DISCOVERING

 STATISTICS  USING  SPSS  

Chapter 19: Logistic regression

Smart Alex’s Solutions

Task 1
A  ‘display  rule’  refers  to  displaying  an  appropriate  emotion  in  a  given  situation.  For  
example,  if  you  receive  a  Christmas  present  that  you  don’t  like,  the  appropriate  
emotional  display  is  to  smile  politely  and  say  ‘Thank  you  
Why  did  you  buy  me  this   Auntie  Kate,  I’ve  always  wanted  a  rotting  cabbage’.  The  
crappy  statistics  textbook  
for  Christmas,  Auntie  Kate?   inappropriate  emotional  display  is  to  start  crying  and  
  scream  ‘Why  did  you  buy  me  a  rotting  cabbage,  you  
selfish  old  bag?’  A  psychologist  measured  children’s  
understanding  of  display  rules  (with  a  task  that  they  could  either  
pass  or  fail),  their  age  (months),  and  their  ability  to  understand  others’  
mental  states  (‘theory  of  mind’,  measured  with  a  false-­‐belief  task  that  
they  could  pass  or  fail).  The  data  are  in  Display.sav.  Can  display  rule  
understanding  (did  the  child  pass  the  test:  yes/no?)  be  predicted  from  the  
theory  of  mind  (did  the  child  pass  the  false-­‐belief  task:  yes/no?),  age  and  their  
interaction?  

The main analysis

To  carry  out  logistic  regression,  the  data  must  be  entered  as  for  normal  regression:  they  are  
arranged  in  the  data  editor  in  three  columns  (one  representing  each  variable).  Open  the  file  
Display.sav.  Looking  at  the  data  editor,  you  should  notice  that  both  of  the  categorical  
variables  have  been  entered  as  coding  variables;  that  is,  numbers  have  been  specified  to  
represent  categories.  For  ease  of  interpretation,  the  outcome  variable  should  be  coded  1  
(event  occurred)  and  0  (event  did  not  occur);  in  this  case,  1  represents  having  display  rule  
understanding,  and  0  represents  an  absence  of  display  rule  understanding.  For  the  false-­‐
belief  task  a  similar  coding  has  been  used  (1  =  passed  the  false-­‐belief  task,  2  =  failed  the  
false-­‐belief  task).  Logistic  regression  is  accessed  by  selecting  
.  Following  this  menu  path  activates  the  main  Logistic  Regression  dialog  box  
shown  in  Figure  1.  

PROFESSOR  ANDY  P  FIELD     1  

DISCOVERING  STATISTICS  USING  SPSS  

 
Figure  1  

The  main  dialog  box  is  very  similar  to  the  standard  regression  option  box.  There  is  a  space  
to  place  a  dependent  (or  outcome)  variable.  In  this  example,  the  outcome  was  the  display  
rule  task,  so  we  can  simply  click  on  this  and  transfer  it  to  the  Dependent  box  by  clicking  on  
.  There  is  also  a  box  for  specifying  the  covariates  (the  predictor  variables).  It  is  possible  to  
specify  both  main  effects  and  interactions  in  logistic  regression.  To  specify  a  main  effect,  
simply  select  one  predictor  (e.g.  age)  and  then  transfer  this  variable  to  the  Covariates  box  by  
clicking  on   .  To  input  an  interaction,  click  on  more  than  one  variable  on  the  left-­‐hand  side  
of  the  dialog  box  (i.e.  highlight  two  or  more  variables)  and  then  click  on    to  move  them  
to  the  Covariates  box.  

For  this  analysis  select  a  Forward:LR  method  of  regression.    

In  this  example  there  is  one  categorical  predictor  variable.  One  of  the  great  things  about  
logistic  regression  is  that  it  is  quite  happy  to  accept  categorical  predictors.  However,  it  is  
necessary  to  ‘tell’  SPSS  which  variables,  if  any,  are  categorical  by  clicking  on    in  the  
main  Logistic  Regression  dialog  box  to  activate  the  dialog  box  in  Figure  2.  

 
Figure  2  

PROFESSOR  ANDY  P  FIELD     2  

DISCOVERING  STATISTICS  USING  SPSS  

The  covariates  are  listed  on  the  left-­‐hand  side,  and  there  is  a  space  on  the  right-­‐hand  side  
in  which  categorical  covariates  can  be  placed.  Simply  highlight  any  categorical  variables  you  
have  (in  this  example  click  on  fb)  and  transfer  them  to  the  Categorical  Covariates  box  by  
clicking  on   .  There  are  many  ways  in  which  you  can  treat  categorical  predictors.  They  
could  be  incorporated  into  regression  by  recoding  them  using  zeros  and  ones  (known  as  
dummy  coding).  Now,  actually,  there  are  different  ways  you  can  arrange  this  coding  
depending  on  what  you  want  to  compare,  and  SPSS  has  several  ‘standard’  ways  built  into  it  
that  you  can  select.  By  default  SPSS  uses  indicator  coding,  which  is  the  standard  dummy  
variable  coding  that  I  explained  in  Chapter  7  (and  you  can  choose  to  have  either  the  first  or  
last  category  as  your  baseline).  To  change  to  a  different  kind  of  contrast  click  on  the  down  
arrow  in  the  Change  Contrast  box.  Select  Indicator  coding  (first).  

Obtaining residuals

To  save  residuals  click  on    in  the  main  Logistic  Regression  dialog  box.  SPSS  saves  each  
of  the  selected  variables  into  the  data  editor.  The  dialog  box  in  Figure  3  gives  several  
options,  and  most  of  these  are  the  same  as  those  in  multiple  regression.  Select  all  of  the  
available  options,  or  as  a  bare  minimum  select  the  same  options  as  shown  in  the  figure.  

 
Figure  3  

Further options

There  is  a  final  dialog  box  that  offers  further  options.  This  box  is  accessed  by  clicking  on  
 in  the  main  Logistic  Regression  dialog  box.  For  the  most  part,  the  default  settings  in  
this  dialog  box  are  fine.  These  options  are  explained  in  the  chapter  and  so  just  make  the  
selections  shown  in  Figure  4.  

PROFESSOR  ANDY  P  FIELD     3  

DISCOVERING  STATISTICS  USING  SPSS  

 
Figure  4  

Interpreting the output

Dependent Variable Encoding

Original Value Internal Value

No 0
Yes 1
 

Categorical Variables Codings

Paramete
Frequency r coding
(1)
False Belief No 29 .000
understanding Yes 41 1.000
 

 
Output  1  

These  tables  tell  us  the  parameter  codings  given  to  the  categorical  predictor  variable.  
Indicator  coding  was  chosen  with  two  categories,  and  so  the  coding  is  the  same  as  the  values  
in  the  data  editor.    

PROFESSOR  ANDY  P  FIELD     4  

DISCOVERING  STATISTICS  USING  SPSS  

Classification Tablea,b

Predicted
Display Rule
understanding Percentage
Observed No Yes Correct
Step 0 Display Rule No 0 31 .0
understanding Yes 0 39 100.0
Overall Percentage 55.7
a. Constant is included in the model.
b. The cut value is .500
 

Variables in the Equation

B S.E. Wald df Sig. Exp(B)

Step 0 Constant .230 .241 .910 1 .340 1.258
 

Variables not in the Equation

Score df Sig.
Step Variables AGE 15.956 1 .000
0 FB(1) 24.617 1 .000
AGE by FB(1) 23.987 1 .000
Overall Statistics 26.257 3 .000
 
Output  1  

For  this  first  analysis  we  requested  a  forward  stepwise  method  and  so  the  initial  model  is  
derived  using  only  the  constant  in  the  regression  equation.  The  above  output  tells  us  about  
the  model  when  only  the  constant  is  included  (i.e.  all  predictor  variables  are  omitted).  The  
log-­‐likelihood  of  this  baseline  model  is  96.124.  This  represents  the  fit  of  the  model  when  the  
most  basic  model  is  fitted  to  the  data.  When  including  only  the  constant,  the  computer  
bases  the  model  on  assigning  every  participant  to  a  single  category  of  the  outcome  variable.  
In  this  example,  SPSS  can  decide  either  to  predict  that  every  child  has  display  rule  
understanding,  or  to  predict  that  all  children  do  not  have  display  rule  understanding.  It  could  
make  this  decision  arbitrarily,  but  because  it  is  crucial  to  try  to  maximize  how  well  the  model  
predicts  the  observed  data  SPSS  will  predict  that  every  child  belongs  to  the  category  in  
which  most  observed  cases  fell.  In  this  example  there  were  39  children  who  had  display  rule  
understanding  and  only  31  who  did  not.  Therefore,  if  SPSS  predicts  that  every  child  has  
display  rule  understanding  then  this  prediction  will  be  correct  39  times  out  of  70  (55.7%).  
However,  if  SPSS  predicted  that  every  child  did  not  have  display  rule  understanding,  then  
this  prediction  would  be  correct  only  31  times  out  of  70  (44.3%).  As  such,  of  the  two  
available  options  it  is  better  to  predict  that  all  children  had  display  rule  understanding  
because  this  results  in  a  greater  number  of  correct  predictions.  The  output  shows  a  
contingency  table  for  the  model  in  this  basic  state.  You  can  see  that  SPSS  has  predicted  that  
all  children  have  display  rule  understanding,  which  results  in  0%  accuracy  for  the  children  
who  were  observed  to  have  no  display  rule  understanding,  and  100%  accuracy  for  those  
children  observed  to  have  passed  the  display  rule  task.  Overall,  the  model  correctly  classifies  
55.7%  of  children.  The  next  part  of  the  output  summarizes  the  model,  and  at  this  stage  this  
entails  quoting  the  value  of  the  constant  (b0),  which  is  equal  to  0.23.  

PROFESSOR  ANDY  P  FIELD     5  

DISCOVERING  STATISTICS  USING  SPSS  

The  final  table  of  the  output  is  labelled  Variables  not  in  the  Equation.  The  bottom  line  of  
this  table  reports  the  residual  chi-­‐square  statistic  as  26.257,  which  is  significant  at  p  <  .001  (it  
labels  this  statistic  Overall  Statistics).  This  statistic  tells  us  that  the  coefficients  for  the  
variables  not  in  the  model  are  significantly  different  from  zero  –  in  other  words,  that  the  
addition  of  one  or  more  of  these  variables  to  the  model  will  significantly  affect  its  predictive  
power.  If  the  probability  for  the  residual  chi-­‐square  had  been  greater  than  .05  it  would  have  
meant  that  none  of  the  variables  excluded  from  the  model  could  make  a  significant  
contribution  to  the  predictive  power  of  the  model.  As  such,  the  analysis  would  have  
terminated  at  this  stage.  

The  remainder  of  this  table  lists  each  of  the  predictors  in  turn  with  a  value  of  Roa’s  
efficient  score  statistic  for  each  one  (column  labelled  Score).  In  large  samples  when  the  null  
hypothesis  is  true,  the  score  statistic  is  identical  to  the  Wald  statistic  and  the  likelihood  ratio  
statistic.  It  is  used  at  this  stage  of  the  analysis  because  it  is  computationally  less  intensive  
than  the  Wald  statistic  and  so  can  still  be  calculated  in  situations  when  the  Wald  statistic  
would  prove  prohibitive.  Like  any  test  statistic,  Roa’s  score  statistic  has  a  specific  distribution  
from  which  statistical  significance  can  be  obtained.  In  this  example,  all  excluded  variables  
have  significant  score  statistics  at  p  <  .001  and  so  all  three  could  potentially  make  a  
contribution  to  the  model.  The  stepwise  calculations  are  relative  and  so  the  variable  that  will  
be  selected  for  inclusion  is  the  one  with  the  highest  value  for  the  score  statistic  that  is  
significant  at  the  .05  level.  In  this  example,  that  variable  will  be  fb  because  it  has  the  highest  
value  of  the  score  statistic.  The  next  part  of  the  output  deals  with  the  model  after  this  
predictor  has  been  added.  

