Data:

Continuous Random Vanables Probabilityof arandomunable

one Expectation -(x) M Variance Variance: covariance

9(x 1)"
=

- Cross Sectional: not dependent on time -

fzix -yad,
=

-samples (y cov(x,y) E)(X- Mx)(y 1y)) G(xy)

- Time Series: sequence over time matters
2
(uxMy)
·

Var(x)
Probability Density function:Cummulative Distribution
=
=

5
-
-

function: f(x)
= =
·
· =

n -

f(x)= rayavara
1
- Panel: Cross sectional + Time series Sf(x)dx
T St. Delv.

1x 255xuxcyx **- planonenuous

=

P(a>X <b) var(x)

· =

-*
=

f(x2) (f(X)) xxy

occoucc-acic.SC,
Var (x)
· =

-
=
·

Rules:
Jof(x)dx 1 Area
S
· = +

Data Collection:
xoxusif
under

RelationshipbetweenXand.
·
f(x) 10 0 xeR curve · var (x) G((X-M)3)
= =

- Experiments, Observational studies f(x) as

write percense function
2(g(X)) 8y70:InverseyRede
- Prospective: design - collect - analyze Var (9(x)) 2)(9(X) Mx)") Txy XLY (Frelependants
Random vanables
0:
=

Discrete
=
- ·
·

1
- Retrospective: collect - design - analyze Properties:
Properties "spread" with
constants

to
covenance
Correlation z-score
Mass function: Lummulative Distribution function, constant
Probability
-
E(b) b -

E EsaNITREEEEi
=
-

Sampling: f(x) P(X x) f(x) P(XIX) Gf(t),

EcaxalEsB **-
1x
Pxy
#I
= +
= =

"Sum of
probability
- Random: choose n individuals from - -(b)

15f(X)=1-ofalirandon,Addalteprobablessetsets
+

Rule:
population with equal chance f(1) f(2) =Six-j, ==x3 jeTy,Y-@NIY If
*
↓ ↑
f(2) f(0)
11(yI]Std. Dev=War(x)
i.e.
+
=

y
+

formula I2abcor(X,Y)
According
-

- Simple: one group with equal chance

to
->

- Stratified: split into groups then do

i
Data Visualization:
spread Distributions Standard Normal:
simple random sample (e.g. gender)
·
· Scatter plot: Histogram:
·

(n 0)
Doesn't exact
- Convenient: choose individuals that are Relationship for 2 show
Range:max-min mean
-

- =
-
·

numeric vanables value of data points Binomial:

·

easy to access (not random) All data must be collected Variance:

-
variance (52=1)
Stem LeafPlot: (P)
·

successprob
-

E(x-e

*
making histogram
-
x
·

before
How frequently values appear Sample:s2 t uals (n) of Distribution:
=

Stats Definitions:
-
-

visualization
Good for
-
# of -

valuesens
showsexact
=

- Population: group we are interested in frequency

"

1) Degrees (v)
E(x
Re. frequency -
-

of f
Population:52 Poisson:
-

Hard to choose stems

-
=

- Parameter: numeric value describing the

- ·

na
population ( M,0,p,N ) ·
Box plot: · Standard Dev. : -
Event rate (1) ·
Chi-squared (X"):
- Sample: subset of a population
*
-

Visualizes are factor in data Sample:5 55u

(k)
=

Degrees
-

f
I
of
Normal:
-

sets ->
- Statistic: known value describing the
·
multiple data
Population:2 v 2
-

compare -
=

sample ( X, S, p, ) ↓ mean (M3 f- Distribution: ①

coup
1 ·

I
-

ofm edian upon her

-> Degrees freedom:v
of n
=
-
1 =
variance (22) -

Degrees of f 12, we -

# of independent pieces i nfo

Probability Definitions:
of
IGR:9-QU
↑
-

available computer variability

-
to

- Statistical Experiment: process that -

Lower whisher:Q-1.S(IGR)
-

Mode:Highest peak (most frequent #)

generates data (different to Experiment as it -

upper whisher:9+1.5(IQR) Not affected byoutliers
*

is about testing data instead of hypothesis)

- Sample Space (S) : set of all possible
-
Outliers:
Values>upmowhisher -
mean: Balance pot one Point Estimate a population Parameter
Estimate
(8) (e.g. M) byfinding a point
Affected by
*
outliers 1

outcomes in statistical experiment Measures

·
of shape:
estimate (E) (e.g. *) which is a
single rake from a statistic (0)(e.g. E 3x, x2, x, =

...
3)
50/50
- Event: subset of a sample space Modality
-Median: Area is split

Statistic (8)
symmetry
-

parameter (8) if:

