings & :
ring'iiÑw exp .
.
resp
.
a :ñnUnñ win

↳ 61 Ñolloi cond

why ? boarder
sample represent ?
.

: e more )

naturalism &
ecological validity - reflects to real - life situations
↳ control variables
over :
manipulate IV Tailor e
experiments
nat .

limited choice of
operation definition
limited ctrl .
over extra nous var .

↳ control over
populations : diverse
pop
.
→ more ind .
diff .

obtain characteristics Taiwan skip unoo


op p
p
-
. .

I
random
assignment Tailor & bias to random
↳ balance btw ctrl naturalism true
. &
,
use
exp .
as much as
possible
↳ settings public manipulate IV. random naturalism
:
, assign &


implementing IV street 1-heather : -
EI in the area Ex bystander
.

exp .

accosting -

specific target

problems : construct
validity e.
g.
is hot sauce can
represent aggression
?

self - selection
outside interferences
↳ behavior
coding :
coding scheme - set of cat to .

classify behaviors
latent content
• manifest 1 - manifest ;
actually observed
latent the action
; intepret
broad ; Iv
• broad 1 narrow
categories -

categories high -
.

Ex friendly
.

smiling touching etc →


, ,
.

narrow
; large number of categories
if too cat can be loads
many cog

. .

GGEZ
Quasi Experiment
Experiment
Torriani ,
Quasi
ctrl extra nous
1 + .
var .

>
exp .

design ñnwvnñÑñ


assignment @ ind level gr level instead can't
random
no c ,
ensure

I
.
.

↳ nonequivalent ctrl member


group design
choose similar
.
: a
gr
. as as
possible
2 Problems with Quasi Experiment
differences
pre
-

existing
↳ test overestimation
use
pre
-
&
post -
test :
prevent
↳ matched lead to statistical
group
: can
regression
↳ AN COVA covariate winnow
equation ctrl equal stat t
: ion & .
as → .

regression
effect iriiniw
iron
large -
we can't collect all covariates
ss selection bias
↳ attrition out nonadherence attention
e.
g. drop ,
e.
g.

3 Patch -

Up Quasi Experiment
not
is if
pretest is
possible -
use ctrl .

group
to rule out
pre
-

existing gr .

Year 2 Year 4

?
D- 20 Moin
pre
-

existing
A. no c intervention )

✗ -1 pre
-

existing no

> 20

1- intervention no

time
ss series
approach

interrupted time series -
baseline → treatment
-
# observations should be rule out nat .

cycle

treatment withdraw /

equivalent time series -


÷

rule out
-

history ,
selection bias maturation ,

work when Itt


carryover
-
are

control series ctrl + time series


group
• -
.

treatment

multiple time series -

multiple group
Covariance
Analysis of : AN COVA

1
Mutiple Regression - Basic
↳ outcomes
multiple predictors e X ) & 1
our
( Y )
how Y
s . mr nanou
; - X
, changes , changes when other ✗ are constant

slope in
simp regression .

↳ least intercept
s
'
Y
square
method
weight ✗ agnñosñiñn linear
: + error
niouqa El Y -

↳ R2 nd regression / SS
"
: radon Y
explained by ✗ total e ,

'
add model ñÉÑuw n' riiwiuloiannv.in
predict : ?
'
R : iiño nonrandom s C R -
R I
,

s relation btw ✗ & Y

s brnooinonial relation wound ga

s relation btw X
,
, Xz .
. . .
.

n

anionic n' noroiñu value §w


^

↳ oinñluw
model
hyp .
: whole model ex ,
predict Y , a. slope cbx , >
intercept
coefficients
regression

over / covariance
apse

area a

SS


total


SS =
[ ( Y -
i
error
SS
regression
multi colline
55$
COV

,
>
arity ; relation btw X, & ✗
a


error

exp design ; IV clairvoyant


MAMMARY
""
×

www.n.uiloi
.

corr .

design ; ,
nun ✗ Taibi

Quasi ;
og group .
at relation oisoiw → if set as COV

→ COV have corr .


with cond .
→ SE iniuwloi

2 ANCOVA
↳ Y bot b. ✗ ✗
Tj
special of µ
ANOVA ANCOVA
+ e s
eij
= = + +

regression
,
.
=
case

is
coding schemes > # level of var
I diff btw self
;l ref

.