In  the  first  step,  false-­‐belief  understanding  (fb)  is  added  to  the  model  as  a  predictor.  As  
such  a  child  is  now  classified  as  having  display  rule  understanding  based  on  whether  they  
passed  or  failed  the  false-­‐belief  task.    

Omnibus Tests of Model Coefficients

Chi-square df Sig.
Step 1 Step 26.083 1 .000
Block 26.083 1 .000
Model 26.083 1 .000
 

Model Summary

-2 Log Cox & Snell Nagelkerke

Step likelihood R Square R Square
1 70.042 .311 .417
 

Classification Tablea

Predicted
Display Rule
understanding Percentage
Observed No Yes Correct
Step 1 Display Rule No 23 8 74.2
understanding Yes 6 33 84.6
Overall Percentage 80.0
a. The cut value is .500
 
Output  3  

PROFESSOR  ANDY  P  FIELD     6  

DISCOVERING  STATISTICS  USING  SPSS  

Output  3  shows  summary  statistics  about  the  new  model  (which  we’ve  already  seen  
contains  fb).  The  overall  fit  of  the  new  model  is  assessed  using  the  log-­‐likelihood  statistic.  In  
SPSS,  rather  than  reporting  the  log-­‐likelihood  itself,  the  value  is  multiplied  by  –2  (and  
sometimes  referred  to  as  –2LL):  this  multiplication  is  done  because  –2LL  has  an  
approximately  chi-­‐square  distribution  and  so  makes  it  possible  to  compare  values  against  
those  that  we  might  expect  to  get  by  chance  alone.  Remember  that  large  values  of  the  log-­‐
likelihood  statistic  indicate  poorly  fitting  statistical  models.  

At  this  stage  of  the  analysis  the  value  of  −2LL  should  be  less  than  the  value  when  only  the  
constant  was  included  in  the  model  (because  lower  values  of  −2LL  indicate  that  the  model  is  
predicting  the  outcome  variable  more  accurately).  When  only  the  constant  was  included,  
−2LL  =  96.124,  but  now  fb  has  been  included  this  value  has  been  reduced  to  70.042.  This  
reduction  tells  us  that  the  model  is  better  at  predicting  display  rule  understanding  than  it  
was  before  fb  was  added.  The  question  of  how  much  better  the  model  predicts  the  outcome  
variable  can  be  assessed  using  the  model  chi-­‐square  statistic,  which  measures  the  difference  
between  the  model  as  it  currently  stands  and  the  model  when  only  the  constant  was  
included.  We  can  assess  the  significance  of  the  change  in  a  model  by  taking  the  log-­‐
likelihood  of  the  new  model  and  subtracting  the  log-­‐likelihood  of  the  baseline  model  from  it.  
The  value  of  the  model  chi-­‐square  statistic  works  on  this  principle  and  is,  therefore,  equal  to  
−2LL  with  fb  included  minus  the  value  of  −2LL  when  only  the  constant  was  in  the  model  
(96.124  −  70.042  =  26.083).  This  value  has  a  chi-­‐square  distribution  and  so  its  statistical  
significance  can  be  easily  calculated.  In  this  example,  the  value  is  significant  at  the  .05  level  
and  so  we  can  say  that  overall  the  model  is  predicting  display  rule  understanding  
significantly  better  than  it  was  with  only  the  constant  included.  The  model  chi-­‐square  is  an  
analogue  of  the  F-­‐test  for  the  linear  regression  sum  of  squares.  In  an  ideal  world  we  would  
like  to  see  a  non-­‐significant  –2LL  (indicating  that  the  amount  of  unexplained  data  is  minimal)  
and  a  highly  significant  model  chi-­‐square  statistic  (indicating  that  the  model  including  the  
predictors  is  significantly  better  than  without  those  predictors).  However,  in  reality  it  is  
possible  for  both  statistics  to  be  highly  significant.  

There  is  a  second  statistic  called  the  step  statistic  that  indicates  the  improvement  in  the  
predictive  power  of  the  model  since  the  last  stage.  At  this  stage  there  has  been  only  one  
step  in  the  analysis,  and  so  the  value  of  the  improvement  statistic  is  the  same  as  the  model  
chi-­‐square.  However,  in  more  complex  models  in  which  there  are  three  or  four  stages,  this  
statistic  gives  you  a  measure  of  the  improvement  of  the  predictive  power  of  the  model  since  
the  last  step.  Its  value  is  equal  to  −2LL  at  the  current  step  minus  −2LL  at  the  previous  step.  If  
the  improvement  statistic  is  significant  then  it  indicates  that  the  model  now  predicts  the  
outcome  significantly  better  than  it  did  at  the  last  step,  and  in  a  forward  regression  this  can  
be  taken  as  an  indication  of  the  contribution  of  a  predictor  to  the  predictive  power  of  the  
model.  Similarly,  the  block  statistic  provides  the  change  in  −2LL  since  the  last  block  (for  use  
in  hierarchical  or  blockwise  analyses).  

Finally,  the  classification  table  at  the  end  of  this  section  of  the  output  indicates  how  well  
the  model  predicts  group  membership.  The  current  model  correctly  classifies  23  children  
who  don’t  have  display  rule  understanding  but  misclassifies  8  others  (i.e.  it  correctly  
classifies  74.2%  of  cases).  For  children  who  do  have  display  rule  understanding,  the  model  

PROFESSOR  ANDY  P  FIELD     7  

DISCOVERING  STATISTICS  USING  SPSS  

correctly  classifies  33  and  misclassifies  6  cases  (i.e.  correctly  classifies  84.6%  of  cases).  The  
overall  accuracy  of  classification  is,  therefore,  the  weighted  average  of  these  two  values  
(80%).  So,  when  only  the  constant  was  included,  the  model  correctly  classified  56%  of  
children,  but  now,  with  the  inclusion  of  fb  as  a  predictor,  this  has  risen  to  80%.  

Variables in the Equation

95.0% C.I.for EXP(B)

B S.E. Wald df Sig. Exp(B) Lower Upper
Step
a FB(1) 2.761 .605 20.856 1 .000 15.812 4.835 51.706
1 Constant -1.344 .458 8.592 1 .003 .261
a. Variable(s) entered on step 1: FB.
 
Output  4  

The  next  part  of  the  output  (Output  4)  is  crucial  because  it  tells  us  the  estimates  for  the  
coefficients  for  the  predictors  included  in  the  model.  It  provides  the  coefficients  and  
statistics  for  the  variables  that  have  been  included  in  the  model  at  this  point  (namely,  fb  and  
the  constant).  The  interpretation  of  this  coefficient  in  logistic  regression  is  that  it  represents  
the  change  in  the  logit  of  the  outcome  variable  associated  with  a  one-­‐unit  change  in  the  
predictor  variable.  The  logit  of  the  outcome  is  simply  the  natural  logarithm  of  the  odds  of  Y  
occurring.  

The  crucial  statistic  is  the  Wald  statistic,  which  has  a  chi-­‐square  distribution  and  tells  us  
whether  the  b  coefficient  for  that  predictor  is  significantly  different  from  zero.  If  the  
coefficient  is  significantly  different  from  zero  then  we  can  assume  that  the  predictor  is  
making  a  significant  contribution  to  the  prediction  of  the  outcome  (Y).  For  these  data  it  
seems  to  indicate  that  false-­‐belief  understanding  is  a  significant  predictor  of  display  rule  
understanding  (note  the  significance  of  the  Wald  statistic  is  less  than  .05).  

We  can  calculate  an  analogue  of  R  using  the  equation  in  the  chapter  (for  these  data,  the  
Wald  statistic  and  its  df  are  20.856  and  1,  respectively,  and  the  original  –2LL  was  96.124).  
Therefore,  R  can  be  calculated  as:  

20.856 − ( 2 ×1)
R=
  96.124  
= .4429

Hosmer  and  Lemeshow’s  measure  ( RL2 )  is  calculated  by  dividing  the  model  chi-­‐square  by  
the  original  −2LL.  In  this  example  the  model  chi-­‐square  after  all  variables  have  been  entered  
into  the  model  is  26.083,  and  the  original  –2LL  (before  any  variables  were  entered)  was  
96.124.  So   RL2  =  26.083/96.124  =  .271,  which  is  different  from  the  value  we  would  get  by  
squaring  the  value  of  R  given  above  (R2  =  .44292  =  .196).  

SPSS  reports  Cox  and  Snell’s  R2  as  .311.  This  is  calculated  from  this  equation:  

PROFESSOR  ANDY  P  FIELD     8  

DISCOVERING  STATISTICS  USING  SPSS  

!
 (−2𝐿𝐿 new − (−2𝐿𝐿 baseline
𝑅!" = 1 − exp  
𝑛

SPSS  reports  −2LL(new)  as  70.04  and  −2LL(baseline)  as  96.124.  The  sample  size,  n,  is  70:  

!
  70.04 − 96.124
𝑅!" = 1 − exp  
70
= 1 − exp −0.3726  
= 1 − 𝑒 !!.!"#$  
= .311  

Nagelkerke’s  adjustment  is  calculated  from:  

!
!
𝑅!"
𝑅! =  
−2𝐿𝐿 baseline
1 − exp −
𝑛
0.311
=  
96.124
1 − exp −
70
0.311
=  
1 − 𝑒 !!.!"!#
0.311
=  
1 − 0.2533
= .416  

As  you  can  see,  there’s  a  fairly  substantial  difference  between  the  two  values!  

The  final  thing  we  need  to  look  at  is  the  odds  ratio,  exp(b)  (Exp(B)  in  the  SPSS  output).  We  
can  interpret  this  in  terms  of  the  change  in  odds.  If  the  value  is  greater  than  1  then  it  
indicates  that  as  the  predictor  increases,  the  odds  of  the  outcome  occurring  increase.  
Conversely,  a  value  less  than  1  indicates  that  as  the  predictor  increases,  the  odds  of  the  
outcome  occurring  decrease.  In  this  example,  we  can  say  that  the  odds  of  a  child  who  has  
false-­‐belief  understanding  also  having  display  rule  understanding  are  15  times  higher  than  
those  of  a  child  who  does  not  have  false-­‐belief  understanding.  

In  the  options,  we  requested  a  confidence  interval  for  exp(b)  and  it  can  also  be  found  in  
Output  4.  The  way  to  interpret  this  confidence  interval  is  to  say  that  if  we  ran  100  
experiments  and  calculated  confidence  intervals  for  the  value  of  exp(b),  then  these  intervals  
would  encompass  the  actual  value  of  exp(b)  in  the  population  (rather  than  the  sample)  on  
95  occasions.  So,  in  this  case,  we  can  be  fairly  confident  that  the  population  value  of  exp(b)  
lies  between  4.84  and  51.71.  However,  there  is  a  5%  chance  that  a  sample  could  give  a  
confidence  interval  that  ‘misses’  the  true  value.  

PROFESSOR  ANDY  P  FIELD     9  

DISCOVERING  STATISTICS  USING  SPSS  

Model if Term Removed

Change in
Model Log -2 Log Sig. of the
Variable Likelihood Likelihood df Change
Step 1 FB -48.062 26.083 1 .000
 

Variables not in the Equation

Score df Sig.
Step Variables AGE 2.313 1 .128
1 AGE by FB(1) 1.261 1 .261
Overall Statistics 2.521 2 .283
 
Output  5  

The  test  statistics  for  fb  if  it  were  removed  from  the  model  are  reported  in  Output  5.  The  
important  thing  to  note  is  the  significance  value  of  the  log-­‐likelihood  ratio.  This  is  highly  
significant  for  this  model  (p  <  .001)  which  tells  us  that  removing  fb  from  the  model  would  
have  a  significant  effect  on  the  predictive  ability  of  the  model  –  in  other  words,  it  would  be  a  
very  bad  idea  to  remove  it!  

