Hypothesis  

Tests  for  
8 One  Sample

Chapter  8                Stat  4570/5570              

Material  from  Devore’s  book  (Ed  8),  and  Cengage
Statistical  Hypotheses
A  statistical  hypothesis:  a  claim about  the  value  of  a  parameter,  
population  characteristic  (could  be  a  combination  of  parameters),  or  
about  the  form  of  an  entire  probability  distribution.  

Examples:  
• H:  µ =  75  cents,  where  µ is  the  true  population  average  of  daily  
per-­student  candy+soda  expenses  in  US  high  schools
• H:  p <  .10,  where  p  is  the  population  proportion  of  defective  
helmets  for  a  given  manufacturer  
• If  µ1 and  µ2 denote  the  true  average  breaking  strengths  of  two  
different  types  of  twine,  one  hypothesis  might  be  the  assertion  
that  µ1   – µ2   =  0,  and  another  is  the  statement    µ1  – µ2 >  5  

2
Null  vs  Alternative  Hypotheses
In  any  hypothesis-­testing  problem,  there  are  always  two  competing  
hypotheses  under  consideration:

1. The  status  quo  (null)  hypothesis  

2. The  research  (alternative)  hypothesis

For  example,  
µ =  .75 versus  µ ≠  .75
p ≥ .10  versus  p <  .10

The  objective  of  hypothesis  testing  is  to  decide,  based  on  sample  
information,  if  the  alternative  hypotheses  is  actually  supported  by  the  
data.
We  usually  do  new  research  to  challenge  the  existing  (accepted)  beliefs.
3
Burden  of  Proof
The  burden   of  proof is  placed  on  those  who  believe  in  the  
alternative  claim.  

In  testing  statistical  hypotheses,  the  problem  will  be  formulated  

so  that  one  of  the  claims  is  initially   favored.  

This  initially  favored  claim  (H0)  will  not  be  rejected  in  favor  of  the  
alternative  claim  (Ha or  H1)  unless  the  sample  evidence  
contradicts  it  and  provides  strong  support  for  the  alternative  
assertion.  
If  the  sample  does  not  strongly  contradict  H0,  we  will  continue  to  
believe  in  the  plausibility  of  the  null  hypothesis.
The  two  possible   conclusions:   1)  reject  H0
2)  fail  to  reject  H0.

4
No  proof…  only  evidence
We  can  never prove  that  a  hypothesis  is  true or  not  true.  

We  can  only  conclude  that  it  is or  is  not supported  by  the  data.

A  test  of  hypotheses is  a  method  for  using  sample  data to  decide  

whether  the  null  hypothesis  should  be  rejected in  favor  of  the  
alternative.  

Thus  we  might  test  the  null  hypothesis    H0:  µ =  .75 against  the  
alternative    Ha: µ ≠  .75.  Only  if  sample  data  strongly  suggests that  µ is  
something  other  than  0.75  should  the  null  hypothesis  be  rejected.  

In  the  absence  of  such  evidence,  H0 should  not  be  rejected,  since  it  is  
still  considered  plausible.  

5
Why  favor  the  null  so  much?
Why  be  so  committed  to  the  null  hypothesis?  
• sometimes  we  do  not  want  to  accept  a  particular  
assertion  unless  (or  until)  data  can  show  strong  support  
• reluctance  (cost,  time)  to  change

Example:  Suppose  a  company  is  considering  putting  a  new  

type  of  coating  on  bearings  that  it  produces.  

The  true  average  wear  life  with  the  current  coating  is  
known  to  be  1000  hours.  With  µ denoting  the  true  average  
life  for  the  new  coating,  the  company  would  not  want  to  
make  any  (costly)  changes  unless  evidence  strongly  
suggested  that  µ exceeds  1000.
6
Hypotheses  and  Test  Procedures
An  appropriate  problem  formulation  would  involve  testing  
H0:  µ =  1000  against  Ha:  µ >  1000.  

The  conclusion  that  a  change  is  justified  is  identified  with  

Ha,  and  it  would  take  conclusive  evidence  to  justify  
rejecting  H0 and  switching  to  the  new  coating.

Scientific  research  often  involves  trying  to  decide  whether  a  

current  theory  should  be  replaced,  or  “elaborated  upon.”

7
Hypotheses  and  Test  Procedures
An  appropriate  problem  formulation  would  involve  testing  
the  hypothesis:

H0:  µ =  1000  against  Ha:  µ >  1000.  