-

- Partition: two subsets that combine to form -

skewness -

Cleft. of skewness
* Not affectedbyoutliers ·
is unbiased Estimator of
sample space (e.g. events A and A’ form S)
- Random Variable: function associating
Testing Errors M8
=

E(8] 0e.g.z[]
=
M
=

each number with an event in the S

Point have
Sampling Distributions:
·
Type I:rejectto when is
it true
·
Estimates have
sampling distributions since they variance

- Law of Large Numbers: as sample size

increases, sample mean converges to
-

P(Commiting Type I error) a

=

Setonin ·
Estimator:
Effecient pointestimater (8) that has small vanance in point
the

population mean - x level

=
of
Significance / Degree of confidence estimate distribution (small rot
- Sampling Distribution: distribution of Ho
2 :Reject
P-value < * Most Estimatur:Statistic that
Effecient same the parameter
sample statistics from many samples
is the
categoryas
-

- Central Limit Theorem: as # of identically -

P-Value>&:full to Reject Ho (E.g. Sample venance is effecient
most estimate
distributed samples increases, the sampling -

Confidence Level 1 -
2 for
population
the vanance
distribution of the mean converges to a Decrease 2:
normal distribution ( M,2/0 ) ⑰,
-

has least
the verance meaning It is
1. Increase sample size
(If sample Size 1130 Assume
normality -> the best statistic getting point
region)
for a
2. Decrease critical region (Increase fTR
estimator (E) to the
estimate
- Paired Samples: look at before and after of II:fail to
·
Type reject to when it is false parameter of interest (8)
the same sample
-
P (Committing Type 11 Error) B =

mean of level;for
Hypothesis: I ANOVA
values
-
can only
b e found when we have Ha 2- factor Experiment faster A , Level;for
- Statistical Hypothesis: assertion about one
↑ accross all K
factor is

ocrmimiting
Statistical Power: Powr=1-B

on
or more populations and a parameter value
-

across all
fac tor A
of Level;for jand K
-> Mean
- 2-Sided Hypothesis: specifies that the P (correctlyrejecting to when it's false)
Error

parameter is exactly equal to a value ·↑Power as can have uptors

- 1-Sided Hypothesis: specifies that the -
The farther Hat is from smaller B
the is

parameter is at least or at most a value Decrease 3:

=

- Null Hypothesis (Ho): hypothesis that we 1. Increase sample size -> Total variation
con accept in the absence of data. We are 2. Increase critical region (Decrease fTR region)
looking to reject the null hypothesis.
Testing Goodness of fit
be
- Alternate Hypothesis (Ha): hypothesis that mean mighti s claimed to 68 in. W 3.6
=

in. Sample size 3 ) =

Looking vanance
at -> X fast
is the opposition of the null. Statistical a) find 2:
evidence and analysis is looking to be used Ho: M 68 = -> Critical Region:M967 and M1>69:
·
Hypotheses:
to support the alternate and reject the null. Ha:M = 68
distributions
Ho: Observed and expected are the same
Ho +Ha all
=

possible outcomes
Reject
to Don'theintto Recorded -> Assume thatt he
null is true

Ha:observed and expected distributions are same

not the
E.g. Mean worse length
is 2 hours:
P(Type/tror) 2
=

P(X26>1n x)
=

=
P(X)69/n 6) 7
+
=

Ho length
Hypotuses Testing:
mean move 2 hourg
=
=
·

P(zc s)+Pl
-

Ha mean move
length 2 hous
67 -be
=
=
z

=a
=

z Given distribution
expected K bins and (ci) Observe
with
frequency per pin
+ distubution a
with bing
E.g. Mean
height boys
of is at most 177cm:
a = 0.095 7x
yy and observed (0,)
=

fuquency per bin

=

Ho follows expected
Ho observed data distribution
mean
height
= =

= 177 cm
b) find B: Ha observal data
does not follow...
=

Ha =mean
height) 177cm fa i l rejectnull (X 68)
to =

m
(Oi-e, wt Reject Hoifi
Gs,e
- P-Value: probability of observing data as 1

E.
5 Statistic
Test x2 w k presella s
B p(671X169(M 7)
1
=
-

=
= -

given
=
=

extreme as the data found while assuming ↑

P(z
=
=
- 1.67) -

P(zz 5) -

null is true Region where Ho is rejected - =

1.67
* Only can use if e.55 for all bins (i) Test statists
0.0475
-

-
0
- Low P-Value (inside critical region)
=
can combine bins to gete, or 0, -
3 0.0475
means sufficiently low probability of getting
=

