analyzing statistical results
TRANSCRIPT
Interpreting Analysis Results
1
2
Chi-square
Gamma
Z and t Statistics
Correlation coefficient and the R2 value
Significance and the p-value
1-sided and 2-sided tests
Our Test Statistics
Variable Types• Nominal, Ordinal, and Interval
Nominal Example: Type of Anesthetic (e.g. Local, General, and Regional)
Ordinal Example: Hospital Trauma Levels (e.g. Level-1, Level-2, Level-3)
Interval Example: Drug Dosage (e.g. mg of substance)
• There are multiple kinds of tests for each type of variable depending on assumptions being made and what is being tested for (e.g. means, variances, association, etc.)
Our focus is on ordinal variablesand testing for associations
3
Does not measure strength of an association, but, rather, it indicates how much evidence there is that the variables are or are not independent
The c2 statistic ranges from 0 to ∞ (always ≥ 0)
The larger the c2 value, the more confident we can be that the variables are not independence
c2 Test of Independence
Example: Assume we have two ordinal variables (Hospital Size and the Relative Cost of the Laryngoscope being used):
H0: There is no association between variables (c2 value is near zero)
H1: There is an association between variables
4
Sample 1
Statistical Independence
5
The c2 Test Statistic
Remember: This number does not reflect the strength of an association. It only provides evidence that an association actually exists.
6
(45%) (40%) (15%)
7753
149
4769
132
18
4926
“Observed Value”
“Expected Value”
Sample 2
The c2 Distribution Okay…so what does c2 = 83 tell us?
With 4 degrees of freedom (d.f.) and an a = 0.10, our benchmark (critical value) for dependence is 7.78
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
0 5 10 15 20 25 30 35 40
d.f. = 4
7.78
Area = a = 0.10
7
The c2 Distribution
8
H0: There is no association between variables (c2 value is near zero)
H1: There is an association between variables
We reject H0 if our c2 value is > our “benchmark” value of 7.78
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
0 5 10 15 20 25 30 35 40
d.f. = 4
7.78
Area = a = 0.10
-0.02
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0 10 20 30 40
d.f. = 5 d.f. = 10 d.f. = 20
Increasing d.f.
Degrees of Freedom = (# of rows – 1) x (# of columns – 1)
Increasing d.f.
3 x 3
5 x 5
9
The c2 Distribution
So, how do we determine the strength and direction of an established association
between variables?
10
Concordant
(x1, y2)
(x2, y3)
X
Y
11
Discordant
(x1, y2)
(x2, y1)
X
Y
12
The Gamma Value (g) Appropriate for ordinal-by-ordinal data
Generally more powerful test statistic than the c2 test statistic (i.e. we are able to detect significant differences with greater ease)
Because ordinal variables can be ordered, we can talk about the direction of association between them (i.e. positive or negative)
13
C = Concordant Pairs
D = Discordant Pairs
lies between -1 and 1, where a value of 0means that there is no relation between the two ordinal variables and | | = 1 represents the strongest associations.
Measure of Association
14
g follows a normal distribution such that our z statistic becomes:
Comparing our z statistic with a, we can determine if the variables are statistically independent
Testing Independence with g
z =s.e.
Standard Error (s.e.) = sample standard deviation / √ n
15
H0: There is no association between variables (g value is near zero)
H1: There is an association between variables
Testing Independence with g
Using g to test for independence does not rely on the degrees of freedom
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
-4 -3 -2 -1 0 1 2 3 4
Assuming an a of 0.10 then our benchmark is z0 = 1.285
If z > z0 = 1.285 then reject H0
Rejection of the H0 means that the variables are not independent (see previous slide)
1.285
Area = a = 0.10
16
Why do we not just use g and forget about c2 since g is a more
powerful statistic?
17
c2 and g An ordinal measure of association (i.e. g ) may equal 0
when the variables are actually statistically dependent
c2 = 437, g = 0
18
Association vs. Causation
If then ?
Stock Market
Hemline Theory = Association without Causation
19
Association vs. Causation Causation can often be obscure or counterintuitive so
how do we tell the difference from Association? Performing controlled comparisons
Increasing the variable resolutions (i.e. the number
of data points)
Sequence Analysis (time becomes a variable)
20