The  conclusion  that  “a  change  is  justified”  is  identified  with  
Ha,  and  it  would  take  conclusive  evidence  to  justify  
rejecting  H0 and  switching  to  the  new  coating.

Scientific  research  often  involves  trying  to  decide  whether  a  

current  theory  should  be  replaced,  or  “elaborated  upon”

8
Hypotheses  and  Test  Procedures
The  word  null  means  “of  no  value,  effect,  or  
consequence,” which  suggests  that  H0 should  be  identified  
with  the  hypothesis  of  no  change  (from  current  opinion),  no  
difference,  no  improvement,  etc.

Example:  10%  of  all  circuit  boards  produced  by  a  certain  

manufacturer  during  a  recent  period  were  defective.  

An  engineer  has  suggested  a  change  in  the  production  

process  in  the  belief  that  it  will  result  in  a  reduced  defective  
rate.    Let  p  denote  the  true  proportion  of  defective  boards  
resulting  from  the  changed  process.    What  does  the  
hypothesis  look  like?

9
Hypotheses  and  Test  Procedures
The  alternative  to  the  null  hypothesis  H0:  θ =  θ0 will  look  like  
one  of  the  following  three  assertions:

1. Ha:  θ ≠  θ0
2. Ha:  θ >  θ0 (in  which  case  the  null  hypothesis  is  θ ≤ θ0)
3. Ha:  θ <  θ0    (in  which  case  the  null  hypothesis  is    θ ≥ θ0)

• The  equality  sign  is  always with  the  null  hypothesis.

• It  is  typically  easier  to  determine  the  alternate  hypothesis  first  then  
the  complementary  statement  is  designated  as  the  null  hypothesis

• The  alternate  hypothesis  is  the  claim  for  which  we  are  seeking  
statistical  proof  
10
Test  Procedures
A  test  procedure is  a  rule,  based  on  sample  data,  for  
deciding  whether  to  reject  H0.  

Example  -­-­ the  circuit  board  problem:  

A  test  of  H0:  p =  .10  versus  Ha:  p <  .10  

We  test  this  on  a  random  sample  of  n =  200  boards.  

How  do  we  use  the  sample  of  200?  

11
Test  Procedures
Testing  procedure  has  two  constituents:  

(1)  a  test  statistic, or  function  of  the  sample  data  which  will  
be  used  to  make  a  decision,  and

(2)  a  rejection  (or  critical)  region consisting  of  those  test  

statistic values  for  which  H0 will  be  rejected  in  favor  of  Ha.  

So  if  we  have  decided  we  can  reject  H0 if  x ≤ 15  – then  the  
rejection  region  consists  of  {0,  1,  2,…,  15}.  Then  H0 will  not
be  rejected  if  x  =  16,  17,.  .  .  ,199,  or  200.

12
Errors  in  Hypothesis  Testing
The  basis  for  choosing  a  particular  rejection  region  lies  in  
consideration  of  the  errors  that  one  might  be  faced  with  in  
drawing  a  conclusion.  

Consider  the  rejection  region  x ≤ 15 in  the  circuit  board  

problem.  Even  when  H0:  p =  .10  is  true,  it  might  happen  
that  an  unusual  sample  results  in  x =  13,  so  that  H0 is  
erroneously  rejected.  

On  the  other  hand,  even  when  Ha:  p <  .10  is  true,  an  
unusual  sample  might  yield  x =  20,  in  which  case  H0 would  
not  be  rejected—again  an  incorrect  conclusion.  
13
Errors  in  Hypothesis  Testing
Definition  
• A  type  I  error  is  when  the  null  hypothesis  is  rejected,  
but  it  is  true.
• A  type  II  error  is  not  rejecting  H0 when  H0 is  false.

This  is  very  similar  in  spirit  to  our  diagnostic  test  examples
• False  negative  test  =  type  I  error
• False  positive  test  =  type  II  error

14
Type  I  error  in  hypothesis  testing
Usually:  Specify  the  largest  value  of    α that  can  be  
tolerated,  and  then  find  a  rejection  region  with  that  α.  

The  resulting  value  of  α is  often  referred  to  as  the  

significance  level  of  the  test.  

Traditional  levels  of  significance  are  .10,  .05,  and  .01,  

though  the  level  in  any  particular  problem  will  depend  on  
the  seriousness  of  a  type  I  error—

The  more  serious  the  type  I  error,  the  smaller  the  

significance  level  should  be.  
15
Example  (Type  I  Error)
Let  µ denote  the  true  average  nicotine  content  of  brand  B  
cigarettes.  The  objective  is  to  test  
Ho:  µ =  1.5  versus  Ha:  µ >  1.5  
based  on  a  random  sample  X1,  X2,.  .  .  ,  X32 of  nicotine  
content.  