·
Test on Categorical Data:
that data given the null is true —> Reject
Null Special Case:Yate's correction ( Table:show and levels
-> fail reject (fTR) region contingency frequency categorical data according
·
to of
variables/factors
·

to

- High P-Value (outside critical region)

means not sufficiently low probability of If 2x2 Table (11) = 2 and 151:2) -> would end up with v1 =
->
Lens
Expected Value Table:
Lets
getting that data given the null is true —> &
Crow;total (row;total)
Fail to Reject Null
x5(1005
251 0.5)
"Yates' correction
-

in
-

=
TestStatistic x
e,j
#valve
conclusion

Pvalue = 0.10 No fredince to

rejectn ull ·
Testing Homogeneity of Categorical Data:(Independence oflevels)

0.05 <Pvalue < 0.10 weak fredince againstthe

null
* Marginal Total values are fixed factors Levels
amton
Is term factors?
.- between
association two
an to

Ho L/R handed
=

does effect
not
-> hij) 5

3
Moderate
fredince

Xt*COi-Zis"
0.01?P-valu <0.05 against the null distribution
the f avourite
of pets Ho They independent
fixed are
=

Rejectto if it
Tests tatistic 12 =

Does effectd istribution Ha Theyare not independent

=

*
evidence null Ha yr
0.001 P.Valu <0.01
strong the
against X2 >
=
·

(171-)CIF1-ces
"name and snacka re independent"
"Type where v
=

* Fix sample size when

collecting the data of handedness does doesn't
P-value <0.001 Very strong evidence the
againsttunull
effect petd istribution l
 · I Sample Vanance Ratio:
Confidence Intervals
Hypothesis Testing Variance (59:

Test Statistic 5 SI/s

=

InternalEstimate s
Centrality
Estimate -
I K(Other point
estimate
spread · I sample:
1)50
(n
52-(n-)e
=

* Assuming true,
-

null is how where Vi n,

=
-

1 S,)52 CIfor2
=

Sample Size: x"2(2,w

livery Is It to observe data as ·
n130: NormallyDistabuted byCLT where v n =
-
1 &
X2 is not
symmetric
"extreme"as what is observed ·
1 <30: Assume normallydistributed ·
I Sample Ratio of
Proportions: order changes!
spread decreases
-

* As r t as ,

Means:
I
mean (x) known Vanance (52): Binary Linear Regressions
Means:

(52)
CIf o r NY/F =

(S) fa ,we a (sp)fopewrite

· +

· I sample known variance

where
I n, =
-

and
1 We nz =
-

1 Ratio
* If = 1 ->
no difference
Response variable is 0 or 1 CIfor u x
z(y)
X
=

-
10 -
M. =
=

statistez
Test Prediction Interval: Prodecting
=

a
observation
next
5/5 null
mean-logistic function:f(x) =
x
· I sample +
Unknown Vanune (5%)
·
Isample + known valance (5%
n <30:CI
for u x
=

tv,
I v(z)
·
Logistic Regression Model:
fitted PIfor future observation (X.) x
=

=(2x) (5) i
Y
where v n 1
(52)
-
=

Non linear Isample +

Unknown vanance V n
=

1
1
· -

->

p =

1 e-
+
(80 3,x,... 8xXu)
+
+

-> cannotfrd n130:CI for M X =

I Ean( PI nextobservation
f or (X0) x
=

(tw,E)(s)N
Y
b;easily
Aprobabilityofgetting
·
I Samples:Looka tDifference in means
·
I mean (x) + Unknown Vananke (54: 1 Proportions:(Binomial)
(v2):* Ifzerosinintervalsee
·

to
known variance

Test
Statistic t X
=
-
M. As X,by 1:p* by e-bi CIfur(1, -

M2) (X,
=
-

X2) =
Es) +w)
-
Success Proportion (P)
X
successes =

-> ↓
s// Unknown Vanance (59): ~

sample size: AP15 and n(1 F) = 5

-

-
odds ofsuccess:
Not he
CI for(M. M2) (X, X-2) -
= -

=
tre · I Sample True Proportion:
(i+si
P2/(1-pe)/same p(1
* can use z-statistic if (P) $12 x p)
CIfor true proportion
-
=

Odds Ratuo:
((ir(=) 1.)e
-
where w
is
=

· Difference in Means known Vanance: factorbywhich

-
px / (1 p1) -

·
Paired Samples:Where
·
I Samples Difference In Proportions:Both
Sample
must meet
Size
mean difference
the odds success
=

of
(X, 1) do
(p, P2) 2a,p.(-) Pu(1-R)
-

X
BtoPa
-

Ma) a tw,z() CIfor true P P2 =

from
changes by,when Ma (M.
+

statistic z
=

CIfor
-

Test
-

moung
=

=
=
-

= 12

(02/n.) (0/na) +

Onlyuse when
do
->
completely I only when ANOVA Ho is rejected
Planned comparisons:
Pairwisecomparison:Ik) pairs
Where null difference in means(a,