Finally,  we  are  told  about  the  variables  currently  not  in  the  model.  First  of  all,  the  residual  
chi-­‐square  (labelled  Overall  Statistics  in  the  output),  which  is  non-­‐significant,  tells  us  that  
none  of  the  remaining  variables  have  coefficients  significantly  different  from  zero.  
Furthermore,  each  variable  is  listed  with  its  score  statistic  and  significance  value,  and  for  
both  variables  their  coefficients  are  not  significantly  different  from  zero  (as  can  be  seen  from  
the  significance  values  of  .128  for  age  and  .261  for  the  interaction  of  age  and  false-­‐belief  
understanding).  Therefore,  no  further  variables  will  be  added  to  the  equation.  

The  next  part  of  the  output  (Output  6)  displays  the  classification  plot  that  we  requested  
in  the  options  dialog  box.  This  plot  is  a  histogram  of  the  predicted  probabilities  of  a  child  
passing  the  display  rule  task.  If  the  model  perfectly  fits  the  data,  then  this  histogram  should  
show  all  of  the  cases  for  which  the  event  has  occurred  on  the  right-­‐hand  side,  and  all  the  
cases  for  which  the  event  hasn’t  occurred  on  the  left-­‐hand  side.  In  other  words,  all  the  
children  who  passed  the  display  rule  task  should  appear  on  the  right  and  all  those  who  failed  
should  appear  on  the  left.  In  this  example,  the  only  significant  predictor  is  dichotomous  and  
so  there  are  only  two  columns  of  cases  on  the  plot.  If  the  predictor  is  a  continuous  variable,  
the  cases  are  spread  out  across  many  columns.  As  a  rule  of  thumb,  the  more  that  the  cases  
cluster  at  each  end  of  the  graph,  the  better.  This  statement  is  true  because  such  a  plot  
would  show  that  when  the  outcome  did  actually  occur  (i.e.,  the  child  did  pass  the  display  
rule  task)  the  predicted  probability  of  the  event  occurring  is  also  high  (i.e.,  close  to  1).  
Likewise,  at  the  other  end  of  the  plot,  it  would  show  that  when  the  event  didn’t  occur  (i.e.,  
when  the  child  failed  the  display  rule  task)  the  predicted  probability  of  the  event  occurring  is  
also  low  (i.e.,  close  to  0).  This  situation  represents  a  model  that  is  correctly  predicting  the  
observed  outcome  data.  If,  however,  there  are  a  lot  of  points  clustered  in  the  centre  of  the  
plot  then  it  shows  that  for  many  cases  the  model  is  predicting  a  probability  of  .5  that  the  
event  will  occur.  In  other  words,  for  these  cases  there  is  little  more  than  a  50–50  chance  that  
the  data  are  correctly  predicted  –  as  such  the  model  could  predict  these  cases  just  as  
accurately  by  simply  tossing  a  coin!  Also,  a  good  model  will  ensure  that  few  cases  are  
misclassified.  In  this  example  there  are  two  Ns  on  the  right  of  the  model  and  one  Y  on  the  

PROFESSOR  ANDY  P  FIELD     10  

DISCOVERING  STATISTICS  USING  SPSS  

left  of  the  model;  these  are  misclassified  cases,  and  the  fewer  of  these  there  are,  the  better  
the  model.  

 
Output  6  

Listing predicted probabilities

SPSS  saved  the  predicted  probabilities  and  predicted  group  memberships  as  variables  in  the  
data  editor  and  named  them  PRE_1  and  PGR_1  respectively.  These  probabilities  can  be  
listed  using  the  Case  Summaries  dialog  box  (see  the  book  chapter).  Output  7  shows  a  
selection  of  the  predicted  probabilities  (because  the  only  significant  predictor  was  a  
dichotomous  variable,  there  will  be  only  two  different  probability  values).  It  is  also  worth  
listing  the  predictor  variables  as  well  to  clarify  from  where  the  predicted  probabilities  come.  

Case Summariesa

False Belief Display Rule Predicted Predicted

Case Number Age in years understanding understanding probability group
1 1 24.00 No No .20690 No
2 5 36.00 No No .20690 No
3 9 34.00 No Yes .20690 No
4 10 31.00 No No .20690 No
5 11 32.00 No No .20690 No
6 12 30.00 Yes Yes .80488 Yes
7 20 26.00 No No .20690 No
8 21 29.00 No No .20690 No
9 29 45.00 Yes Yes .80488 Yes
10 31 41.00 No Yes .20690 No
11 32 32.00 No No .20690 No
12 43 56.00 Yes Yes .80488 Yes
13 60 63.00 No Yes .20690 No
14 66 79.00 Yes Yes .80488 Yes
Total N 14 14 14 14 14
a. Limited to first 100 cases.

 
Output  7  

We  found  from  the  model  that  the  only  significant  predictor  of  display  rule  
understanding  was  false-­‐belief  understanding.  This  could  have  a  value  of  either  1  (pass  the  
false-­‐belief  task)  or  0  (fail  the  false-­‐belief  task).  These  values  tells  us  that  when  a  child  

PROFESSOR  ANDY  P  FIELD     11  

DISCOVERING  STATISTICS  USING  SPSS  

doesn’t  possess  second-­‐order  false-­‐belief  understanding  (fb  =  0,  No),  there  is  a  probability  of  
.2069  that  they  will  pass  the  display  rule  task,  approximately  a  21%  chance  (1  out  of  5  
children).  However,  if  the  child  does  pass  the  false-­‐belief  task  (fb  =  1,  yes),  there  is  a  
probability  of  .8049  that  they  will  pass  the  display  rule  task,  an  80.5%  chance  (4  out  of  5  
children).  Consider  that  a  probability  of  0  indicates  no  chance  of  the  child  passing  the  display  
rule  task,  and  a  probability  of  1  indicates  that  the  child  will  definitely  pass  the  display  rule  
task.  Therefore,  the  values  obtained  provide  strong  evidence  for  the  role  of  false-­‐belief  
understanding  as  a  prerequisite  for  display  rule  understanding.  

Assuming  we  are  content  that  the  model  is  accurate  and  that  false-­‐belief  understanding  
has  some  substantive  significance,  then  we  could  conclude  that  false-­‐belief  understanding  is  
the  single  best  predictor  of  display  rule  understanding.  Furthermore,  age  and  the  interaction  
of  age  and  false-­‐belief  understanding  do  not  significantly  predict  display  rule  understanding.  
As  a  homework  task,  why  not  rerun  this  analysis  using  the  forced  entry  method  of  analysis  –  
how  do  your  conclusions  differ?  

This  conclusion  is  fine  in  itself,  but  to  be  sure  that  the  model  is  a  good  one,  it  is  important  
to  examine  the  residuals.    

Task 2
Are  there  any  influential  cases  or  outliers  in  the  model  for  Task  1?  

To  answer  this  question  we  need  to  look  at  the  model  residuals.  The  main  purpose  of  
examining  residuals  in  logistic  regression  is  to  (1)  isolate  points  for  which  the  model  fits  
poorly,  and  (2)  isolate  points  that  exert  an  undue  influence  on  the  model.  To  assess  the  
former  we  examine  the  residuals,  especially  the  Studentized  residual,  standardized  residual  
and  deviance  statistics.  All  of  these  statistics  have  the  common  property  that  95%  of  cases  in  
an  average,  normally  distributed  sample  should  have  values  which  lie  within  ±1.96,  and  99%  
of  cases  should  have  values  that  lie  within  ±2.58.  Therefore,  any  values  outside  of  ±3  are  
cause  for  concern  and  any  outside  of  about  ±2.5  should  be  examined  more  closely.  To  assess  
the  influence  of  individual  cases  we  use  influence  statistics  such  as  Cook’s  distance  (which  is  
interpreted  in  the  same  way  as  for  linear  regression:  as  a  measure  of  the  change  in  the  
regression  coefficient  if  a  case  is  deleted  from  the  model).  Also,  the  value  of  DFBeta,  which  is  
a  standardized  version  of  Cook’s  statistic,  tells  us  something  of  the  influence  of  certain  cases  
–  any  values  greater  than  1  indicate  possible  influential  cases.  Additionally,  leverage  
statistics  or  hat  values,  which  should  lie  between  0  (the  case  has  no  influence  whatsoever)  
and  1  (the  case  exerts  complete  influence  over  the  model),  tell  us  about  whether  certain  
cases  are  wielding  undue  influence  over  the  model.  The  expected  value  of  leverage  is  
defined  as  for  linear  regression.  If  you  request  these  residual  statistics,  SPSS  saves  them  in  as  
new  columns  in  the  data  editor.    

The  basic  residual  statistics  for  this  example  (Cook’s  distance,  leverage,  standardized  
residuals  and  DFBeta  values)  show  little  cause  for  concern  (Output  8).  Note  that  all  cases  

PROFESSOR  ANDY  P  FIELD     12  

DISCOVERING  STATISTICS  USING  SPSS  

have  DFBetas  less  than  1  and  leverage  statistics  (LEV_1)  close  to  the  calculated  expected  
value  of  0.03.  There  are  also  no  unusually  high  values  of  Cook’s  distance  (COO_1)  which,  all  
in  all,  means  that  there  are  no  influential  cases  having  an  effect  on  the  model.  Cook’s  
distance  is  an  unstandardized  measure  and  so  there  is  no  absolute  value  at  which  you  can  
say  that  a  case  is  having  an  influence,  Instead,  you  should  look  for  values  of  Cook’s  distance  
which  are  particularly  high  compared  to  the  other  cases  in  the  sample.  However,  Stevens  
(2002)  suggests  that  a  value  greater  than  1  is  problematic.  About  half  of  the  leverage  values  
are  a  little  high  but  given  that  the  other  statistics  are  fine,  this  is  probably  no  cause  for  
concern.  The  standardized  residuals  all  have  values  within  ±2.5  and  predominantly  have  
values  within  ±2,  and  so  there  seems  to  be  very  little  here  to  concern  us.    

Case Summariesa

Analog of
Cook's
influence Leverage Normalized DFBETA for DFBETA
Case Number statistics value residual constant for FB(1)
1 1 .00932 .03448 -.51075 -.04503 .04503
2 2 .00932 .03448 -.51075 -.04503 .04503
3 3 .00932 .03448 -.51075 -.04503 .04503
4 4 .00932 .03448 -.51075 -.04503 .04503
5 5 .00932 .03448 -.51075 -.04503 .04503
6 6 .13690 .03448 1.95789 .17262 -.17262
7 7 .00932 .03448 -.51075 -.04503 .04503
8 8 .00932 .03448 -.51075 -.04503 .04503
9 9 .13690 .03448 1.95789 .17262 -.17262
10 10 .00932 .03448 -.51075 -.04503 .04503
11 11 .00932 .03448 -.51075 -.04503 .04503
12 12 .00606 .02439 .49237 .00000 .03106
13 13 .00932 .03448 -.51075 -.04503 .04503
14 14 .10312 .02439 -2.03101 .00000 -.12812
15 15 .00932 .03448 -.51075 -.04503 .04503
16 16 .00932 .03448 -.51075 -.04503 .04503
17 17 .13690 .03448 1.95789 .17262 -.17262
18 18 .00932 .03448 -.51075 -.04503 .04503
19 19 .00606 .02439 .49237 .00000 .03106
20 20 .00932 .03448 -.51075 -.04503 .04503
21 21 .00932 .03448 -.51075 -.04503 .04503
22 22 .00606 .02439 .49237 .00000 .03106
23 23 .00606 .02439 .49237 .00000 .03106
24 24 .10312 .02439 -2.03101 .00000 -.12812
25 25 .00932 .03448 -.51075 -.04503 .04503
26 26 .00932 .03448 -.51075 -.04503 .04503
27 27 .00932 .03448 -.51075 -.04503 .04503
28 28 .00932 .03448 -.51075 -.04503 .04503
29 29 .00606 .02439 .49237 .00000 .03106
30 30 .00932 .03448 -.51075 -.04503 .04503
31 31 .13690 .03448 1.95789 .17262 -.17262
32 32 .00932 .03448 -.51075 -.04503 .04503
33 33 .00606 .02439 .49237 .00000 .03106
34 34 .00606 .02439 .49237 .00000 .03106
35 35 .00606 .02439 .49237 .00000 .03106
36 36 .00606 .02439 .49237 .00000 .03106
37 37 .10312 .02439 -2.03101 .00000 -.12812
38 38 .00606 .02439 .49237 .00000 .03106
39 39 .13690 .03448 1.95789 .17262 -.17262
40 40 .10312 .02439 -2.03101 .00000 -.12812
41 41 .00606 .02439 .49237 .00000 .03106
42 42 .00606 .02439 .49237 .00000 .03106
43 43 .00606 .02439 .49237 .00000 .03106
44 44 .00606 .02439 .49237 .00000 .03106
45 45 .00606 .02439 .49237 .00000 .03106
Total N 45 45 45 45 45
a. Limited to first 100 cases.