Suppose  the  distribution  of  nicotine  content  is  known  to  be  
normal  with  σ =  .20.  

Then  X is  normally  distributed  with  mean  value  µx =  µ and  

standard  deviation  σx =  .20/ = .0354.

16
Example  (Type  I  Error) cont’d

Rather  than  use  X itself  as  the  test  statistic,  let’s  

standardize  X,  assuming  that  H0 is  true.

Test  statistic:  Z  =

Z  expresses  the  distance  between  X and  its  expected  

value  (when  H0 is  true) as  some  number  of  standard  
deviations  of  the  sample  mean.  

17
Example  (Type  I  Error) cont’d

As  Ha:  µ >  1.5,  the  form  of  the  rejection  region  is  z  ≥ c.  
What  is  c  so  that  α =  0.05?

When  H0 is  true,  Z  has  a  standard  normal  distribution.  Thus

α =  P(type  I  error)  =  P(rejecting  H0 when  H0 is  true)

=  P(Z  ≥ c  when  Z  ~  N(0,  1))

The  value  c  must  capture  upper-­tail  area  .05  under  the  z  

curve.  So,   c =  z.05  =  1.645.

18
Case  I:  Testing  means  of  a  normal  population  with  known  σ

Null  hypothesis:  H0  :  µ =  µ0

Test  statistic  value  :                                                              

Alternative  Hypothesis Rejection   Region   for  Level   α Test

19
Case  I:  Testing  means  of  a  normal  population  with  known  σ

Rejection  regions   for  z tests:  (a)  upper-­tailed   test;;  (b)  lower-­tailed   test;;  (c)  two-­tailed   test
20
Type  II  Error  Example  
A  certain  type  of  automobile  is  known  to  sustain  no  visible  
damage  25%  of  the  time  in  10-­mph  crash  tests.  A  modified  
bumper  design  has  been  proposed  in  an  effort  to  increase  
this  percentage.  

Let  p  denote  the  proportion  of  all  10-­mph  crashes  with  this  
new  bumper  that  result  in  no  visible  damage.  

How  do  we  examine  a  hypothesis  test  for  n  =  20  

independent   crashes  with  the  new  bumper  design?

21
Type  II  Error  Example  1 cont’d

The  accompanying  table  displays  β for  selected  values  of  p  

(each  calculated  for  the  rejection  region  R8).  

Clearly,  β decreases  as  the  value  of  p  moves  farther  to  the  
right  of  the  null  value  .25.  

Intuitively,  the  greater  the  departure  from  H0,  the  less  likely  
it  is  that  such  a  departure  will  not  be  detected.  
Thus,  1-­ β is  often  called  the  “power  of  the  test”
22
 Errors  in  Hypothesis  Testing
We  can  also  obtain  a  smaller  value  of  α -­-­ the  probability  that  
the  null  will  be  incorrectly  rejected  – by  decreasing  the  size  of  
the  rejection  region.  

However,  this  results  in  a  larger  value  of  β for  all  parameter  
values  consistent  with  Ha.

No  rejection  region  that  will  simultaneously  make  both  α

and  all  β ’s  small.    A  region  must  be  chosen  to  strike  a  
compromise  between  α and β.

23
Case  II:  Large  sample  tests  for  means

When  the  sample  size  is  large,  the  z  tests  for  case  I  are  
easily  modified  to  yield  valid  test  procedures  without  
requiring  either  a  normal  population  distribution  or  
known  σ.

Earlier  we  used  the  key  result  to  justify  large-­sample  

confidence  intervals:  
A  large  n  (>40)  implies  that  the  standardized  variable

has  approximately  a  standard  normal  distribution.

24
Case  III:  Testing  means  of  a  
Normal population  with  unknown  σ,  and  small  n

The  One-­Sample  t  Test

Null  hypothesis:  H0:  µ =  µ0

Test  statistic  value:                                          

Alternative  Hypothesis Rejection   Region   for  a  Level   α

Test

25
CI  and  Hypotheses cont’d
Rejection  regions  have  a  lot  in  common  with  confidence  intervals.

Source:  s hex.org
26
Proportions:  Large-­Sample  Tests
The  estimator                            is  unbiased                                ,  has  
approximately  a  normal  distribution,  and  its  standard  
deviation  is

When  H0 is  true,                                and                                                                so          

does  not  involve  any  unknown  parameters.  It  then  follows  
that  when  n  is  large  and  H0 is  true,  the  test  statistic  

has  approximately  a  standard  normal  distribution.