One factor experiments original Ho is rejected

=
·

Randomized ·

Cluster levels
Design companion of t wo levels together into contrasts
- Factor: Variable that separates the conditions -
a Priori:Lookat test
data before ANOVA before performing the ANOVAtest, using
- Levels/Treatments: Value of the factor and identifytwo levels that
· Difference in Means Unknown Variance
may
have
differences
qualitative analysis (E.g. Lookat box plots
+ E.g. Looka tbox plot before doing ANOVA Test
"
E.g.
factor:coffeeSoTim
Hr:Ms MT
=

-
Linear Contrasts:Aggregate
our groups
(X, xu) -

do Horton Ha.. mpleT-Test:Unknown population Vanance

t
-

Test Statistic
=

(s?/n.) ((/nr) +
- Within Sample Variance: Variance within one Ho(i,j):M,=Mj -

M,- Mj 0
=

w =

51(,M) where iCi= 0

level itself The more pairwise compansons done without
a prior;
- Between Sample Variance: Variance between C: indicating the
coefficient sideand length

when=lni-les
higher
the the chance of
Type I error.
of mean
every combination of two levels
the

* rpairwise -
P1 orcore Type o 1is =

·
Ho:Zi((, M.) ·Ha:x (C, M;) 0
·
Assumptions:for the
populations of the K levels =
0

Test:Looka tpotential differences after

necrest integer Tukey's
=>

E.g. Room 1,2,3,5 vs Room 4

-

Independent One factor experimentwith no

ANOVA Test (onlyposterior knowledge)
repeated measure Increases Probability ofType II Video (Higher p-values)
(1M, (1Mz (1)M.
tatoes
or w
-
Have no outliers blocking =
+ +

Ho(i,j):M,=Mj -

M,- Mj 0
=

- NormallyDistributed.The sample data for

where
QC 0
=

distributed with
1.
mm
studentized Range Standard fror:
Ho:M, M2 1y M5 4Ma 0
normally
- =

level
+ + +

each
-

is
·
Difference in Means for Paired Samples: Viardns =

common vanance wh
individual Mrand - SRSE M- (, h) where
E.g. Room 1,2 Room 3.5:Not participating
=

mean +

sample size vs

ado
of
groups i, j
statistic
Test to v n
with = -
1 ·
Hypothesis; Test
Bartlett
Test with w (1)M (1M2 ( 1M3
=
+
+ -

5Ma
+ +

(1)M

where = mean difference

Ho:M =
12 ... Mk = 2.
a-staxstcSRSaretmentin t where Y,and j
-
where

Ho:M, M2M, +
i Ci
-
0
=

Ms 0
=

SSW:
Ha: At least two of
34.... Mr3 equal
aren't sum squared contrastw
~

scrincal
of
due to

valuefromturettableinrefore
error

leasto ne M,is different

At subset
A of SSA

Statistical Model: Mull assumptions: Independanto f; and W DOf

ofError (N k)
=
-

·
normally
Proportions : hisa re independent and sNull Hypothesis:"Difference between;andi s
and
Yij M
= +
xi +

Gij distributed with mean & M;I f9>9*:RejectHosignificant"

Ho:M =
* Large sample size: 1825 and n (1 p) = 5 variance &
S.
-

for all is Interested

in Jusesame *): a
-

Yij:value of level; observation j size level; 1The contrast

is significant
I
Proportion:"I proportion Z-Test"
of

Bet's
,

↑original
·
·
Test:Test Homogeneity Vanance of
explaining
in whythe

-
M: Population grand mean (mean a ll
of
Mi)
-
F- Test: Ho was rejected"

porpo
=
Ho:T Ha:Atleastone
Rejecttoe
J different
2): Effect l evel 1 (How data points change ... is
fw ssw
of
Statistic
Test Test Statistic
-

z =

3 on the Grand
=

bya common factor, as a

1
mean
1. find K Sample Variances:59,S,,,..., SR
-

2.Si 0
=
of the level
result you are on
5i5j(Yij 5,)2
2. find Pooled Vanance Estimate:
s
-