 
 

PROFESSOR  ANDY  P  FIELD     13  

DISCOVERING  STATISTICS  USING  SPSS  

Case Summariesa

Analog of
Cook's
influence Leverage Normalized DFBETA for DFBETA
Case Number statistics value residual constant for FB(1)
1 46 .10312 .02439 -2.03101 .00000 -.12812
2 47 .00606 .02439 .49237 .00000 .03106
3 48 .00932 .03448 -.51075 -.04503 .04503
4 49 .00932 .03448 -.51075 -.04503 .04503
5 50 .10312 .02439 -2.03101 .00000 -.12812
6 51 .00606 .02439 .49237 .00000 .03106
7 52 .00606 .02439 .49237 .00000 .03106
8 53 .00606 .02439 .49237 .00000 .03106
9 54 .00606 .02439 .49237 .00000 .03106
10 55 .00606 .02439 .49237 .00000 .03106
11 56 .00606 .02439 .49237 .00000 .03106
12 57 .10312 .02439 -2.03101 .00000 -.12812
13 58 .00606 .02439 .49237 .00000 .03106
14 59 .00606 .02439 .49237 .00000 .03106
15 60 .13690 .03448 1.95789 .17262 -.17262
16 61 .00606 .02439 .49237 .00000 .03106
17 62 .00606 .02439 .49237 .00000 .03106
18 63 .00606 .02439 .49237 .00000 .03106
19 64 .00606 .02439 .49237 .00000 .03106
20 65 .00606 .02439 .49237 .00000 .03106
21 66 .00606 .02439 .49237 .00000 .03106
22 67 .00606 .02439 .49237 .00000 .03106
23 68 .10312 .02439 -2.03101 .00000 -.12812
24 69 .00606 .02439 .49237 .00000 .03106
25 70 .00606 .02439 .49237 .00000 .03106
Total N 25 25 25 25 25
a. Limited to first 100 cases.

 
Output  8  

You  should  note  that  these  residuals  are  slightly  unusual  because  they  are  based  on  a  
single  predictor  that  is  categorical.  This  is  why  there  isn’t  a  lot  of  variability  in  the  values  of  
the  residuals.  Also,  if  substantial  outliers  or  influential  cases  had  been  isolated,  you  would  
not  be  justified  in  eliminating  these  cases  to  make  the  model  fit  better.  Instead  these  cases  
should  be  inspected  closely  to  try  to  isolate  a  good  reason  why  they  were  unusual.  It  might  
simply  be  an  error  in  inputting  data,  or  it  could  be  that  the  case  was  one  which  had  a  special  
reason  for  being  unusual:  for  example,  the  child  had  found  it  hard  to  pay  attention  to  the  
false-­‐belief  task  and  you  had  noted  this  at  the  time  of  the  experiment.  In  such  a  case,  you  
may  have  good  reason  to  exclude  the  case  and  duly  note  the  reasons  why.  

Task 3
Piff,  Stancato,  Côté,  Mendoza-­‐Dentona,  and  Keltner  (2012)  showed  that  people  of  a  higher  
social  class  are  more  unpleasant.  In  the  first  study  in  their  paper  they  observed  the  behaviour  
of  drivers:  they  classified  social  class  by  the  type  of  car  (Vehicle)  on  a  5-­‐point  scale.  They  then  
observed  whether  the  drivers  cut  in  front  of  other  cars  at  a  busy  intersection  (Vehicle_Cut).  
The  data  are  in  Piff  et  al.  (2012)  Vehicle.sav.  Do  a  logistic  regression  to  see  whether  social  
class  predicts  whether  or  not  a  driver  cut  in  front  of  other  vehicles.  

PROFESSOR  ANDY  P  FIELD     14  

DISCOVERING  STATISTICS  USING  SPSS  

There  is  only  one  predictor  variable  (Vehicle)  in  this  data  set,  so  I  selected  the  forced  entry  
(Enter)  method  of  regression.  

Interpreting the output

Step 0

 
Output  9  

 
Output  10  

 
Output  11  

 
Output  12  

 
Output  13  

The  above  output  tells  us  about  the  model  when  only  the  constant  is  included.  When  
including  only  the  constant,  SPSS  bases  the  model  on  assigning  every  participant  to  a  single  
category  of  the  outcome  variable.  In  this  example,  SPSS  can  decide  either  to  predict  that  all  
participants  cut  off  other  vehicles  at  intersections,  or  that  all  participants  did  not  cut  off  

PROFESSOR  ANDY  P  FIELD     15  

DISCOVERING  STATISTICS  USING  SPSS  

other  vehicles  at  intersections.  It  could  make  this  decision  arbitrarily,  but  because  it  is  crucial  
to  try  to  maximize  how  well  the  model  predicts  the  observed  data  SPSS  will  predict  that  all  
participants  (across  all  status  levels)  belong  to  the  category  in  which  most  observed  cases  
fell.  In  this  example  there  were  34  participants  who  did  cut  off  other  vehicles  at  
intersections  and  240  who  did  not.  Therefore,  if  SPSS  predicts  that  all  participants  did  not  
cut  off  other  vehicles  then  this  prediction  will  be  correct  240  times  out  of  274  (i.e.  87.6%).  
However,  if  SPSS  predicted  that  all  participants  did  cut  off  other  vehicles,  then  this  
prediction  would  be  correct  only  34  times  out  of  274  (12.4%).  As  such,  of  the  two  available  
options  it  is  better  to  predict  that  all  participants  did  not  cut  off  other  vehicles  because  this  
results  in  a  greater  number  of  correct  predictions.  Output    shows  a  contingency  table  for  the  
model  in  this  basic  state.  You  can  see  that  SPSS  has  predicted  that  all  participants  did  not  cut  
off  other  vehicles,  which  results  in  0%  accuracy  for  those  who  did  cut  off  other  vehicles,  and  
100%  accuracy  for  those  who  did  not.  Overall,  the  model  correctly  classifies  87.6%  of  
participants.  Output    summarizes  the  model,  and  at  this  stage  this  entails  quoting  the  value  
of  the  constant  (b0),  which  is  equal  to  −1.95.  

The  final  table  of  the  output  (Output  )  is  labelled  Variables  not  in  the  Equation.  The  
bottom  line  of  this  table  reports  the  residual  chi-­‐square  statistic  as  4.01  which  is  only  just  
significant  at  p  <  .05  (it  labels  this  statistic  Overall  Statistics).  This  statistic  tells  us  that  the  
coefficient  for  the  variable  not  in  the  model  is  significantly  different  from  zero  –  in  other  
words,  that  the  addition  of  this  variable  to  the  model  will  significantly  affect  its  predictive  
power.  If  the  probability  for  the  residual  chi-­‐square  had  been  greater  than  .05  it  would  have  
meant  that  the  variable  excluded  from  the  model  could  not  make  a  significant  contribution  
to  the  predictive  power  of  the  model.  As  such,  the  analysis  would  have  terminated  at  this  
stage.  

Step 1

The  next  part  of  the  output  deals  with  the  model  after  the  predictor  variable  (Vehicle)  has  
been  added  to  the  model.  As  such,  a  person  is  now  classified  as  either  cutting  off  other  
vehicles  at  an  intersection  or  not,  based  on  the  type  of  vehicle  they  were  driving  (as  a  
measure  of  social  status).    

 
Output  2  

 
Output  3  

PROFESSOR  ANDY  P  FIELD     16  

DISCOVERING  STATISTICS  USING  SPSS  

 
Output  4  

 
Output  5  

 The  above  output  shows  summary  statistics  about  the  new  model.  The  overall  fit  of  the  
new  model  is  significant  because  the  Model  chi-­‐square  in  the  table  labelled  Omnibus  Tests  of  
Model  Coefficients  (Output  )  is  significant,  χ2(1)  =  4.16,  p  =  .041.  Therefore,  the  model  that  
included  the  variable  Vehicle  predicted  whether  or  not  participants  cut  off  other  vehicles  at  
intersections  better  than  the  model  when  only  the  constant  was  included.    

The  classification  table  at  the  end  of  this  section  of  the  output  (Output  )  indicates  how  
well  the  model  predicts  group  membership.  In  step  1,  the  model  correctly  classifies  240  
participants  who  did  not  cut  off  other  vehicles  and  does  not  misclassify  any  (i.e.  it  correctly  
classifies  100%  of  cases).  For  participants  who  do  did  cut  off  other  vehicles,  the  model  
correctly  classifies  0  and  misclassifies  34  cases  (i.e.  correctly  classifies  0%  of  cases).  The  
overall  accuracy  of  classification  is,  therefore,  the  weighted  average  of  these  two  values  
(87.6%).  Therefore,  the  accuracy  is  no  different  than  when  only  the  constant  was  included  in  
the  model.    

 
Output  6  

The  significance  of  the  Wald  statistic  in  Output    is  .047,  which  is  less  than  .05.  Therefore,  
we  can  conclude  that  the  status  of  the  vehicle  the  participant  was  driving  significantly  
predicted  whether  or  not  they  cut  off  another  vehicle  at  an  intersection.    

The  final  thing  we  need  to  look  at  is  exp  b  (Exp(B)  in  Output  ).  We  can  interpret  exp  b  in  
terms  of  the  change  in  odds.  If  the  value  is  greater  than  1  then  it  indicates  that  as  the  
predictor  increases,  the  odds  of  the  outcome  occurring  increase.  Conversely,  a  value  less  

PROFESSOR  ANDY  P  FIELD     17  

DISCOVERING  STATISTICS  USING  SPSS  

than  1  indicates  that  as  the  predictor  increases,  the  odds  of  the  outcome  occurring  
decrease.  In  this  example,  the  exp  b  for  vehicle  in  step  1  is  1.441,  which  is  greater  than  1,  
thus  indicating  that  as  the  predictor  (vehicle)  increases,  the  value  of  the  outcome  also  
increases,  that  is,  the  value  of  the  categorical  variable  moves  from  0  (did  not  cut  off  vehicle)  
to  1  (cut  off  vehicle).  In  other  words,  drivers  of  vehicles  of  a  higher  status  were  more  likely  
to  cut  off  other  vehicles  at  intersections.  

Task 4
In  their  second  study,  Piff  et  al.  (2012)  again  observed  the  behaviour  of  drivers  and  classified  
social  class  by  the  type  of  car  (Vehicle).  However,  they  observed  whether  the  drivers  cut  off  a  
pedestrian  at  a  crossing  (Pedestrian_Cut).  The  data  are  in  Piff  et  al.  (2012)  Pedestrian.sav.  
Do  a  logistic  regression  to  see  whether  social  class  predicts  whether  or  not  a  driver  prevents  
a  pedestrian  from  crossing.  

As  in  the  previous  task,  there  is  only  one  predictor  variable  (Vehicle),  so  I  selected  the  forced  
entry  (Enter)  method  of  regression.  

