Quantitative Research Techniques

Assignment 3 Hints

residuals  above  2,  so  that’s  as  we  expect,  but  3%  of  cases  with  residuals  above  2.5  (we’d  
gggggggggggggggggggggggggggggggggggggggggggggggggggggggggggg
gggggggggggggggggggggggggggggggggggggg
expect  only  1%),  which  indicates  possible  outliers.    

gggggggggggggggggggggggggggggggggggggggggggggggggggggggg
Normality  of  errors:  The  histogram  reveals  a  skewed  distribution,  indicating  that  the  
gggggggggggggggggggggggbbbbbbbbbbbbbbbBBBBBBBBBBBBBBBBBBB
normality  of  errors  assumption  has  been  broken.  The  normal  P–P  plot  verifies  this  because  
BBBBBB
the  dashed  line  deviates  considerably  from  the  straight  line  (which  indicates  what  you’d  get  
BBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB
from  normally  distributed  errors).  

BBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB
 Homoscedasticity  and  independence  of  errors:  The  scatterplot  of  ZPRED  vs.  ZRESID  does  not  
show  a  random  pattern.  There  is  a  distinct  funnelling,  indicating  heteroscedasticity.  
gggggggggggggggggggggggggggggggggggggggggggggggggggggggggggg
However,  the  Durbin–Watson  statistic  does  fall  within  Field’s  recommended  boundaries  of  
gggggggggggggggggggggggggggggggggggggg
1–3,  which  suggests  that  errors  are  reasonably  independent.    
gggggggggggggggggggggggggggggggggggggggggggggggggggggggg
gggggggggggggggggggggggbbbbbbbbbbbbbbbBBBBBBBBBBBBBBBBBBB
Multicollinearity:  For  the  age  and  experience  variables  in  the  model,  VIF  values  are  above  10  
BBBBBB
(or  alternatively,  tolerance  values  are  all  well  below  0.2),  indicating  multicollinearity  in  the  
BBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB
data.  In  fact,  the  correlation  between  these  two  variables  is  around  .9!  So,  these  two  
gggggggggggggggggggggggggggggggggggggggggggggggggggggggggggg
variables  are  measuring  very  similar  things.  Of  course,  this  makes  perfect  sense  because  the  
gggggggggggggggggggggggggggggggggggggg
older  a  model  is,  the  more  years  she  would’ve  spent  modelling!  So,  it  was  fairly  stupid  to  
measure  both  of  these  things!  This  also  explains  the  weird  result  that  the  number  of  years  
gggggggggggggggggggggggggggggggggggggggggggggggggggggggg
gggggggggggggggggggggggbbbbbbbbbbbbbbbBBBBBBBBBBBBBBBBBBB
spent  modelling  negatively  predicted  salary  (i.e.  more  experience  =  less  salary!):  in  fact  if  
BBBBBB
you  do  a  simple  regression  with  experience  as  the  only  predictor  of  salary  you’ll  find  it  has  
BBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB
the  expected  positive  relationship.  This  hopefully  demonstrates  why  multicollinearity  can  
bias  the  regression  model.  
BBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB
BBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB
All  in  all,  several  assumptions  have  not  been  met  and  so  this  model  is  probably  fairly  
unreliable.  

Task 5
A  study  was  carried  out  to  explore  the  relationship  between  Aggression  and  several  
potential  predicting  factors  in  666  children  who  had  an  older  sibling.  Variables  
measured  were  Parenting_Style  (high  score  =  bad  parenting  practices),  
Computer_Games  (high  score  =  more  time  spent  playing  computer  games),  Television  
(high  score  =  more  time  spent  watching  television),  Diet  (high  score  =  the  child  has  a  
good  diet  low  in  harmful  additives),  and  Sibling_Aggression  (high  score  =  more  
aggression  seen  in  their  older  sibling).  Past  research  indicated  that  parenting  style  and  
sibling  aggression  were  good  predictors  of  the  level  of  aggression  in  the  younger  child.  
All  other  variables  were  treated  in  an  exploratory  fashion.  The  data  are  in  the  file  Child  
Aggression.sav.  Analyse  them  with  multiple  regression.  

We  need  to  conduct  this  analysis  hierarchically,  entering  parenting  style  and  sibling  
aggression  in  the  first  step  (forced  entry)  and  the  remaining  variables  in  a  second  step  
(stepwise):  
    

Based  on  the  final  model  (which  is  actually  all  we’re  interested  in)  the  following  variables  
predict  aggression:  
 ! Parenting  style  (b  =  0.062,  β  =  0.194,  t  =  4.93,  p  <  .001)  significantly  predicted  
aggression.  The  beta  value  indicates  that  as  parenting  increases  (i.e.  as  bad  practices  
increase),  aggression  increases  also.  
! Sibling  aggression  (b  =  0.086,  β  =  0.088,  t  =  2.26,  p  <  .05)  significantly  predicted  
aggression.  The  beta  value  indicates  that  as  sibling  aggression  increases  (became  
more  aggressive),  aggression  increases  also.  
! Computer  games  (b  =  0.143,  β  =  0.037,  t  =  3.89,  p  <  .001)  significantly  predicted  
aggression.  The  beta  value  indicates  that  as  the  time  spent  playing  computer  games  
increases,  aggression  increases  also.  
! Good  diet  (b  =  –0.112,  β  =  –0.118,  t  =  –2.95,  p  <  .01)  significantly  predicted  
aggression.  The  beta  value  indicates  that  as  the  diet  improved,  aggression  
decreased.  