Where MSE
=

Gij: Deviation of level i, observation; from

wk(ni
-
n -

k
Spooled
=
-

1)s. S
the mean of level ;
3. Find statistic b:
Bartlett
- % variability explained by
of
Wy:
contrast
-W
·
Upper-Tailed at
- Test Sig. Level 2: ( k)
Orthogonality:for two contrastsWas and Way,
-

storica ((sp((x X(sR(ux y

z
b
-

...
both SSWa and SSW independent
subsets of SSA.
Difference
are
None of the
Proportions:"2-proportion
·
->
in z-Test" Ho: 1 =... Ck
= 0
=

levels have an Cn0 Overlap SSW)

of

4. Find Critical Value br(, n1, nk)

(Ha:
least effecton the
p. Pr At o ne di o
. .
.,

TestS tatistic z
-

forWa # C,M,and WD
=
diM,:
=
observed value
=

size oflevel

n,bk(2,n.)7
↑
(x!IXz)(1 x,Ix2) Intheend
-

Ho:M, =... Mk
=

bx(h,n,,..., nr) = 5.i (C,di)/ni 0

=
->
means
Orthogonal
Ha:Atleasto ne M,i s different ↑ K-1
Each from * most
At orthogonal Linear Contrasts:

befoundfromBartlett
tale
Batitsee
can directly
of level;
·
Analysis Variance:Samplmean
of
Grand mean
SSA SSW1+SSWe .. SSWk
+

-
1
↑
=

-
Reject Ho Mr)
similateon
b <bx(x, n,
Repeated
if
-
...,
One factor ANOVAuth Measures
Accounts for sample size leveli
of
f.T.R Ho ifbbx(f,ne...., nk)
·

Observations each other test

Whatwe to
want achelve
are not
completely independentof
Variance:Ho and Ha be
must & Datapoint;for level; as if Allows to do ANOVA E.g. Take 10 subjects and
gluethree treatments each
to -> 30 data

Vanunce (22) notSD(4)

F.5.RNo ifs H:M, =

MrHa:Atleast1mean is different
where k # =

with ...

Treatments
of

· I Sample Vanance: ·
One Way ANOVA: ·
Statistical Model:
N Overall ofobservation- k =# of levels M Total mean
- =
=
-

ofTreatment on
Yij 2i Effect mean
M 2i B Gij
- =

"

Ch-1S
=
+ + +

TestStatistic x =

v
with n -

Bi Effect
=
of
subject; on mean

Gij Deviation
=
of level i, observation; from the mean of level;

!
same
* I X,=0 and Bj 0 =
·
Assumptions: as one
facter fo

Ho: 2, =

... dk
=
-

ericity:Vanance of the

differences are equal for all

Ha:Atleasto n e 2;different pairs of treatments:
Explains vanability in the one-wayANOVA
(52 52) (V?- 55) (0 55)
-
=
=
-
 Simple Linear Regression
Variance (ANOVA): Only Outlier Analysis:
·
Analysis of for B ·

·
variables:
easier find SSR and
to SSE and is easier implement
to - Anscombe's Quartert:
-

Response vanable (Dependent) only 1 with software

& plots with same linear regression line, vanances, and
-
Explanatoryvariable (Independent) can have multiple Divide byDegrees f reedom
of
Ratio variance
of correlation, but that lookve r y different
) must data
looka t

· Determenistic Relationship: When response vanable directly Outlier:observation that

is substantiallydifferentto
-

Celcius to farenheit other observations

relates to explanatoryvariable E.g.
date
Y intercept
Error RV
1. Daten error
measurement E.g. Typo when
entering
Line of Best fit: Slope
-

Remove / correctthe outlier

↑
·

↑
v 5 value
n
-> predicted
process)
simple Linear Regression:Y B0 B,X G
2. Sampling error (notparto fnormal E.g. Selecting
- =
+
+

participants are
that not a fit for the population of interest

Remove / correctthe outlier

-

fitted Regression y bo
:
=

b, X
+

3. Natural variation (occurs bychance f rom an

not error)
↑
is the predicted I near
Hypotnesis Test:
(log, 12)
-

regression keep outlier butconsider data transformation

*

bo pointestimators
o f8 and Reject Ho If f >
Defect Outliers:
-
and b, are unblused B, Ho:3, =0 7 ( 1,v=n z, 1)
=
-

sample
n= size /
More data better
estimates
Ha: 5, F0 This is
only an upper-acted sided test1. fit data
to a regression line y=bo +b,x
y
Residual / Error:
atrue manone
·