Step 0

 
Output  7  

 
Output  8  

 
Output  9  

PROFESSOR  ANDY  P  FIELD     18  

DISCOVERING  STATISTICS  USING  SPSS  

 
Output  10  

 
Output  11  

The  above  output  tells  us  about  the  model  when  only  the  constant  is  included.  When  
including  only  the  constant,  SPSS  bases  the  model  on  assigning  every  participant  to  a  single  
category  of  the  outcome  variable.  In  this  example  there  were  54  participants  who  did  cut  off  
pedestrians  at  intersections  and  98  who  did  not.  Therefore,  if  SPSS  predicts  that  all  
participants  did  not  cut  off  pedestrians  then  this  prediction  will  be  correct  98  times  out  of  
152  (i.e.  64.5%).  However,  if  SPSS  predicted  that  all  participants  did  cut  off  other  vehicles,  
then  this  prediction  would  be  correct  only  54  times  out  of  152  (35.5%).  As  such,  of  the  two  
available  options  it  is  better  to  predict  that  all  participants  did  not  cut  off  pedestrians  
because  this  results  in  a  greater  number  of  correct  predictions.  Output    shows  a  contingency  
table  for  the  model  in  this  basic  state.  You  can  see  that  SPSS  has  predicted  that  all  
participants  did  not  cut  off  pedestrians,  which  results  in  0%  accuracy  for  those  who  did  cut  
off  pedestrians,  and  100%  accuracy  for  those  that  did  not.  Overall,  the  model  correctly  
classifies  64.5%  of  participants.  Output    summarizes  the  model,  and  at  this  stage  this  entails  
quoting  the  value  of  the  constant  (b0),  which  is  equal  to  −0.596.  

The  final  table  of  the  output  (Output  )  is  labelled  Variables  not  in  the  Equation.  The  
bottom  line  of  this  table  reports  the  residual  chi-­‐square  statistic  as  4.77,  which  is  significant  
at  p  <  .05  (it  labels  this  statistic  Overall  Statistics).  This  statistic  tells  us  that  the  coefficient  
for  the  variable  not  in  the  model  is  significantly  different  from  zero  –  in  other  words,  that  
the  addition  of  this  variable  to  the  model  will  significantly  affect  its  predictive  power.  If  the  
probability  for  the  residual  chi-­‐square  had  been  greater  than  .05  it  would  have  meant  that  
the  variable  excluded  from  the  model  could  not  make  a  significant  contribution  to  the  
predictive  power  of  the  model.  As  such,  the  analysis  would  have  terminated  at  this  stage.  

Step 1

The  next  part  of  the  output  deals  with  the  model  after  the  predictor  variable  (Vehicle)  has  
been  added  to  the  model.  As  such,  a  person  is  now  classified  as  either  cutting  off  
participants  at  an  intersection  or  not,  based  on  the  type  of  vehicle  they  were  driving  (as  a  
measure  of  social  status).    

PROFESSOR  ANDY  P  FIELD     19  

DISCOVERING  STATISTICS  USING  SPSS  

 
Output  12  

 
Output  13  

 
Output  14  

 
Output  15  

The  above  output  shows  summary  statistics  about  the  new  model.  The  overall  fit  of  the  
new  model  is  significant  because  the  Model  chi-­‐square  in  the  table  labelled  Omnibus  Tests  of  
Model  Coefficients  (Output  )  is  significant,  χ2(1)  =  4.86,  p  =  .028.  Therefore,  the  model  that  
included  the  variable  Vehicle  predicted  whether  or  not  participants  cut  off  pedestrians  at  
intersections  better  than  the  model  when  only  the  constant  was  included.    

The  classification  table  at  the  end  of  this  section  of  the  output  (Output  )  indicates  how  
well  the  model  predicts  group  membership.  In  step  1,  the  model  correctly  classifies  91  
participants  who  did  not  cut  off  pedestrians  and  misclassifies  7  who  did  (i.e.,  it  correctly  
classifies  92.9%  of  cases).  For  participants  who  do  did  cut  off  other  vehicles,  the  model  
correctly  classifies  6  and  misclassifies  48  cases  (i.e.  correctly  classifies  11.1%  of  cases).  The  
overall  accuracy  of  classification  is,  therefore,  the  weighted  average  of  these  two  values  
(63.8%).  So,  when  only  the  constant  was  included,  the  model  correctly  classified  64.5%  
participants,  but  now,  with  the  inclusion  of  Vehicle  as  a  predictor,  this  has  actually  
decreased  slightly  to  63.8%.  

PROFESSOR  ANDY  P  FIELD     20  

DISCOVERING  STATISTICS  USING  SPSS  

 
Output  16  

The  significance  of  the  Wald  statistic  in  Output    is  .031,  which  is  less  than  .05.  Therefore,  
we  can  conclude  that  the  status  of  the  vehicle  the  participant  was  driving  significantly  
predicted  whether  or  not  they  cut  off  pedestrians  at  intersections.    

The  final  thing  we  need  to  look  at  is  exp  b  (Exp(B)  in  Output  ).  We  can  interpret  exp  b  in  
terms  of  the  change  in  odds.  If  the  value  is  greater  than  1  then  it  indicates  that  as  the  
predictor  increases,  the  odds  of  the  outcome  occurring  increase.  Conversely,  a  value  less  
than  1  indicates  that  as  the  predictor  increases,  the  odds  of  the  outcome  occurring  
decrease.  In  this  example,  the  exp  b  for  vehicle  in  step  1  is  1.495,  which  is  greater  than  1,  
thus  indicating  that  as  the  predictor  (vehicle)  increases,  the  value  of  the  outcome  also  
increases,  that  is,  the  value  of  the  categorical  variable  moves  from  0  (did  not  cut  off  
pedestrian)  to  1  (cut  off  pedestrian).  In  other  words,  drivers  of  vehicles  of  a  higher  status  
were  more  likely  to  cut  off  pedestrians  at  intersections.    

Task 5
Four  hundred  and  sixty-­‐seven  lecturers  completed  questionnaire  measures  of  Burnout  
(burnt  out  or  not),  Perceived  Control  (high  score  =  low  perceived  control),  Coping  Style  
(high  score  =  high  ability  to  cope  with  stress),  Stress  from  Teaching  (high  score  =  
teaching  creates  a  lot  of  stress  for  the  person),  Stress  from  Research  (high  score  =  
research  creates  a  lot  of  stress  for  the  person)  and  Stress  from  Providing  Pastoral  
Care  (high  score  =  providing  pastoral  care  creates  a  lot  of  stress  for  the  person).  
Cooper,  Sloan,  and  Williams’  (1988)  model  of  stress  indicates  that  perceived  control  
and  coping  style  are  important  predictors  of  burnout.  The  remaining  predictors  were  
measured  to  see  the  unique  contribution  of  different  aspects  of  a  lecturer’s  work  to  
their  burnout.  Conduct  a  logistic  regression  to  see  which  factors  predict  burnout.  The  
data  are  in  Burnout.sav.  

Test

The  analysis  should  be  done  hierarchically  because  Cooper’s  model  indicates  that  perceived  
control  and  coping  style  are  important  predictors  of  burnout.  So,  these  variables  should  be  
entered  in  the  first  block.  The  second  block  should  contain  all  other  variables,  and  because  
we  don’t  know  anything  much  about  their  predictive  ability,  we  should  enter  them  in  a  
stepwise  fashion  (I  chose  Forward:  LR).  

PROFESSOR  ANDY  P  FIELD     21  

DISCOVERING  STATISTICS  USING  SPSS  

SPSS output

Omnibus Tests of Model Coefficients

Chi-square df Sig.
Step 1 Step 165.928 2 .000
Block 165.928 2 .000
Model 165.928 2 .000
 

Model Summary

-2 Log Cox & Snell Nagelkerke

Step likelihood R Square R Square
1 364.179 .299 .441
 

Variables in the Equation

95.0% C.I.for EXP(B)

B S.E. Wald df Sig. Exp(B) Lower Upper
Step
a
LOC .061 .011 31.316 1 .000 1.063 1.040 1.086
1 COPE .083 .009 77.950 1 .000 1.086 1.066 1.106
Constant -4.484 .379 139.668 1 .000 .011
a. Variable(s) entered on step 1: LOC, COPE.
 
Output  17  

At  step  1,  the  overall  fit  of  the  model  is  significant,  χ2(2)  =  165.93,  p  <  .001.  The  model  
accounts  for  29.9%  or  44.1%  of  the  variance  in  burnout  (depending  on  which  measure  of  R2  
you  use).  

At  step  2,  the    overall  fit  of  the  model  is  significant  after  both  the  first  new  variable  
(teaching),  χ2(3)  =  193.34,  p  <  .001,  and  second  new  variable  (pastoral)  have  been  entered,  
χ2(4)  =  205.40,  p  <  .001.  The  final  model  accounts  for  35.6%  or  52.4%  of  the  variance  in  
burnout  (depending  on  which  measure  of  R2  you  use).  

PROFESSOR  ANDY  P  FIELD     22  

DISCOVERING  STATISTICS  USING  SPSS  

Omnibus Tests of Model Coefficients

Chi-square df Sig.
Step 1 Step 27.409 1 .000
Block 27.409 1 .000
Model 193.337 3 .000
Step 2 Step 12.060 1 .001
Block 39.470 2 .000
Model 205.397 4 .000
 

Model Summary

-2 Log Cox & Snell Nagelkerke

Step likelihood R Square R Square
1 336.770 .339 .500
2 324.710 .356 .524
 

Variables in the Equation

95.0% C.I.for EXP(B)

B S.E. Wald df Sig. Exp(B) Lower Upper
Step 1a LOC .092 .014 46.340 1 .000 1.097 1.068 1.126
COPE .131 .015 76.877 1 .000 1.139 1.107 1.173
TEACHING -.083 .017 23.962 1 .000 .921 .890 .952
Constant -1.707 .619 7.599 1 .006 .181
Step 2b LOC .107 .015 52.576 1 .000 1.113 1.081 1.145
COPE .135 .016 75.054 1 .000 1.145 1.110 1.181
TEACHING -.110 .020 31.660 1 .000 .896 .862 .931
PASTORAL .044 .013 11.517 1 .001 1.045 1.019 1.071
Constant -3.023 .747 16.379 1 .000 .049
a. Variable(s) entered on step 1: TEACHING.
b. Variable(s) entered on step 2: PASTORAL.

 
Output  18  

In  terms  of  the  individual  predictors  we  could  report  as  follows:  

•     B   95%  CI  for  Exp(B)  

(SE)  

    Lower   Exp(B)   Upper  

Step  1          

• Constant  
  –4.48**  
     
(0.38)  

• Perceived  
  0.06**   1.04   1.06   1.09  
Control   (0.01)  

• Coping  
  Style   0.08**   1.07   1.09   1.11  
(0.01)  

Final          

• Constant  
 
–3.02**        

PROFESSOR  ANDY  P  FIELD     23  

DISCOVERING  STATISTICS  USING  SPSS  

(0.75)  

• Perceived  
  0.11**   1.08   1.11   1.15  
Control   (0.02)  

• Coping  
  Style   0.14**   1.11   1.15   1.18  
(0.02)  

• Teaching  
  Stress   –0.11**   0.86   0.90   0.93  
(0.02)  

• Pastoral  
  Stress   0.04*   1.02   1.05   1.07  
(0.01)  
2   2
Note:  R  =  .36  (Cox  and  Snell),  .52  (Nagelkerke).  Model  χ (4)  =  205.40,  p  <  .001.  *p  <  .01,  **p  <  .001.  

It  seems  as  though  burnout  is  significantly  predicted  by  perceived  control,  coping  style  (as  
predicted  by  Cooper),  stress  from  teaching  and  stress  from  giving  pastoral  care.  The  Exp(B)  
and  direction  of  the  beta  values  tell  us  that,  for  perceived  control,  coping  ability  and  pastoral  
care,  the  relationships  are  positive.  That  is  (and  look  back  to  the  question  to  see  the  
direction  of  these  scales,  i.e.,  what  a  high  score  represents),  poor  perceived  control,  poor  
ability  to  cope  with  stress  and  stress  from  giving  pastoral  care  all  predict  burnout.  However,  
for  teaching,  the  relationship  if  the  opposite  way  around:  stress  from  teaching  appears  to  be  
a  positive  thing  as  it  predicts  not  becoming  burnt  out!        