. .
& .

k
dummy coding : #
dummy var .
= -
1 I diff btw lect
. .

;
ref .

Ex level ;
.
3 control ,
lecture & self -

learning ☒ of ref .
level

K -
1 =
3-1 a 2
; lecture → ✗
lect
=
1 ✗
s -1
=
0
; Y a
bot b. ✗ lect + balls -1?
1. oui e

self l >
balls -1
o.i.io;
-

.
→ ✗
lect
= 0 ✗
s -1
=
1 : Y a
bot b ✗ lect
,
+
.

ref balls -1 ?
>
c- Ctrl . ) * *
.
→ ✗
leaf
= 0 ✗
s -1
=
0
;
Y -
bot b ✗ lect
,
+


simple effect coding compare each IV. to ref level with intercept : . c
grand means

↳ deviation
coding how each slope diff from grand : . mean

↳ Helmert coding btw cond another cond.lu min series


compare
: . &

E¥ .
I V5 2+3+4
,
2 VS 3+4

↳ covariates
ion window
explain → error
cmserror ) t → FT
" how covariates works ?
Ex driving study
.
; IV
a
car size
,
DV a
driving performance ,
covariates =

driving exp .

variation

case 1 -

reducing error

small

medium

riot do :) ai
-
S TTT ANOVA
sig
• -

large
,


-
ANCOVA ion er ror oonuisniow → FT ( Power T )

dealing with non -

orthogonal predictors
↳ sstñ
Ss
type -11 ring
unique


case 2 - control diff .
an covariates

rim Éwoiou
existing diff I nÑño=Tw '

pre
- -
.

serious effect n' uñoss

s
-
set
equal on covariates

Ex .
annual driving exp .
=
s
yrs .

is estimated marginal means I


adjusted Y mean e em means )

↳ based model Ex 90 pretest ion set


gr
.
means on
regression .
COV e as
equal

>s issues with ANCOUA


↳ covariates Iou error t.FI but also reduce F when IV Tailor.ba .
ranks 1 Nuri Itt

ñu DU moi COV Ñwñori'D inf.nu DV omw

↳ random
assignment expect : in CoV
iriiñwnngr .

.
COV should be
independent from IV

COV rosin n' on treatment e insertion Ñw DV I


nonequivalent gr
.

design : Yair;
expected value e
pre &
post , inn:7ai7ñ random
assign .

ion COV oonñ conclude Tailor in diff is from treatment


gr .
.

C
0,2 a J
? =
. . .
= J
? I

↳ ANCOVA
assumption ANOVA
assumptions homo of normality independence
'
: -

.
s
, ,

linear him'iñwunw
relationships btw Y & COV -
if not ; ns..nu COV
Ex randomized

homogeneity of regression slope iioir.gr mirin


.

block design
.

test
ñnfai
. . .

+
sig .
; ion
Ctj 00h11m run

Yij a
µ +
Tj + COV +
CTJ +
eij
'

✗ %
regression
n

a
Moderation
Analysis
1 Moderation senior detail rin.uisgr.rioldri.la
.

Factorial
↳ btw moderation
if ✗ & W both cat →

relationship
are

depends

X Y ANOVA
.

on
-

↳ rinunn
boundary cond .
vos
relationship - ✗ ñu Y o :& www.winniouwnrniluwu-n
↳ ✗ a cont Variables I uñwloi
simple effect vos X or W

• a- ; ninñryñn " Usnoiwannno'T e HÑonÑ1mñÑonwo ] I

/
.

w = cont .
or cat .
variables
Y =
bot b. ✗ +
baw +
b. WX

2
Categorical × Continuous
coefficients comp group

regression
.


+

diff b
Y when X
=

W O }
.

&
when
-

diff btw ref ✗ =
O
>
gr .gr
- -

b
. . .

"
ref
,
-

j .

group
→ to slope bn
y bot b. ✗ baw b. WX
=

= + +

°
from ref

of ✗
s
changes in
slope .

group
slope cref.gr .
)

↳ moderation leads to simple slopes -

slope from each cond . are non -

parallel
3
Continuous ✗ Continuous
↳ if ✗ & W are ④
,
XW tend to TTT →
"

multi colline arity


"

& oinoiwiniuwloi

interaction ;

>

slope
changes

↳ solved
by mean
centering c ✗ -
I 1 iiño
put on model e W
,
X )

GGEZ
Experimental Design
and Analysis

You might also like