RejectNo: 3, 70 with statistically

i!
-

a
e;
2. find St.dev. ofresiduals
yi sufferent slope

↑
value

y, yi
·

st
e, 8 Population
I
standard
=

-
y,)"
if Sresiduals 2(3,
=

RejectHo:Rethink liner
-

deviation fall to
regression
=
-

&Population
Predicted oneinof appropriate for data
variance n -

2
observed is
Standard deviation
Optional S sample
* Line
Quality of fit:
=

residual values · ·
Standardized Residual:
=Sample variance
-

#2: Negative and positive values can canal out and

give
cov(X,Y)
a sum of zero even when line
the is nota
good fit
-

Population correlation coeff: =

SRi 2:/sv =

xx5y
2
-

2.: Easier analyze

to and compute compared to [le;) [1,
bounded between +
1] X and 9 St.dev.
Like the -score for residuals
* Bigger value is worse SXY
Estimated Correlation Coeff: 2
* Potential outlier:(SRil>3
SxxSyx
-

·
Ordinary Squares
Least Method:
* As (r) -> 1:Association between a nd
X
·
Studentized Residual:
y,and y Y increases
[i, e, e, y, b0 b,X;
=

min given Calcula

= +

points lineis
-

sign of a tells the direction of association StRi 2:/s ->

=
without
Derivative to
set zero is used to find optimal point Mighthave different
same slope estimate but R-values
* Potential outlier:(SAR,>3
b, =

ux.ilxixi)-(xix.,ee E(x, -

*i(X,- x)2
x) is
xy
= -

Coeff
~
ofDetermination:re=1 -

SSfrom 21, +
1]

Proportion variation
of on thatis explained
Y
·

Categorical vanables:(i.e. not numerical)

b (,Yi b1((i(xi)
model
-

b,
by
the regression
Dummy /
-
y Use Indicator Variables
=

-
=

Higher re -> better

fitted model

Better
=

determining
at how good of a fit the #ofcategorical -> n-1Indicator Vars.

SxX=5i(xitSYyilyiSxyzilxiLeyi) -
model

correlation
-
is compared

causation:Good strand
a
to a since error is squared
E.g. Yeast Type EA, B, 25
=

does not in Y Yeast Type Ia =

2B =
·
B, confidence Interval: slope allways mean an increase in X-> increase

A 1 ①

syy bsnipotentionables Assumptions

5i(3, -5,) of Linear B
Regression:
Ss) -
· o 1

g2 C
=
=

Defaultcase is yeast
n -

2
n 2
1. Linearity:
-

& Estimate bo, and b, Data should look

reasonably linear
where the a
of Vanance of Residuals (52) -
Check using scatter plot with X
and * -

A 56YastA
=
-

z bfyetre
=

for
CI B,:b, tya xxwhere w = n -

2 Residuals / Errors (ei):

2. NormallyDistributed -
model:Y Po
=

Bp4(PH)
+
+
B12A 5323 +

2 N(1 0,52)
=

·
B, Hypothesis Test:slope
- checkusing goodness-of-fit test, plot data,

Ho:(810 0 bi -

(8.0 Normal (G-a plots

quantile plot
Statistic t
-

Test
=

3. Constant Variance:
-
Hai (B,)a = 0 S/Sxx vanance around true points
shouldnot
charge drastically -v
n
=
-
k -
1

similar valance regardless x-value

of thatdon't
follow
a clear pattern .

... .......and size

.

9. =cpdence
Rejective to does not
always gaurantee a linear Slope E.g. Sinusoidal of
residuals are independent one another no
with trend
Rejectsince
Bo confidence Interval:y-intercept
same
-
checkusing residual is (Hr(ful:BPH1 Bz BB 0 ->

==
-
·
=
=
=

B,
residual-='5 = Palu <0.05
=
as for CI
- observationorder plot
. .
==
.

.
= =
=
==
=

(H0)PH:BpH 0 ->
Reject 2 0.001
:
=
at

(nsxx) xP
=

for
CI 3:bo tax where v n
=

Observation
-

(H0)1:BA 0
= -> ax
Reject
+
0.001
=

(H0)B:BB 0 Reject at 1 0.05

=
=

Two-factor experiment ANOVA

-
->

·
B Hypothesis Test:1-Intercept
-

(H0)0 : B0 0
=
-> Rejectatc 0.001
=

s (amp)
=

Ho:8. (). =
bo -

(80). -

Interpretation:
-

Ha: 3. F (80 teststatistic t =

5 Samp =
-

least 1 21 1,2p 0
- = =

yE
-
- 161.897 54.299(p+1) 89.998
+ +

s
[ix."/n(Sxx) -

YeastB ->
2A=8,25 1
=

15
-
- 161.89) 59.294
+
(PH) +
24.166

- yi 161.897 59.294(PH)
①->
0
Gives information 2= 0,2p
sum
= - =

t here
- +
=

on if Is
DOf
of abn
=
-

1
a baseline value when X =0.