Task 6
An  HIV  researcher  explored  the  factors  that  influenced  condom  use  with  a  new  partner  
(relationship  less  than  1  month  old).  The  outcome  measure  was  whether  a  condom  
was  used  (Use:  condom  used  =  1,  not  used  =  0).  The  predictor  variables  were  mainly  
scales  from  the  Condom  Attitude  Scale  (CAS)  by  Sacco,  Levine,  Reed,  and  Thompson  
(1991):  Gender;  the  degree  to  which  the  person  views  their  relationship  as  ‘safe’  from  
sexually  transmitted  disease  (Safety);  the  degree  to  which  previous  experience  
influences  attitudes  towards  condom  use  (Sexexp);  whether  or  not  the  couple  used  a  
condom  in  their  previous  encounter,  1  =  condom  used,  0  =  not  used,  2  =  no  previous  
encounter  with  this  partner  (Previous);  the  degree  of  self-­‐control  that  a  person  has  
when  it  comes  to  condom  use  (Selfcon);  the  degree  to  which  the  person  perceives  a  
risk  from  unprotected  sex  (Perceive).  Previous  research  (Sacco,  Rickman,  Thompson,  
Levine,  &  Reed,  1993)  has  shown  that  gender,  relationship  safety  and  perceived  risk  
predict  condom  use.  Carry  out  an  analysis  using  Condom.sav  to  verify  these  previous  
findings,  and  to  test  whether  self-­‐control,  previous  usage  and  sexual  experience  
predict  condom  use.  

The  correct  analysis  was  to  run  a  hierarchical  logistic  regression  entering  Perceive,  Safety  
and  Gender  in  the  first  block  and  Previous,  Selfcon  and  Sexexp  in  a  second.  I  used  forced  
entry  on  both  blocks,  but  you  could  choose  to  run  a  forward  stepwise  method  on  block  2  

PROFESSOR  ANDY  P  FIELD     24  

DISCOVERING  STATISTICS  USING  SPSS  

(either  strategy  is  justified).  For  the  variable  Previous  I  used  an  indicator  contrast  with  ‘No  
condom’  as  the  base  category.  

Block  0  

The  output  of  the  logistic  regression  will  be  arranged  in  terms  of  the  blocks  that  were  
specified.  In  other  words,  SPSS  will  produce  a  regression  model  for  the  variables  specified  in  
block  1,  and  then  produce  a  second  model  that  contains  the  variables  from  both  blocks  1  
and  2.  The  results  from  block  1  are  shown  below.  In  this  analysis  we  forced  SPSS  to  enter  
Perceive,  Safety  and  Gender  into  the  regression  model  first.  First,  the  output  tells  us  that  
100  cases  have  been  accepted,  that  the  dependent  variable  has  been  coded  0  and  1  
(because  this  variable  was  coded  as  0  and  1  in  the  data  editor,  these  codings  correspond  
exactly  to  the  data  in  SPSS).    

Case Processing Summary

a
Unweighted Cases N Percent
Selected Cases Included in Analysis 100 100.0
Missing Cases 0 .0
Total 100 100.0
Unselected Cases 0 .0
Total 100 100.0
a. If weight is in effect, see classification table for the total
number of cases.
 

Dependent Variable Encoding

Original Value Internal Value

Unprotected 0
Condom Used 1
 

Categorical Variables Codings

Parameter coding
Frequency (1) (2)
Previous No Condom 50 .000 .000
Use with Condom used 47 1.000 .000
Partner First Time with partner 3 .000 1.000
 

Classification Tablea,b

Predicted

Condom Use
Condom Percentage
Observed Unprotected Used Correct
Step 0 Condom Use Unprotected 57 0 100.0
Condom Used 43 0 .0
Overall Percentage 57.0
a. Constant is included in the model.
b. The cut value is .500
 
Output  19  

Block 1

The  next  part  of  the  output  tells  us  about  block  1:  as  such  it  provides  information  about  the  
model  after  the  variables  Perceive,  Safety  and  Gender  have  been  added.  The  first  thing  to  
note  is  that  −2LL  has  dropped  to  105.77,  which  is  a  change  of  30.89  (which  is  the  value  given  

PROFESSOR  ANDY  P  FIELD     25  

DISCOVERING  STATISTICS  USING  SPSS  

by  the  model  chi-­‐square).  This  value  tells  us  about  the  model  as  a  whole,  whereas  the  block  
tells  us  how  the  model  has  improved  since  the  last  block.  The  change  in  the  amount  of  
information  explained  by  the  model  is  significant  (χ2(3)  =  30.89,  p  <  .001)  and  so  using  
perceived  risk,  relationship  safety  and  gender  as  predictors  significantly  improves  our  ability  
to  predict  condom  use.  Finally,  the  classification  table  shows  us  that  74%  of  cases  can  be  
correctly  classified  using  these  three  predictors.  

Omnibus Tests of Model Coefficients

Chi-square df Sig.
Step 1 Step 30.892 3 .000
Block 30.892 3 .000
Model 30.892 3 .000
 

Model Summary

-2 Log Cox & Snell Nagelkerke

Step likelihood R Square R Square
1 105.770 .266 .357
 

Classification Tablea

Predicted

Condom Use
Condom Percentage
Observed Unprotected Used Correct
Step 1 Condom Use Unprotected 45 12 78.9
Condom Used 14 29 67.4
Overall Percentage 74.0
a. The cut value is .500

 
Output  20  

Hosmer  and  Lemeshow’s  goodness-­‐of-­‐fit  test  statistic  tests  the  hypothesis  that  the  
observed  data  are  significantly  different  from  the  predicted  values  from  the  model.  So,  in  
effect,  we  want  a  non-­‐significant  value  for  this  test  (because  this  would  indicate  that  the  
model  does  not  differ  significantly  from  the  observed  data).  In  this  case  (χ2(8)  =  9.70,  p  =  
.287)  it  is  non-­‐significant,  which  is  indicative  of  a  model  that  is  predicting  the  real-­‐world  data  
fairly  well.  

Hosmer and Lemeshow Test

Step Chi-square df Sig.

1 9.700 8 .287
 
Output  21  

Output  34,  labelled  Variables  in  the  Equation,  then  tells  us  the  parameters  of  the  model  
for  the  first  block.  The  significance  values  of  the  Wald  statistics  for  each  predictor  indicate  
that  both  perceived  risk  (Wald  =  17.78,  p  <  .001)  and  relationship  safety  (Wald  =  4.54,  p  <  
.05)  significantly  predict  condom  use.  Gender,  however,  does  not  (Wald  =  0.41,  p  >  .05).  

PROFESSOR  ANDY  P  FIELD     26  

DISCOVERING  STATISTICS  USING  SPSS  

Variables in the Equation

95.0% C.I.for EXP(B)

B S.E. Wald df Sig. Exp(B) Lower Upper
Step
a
PERCEIVE .940 .223 17.780 1 .000 2.560 1.654 3.964
1 SAFETY -.464 .218 4.540 1 .033 .629 .410 .963
GENDER .317 .496 .407 1 .523 1.373 .519 3.631
Constant -2.476 .752 10.851 1 .001 .084
a. Variable(s) entered on step 1: PERCEIVE, SAFETY, GENDER.
 
Output  22  

The  odds  ratio  for  perceived  risk  (Exp(B)  =  2.56,  CI0.95  =  [1.65,  3.96])  indicates  that  if  the  
value  of  perceived  risk  goes  up  by  1,  then  the  odds  of  using  a  condom  also  increase  (because  
Exp(B)  is  greater  than  1).  The  confidence  interval  for  this  value  ranges  from  1.65  to  3.96,  so  
we  can  be  very  confident  that  the  value  of  Exp(B)  in  the  population  lies  somewhere  between  
these  two  values.  What’s  more,  because  both  values  are  greater  than  1  we  can  also  be  
confident  that  the  relationship  between  perceived  risk  and  condom  use  found  in  this  sample  
is  true  of  the  whole  population.  In  short,  as  perceived  risk  increase  by  1,  people  are  just  over  
twice  as  likely  to  use  a  condom.  

The  odds  ratio  for  relationship  safety  (Exp(B)  =  0.63,  CI0.95  =  [0.41,  0.96])  indicates  that  if  
the  relationship  safety  increases  by  one  point,  then  the  odds  of  using  a  condom  decrease  
(because  Exp(B)    is  less  than  1).  The  confidence  interval  for  this  value  ranges  from  0.41  to  
0.96,  so  we  can  be  very  confident  that  the  value  of  Exp(B)  in  the  population  lies  somewhere  
between  these  two  values.  In  addition,  because  both  values  are  less  than  1  we  can  be  
confident  that  the  relationship  between  relationship  safety  and  condom  use  found  in  this  
sample  would  be  found  in  95%  of  samples  from  the  same  population.  In  short,  as  
relationship  safety  increases  by  one  unit,  subjects  are  about  1.6  times  less  likely  to  use  a  
condom.  

The  odds  ratio  for  gender  (Exp(B)  =  1.37,  CI0.95  =  [0.52,  3.63])  indicates  that  as  gender  
changes  from  0  (male)  to  1  (female),  then  the  odds  of  using  a  condom  increase  (because  
Exp(B)    is  greater  than  1).  However,  the  confidence  interval  for  this  value  crosses  1,  which  
limits  the  generalizability  of  our  findings  because  the  value  of  Exp(B)  in  other  samples  (and  
hence  the  population)  could  indicate  either  a  positive  (Exp(B)  >  1)  or  negative  (Exp(B)  <  1)  
relationship.  Therefore,  gender  is  not  a  reliable  predictor  of  condom  use.  

A  glance  at  the  classification  plot  (Output  35)  brings  not  such  good  news  because  a  lot  of  
cases  are  clustered  around  the  middle.  This  indicates  that  the  model  could  be  performing  
more  accurately  (i.e.,  the  classifications  made  by  the  model  are  not  completely  reliable).    

PROFESSOR  ANDY  P  FIELD     27  

DISCOVERING  STATISTICS  USING  SPSS  

 
Output  23  

Block 2

The  output  below  shows  what  happens  to  the  model  when  our  new  predictors  are  added  
(previous  use,  self-­‐control  and  sexual  experience).  This  part  of  the  output  describes  block  2,  
which  is  just  the  model  described  in  block  1  but  with  a  new  predictors  added.  So,  we  begin  
with  the  model  that  we  had  in  block  1  and  we  then  add  Previous,  Selfcon  and  Sexexp  to  it.  
The  effect  of  adding  these  predictors  to  the  model  is  to  reduce  the  –2  log-­‐likelihood  to  87.97  
(a  reduction  of  48.69  from  the  original  model  as  shown  in  the  model  chi-­‐square  and  an  
additional  reduction  of  17.80  from  the  reduction  caused  by  block  1  as  shown  by  the  block  
statistics).  This  additional  improvement  of  block  2  is  significant  (χ2(4)  =  17.80,  p  <  .01),  which  
tells  us  that  including  these  three  new  predictors  in  the  model  has  significantly  improved  our  
ability  to  predict  condom  use.  The  classification  table  tells  us  that  the  model  is  now  correctly  
classifying  78%  of  cases.  Remember  that  in  block  1  there  were  74%  correctly  classified  and  
so  an  extra  4%  of  cases  are  now  classified  (not  a  great  deal  more  –  in  fact,  examining  the  
table  shows  us  that  only  four  extra  cases  have  now  been  correctly  classified).  

PROFESSOR  ANDY  P  FIELD     28  

DISCOVERING  STATISTICS  USING  SPSS  

Omnibus Tests of Model Coefficients

Chi-square df Sig.
Step 1 Step 17.799 4 .001
Block 17.799 4 .001
Model 48.692 7 .000
 

Model Summary

-2 Log Cox & Snell Nagelkerke

Step likelihood R Square R Square
1 87.971 .385 .517
 

Hosmer and Lemeshow Test

Step Chi-square df Sig.

