p-­Values  and  significance  levels  (false  positive  or  false  alarm  rates)  
 
Let's  say  123  people  in  the  class  toss  a  coin.  Call  it  "Coin  A."    There  are  65  heads.      
 
Then  they  toss  another  coin.  Call  it  "Coin  B."    There  are  72  heads.  
 
Unbeknownst  to  the  class,  one  of  these  coins  is  biased  towards  heads  and  one  is  fair.  
We  should  expect  around  half  heads  with  the  fair  coin  and  more  for  the  biased  coin.    
But  of  course  it  is  possible  that  we  could  get  72  heads  with  a  fair  coin  and  it  is  possible  
to  get  65  heads  with  a  biased  coin.  So  knowing  this  doesn't  really  tell  us  which  one  is  
which.    
 
Our  null  hypothesis  is  that  the  coins  are  fair  –  this  means  that  the  probability  of  getting  
heads  is  50%.  
 
Our  alternative  hypothesis  is  that  the  coins  are  biased  towards  heads.    This  means  that  the  
probability  of  getting  heads  is  >50%.  
 
We  want  to  see  if  we  can  reject  the  null  hypothesis  based  on  the  data,  so  we  generate  a  null  
distribution.    We  use  the  computer  to  simulate  123  people  tossing  fair  coins  and  count  how  
many  heads  there  were  and  repeat  this  1000  times.    This  gives  us  the  following  histogram.    
Notice  that  the  x  axis  in  this  case  is  counts  of  heads  instead  of  fraction  of  heads  (p_hat).  If  
we  know  how  many  total  coin  flips  there  are  (123  here)  it  is  easy  to  convert  counts  to  
fraction  and  back  again:  
 
 
 

Now  we  point  out  where  our  data  falls:  

 
 
Should  we  reject  our  null  hypothesis  in  each  of  these  cases?  
 
The  decision  to  reject  the  null  hypothesis  depends  on  a  cutoff.    We  need  to  decide  on  an  
acceptable  false  positive  rate,  also  called  a  significance  level.    If  the  probability  of  getting  
our  statistic  or  more  extreme  is  less  than  that  significance  level  cutoff,  then  we  will  reject  
the  null  hypothesis.  
 
So  we  need  to  decide  on  a  significance  level  cutoff  (acceptable  false  positive  rate)  and  then  
see  if  the  p-­‐values  for  our  actual  data  are  more  or  less  than  this  cutoff.      
 
The  significance  level  is  the  threshold  p-­‐value.    If  the  p-­‐value  for  the  data  is  less  than  this  
significance  level,  then  we  will  reject  the  null  hypothesis.  
 
Normally,  you  would  decide  on  a  significance  level  at  which  to  reject  the  null  hypothesis.    
There  is  a  convention  to  have  a  significance  level  of  5%,  but  this  is  ultimately  an  arbitrary  
choice.    We  want  to  pick  a  threshold  low  enough  that  we  think  we  would  be  unlikely  to  get  
our  data  result  if  the  null  hypothesis  were  true.  
 
 
 
 
  

 
Let’s  first  calculate  the  p-­‐values  for  each  of  our  samples.  
 
First,  for  Coin  A:  

 
 
So  if  the  null  hypothesis  were  true,  we  would  get  65  heads  or  more  25%  of  the  time.  
 
  

Now  for  Coin  B:

 
If  the  null  hypothesis  were  true,  we  would  get  72  heads  or  more  3%  of  the  time.  
 
So  what  to  do?    We  need  to  pick  a  significance  level.    Remember  that  we  don’t  know  which  
of  our  coins  is  fair  and  which  is  biased.    In  fact,  the  students  who  flipped  them  didn't  know  
if  either  were  fair  or  either  were  biased!  So  we  need  to  pick  a  single  significance  level  that  
we’ll  use  to  test  all  our  samples.    In  a  real  situation,  the  decision  of  what  significance  level  
to  use  should  be  made  before  you  see  your  data.    
 
Remember  that  the  significance  level  is  the  p-­‐value  cutoff  you  use  to  decide  whether  the  
data  is  in  agreement  with  the  null  hypothesis  or  whether  it  disagrees.  If  the  p-­‐value  of  your  
data  is  less  than  the  cutoff  then  your  data  disagrees  with  the  null  hypothesis  (you  reject  the  
null  hypothesis).    If  the  p-­‐value  of  your  data  is  more  than  the  cutoff,  then  your  data  is  in  
agreement  with  the  null  hypothesis  (you  fail  to  reject  the  null  hypothesis).    
 
Agrees  with  =  fail  to  reject  =  consistent  with      the  null  hypothesis  
Disagrees  with  =  reject  =  inconsistent  with            the  null  hypothesis  
 
Let’s  explore  the  consequences  of  this  choice.  
 
  

Before  we  do  this,  I'll  tell  you  which  coin  is  which...  
 
Coin  A  is  biased  and  Coin  B  is  fair!    
 
Let’s  say  we  choose  a  significance  level  of  0.3      
 

 
 
The  red  region  is  now  the  region  where  we  reject  our  null  hypothesis.    With  this  
significance  level,  we  would  correctly  reject  the  null  hypothesis  for  the  biased  coin  but  
incorrectly  reject  it  for  the  fair  coin  (Type  I  error).  
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  

Let’s    say  we  choose  significance  level  of  0.01  

 
Now  we  correctly  fail  to  reject  the  null  hypothesis  for  the  fair  coin,  but  we  also  fail  to  reject  
it  for  the  biased  coin  (Type  II  error).  
 
  

How  about  a  significance  level  of  0.15?  

 
Here  we  are  wrong  about  both  coins!    We  fail  to  reject  for  the  biased  coin  (Type  II)  and  
reject  for  the  fair  coin  (Type  I).      
 
Ultimately,  when  we  are  doing  science,  we  need  to  make  a  statement  about  our  data,  so  we  
have  to  pick  a  single  significance  level.      
 
Let’s  choose  the  standard  0.05  significance  level  cutoff.    Then  we  would  make  a  statement  
like  this:  
 
We  did  two  experiments  with  two  different  coins    In  the  first,  we  got  65  heads  and  in  the  
second  we  got  72  heads.    The  p-­‐value  for  65  heads  is  0.25.    The  p-­‐value  for  72  heads  is  0.03.    
We  chose  a  significance  level  (a  p-­‐value  cutoff)  of  0.05.    Based  on  this  significance  level,  we  
fail  to  reject  the  null  hypothesis  for  the  coin  that  gave  us  65  heads.    We  reject  the  null  
hypothesis  for  the  coin  that  gave  us  72  heads.    Therefore,  our  data  is  consistent  with  the  coin  
that  gave  us  65  heads  being  fair  and  consistent  with  the  coin  that  gave  us  72  heads  being  
biased.  
 
When  we  do  a  real  experiment,  we  wouldn't  know  whether  the  coins  were  fair  or  biased.  So  
we  could  get  unlucky  (as  I  set  us  up  to  be  here)  and  just  be  completely  wrong!    Setting  a  
significance  level  lets  us  control  the  tradeof  between  false  alarms  (false  positives,  Type  I  
errors)  and  missed  opportunities  (false  negatives,  Type  II  errors).