changing yeastchanges the Y-intercept

* Interactions between
categorical and numeric
data will
cause slope to
change Br(PH) (27) ...
E.g.
+

...
 |
Multiple Linear Regression ·
ANOVA:Only for 5, ..., BX upour tall test
·
Quality of fit:

R2=SSR=1-30
it
Model:Kexplanatoryvanables ## of a coeffs Coeff Determination:
Ho:B p2 Pn Rejectto
· -> -
of
0
= =
- =

...
=

Multiple Linear Regression:Y 8 =

+8,x, ... Pkxk +

E
Ha: At leasto ne B,0 osen. GluS proportion ofvanabilityin data thatan
-

-Model be explained the regression model

by
-

fitted Regression (predictal): bo +b,X,

=
+
.. brXk
+
Significant:At leastone B, 0 v.i =
1, .
.
.
.,
k

Vi k (explanatory rars.) Ve n k- 1
=

As # o fexplanatory
vanables increases:R2-]
-
=

only var
↑ one response
data points:
A
9(X2, X2i,x,,,..., **, Y:)3 Because SS-
decreases SSTis constant
- -

i = 1,2, n as
...,

- captured
where n>Ki is each data
More vanability gets -> SSR -> SSEN
point
(E.g. a person) and
-

Adding more explanatoryvanables will push R2-1

↓i s a explanatoryvanable (e.g. height, weight, ...) E.g. 13 seeds are tested 3
with
types of fertilizer different
with
even if variables
the are not
statisticallysignificant
* Assume every explanatory vanable (K) is
independant R output:
Leads to model
overfitted and overlycomplex
-
fitted Model: Y, bo +b,X,i
=

b2X2i
+ +

.. +

baxki

Adjusted Dr.Rad
ssESui
Each b,is =
1
Linear
-

· Model Assumptions:unbiased estimate

of B
fort test
1. Have a power of 1for each term (xi, xe,X,,...): use If the Decrease in SSG
as more vars are added IOf
the Decreasei n
-y b b,X b2Xi b,X;
-
does not match increasei n DOF ->
Rady
brxi
3.)
=
+
+

3
+ +

..
poorer fit
+

meaning

6
Explanatory a

(SE) I)
a
still considered a liner model since each farm is added (nearly vanubleg Radi =

1 -

full
=

vs. Partial Model:

-
can transform each form to to
get a linear model
+ +
y b! b,X. bz bux, . . bux
' + +
checki fcertain explanaters the
=
+

vars, should be in

sR i v
n

f=
fitted model partial models
y bo bynot
or not
by forming
x linear model fitted model: I bo
=

a
b, X, bXz byX
+ =
+ + +

-
Use ANOVAon full and partial models compare
to

* 1. Have no interaction eachother

with (Independance):Estimate:bo, b,..., bk
which is the better representation of the data
· Std. error:C I bi =(t *(x)SE;
=

* Lookfor:
(g bu b,x, bzXz b2X, Xa
=
+
+
+

·t-value:t bi (P.)0 / SE, Radyi s

6
1.Which
= -

higher
still considered a linear model (from table)
· P-value:Probability of statistic 2. Which model is more simple
transform a linear model
to get
E.g.full:y 30 3,4,
-

can each term 32Xz 33X3

X, at 0.001level
+

and X2: significant

+ +
=

y bi b,x +beXn beXs' XS x,Xz 0.1level

Partal:Y B0 3,X, + B3X3
+

X3:Notsignificant
at
=
=

+
+ where =

·
Assumptions of Multiple Linear Reg.:
·
Sval
we (Residual Sz) st
squares Method:
=

· Ordinaryleast
-

Lineauty -

Normality of Residuals
statistic
min(e,z (y, =
-

5,)u (yi- (b b,X,i

+

+... +

bxxvi)" SSz =
-

(Ho)nx: B =
pe B = 0
= -> use -
constant vanance -
Independance of Errors
testthis
=

to

NoMulticollneanannot
*
B.

3
- (H.Jo :

0
=

General MLTEquation

!
Y bx
-> to
b e correlated each other
e
((01:B
- = +

-
0
=

to statistics or else
they will compete for same statistical effect

([],b (2), Ex== x -].e ()

use
Plot:checkfor
Matux multicolinearity
HoCz:32 =
·

where y = =

x -
to fetalel
(should lookrandom)
=

Avoid correlation (r) > 10.71

* Interaction Term: E.g. B,X, X, ...

can do (Ho3. Since iti s at test, of
not test
...

