Transcript Document
Intermediate Applied Statistics
STAT 460
Lecture 20, 11/19/2004
Instructor:
Aleksandra (Seša) Slavković
[email protected]
TA:
Wang Yu
[email protected]
Revised schedule
Nov 8 lab on 2-way ANOVA
Nov 10 lecture on two-way
ANOVA and blocking
Nov 12 lecture repeated measure
and review
Post HW9
Nov 15 lab on repeated measures
Nov 17 lecture on categorical
data/logistic regression
HW9 due
Post HW10
Nov 19 lecture on categorical
data/logistic regression
Nov 22 lab on logistic regression
& project II introduction
No class
No class
Thanksgiving
Thanksgiving
Nov 29 lab
Dec 1 lecture
HW10 due
Post HW11
Dec 3 lecture and Quiz
Dec 6 lab
Dec 8 lecture
HW 11 due
Dec 10 lecture & project II
due
Dec 13 Project II due
Last lecture
Categorical Data
This lecture
Categorical Data/Response (ch. 18,19,20)
Odds
Review: Categorical Variable
Notation:
Population proportion = = sometimes we use p
Population size = N
Sample proportion =
= X/n = # with trait / total #
Sample size = n
ˆ
The Rule for Sample Proportions
ˆ
If numerous samples of size n are taken, the frequency curve of the
sample proportions ( ‘s) from the various samples will be
approximately normal with the mean and standard deviation
(1 )
ˆ ~ N( , (1- )/n )
n
One-sample approximate z test and
z-interval for π.
These tests can be extended to test the difference in
parameters π between two groups.
Difference between proportions
These tests can be extended to test the difference in
parameters π between two groups.
Warning:
z-tests for proportions are based on an
approximation. They don’t work for small
samples. It is often said that n is large
enough if
nˆ and n(1- ˆ ) are both greater than 5
Because of improved computing power,
an exact test based on the binomial
distribution rather than the normal is now
available in most software.
Analysis Grid (ref. Handout)
Quantitative
Explanatory
Discrete
Explanatory
Both
Quantitative
Outcome
Regression
ANOVA
Regression
(ANCOVA)
Discrete
Outcome
Logistic
Regression
Chi-Square
Logistic
Test of
Regression
Independence
Contingency Table
A statistical tool for summarizing and displaying results for
categorical variables
A two-way table if for two categorical variables
2x2 Table, for two categorical variables, each with two
categories
Place the counts of each combination of the two variables in
the appropriate cells of the table.
Exploratory variable as labels for the rows, response variable
as labels for the columns.
Example
A university offers only two degree programs: English and
Computer Science. Admission is competitive and there is a
suspicion of discrimination against women in the admission
process. Here is a two-way table of all applicants by sex and
admission status:
These data show an association between the sex of the applicants
and their success in obtaining admission.
Male
Female
Total
Admit
35
20
55
Deny
45
40
85
Total
80
60
140
Marginal & Conditional Distributions
Marginal Distributions:
Exploratory Variable: add up values for the rows; take away
response variable
In our example distribution is: 55, 85, 140
Observed proportions:
‘admit’ = 55/140 = 0.39
‘deny’ = 85/140 = 0.61
NOTE: they add up to 1
Response Variable: add up values for the columns; take away
exploratory variable
In our example distribution is?
Observed proportions are:
Do they add up to 1?
Marginal & Conditional Distributions
Conditional Distribution:
Conditional percentages; what percent of a particular row or a
column a count in a cell is.
Conditional distribution of gender for those admitted:
% of admitted who are male = 35/55 = 0.63 = 63%
% of admitted who are female = ?
What is:
% of male applicants admitted = ?
% of female applicants admitted = ?
Statistical Significance
An observed relationship is statistically significant if the chances of
observing the relationship in the sample when there is no actual
relationship in the population are small (usually less than 5%)
In other words, a relationship is statistically significant if that
relationship is stronger than 95% of the relationships we would
expect to see just by chance.
If we say that there was no statistically significant relationship
found, that does not mean that there is no relationship at all!
Warnings:
If a sample size is small, strong relationships may not achieve
significance
If a sample size is large, even minor relationships could achieve
significance but these might not then have practical importance
Chi-Squared Test (2 Test)
A Chi-Squared Test for independence
The Chi-Squared Statistics (2 ) for contingency table.
Follows 2 distribution
Skewed to the right
Min = 0, Max = infinity
As the strength of observed relationship in the sample increase, the statistic
increases.
It combines info about a strength of the relationship and the sample size into a
one number
Can be calculated for any size contingency table
For 2 x 2 table: if 2 > 3.84 then we have a statistically significant relationship
We either show (2 > 3.84) or fail to show significant relationship (if 2 <
3.8); we either reject (2 > 3.84 ) or fail to reject (2 < 3.84) the claim of
independence between two variables that is our null hypothesis.
H0: variables are independent
HA: variabls are NOT independent
2
The chi-squared distribution with k-1 degrees
of freedom acts as though it was the sum the
squares of k-1 independent Normal(0,1)
distributions. (Not that you need to know.)
See table on pages 1100-1101 in textbook.
