Page 1 Lecture 7 – MEEN 260 Dr. R.

Tafreshi

Lecture 7: Statistical Data Analysis – Student’s t distribution

Lecture Outline:
 Normal Probability Distributions
 To understand the difference between Two-Sided vs. One-Sided confidence interval
 To understand where to use Student’s t distribution
o Confidence Interval for Small Samples
 Comparing Two Sample Sets
 Examples

Announcements / Reminders / Discussions:

 Conference Papers…

Reading Assignments:
 Chapter 3 of textbook by Beckwith
o Specially additional problem solving examples in the textbook:
Examples 3.4 – 3.10

Next Lecture: Propagation of Uncertainty

Chi-Squared Distribution
Page 2 Lecture 7 – MEEN 260 Dr. R. Tafreshi

Review:
Note: The textbook (and many statistics textbooks) use the notation zc /2
Example: zc /2 = z0.475 c: Confidence interval
This refers to the value of z when:
zc/2


0
p( z )dz  c / 2  0.95 / 2  0.475

  zc/2
P(   zc /2  x    zc /2 )  
  zc/2
p( x)dx

– 1 = 68.3%
– 1.96  = 95.0%

– 2 = 95.5%
– 2.575  = 99.0%

– 3 = 99.73%
– 6 = 99.9999998%

From: http://www.answers.com/topic/normal-distribution

Two-Sided vs. One-Sided confidence interval

“I am 95% sure the average students’ age at TAMUQ is 20.5 ± 2.5 yrs”
18    23 Two-sided confidence interval
s s
x  zc /2 x    x  zc /2 x
n n

“I am 97.5% sure the average students‘ age at TAMUQ is at least 18 yrs”

  18 One-sided confidence interval
s
xz x 
n
Page 3 Lecture 7 – MEEN 260 Dr. R. Tafreshi

Samples
 Samples should be representative of the complete set (the population) of values
from which it has been chosen
 The larger the random sample the more representative of the population it is likely
to be
 What if we can’t afford to take ~100 samples?
o Time
o $$$
o Opportunity

 For small samples (<30) our sampling assumptions no longer are valid
 Solution: “Student’s t” distribution

Student’s t distribution
 Early 1900s Guiness brewery hired top
college graduates in statistics and
biochemistry to improve their manufacturing
process
 At Guiness, William Gosset developed a
way to test the quality of their product with
only a few samples, but because of his
employer’s desire for secrecy, he published his work under the name of “Student”.

 The t-test assumes that the underlying population follows a Gaussian distribution,
but accounts for the fact that when using S x » s with only a few samples, we are
underestimating the std. dev.
 Thus the values in the t-table are greater than those in the z-table.

 The t-statistic depends not only on the assumed confidence interval, but on the
number of samples taken. The “degrees of freedom” of the t-statistic is ν = n −1.
Page 4 Lecture 7 – MEEN 260 Dr. R. Tafreshi

Confidence Interval for Small Samples

 Student’s t distribution for calculating confidence intervals
S S
x  t /2,v x    x  t /2,v x
n n
o Where t /2,v is the t-statistic that we get from the tables

o What is ν?
 The degrees of freedom equals sample size minus 1:
ν = n −1

o What is  ?
 The significance level, the complement of the confidence interval:
 =1–c
Page 5 Lecture 7 – MEEN 260 Dr. R. Tafreshi

Example – Prob. 3.22 of textbook

The manufacturer of inexpensive outdoor thermometers checks a sample of nine against a
68 F standard. The following results were obtained:
68.5 67.5 67 69 68 67 67.5 69 69

 Calculate the range within which the population mean is expected to lie with a
confidence level of 95%

Solution:
 Given: x = 68.06, S x = 0.846, n = 9,  = 0.05
 From table …
S S
 Confidence Interval: x  t /2,v x    x  t /2,v x
n n

… 67.41    68.71

Table 3.5 of textbook: Student's t-Distribution (Values of t ,v )