27
Proportions:  Large-­Sample  Tests
Alternative  Hypothesis Rejection  Region

Ha:  p  >  p0 z  ≥ zα (upper-­tailed)

Ha:  p  <  p0   z  ≤ –zα (lower-­tailed)  

Ha:  p  ≠  p0 either  z  ≥ zα/2  

or  z  ≤ –zα/2 (two-­tailed)

These  test  procedures  are  valid  provided  that  np0 ≥ 10 and

n(1  – p0)  ≥ 10.

28
Example  
Natural  cork  in  wine  bottles  is  subject  to  deterioration,  and  
as  a  result  wine  in  such  bottles  may  experience  
contamination.  

The  article  “Effects  of  Bottle  Closure  Type  on  Consumer  

Perceptions  of  Wine  Quality” (Amer. J. of  Enology  and  
Viticulture, 2007:  182–191)  reported  that,  in  a  tasting  of  
commercial  chardonnays,  16  of  91  bottles  were  considered  
spoiled  to  some  extent  by  cork-­associated  characteristics.

Does  this  data  provide  strong  evidence  for  concluding  that  

more  than  15%  of  all  such  bottles  are  contaminated  in  this  
way?    Use  a  significance  level  equal  to  0.10.

29
P-­Values
The  P-­value  is  a  probability  of  observing  values  of  the  test  
statistic  that  are  as  contradictory  or  even  more  
contradictory  to  H0  as  the  test  statistic  obtained  in  our  
sample.

• This  probability  is  calculated  assuming  that  the  null

hypothesis  is  true.
• Beware:  The  P-­value  is  not  the  probability  that  H0  
is  true,  nor  is  it  an  error  probability!
• The  P-­value  is  between  0  and  1.

30
Example
Urban  storm  water  can  be  contaminated  by  many  sources,  
including  discarded  batteries.  When  ruptured,  these  batteries  
release  metals  of  environmental  significance.

The  article  “Urban  Battery  Litter” (J.  of  Environ.  Engr., 2009:  
46–57)  presented  summary  data  for  characteristics  of  a  
variety  of  batteries  found  in  urban  areas  around  Cleveland.

A  sample  of  51  Panasonic  AAA  batteries  gave  a  sample  mean  

zinc  mass  of  2.06g  and  a  sample  standard  deviation  of  
0.141g.

Does  this  data  provide  compelling  evidence  for  

concluding  that  the  population  mean  zinc  mass  exceeds  
2.0g?
31
P-­Values
More  generally,  the  smaller  the  P-­value,  the  more  
evidence  there  is  in  the  sample  data  against  the  null  
hypothesis  and  for  the  alternative  hypothesis.  

The  p-­value  measures  the  “extremeness” of  the  sample.

That  is,  H0  should  be  rejected  in  favor  of  Ha when  the  P-­
value  is  sufficiently  small  (such  large  sample  statistic  is  
unlikely  if  the  null  is  true).  

So  what  constitutes  “sufficiently  small”?

What  is  “extreme” enough?

32
Decision  rule  based  on  the  P-­value
Select  a  significance  level  α (as  before,  the  desired  type  I  
error  probability),  then   α defines  the  rejection  region.

Then  the  decision  rule  is:

reject  H0  if  P-­value  ≤ α

do  not reject  H0  if  P-­value  >  α

Thus  if  the  P-­value  exceeds  the  chosen  significance  level,  

the  null  hypothesis  cannot  be  rejected  at  that  level.

Note,  the  P-­value  can  be  thought  of  as  the  smallest  
significance  level at  which  H0 can  be  rejected.
33
P-­Values
In  the  previous  example,  we  calculated  the  P-­value  =  
.0012.  Then  using  a  significance  level  of  .01,  we  would  
reject  the  null  hypothesis  in  favor  of  the  alternative  
hypothesis  because  .0012  ≤ .01.

However,  suppose  we  select  a  significance  level  of  0.001,  

which  requires  far  more  substantial  evidence  from  the  data  
before  H0  can  be  rejected.  In  that  case  we  would  not  reject  
H0  because  .0012  >  .001.

This  is  why  we  cannot  change  significance  level  after  we  
see  the  data  – NOT  ALLOWED  though  tempting!

34
P-­Values  for  z Tests
The  calculation  of  the  P-­value  depends  on  whether  the  test  
is  upper-­,  lower-­,  or  two-­tailed.  