The  only  factor  not  to  predict  aggression  was:  

" Television  (b  if  entered  =  0.032,  t  =  0.72,  p  >  .05)  did  not  significantly  predict  
aggression.  

Based  on  the  standardized  beta  values,  the  most  substantive  predictor  of  aggression  was  
actually  parenting  style,  followed  by  computer  games,  diet  and  then  sibling  aggression.  

R2  is  the  squared  correlation  between  the  observed  values  of  aggression  and  the  values  
of  aggression  predicted  by  the  model.  The  values  in  this  output  tell  us  that  sibling  aggression  
and  parenting  style  in  combination  explain  5.3%  of  the  variance  in  aggression.  When  
computer  game  use  is  factored  in  as  well,  7%  of  variance  in  aggression  is  explained  (i.e.  an  
additional  1.7%).  Finally,  when  diet  is  added  to  the  model,  8.2%  of  the  variance  in  aggression  
is  explained  (an  additional  1.2%).  With  all  four  of  these  predictors  in  the  model  still  less  than  
half  of  the  variance  in  aggression  can  be  explained.  

The  Durbin–Watson  statistic  tests  the  assumption  of  ‘independence  of  errors’,  which  
means  that  for  any  two  observations  (cases)  in  the  regression,  their  residuals  should  be  
uncorrelated  (or  independent).  In  this  output  the  Durbin–Watson  statistic  falls  within  the  
recommended  boundaries  of  1–3,  which  suggests  that  errors  are  reasonably  independent.  

The  scatterplot  helps  us  to  assess  both  homoscedasticity  and  independence  of  errors.  The  
scatterplot  of  ZPRED  vs.  ZRESID  does  show  a  random  pattern  and  so  indicates  no  violation  of  
the  independence  of  errors  assumption.  Also,  the  errors  on  the  scatterplot  do  not  funnel  
out,  indicating  homoscedasticity  of  errors,  thus  no  violations  of  these  assumptions.  

 
gggggggggggggggggggggggggggggggggggggggggggggggggggggggggggg
gggggggggggggggggggggggggggggggggggggg
Task 6
gggggggggggggggggggggggggggggggggggggggggggggggggggggggg
gggggggggggggggggggggggbbbbbbbbbbbbbbbBBBBBBBBBBBBBBBBBBB
BBBBBB
Repeat  the  analysis  in  Labcoat  Leni’s  Real  Research  8.1  using  bootstrapping  for  the  
BBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB
confidence  intervals.  What  are  the  confidence  intervals  for  the  regression  parameters?  

 
BBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB
 These  results  show  that  after  controlling  for  age,  grade  and  gender,  narcissism  
significantly  predicted  the  Facebook  profile  picture  ratings  over  and  above  extroversion,  b  =  
0.37  [0.105,  0.23],  p  =  .01.  

Task 7
Coldwell,  Pike  and  Dunn  (2006)  investigated  whether  household  chaos  predicted  
children’s  problem  behaviour  over  and  above  parenting.  From  118  families  they  
recorded  the  age  and  gender  of  the  younger  sibling  (Child_age  and  Child_gender).  
They  then  interviewed  the  child  about  their  relationship  with  their  mum  using  the  
Berkeley  Puppet  Interview  (BPI),  which  measures  (1)  warmth/enjoyment  
(Child_warmth),  and  (2)  anger/hostility  (Child_anger).  Higher  scores  indicate  more  
anger/hostility  and  warmth/enjoyment,  respectively.  Each  mum  was  interviewed  
about  their  relationship  with  the  child  resulting  in  scores  for  relationship  positivity  
(Mum_pos)  and  relationship  negativity  (Mum_neg).  Household  chaos  (Chaos)  was  
assessed  using  the  Confusion,  Hubbub,  and  Order  Scale.  The  outcome  variable  was  the  
child’s  adjustment  (sdq):  the  higher  the  score,  the  more  problem  behaviour  the  child  is  
reported  to  be  displaying.  The  data  are  in  the  file  Coldwell  et  al.  (2006).sav.  Conduct  a  
hierarchical  regression  in  three  steps:  (1)  enter  child  age  and  gender;  (2)  add  the  
variables  measuring  parent–child  positivity,  parent–child  negativity,  parent–child  
warmth  and  parent–child  anger;  (3)  add  chaos.  Is  household  chaos  predictive  of  
children’s  problem  behaviour  over  and  above  parenting?  

Looking  at  the  output  tables  above,  we  can  conclude  that  household  chaos  significantly  
predicted  younger  sibling’s  problem  behaviour  over  and  above  maternal  parenting,  child  age  
and  gender,  t(88)  =  2.09,  p  <  .05.  The  positive  standardized  beta  value  (.218)  indicates  that  
there  is  a  positive  relationship  between  household  chaos  and  child’s  problem  behaviour.  In  
other  words,  the  higher  the  level  of  household  chaos,  the  more  problem  behaviours  the  
child  displayed.  The  value  of  R2  (.11)  tells  us  that  household  chaos  accounts  for  11%  of  the  
variance  in  child  problem  behaviour.