30 1.301 1.697 2.042 2.75

Page 6 Lecture 7 – MEEN 260 Dr. R. Tafreshi

Example – Prob. 3.22 (modified) of textbook

The manufacturer of inexpensive outdoor thermometers checks a sample of nine against a
68 F standard. The following results were obtained:
68.5 67.5 67 69 68 67 67.5 69 69

 Calculate the lower limit above which the population mean is expected to lie with a
confidence level of 99%

Solution:
 Given: x = 68.06, s x = 0.846, n = 9,  = 0.01

 From table …

sx
 Confidence Interval: x  t ,v  67.24  
n

 See page 57 of your textbook for more details about one-sided vs. two-sided
confidence interval

Summary of distributions:
 Confidence interval of all possible measurements
x
  zc/2  x    zc/2 ; z


 Confidence interval containing the “true value” (population mean)

s s x 
x  zc /2 x    x  zc /2 x ; z
n n sx / n

 Confidence interval containing the “true value” (population mean)

s s x 
x  t /2,v x    x  t /2,v x ; t
n n sx / n

 One-sided confidence intervals,

s sx
o e.g. xz x  ;   x  t ,v
n n
Page 7 Lecture 7 – MEEN 260 Dr. R. Tafreshi

Comparing Two Sample Sets

 How do we compare two sets of sample to see if they are different (statistically)?

 Examples:
o Compare the quality of manufactured goods between day-shift workers and
night-shift workers
o Compare the quality of products between two different companies

 Assume:
o x1 , x2 : the sample means
o s1 , s2 : the sample deviations
o n1 , n2 : the number of samples in each set

 Use a t-test where:

2
 ( s1 ) 2 ( s2 ) 2 
 n  n 
x1  x2
t x1  x2  ; v  1 2 2 
2
( s1 ) 2 ( s2 ) 2  ( s1 ) 
2
 ( s2 ) 2 
  n   
n1 n2  1    n2 
n1  1 n2  1
(v is rounded down to the nearest integer)

 To verify if the sets are significantly different (statistically):

If t x1  x2 > t /2,v They are different with a confidence level (1-  )
Page 8 Lecture 7 – MEEN 260 Dr. R. Tafreshi

Example: assume two manufacturing machines with products with

x1 = 1076.75 ( s1 )2 = 29.30 n1 = 12
x2 = 1072.33 ( s2 )2 = 26.24 n2 = 12
2
 ( s1 ) 2 ( s2 ) 2 
 n  n 
x1  x2
t x1  x2  ; v  1 2 2 
2
( s1 ) 2 ( s2 ) 2  ( s1 ) 
2
 ( s2 ) 2 
  n   
n1 n2  1    n2 
n1  1 n2  1
1076.75  1072.33
t x1  x2  texp   2.055 ;
(29.30)2 (26.24)2

12 12
2
 (29.30) 2 (26.24) 2 
 12  12 
v    21.9
2 2 2
 (29.30)   (26.24) 2 
 12   
    12 
11 11

 Compare texp and t /2,v for  = 0.10:

texp  2.055 > t0.10/2,21 =1.721
o Statistically different with 10% significance level
 (We are 90% sure they are different)

But for  = 0.05:

texp  2.055 < t0.05/2,21 =2.080
o Not statistically different with 5% significance level
 (We cannot be 95% sure that they are different)
Research Example:

“Predicting Epileptic Seizures in Scalp EEG Based on a Variational Bayesian Gaussian Mixture Model of Zero-Crossing
Intervals”, Ali Shahidi Zandi, Reza Tafreshi, Manouchehr Javidan, and Guy A. Dumont, IEEE TBME Journal, 2013.

* Some material based on original slides by Drs. Reza Langari and Bryan Rasmussen, TAMU MEEN.
Page 9 Lecture 7 – MEEN 260 Dr. R. Tafreshi

Table 3.2