Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
STAB22 Statistics I
Lecture 13
1
Principles of Experimental
Design
1.
2.
Control: control sources of variation,
besides factors, by making conditions as
similar as possible for all treatment groups
Randomize: helps equalize effects of
unknown/uncontrollable sources of variation
Note: Randomization does not eliminate the
effects of these sources, but tries to spread
them out across the treatment levels so that we
can see past them
2
Principles of Experimental
Design
3.
4.
Replicate: get several measurements of
response for each treatment
Blocking: for variables we can identify but
cannot control and which affect response,
divide subjects into groups of same variable
values (a.k.a. blocks) and randomize within
each block
Removes much of the variability due to the
difference among the blocks.
3
Experimental Designs
Completely Randomized Design (CRD):
All experimental units are allocated at
random among all treatments
Randomized Block Design (RBD):
Random assignment of units to treatments is
carried out separately within each block.
4
Blocking
Assume 36 adult & 6 child participants in
previous insomnia experiment
If you just randomize, you could get all children in
one treatment
How would you block age?
5
Experiments
Placebo: “fake” treatment designed to look
like a real one, used when just knowledge of
receiving any treatment can affect response
E.g. Used to test effectiveness of pain medication
For comparing results, often use current
standard treatment as baseline
Subjects getting placebo/standard treatment
called control group
6
Blinding
Knowledge of assigned treatment can often
influence the assessment of the response
Two classes of individuals can affect experiment
Those who influence results (e.g. subjects, nurses)
Those who evaluate results (e.g. physicians, judges)
Blinding avoids bias from knowing treatment
Single-blind: every individual in either one of the
two classes doesn’t know treatments
Double-blind: every individual in both of the two
classes doesn’t know treatments
7
Experiments
Golden standard for experiments:
randomized
double-blind
● comparative
● placebo-controlled.
Even so, could have confounding problems
Confounding variable: variable associated
with both factor & response
Cannot tell whether effect on response is caused
by factor or confounding variable
E.g. Subject’s weight might be is associated with both
insomnia & diet (veg/non-veg)
8
Probability
Many things in life are random (uncertain): roll of
a die, tomorrow’s weather, etc
Don’t know for certain what will happen, but
Could know how likely is something to happen
Probability describes how likely is “something”
(event) to happen.
Probabilities take values between 0 and 1:
A value of 1 means event is sure to happen
A value of 0 means event is not going to happen
9
Probability Terminology
Random phenomenon: process whose result is not
known beforehand
Trial: particular realization of random phenomenon
E.g. Particular roll of a standard die
Outcome: basic result of a trial that cannot be
broken down to simpler results
E.g. Rolling a standard (6-sided) die
E.g. “Rolling a 1” is an outcome
Sample Space: collection of all possible outcomes
Sample space usually denoted by S
E.g. for rolling standard die, S={1, 2, 3, 4, 5, 6}
10
Events
Event: collection of one or more outcomes
Usually denoted by capital letters (A, B, etc)
E.g. A = {2, 4, 6} = “Rolling an even number”
Events can contain just single outcome, e.g. A = {1}
Probabilities are assigned to events
Example: Student’s course grade is random
phenomenon w/ sample space S={A, B, C, D, F}
Are the following events of outcomes?
“Getting a passing grade”
“Getting a failing grade”
11
Empirical Probability
In short term, unpredictable
In long run, predictable
Limiting rel. frequency value
called empirical probability
of event A, denoted by P(A)
0.8
0.6
0.4
Rel. Freq. {Heads}
0.2
Relative Frequency of
# times A occurs
event A:
total # of trials
1.0
Observe sequence of coin tosses (trials) &
count # of times of event A={Heads}
0.0
0
200
400
600
800
# of trials
1000
12
Law of Large Numbers (LLN)
LLN guarantees relative frequency will
eventually settle on specific value
LLN holds if trials are independent (i.e. outcome
of one trial does not influence that of another)
LLN does not apply in the short-run
E.g. Assume you flip fair coin, i.e. P(Heads)=1/2,
and first 10 outcomes are “Tails”.
Is next flip more likely to be “Heads”? NO
LLN only applies in the (infinitely) long run
13
Other Ways to Assign
Probabilities
Theoretical Probability: Sometimes P(A)
can be deduced from mathematical model
For equally likely outcomes (e.g. rolling fair die)
# of outcomes in A
probability is: P A
total # outcomes
Personal / Subjective Probability: Assign
P(A) based on personal beliefs
Typically expert’s views (e.g. sports analyst’s
chances of Maple Leafs winning Stanley Cup)
Most general case (can always be used)
14
Example
Consider rolling 2 fair dice: find (theoretical)
probability of their sum being 4
15