5

Significance, Partial Tests, Degrees of Freedom

  • Upload
    barry

  • View
    77

  • Download
    1

Embed Size (px)

DESCRIPTION

My original notes on the three (and a half) tests for multiple regression, followed by a more simple outline and comments on degrees of freedom.

Citation preview

Page 1: Significance, Partial Tests, Degrees of Freedom
Page 2: Significance, Partial Tests, Degrees of Freedom
Page 3: Significance, Partial Tests, Degrees of Freedom

     

Page 4: Significance, Partial Tests, Degrees of Freedom

That  was  a  more  or  less  complete  exposition  of  the  problem.    Here’s  a  simpler  account.    All  cases  assume  we  want  95%  confidence  (i.e.  5%  significance  level).    

TEST  1:  SIGNIFICANT  OVERALL  REGRESSION  Objective:  determine  whether  the  full  model  is  any  better  than  just  using  the  average  of  y  values  and  calling  it  a  day.  Ho:  B1=B2=…=Bk=0  Ha:  at  least  one  of  those  slopes  =/=  0  Test  Stat  (F  stat):  F=  MSR/MSE  Critical  Value:  F(k,  n-­‐‑k-­‐‑1,  .95)    TEST  2:  PARTIAL  F  TEST  Objective:  determine  whether  a  full  model  could  be  improved  by  lopping  off  one  of  the  variables.  Ho:  B*=0  Ha:  B*=/=0  F  Stat:  Extra  SS  (B*)/MSE  or  T  Stat:  Bhat*/SE  Bhat*  Critical  F  Value:  F  (1,  n-­‐‑p-­‐‑2,  .95)  Critical  T  Value:  T  (n-­‐‑p-­‐‑2,  .025)  assuming  a  two-­‐‑tailed  test  (where  Ho:  something=something  else)    TEST  3:  MULTIPLE  PARTIAL  F  TEST  Objective:  determine  whether  we  can  cut  a  bunch  of  stuff  and  still  have  a  workable  model.  First  Assumption:  Full  Model  is  B1,  B2,  …  ,  Bp,  B*1,  …  ,  B*k    (along  with  y-­‐‑int  and  error,  etc)  Ho:  B*1=  …    =  B*k  =  0  Ha:  at  least  one  of  the  *  slopes  is  not  zero.  F  Stat:  (All  the  extra  sum-­‐‑of-­‐‑square  value  for  the  *  terms  combined/k)/MSE  Critical  Value:  F  (k,  n-­‐‑p-­‐‑k-­‐‑1,  .95)      

Page 5: Significance, Partial Tests, Degrees of Freedom

 AN  EASY  WAY  TO  REMEMBER  THE  DEGREES  OF  FREEDOM  FOR  ALL  THREE  F  TESTS:    

1. Look  at  your  full  model.    How  many  variables  have  a  big  letter  B  next  to  them?  Write  that  number  down.  I’m  including  the  y-­‐‑intercept  here.    If  it  has  a  “B”  it  gets  counted.  

2. Look  at  your  n  size.  Write  that  number  down.  3. Subtract  (1)  from  (2).    That’s  your  DENOMINATOR  DF.  4. Look  at  the  variables  listed  in  your  null  hypothesis.    The  ones  that  Ho  

claims  equal  zero.    Count  ‘em.    That  number  is  your  NUMERATOR  DF.  5. Slap  a  .95  on  the  end  of  that  thing  for  good  measure.    Done  deal.