1 9.186 8 .327
 

Classification Tablea

Predicted

Condom Use
Condom Percentage
Observed Unprotected Used Correct
Step 1 Condom Use Unprotected 47 10 82.5
Condom Used 12 31 72.1
Overall Percentage 78.0
a. The cut value is .500
 
Output  24  

The  section  labelled  Variables  in  the  Equation  (Output  37)  now  contains  all  predictors.  
This  part  of  the  output  represents  the  details  of  the  final  model.  The  significance  values  of  
the  Wald  statistics  for  each  predictor  indicate  that  both  perceived  risk  (Wald  =  16.04,  p  <  
.001)  and  relationship  safety  (Wald  =  4.17,  p  <  .05)  still  significantly  predict  condom  use  and,  
as  in  block  1,  gender  does  not  (Wald  =  0.00,  p  >  .05).  We  can  now  look  at  the  new  predictors  
to  see  which  of  these  has  some  predictive  power.    

Variables in the Equation

95.0% C.I.for EXP(B)

B S.E. Wald df Sig. Exp(B) Lower Upper
Step
a PERCEIVE .949 .237 16.038 1 .000 2.583 1.623 4.109
1 SAFETY -.482 .236 4.176 1 .041 .617 .389 .980
GENDER .003 .573 .000 1 .996 1.003 .326 3.081
SEXEXP .180 .112 2.614 1 .106 1.198 .962 1.490
PREVIOUS 4.032 2 .133
PREVIOUS(1) 1.087 .552 3.879 1 .049 2.965 1.005 8.747
PREVIOUS(2) -.017 1.400 .000 1 .990 .983 .063 15.287
SELFCON .348 .127 7.510 1 .006 1.416 1.104 1.815
Constant -4.959 1.146 18.713 1 .000 .007
a. Variable(s) entered on step 1: SEXEXP, PREVIOUS, SELFCON.

 
Output  25  

Previous  use  has  been  split  into  two  components  (according  to  whatever  contrasts  were  
specified  for  this  variable).  Looking  at  the  very  beginning  of  the  output,  we  are  told  the  
parameter  codings  for  Previous(1)  and  Previous(2).  You  can  tell  by  remembering  the  rule  

PROFESSOR  ANDY  P  FIELD     29  

DISCOVERING  STATISTICS  USING  SPSS  

from  contrast  coding  in  ANOVA  which  groups  are  being  compared:  that  is,  we  compare  
groups  with  codes  of  0  against  those  with  codes  of  1.  From  the  output  we  can  see  that  
Previous(1)  compares  the  condom  used  group  against  the  no  condom  used  group,  and  
Previous(2)  compares  the  base  category  of  first  time  with  partner  against  the  other  two  
categories.  Therefore  we  can  tell  that  previous  use  is  not  a  significant  predictor  of  condom  
use  when  it  is  the  first  time  with  a  partner  compared  to  when  it  is  not  the  first  time  (Wald  =  
0.00,  p  <  .05).  However,  when  we  compare  the  condom  used  category  to  the  other  
categories  we  find  that  using  a  condom  on  the  previous  occasion  does  predict  use  on  the  
current  occasion  (Wald  =  3.88,  p  <  .05).  

Of  the  other  new  predictors  we  find  that  self-­‐control  predicts  condom  use  (Wald  =  7.51,  p  
<  .01)  but  sexual  experience  does  not  (Wald  =  2.61,  p  >  .05).  

The  odds  ratio  for  perceived  risk  (Exp(B)  =  2.58,  CI0.95  =  [1.62,  4.106])  indicates  that  if  the  
value  of  perceived  risk  goes  up  by  1,  then  the  odds  of  using  a  condom  also  increase.  What’s  
more,  because  the  confidence  interval  doesn’t  cross  1  we  can  also  be  confident  that  the  
relationship  between  perceived  risk  and  condom  use  found  in  this  sample  is  true  of  the  
whole  population.  As  perceived  risk  increases  by  1,  people  are  just  over  twice  as  likely  to  use  
a  condom.  

The  odds  ratio  for  relationship  safety  (Exp(B)  =  0.62,  CI0.95  =  [0.39,  0.98])  indicates  that  if  
the  relationship  safety  decreases  by  one  point,  then  the  odds  of  using  a  condom  increase.  
The  confidence  interval  does  not  cross  1  so  we  can  be  confident  that  the  relationship  
between  relationship  safety  and  condom  use  found  in  this  sample  would  be  found  in  95%  of  
samples  from  the  same  population.  As  relationship  safety  increases  by  one  unit,  subjects  are  
about  1.6  times  less  likely  to  use  a  condom.  

The  odds  ratio  for  gender  (Exp(B)  =  1.00,  CI0.95  =  [0.33,  3.08])  indicates  that  as  gender  
changes  from  0  (male)  to  1  (female),  then  the  odds  of  using  a  condom  do  not  change  
(because  Exp(B)  is  equal  to  1).  The  confidence  interval  crosses  1,  therefore  gender  is  not  a  
reliable  predictor  of  condom  use.  

The  odds  ratio  for  previous  use  (1)  (Exp(B)  =  2.97,  CI0.95  =  [1.01,  8.75])  indicates  that  if  the  
value  of  previous  usage  goes  up  by  1  (i.e.,  changes  from  not  having  used  one  or  being  the  
first  time  to  having  used  one),  then  the  odds  of  using  a  condom  also  increase.  What’s  more,  
because  the  confidence  interval  doesn’t  cross  1  we  can  also  be  confident  that  this  
relationship  is  true  in  the  whole  population.  If  someone  used  a  condom  on  their  previous  
encounter  with  this  partner  (compared  to  if  they  didn’t  use  one,  or  if  it  is  their  first  time)  
then  they  are  three  times  more  likely  to  use  a  condom.  For  previous  use  (2)  the  odds  ratio  
(Exp(B)  =  0.98,  CI0.95  =  [0.06,  15.29])  indicates  that  if  the  value  of  previous  usage  goes  up  by  1  
(i.e.,  changes  from  not  having  used  one  or  having  used  one  to  being  their  first  time  with  this  
partner),  then  the  odds  of  using  a  condom  do  not  change  (because  the  value  is  very  nearly  
equal  to  1).  What’s  more,  because  the  confidence  interval  crosses  1  we  can  tell  that  this  is  
not  a  reliable  predictor  of  condom  use.  

The  odds  ratio  for  self-­‐control  (Exp(B)  =  1.42,  CI0.95  =  [1.10,  1.82])  indicates  that  if  self-­‐
control  increases  by  one  point,  then  the  odds  of  using  a  condom  increase  also.  The  

PROFESSOR  ANDY  P  FIELD     30  

DISCOVERING  STATISTICS  USING  SPSS  

confidence  interval  does  not  cross  1,  so  we  can  be  confident  that  the  relationship  between  
relationship  safety  and  condom  use  found  in  this  sample  would  be  found  in  95%  of  samples  
from  the  same  population.  As  self-­‐control  increases  by  one  unit,  subjects  are  about  1.4  times  
more  likely  to  use  a  condom.  

The  odds  ratio  for  sexual  experience  (Exp(B)  =  1.20,  CI0.95  =  [0.95,  1.49])  indicates  that  as  
sexual  experience  increases  by  one  unit,  then  the  odds  of  using  a  condom  increase  slightly.  
However,  the  confidence  interval  crosses  1,  therefore  sexual  experience  is  not  a  reliable  
predictor  of  condom  use.  

A  glance  at  the  classification  plot  (Output  38)  brings  good  news  because  a  lot  of  cases  
that  were  clustered  in  the  middle  are  now  spread  towards  the  edges.  Therefore,  overall  this  
new  model  is  more  accurately  classifying  cases  compared  to  block  1.    

 
Output  26  

Task 7
How  reliable  is  the  model  for  Task  6?      

Multicollinearity  can  affect  the  parameters  of  a  regression  model.  Logistic  regression  is  
equally  prone  to  the  biasing  effect  of  collinearity,  and  it  is  essential  to  test  for  collinearity  
following  a  logistic  regression  analysis  (see  the  book  for  details  of  how  to  do  this).  The  
results  of  the  analysis  are  shown  below.  Menard  (1995)  suggests  that  tolerance  values  below  
0.1    indicate  a  serious  collinearity  problem;  from  the  first  table  in  Output  39  we  can  see  that  
the  tolerance  values  for  all  variables  are  close  to  1,  much  larger  than  this  criterion.  Myers  
(1990)  also  suggests  that  a  VIF  value  greater  than  10  is  cause  for  concern,  and  in  these  data  
the  values  are  all  less  than  this  criterion.    

Output  39  also  shows  a  table  labelled  Collinearity  Diagnostics.  In  this  table,  we  are  given  
the  eigenvalues  of  the  scaled,  uncentred  cross-­‐products  matrix,  the  condition  index  and  the  

PROFESSOR  ANDY  P  FIELD     31  

DISCOVERING  STATISTICS  USING  SPSS  

variance  proportions  for  each  predictor.  If  any  of  the  eigenvalues  in  this  table  are  much  
larger  than  others  then  the  uncentred  cross-­‐products  matrix  is  said  to  be  ill-­‐conditioned,  
which  means  that  the  solutions  of  the  regression  parameters  can  be  greatly  affected  by  
small  changes  in  the  predictors  or  outcome.  In  plain  English,  these  values  give  us  some  idea  
as  to  how  accurate  our  regression  model  is:  if  the  eigenvalues  are  fairly  similar  then  the  
derived  model  is  likely  to  be  unchanged  by  small  changes  in  the  measured  variables.  The  
condition  indexes  are  another  way  of  expressing  these  eigenvalues  and  represent  the  square  
root  of  the  ratio  of  the  largest  eigenvalue  to  the  eigenvalue  of  interest  (so,  for  the  dimension  
with  the  largest  eigenvalue,  the  condition  index  will  always  be  1).  For  these  data  the  
condition  indexes  are  all  relatively  similar,  showing  that  a  problem  is  unlikely  to  exist.  

Coefficientsa

Collinearity Statistics
Model Tolerance VIF
1 Perceived Risk .849 1.178
Relationship Safety .802 1.247
GENDER .910 1.098
2 Perceived Risk .740 1.350
Relationship Safety .796 1.256
GENDER .885 1.130
Previous Use with Partner .964 1.037
Self-Control .872 1.147
Sexual experience .929 1.076
a. Dependent Variable: Condom Use
 
Collinearity Diagnosticsa

Variance Proportions
Previous
Condition Perceived Relationship Use with Sexual
Model Dimension Eigenvalue Index (Constant) Risk Safety GENDER Partner Self-Control experience
1 1 3.137 1.000 .01 .02 .02 .03
2 .593 2.300 .00 .02 .10 .55
3 .173 4.260 .01 .55 .76 .08
4 9.728E-02 5.679 .98 .40 .13 .35
2 1 5.170 1.000 .00 .01 .01 .01 .01 .01 .01
2 .632 2.860 .00 .02 .06 .43 .10 .00 .02
3 .460 3.352 .00 .03 .10 .01 .80 .00 .00
4 .303 4.129 .00 .07 .01 .24 .00 .00 .60
5 .235 4.686 .00 .04 .34 .17 .05 .50 .00
6 .135 6.198 .01 .61 .40 .00 .00 .47 .06
7 6.510E-02 8.911 .98 .23 .08 .14 .03 .03 .31

 
a. Dependent Variable: Condom Use

Output  27  

The  final  step  in  analysing  this  table  is  to  look  at  the  variance  proportions.  The  variance  of  
each  regression  coefficient  can  be  broken  down  across  the  eigenvalues,  and  the  variance  
proportions  tell  us  the  proportion  of  the  variance  of  each  predictor’s  regression  coefficient  
that  is  attributed  to  each  eigenvalue.  These  proportions  can  be  converted  to  percentages  by  
multiplying  them  by  100  (to  make  them  more  easily  understood).  In  terms  of  collinearity,  we  
are  looking  for  predictors  that  have  high  proportions  on  the  same  small  eigenvalue,  because  
this  would  indicate  that  the  variances  of  their  regression  coefficients  are  dependent  (see  the  
main  textbook  for  more  detail).  Again,  no  variables  appear  to  have  similarly  high  variance  
proportions  for  the  same  dimensions.  The  result  of  this  analysis  is  pretty  clear-­‐cut:  there  is  
no  problem  of  collinearity  in  these  data.    