Goal is to find that minimizes SSE T Non-Interaction Terms:Main Effect terms

c
f(v,
-
=
-

3,v 9) 30.98 p-value 0.996x10-

- = =
-

= =

#o fInteraction Terms:Based # main effect

b (X ix)"X y Use R to solve reject (Ho(full
on of
+
-

- 0.001 We can
Since Rake < * of
=

model is significant Types:

thatthe
-

and say
-

ococoico-Reject
1. first
O rder:2-Vanables interaction
* (Hobo (HO), (Hoben Sie Revale 20001:
interaction
Vanance Covanance Matux:
(x +x) 2. Second Order:3-vanables
-
-

So, 3,32 are

significantlynon-zero coefficients :
: :
-
Vanance:(ii Diagonals second order
E.g. order
first
=

ME
-
fall to reject (Ho). Since p-valu > 0.001:
Covanance:Cij -mmm
y 30 3,X, 12Xz 33X3 5,2X1Xz 313X,X3 B2yX2X3 P23X,x2X3
-
+
+

↑
= + + + +

linley coefficiento f
+

is a zero

vanance of bi C,,52 =

covariance of
bibi C,j52
=

Do not an ME
fit in Interaction term if it's not significant
-Partial Model:X an
-

is not significant since Rake

-

significantME
A
mayappear insignificant since interaction
vanance
e
f(y,- y)
is very
high fall to reject
Ho.B 0
=

(52=SSt= multicolinearity significance

forms can lead to -> Dilutes
of
residuals n -
k -

↑
1 Re-run with partial Model Yebotb, xi+bXz: -

Picking to use a interaction forms

model with

· B,confidence Interval: Degrees of freedom -

Individual P-values are
significant(<0.001)
still depends on High Rad,Value, complexity(interperability)

B,is:bil S Ciiwhere
-

Total Pavalue is even more significantsmallers RCBD ANOVA:same SSA, SSB, SSE, SSTas Repeated Measures
for
CI tv,c/2 Wn =
- k- 1
·

Higher Adjusted R- Lower Multiok R2

·
S. Hypothesis Test: - > for Hol
because vanability
* should remove
Ho: B (i)0
=

bi -

(5,)0 is explained
by
statistic t
Test from the
Xy
=
one fewer variable
Ha:Bi E(B:). Sci, Is for Ho2
⑳ model
standard
errow
Multi-factor experiment · RCBD Assumptions: Two -
factor experiment
-

1
for vanation from an additional factor
Equality variances of
-

Account for interaction between facture

Account

Repeated
One factor ANOVA with Normalit samethong interaction:value factor influences
-

Measures ·momized
B)ockDesign:
of

-
No Outliers value of faster 2
-> bofsubjects or Blocks Assume factor 1 and I are Independant Positive or
Negative
Independence of Observations
stratify bythe additional factor
-

Two-WayANOVA:
-

-> Blocks are Independent

original factor
·

k = # of Treatments Randomize subjects into treatments

-
for ~
a: #Levels in factor A -
b:Levels in factor B
·subjects are randomlyassigned
* Randomized complete BlockDesign (RCBD): treatments
within a block
-

Yisk: Observation K of combination (i,j)

Leval;o ffactor A, Levels f actor B
Assign each Treatmentonce to each Block * factors have no interaction effect of

-
observations in each
n # of (i,j) (n stays same
Why
Statistical Model; M Total
·
Use RCBD:
(a)(b):
mean
- =

# of groups
·
-

-
Si Effect ofTreatmentonM
=
More vanability can be explained
(a)(b)(n):#
-

- total observations
of

Yjj m 2i Bj Gij Bj Effectof Block;o n M since SSB takes from SSE

=
=

away
-

Statistical Model:
+ + +

·
Interaction Effect on M
-

Gij Random
=
hose -

Increases Power:can get same &

Ho1:2, =... 6k =0 -> =

has
Treatment no effect
conclusions with
smaller samplesize ↑ ijk m 2,
=
+ +

3i +

(B) ii Gijk +

nisample size
Haz:Atleastone 2 is different
room;
Interaction
hassome effectH ol:
2, ... La 0
2i, 35:Effects
offactor and B
=

Treatment
-

- -
A on M
Tis =datapointsfrom room
How:B, =

... Bp 0 =
Blockhas
->
no effect
Ha(A):Atleasto ne aisn't equal Effect -
((ixi 0
=

HOLB):5, . . . Bn 0
H0(AB):(48)1,1 (B)a,b
=

0
room i
E!jBj
= =
... =

0
=

Yi= mean of
effect
-

- > Blockhas
Haz:Atleasto ne is different some
=

Ha(B):Atleast one Bisnitequal HalAB):Atleast one (dB) i,j isn't

equal
Y.grand
mean
-

50*; (< B) is 0
=