0.3
0.25
df=3
0.2
df=4
0.15
df=5
0.1
df=10
0.05
7.5
7.0
6.5
6.0
5.5
5.0
4.5
4.0
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
0
You Must Know:
How to calculate 2 statistic
Compute the expected numbers
Compare the expected and observed numbers
Compute the 2 statistic
How to compare it to 3.84 for 2x2 tables
How to make proper conclusion about statistical relationship
and in general about the question of interest for any two-way
and k-way tables.
2
Calculate
obs
2
Det ermine crit
Decision
based on and df
2
2
if obs
crit
don't reject H0
For our example:
Computing 2 statistic:
Expected number = the number of counts (individuals) that we expect to fall in a particular
cell = (row total)(column total)/(table total)
Expected number of admitted male students = (55 x 80)/140 = 31.42
Expected number of admitted female students = ?
Observed number = the number of counts in the cell
Observed number of admitted male students = 35
Observed number of admitted female students = ?
Compare the observed and expected number :
( observed – expected)2/(expected number)
For male students: (35 - 31.42)2/(31.42) = 0.41
For female students: = ?
Compute the statistic = Sum all the above calculated numbers for all the cells
In our case 2 = 1.58
Compare it to 3.84
Is it statistically significant? Are admission decisions independent of the gender?
Relative Risk, Increased Risk, Odds Ratio
Quantifications of the chances of a particular outcome and
how do these chances change
What are the chances that a randomly selected individual
would fall into a particular category for a categorical variable.
There are two basic ways to express these chances:
Proportions = expressing one category as a proportion
of the total
Proportion of admitted students who are female =
20/55 = 0.36
Odds = comparing one category to another
Odds of being admitted = 55 to 85 = 55/85 to 1
Expressing Proportions & Odds
There are 4 equivalent ways to express proportions:
Percent = Proportion = Probability = Risk
36% (percent) of all admitted students are females
The proportion of females admitted is 0.36
The probability that a female would be admitted is 0.36
The risk for a female to be admitted is 0.36
Odds = expressed by reducing the numbers with and without a
characteristic we are interested in to the smallest possible whole number:
The odds of being admitted = 55 to 85 = 7 to 11 = 7/11 to 1
Going back and forth between proportions and odds:
If the proportion has value p then the odds are: /(1- ) to 1
If the odds of having a characteristic are a to b, then the proportion with
the characteristic is a/(a+b)
Generalized forms for the expressions:
Percentage with the characteristic = (number with the
characteristic/total) x 100%
Proportion with the characteristic = (number with the
characteristic/total)
Probability of having the characteristics = (number with the
characteristic/total)
Risk of having the characteristic = (number with the
characteristic/total)
Odds of having the characteristic = (number with the
characteristic/number without characteristics) to 1
= /(1- )
Types of Risk: Relative risk & Increased Risk
Relative risk = the ratio of the risks for each category of the exploratory variable
Relative risk of being a female based on whether you are rejected or accepted:
Risk for being rejected if you are female = 40/85 = 0.47
Risk of being accepted if you are female = 20/55 = 0.36
Relative risk = 0.47/0.36 = 1.31 to 1
What does this mean?
What does a relative risk of 1 mean?
Increased Risk = usually, the percent increase in risk
Increased risk = (change in risk/original risk) x 100%
Change in risk = 0.47 – 0.36 = 0.11
Original risk = Baseline risk = 0.36
Increased risk = 0.11/0.36x 100% = 0.31 = 31%
There is a 23% increase in the chances of females to be rejected
Increased risk = (relative risk – 1.0) x 100%
Increased risk = (1.31 – 1.0) x 100% = 31%
Odds Ratio
First calculate the odds of having a characteristic versus not
having it:
Odds for female being admitted = 20/35 =0.571429
Odds for female being rejected = 40/45= 0.888889
Then take the ratio of these odds:
Odds ratio = 0.888889/ 0.571429 = 1.5556
Not too close to 1.31, but sometimes it can be close to relative risk
Odds ratio = (upper left * lower right)/(upper right * lower left)
Sometimes you need to reverse denominator and numerator so
that the ratio is greater than 1 (easier to interpret)
Misleading items about Risk/Odds
The baseline risk is missing
The time period of the risk is not identified
The reported risk is not necessarily your risk (relative risk vs. your
risk)
Retrospective vs. Prospective study
Prospective: take a random sample and record success and failure
in the future
Retrospective: take a random sample and record success and
failure that happened in the past
In retrospective study you can meaningfully interpret odds ratio, but
not individual odds
Simpson’s Paradox
Lurking variable = A variable that changes the nature of
association even reverses direction of relationship
between two other variables.
A nature of association changes due to a lurking
variable
In our example we didn’t consider type of a program
(major) as a variable. What happens if we do, and if
construct two separate tables, one for each major?
Example of Simpson’s Paradox
Computer Science admits each 50% of males and females
English takes ¼ of both males and females
Now there doesn’t seem to be an association between sex and
admission decision in either program
Hence, type of program was a lurking variable
Computer Science
English
Male
Female
Admit
30
10
Deny
30
Total
60
Male
Female
Admit
5
10
10
Deny
15
30
20
Total
20
40
Commands in SAS
To create contingency tables, calculate
chi-square statistic, etc…
Statistics/Table Analysis
To run the logistic regression
Statistics/Regression/Logistic
Next
Lab Monday
Categorical Data,
Logistic Regression -- we will work through the lab
together and learn about logistic regression
Project II