Each  of  these  is  the  probability  of  getting  a  value  at  least  as  
extreme  as  what  was  obtained  (assuming  H0 true).

35
P-­Values  for  z Tests
The  three  cases  are  illustrated  in  Figure  8.9.

36
P-­Values  for  z Tests cont’d

37
Example
The  target  thickness  for  silicon  wafers  used  in  a  certain  
type  of  integrated  circuit  is  245  µ m.  

A  sample  of  50  wafers  is  obtained  and  the  thickness  of  
each  one  is  determined,  resulting  in  a  sample  mean  
thickness  of  246.18  µ m  and  a  sample  standard  deviation  of  
3.60  µ m.  

Does  this  data  suggest  that  true  average  wafer  thickness  is
something  other  than  the  target  value?    Use  a  significance  
level  of  .01.

38
P-­Values  for  t Tests
Just  as  the  P-­value  for  a  z test  is  the  area  under  the  z
curve,  the  P-­value  for  a  t  test  will  be  the  area  under  the  t-­
curve.

The  number  of  df  for  the  one-­sample  t  test  is  n  – 1.

39
P-­Values  for  t Tests cont’d

40
P-­Values  for  t Tests
The  table  of  t  critical  values  used  previously  for  confidence  
and  prediction  intervals  doesn’t  contain  enough  information  
about  any  particular  t  distribution  to  allow  for  accurate  
determination  of  desired  areas.  

There  another  t  table  in  Appendix  Table  A.8,  one  that  

contains  a  tabulation  of  upper-­tail  t-­curve  areas.  But  we  can  
also  use  other  tables  to  get  an  approximation  of  the  p-­value  
(software  is  the  best).

41
More  on  Interpreting  P-­values

42
How  are  p-­values  distributed? cont’d

Figure  below  shows  a  histogram  of  the  10,000  P-­values  from  a  simulation  
experiment  under  a  null  μ =  20  (with  n  =  4  and  σ =  2).  

When  H0 is  true,  the  probability  distribution  of  the  P-­value  is  a  uniform  
distribution  on  the  interval  from  0  to  1.

43
Example cont’d

About  4.5%  of  these  P-­values  are  in  the  first  class  interval  
from  0  to  .05.  

Thus  when  using  a  significance  level  of  .05,  the  null  

hypothesis  is  rejected  in  roughly  4.5%  of  these  10,000  
tests.  

If  we  continued  to  generate  samples  and  carry  out  the  test  
for  each  sample  at  significance  level  .05,  in  the  long  run  5%  
of  the  P-­values  would  be  in  the  first  class  interval.  

44
Example   cont’d

A  histogram  of  the  P-­values  when  we  simulate  under  an  alternative  
hypothesis.  There  is  a  much  greater  tendency  for  the  P-­value  to  be  
small  (closer  to  0)  when   µ =  21  than  when  µ =  20.

(b)  μ =  21
45
Example   cont’d

Again  H0 is  rejected  at  significance  level  .05  whenever

the  P-­value  is  at  most  .05  (in  the  first  bin).  

Unfortunately,  this  is  the  case  for  only  about  19%  of  the  
P-­values.  So  only  about  19%  of  the  10,000  tests  correctly
reject  the  null  hypothesis;;  for  the  other  81%,  a  type  II  error  
is  committed.  

The  difficulty  is  that  the  sample  size  is  quite  small  and  21  is  
not  very  different  from  the  value  asserted  by  the  null  
hypothesis.

46
Example   cont’d

Figure  below  illustrates  what  happens  to  the  P-­value  when  

H0 is  false  because  µ =  22.

(c)  μ =  22
47
Example   cont’d

The  histogram  is  even  more  concentrated  toward  values

close  to  0  than  was  the  case  when  µ =  21.

In  general,  as  µ moves  further  to  the  right  of  the  null  value  
20,  the  distribution  of  the  P-­value  will  become  more  and  
more  concentrated  on  values  close  to  0.  

Even  here  a  bit  fewer  than  50%  of  the  P-­values  are  smaller  
than  .05.  So  it  is  still  slightly  more  likely  than  not  that  the  
null  hypothesis  is  incorrectly  not  rejected.  Only  for  values  of  
µ much  larger  than  20  (e.g.,  at  least  24  or  25)  is  it  highly  
likely  that  the  P-­value  will  be  smaller  than  .05  and  thus  give  
the  correct  conclusion.  
48
Statistical  Versus  Practical  Significance
When  using

one  must  be  especially  careful  – with  large  n,  what  

happens  to  z?    How  does  this  affect  hypothesis  testing?

49
R  code

50