Residuals  should  be  checked  for  influential  cases  and  outliers.  As  a  brief  guide,  the  output  
lists  cases  with  standardized  residuals  greater  than  2.  In  a  sample  of  100,  we  would  expect  
around  5–10%  of  cases  to  have  standardized  residuals  with  absolute  values  greater  than  

PROFESSOR  ANDY  P  FIELD     32  

DISCOVERING  STATISTICS  USING  SPSS  

this.  For  these  data  we  have  only  four  cases  and  only  one  of  these  has  an  absolute  value  
greater  than  3  (Output  40).  Therefore,  we  can  be  fairly  sure  that  there  are  no  outliers.  

Casewise Listb

Selected Observed Predicted Temporary Variable

a
Case Status Condom Use Predicted Group Resid ZResid
41 S U** .891 C -.891 -2.855
53 S U** .916 C -.916 -3.294
58 S C** .142 U .858 2.455
83 S C** .150 U .850 2.380
a. S = Selected, U = Unselected cases, and ** = Misclassified cases.
b. Cases with studentized residuals greater than 2.000 are listed.
 
Output  28  

Task 8
Using  the  final  model  of  condom  use  in  Task  6,  what  are  the  probabilities  that  
participants  12,  53  and  75  will  use  a  condom?  

The  values  predicted  for  these  cases  will  depend  on  exactly  how  you  ran  the  analysis  (and  
the  parameter  coding  used  on  the  variable  Previous).  Therefore,  your  answers  might  differ  
slightly  from  mine.  

Case Summariesa

Case Predicted
Number Value Predicted Group
12 12 .49437 Unprotected
53 53 .88529 Condom Used
75 75 .37137 Unprotected

a. Limited to first 100 cases.

 
Output  29  

Task 9
A  female  who  used  a  condom  in  her  previous  encounter  with  her  new  partner  scores  2  
on  all  variables  except  perceived  risk  (for  which  she  scores  6).  Use  the  model  to  
estimate  the  probability  that  she  will  use  a  condom  in  her  next  encounter.  

Step  1.  Logistic  regression  equation:  

1
  P(Y ) =  
1 + e− Z

where  

  Z = b0 + b1 X1 + b2 X 2 + ... + bn X n  

PROFESSOR  ANDY  P  FIELD     33  

DISCOVERING  STATISTICS  USING  SPSS  

Step  2.  Use  the  values  of  b  from  the  SPSS  output  (final  model  and  the  values  of  X  for  each  
variable  (from  the  question)  to  construct  the  following  table:  

Variable   bi   Xi   bi  Xi  

Gender   0.0027   1   0.0027  

Safety   −0.4823   2   −0.9646  

Sexexp   0.1804   2   0.3608  

Previous  (1)   1.0870   1   1.0870  

Previous  (2)   −.0167   0   0  

Selfcon   0.3476   2   0.6952  

Perceive   0.9489   6   5.6934  

Step  3.  Place  the  values  of  bi  Xi  into  the  equation  for  z  (remembering  to  include  the  
constant):  

z = −4.6009 + 0.0027 − 0.9646 + 0.3608 + 1.0870 + 0 + 0.6952 + 5.6934

= 2.2736  

Step  4.  Replace  this  value  of  z  into  the  logistic  regression  equation:  

1 1
P(Y ) = =
1+ e −Z
1 + e−2.2736  
= .9067

Therefore,  there  is  a  91%  chance  that  she  will  use  a  condom  on  her  next  encounter.  

Task 10
At  the  start  of  the  chapter  we  looked  at  whether  the  type  of  instrument  a  person  plays  
is  connected  to  their  personality.  A  musicologist  got  200  singers  and  guitarists  from  
bands.  She  noted  the  Instrument  they  played    (Singer,  Guitar),  and  measured  two  
personality  variables  in  each:  Extroversion  and  Agreeableness.  Conduct  a  logistic  

PROFESSOR  ANDY  P  FIELD     34  

DISCOVERING  STATISTICS  USING  SPSS  

regression  to  see  which  of  these  variables  (ignore  the  interaction)  predicts  which  
instrument  a  person  plays.  Data  are  in  Sing  or  Guitar.sav.  

Logistic  regression  is  located  in  the  regression  menu  accessed  by  selecting  
.  Following  this  menu  path  activates  the  main  Logistic  
Regression  dialog  box  shown  in  Figure  5.  Complete  the  dialog  boxes  as  shown  in  Figures  5–7  
(see  the  book  chapter  for  more  information  on  how  to  complete  these  boxes).  

 
Figure  5  

 
Figure  6  

PROFESSOR  ANDY  P  FIELD     35  

DISCOVERING  STATISTICS  USING  SPSS  

 
Figure  7  

Interpreting the output

PROFESSOR  ANDY  P  FIELD     36  

DISCOVERING  STATISTICS  USING  SPSS  

 
Output  30  

For  this  first  analysis  we  requested  a  forward  stepwise  method  and  so  the  initial  model  is  
derived  using  only  the  constant  in  the  regression  equation.  Output  42  tells  us  about  the  
model  when  only  the  constant  is  included  (i.e.,  all  predictor  variables  are  omitted).  The  log-­‐
likelihood  of  this  baseline  model  is  271.957.  This  represents  the  fit  of  the  model  when  the  
most  basic  model  is  fitted  to  the  data.  When  including  only  the  constant,  the  computer  
bases  the  model  on  assigning  every  participant  to  a  single  category  of  the  outcome  variable.  
The  output  shows  a  contingency  table  for  the  model  in  this  basic  state.  You  can  see  that  
SPSS  has  predicted  that  all  participants  are  singers,  which  results  in  0%  accuracy  for  the  
participants  who  played  the  guitar,  and  100%  accuracy  for  those  participants  who  were  
singers.  Overall,  the  model  correctly  classifies  53.8%  of  participants.  The  next  part  of  the  
output  summarizes  the  model,  and  at  this  stage  this  entails  quoting  the  value  of  the  
constant  (b0),  which  is  equal  to  −0.153.  

The  final  table  of  the  output  is  labelled  Variables  not  in  the  Equation.  The  bottom  line  of  
this  table  reports  the  residual  chi-­‐square  statistic  as  115.231  which  is  significant  at  p  <  .001  
(it  labels  this  statistic  Overall  Statistics).  This  statistic  tells  us  that  the  coefficients  for  the  
variables  not  in  the  model  are  significantly  different  from  zero  –  in  other  words,  that  the  
addition  of  one  or  more  of  these  variables  to  the  model  will  significantly  affect  its  predictive  
power.  If  the  probability  for  the  residual  chi-­‐square  had  been  greater  than  .05  it  would  have  
meant  that  none  of  the  variables  excluded  from  the  model  could  make  a  significant  
contribution  to  the  predictive  power  of  the  model.  As  such,  the  analysis  would  have  
terminated  at  this  stage.  

The  remainder  of  this  table  lists  both  of  the  predictors  in  turn  with  a  value  of  Roa’s  
efficient  score  statistic  for  each  one  (column  labelled  Score).  In  large  samples  when  the  null  
hypothesis  is  true,  the  score  statistic  is  identical  to  the  Wald  statistic  and  the  likelihood  ratio  
statistic.  It  is  used  at  this  stage  of  the  analysis  because  it  is  computationally  less  intensive  
than  the  Wald  statistic  and  so  can  still  be  calculated  in  situations  when  the  Wald  statistic  
would  prove  prohibitive.  Like  any  test  statistic,  Roa’s  score  statistic  has  a  specific  distribution  
from  which  statistical  significance  can  be  obtained.  In  this  example,  both  excluded  variables  
have  significant  score  statistics  at  p  <  .001  and  so  both  could  potentially  make  a  contribution  
to  the  model.  The  stepwise  calculations  are  relative  and  so  the  variable  that  will  be  selected  
for  inclusion  is  the  one  with  the  highest  value  for  the  score  statistic  that  is  significant  at  the  
.05  level.  In  this  example,  that  variable  will  be  Extroversion  because  it  has  the  highest  value  
of  the  score  statistic.  The  next  part  of  the  output  deals  with  the  model  after  this  predictor  
has  been  added.  

PROFESSOR  ANDY  P  FIELD     37  

DISCOVERING  STATISTICS  USING  SPSS  

In  the  first  step,  Extroversion  is  added  to  the  model  as  a  predictor.  As  such,  a  participant  
is  classified  as  being  a  singer  based  on  how  extroverted  they  are.    In  the  second  step,  
Agreeableness  is  added  to  the  model  as  a  predictor.  As  such  a  participant  is  classified  as  
being  a  singer  based  on  how  agreeable  they  are.    

 
Output  31  

Output  43  shows  summary  statistics  about  the  two  new  models.  The  overall  fit  of  the  
new  models  is  assessed  using  the  log-­‐likelihood  statistic.  In  SPSS,  rather  than  reporting  the  
log-­‐likelihood  itself,  the  value  is  multiplied  by  –2  (and  sometimes  referred  to  as  –2LL):  this  
multiplication  is  done  because  –2LL  has  an  approximately  chi-­‐square  distribution  and  so  
makes  it  possible  to  compare  values  against  those  that  we  might  expect  to  get  by  chance  
alone.  Remember  that  large  values  of  the  log-­‐likelihood  statistic  indicate  poorly  fitting  
statistical  models.  

The  value  of  −2  log-­‐likelihood  for  each  new  model  should  be  less  than  the  value  for  the  
previous  model  (because  lower  values  of  −2LL  indicate  that  the  model  is  predicting  the  
outcome  variable  more  accurately).  When  only  the  constant  was  included,  −2LL  =  271.96.  
When  extroversion  was  included  (step  1)  this  value  was  reduced  to  83.41,  which  tells  us  that  
the  model  is  better  at  predicting  instrument  played  than  it  was  before  Extroversion  was  
added.  Finally,  when  Agreeableness  was  added  to  the  model  (step  2),  this  value  of  −2LL  was  
reduced  even  further  to  46.78,  which  tells  us  that  the  model  was  better  at  predicting  which  
instrument  participants’  played  when  both  predictors  were  included.    The  question  of  how  
much  better  the  models  predicts  the  outcome  variable  can  be  assessed  using  the  model  chi-­‐
square  statistic,  which  measures  the  difference  between  the  model  as  it  currently  stands  
and  the  model  when  only  the  constant  was  included.  We  can  assess  the  significance  of  the  

PROFESSOR  ANDY  P  FIELD     38  

DISCOVERING  STATISTICS  USING  SPSS  

change  in  a  model  by  taking  the  log-­‐likelihood  of  the  new  model  and  subtracting  the  log-­‐
likelihood  of  the  baseline  model  from  it.  The  value  of  the  model  chi-­‐square  statistic  when  
only  Extroversion  is  included  is  271.96  −  83.41  =  188.55  and  the  value  of  the  model  chi-­‐
square  statistic  when  Agreeableness  is  added  to  the  model  is  83.41  −  46.78  =  36.63;  these  
values  are  both  significant.  Therefore,  the  model  that  included  extroversion  predicted  
instrument  played  significantly  better  than  the  model  when  only  the  constant  was  included,  
and  the  model  that  included  both  agreeableness  and  extroversion  predicted  instrument  
played  better  than  when  only  extroversion  and  the  constant  were  included  in  the  model.    

The  classification  table  in  Output  43  indicates  how  well  the  model  predicts  group  
membership.  In  step  1,  the  model  correctly  classifies  106  participants  who  are  singers  and  
does  not  misclassify  any  (i.e.,  it  correctly  classifies  100%  of  cases).  For  participants  who  play  
guitar,  the  model  correctly  classifies  84  and  misclassifies  7  cases  (i.e.,  it  correctly  classifies  
92.3%  of  cases).  The  overall  accuracy  of  classification  is,  therefore,  the  weighted  average  of  
these  two  values  (96.4%).  So,  when  only  the  constant  was  included,  the  model  correctly  
classified  53.8%  of  children,  but  now,  with  the  inclusion  of  extroversion  as  a  predictor,  this  
has  risen  to  96.4%.  In  step  2,  the  model  correctly  classifies  103  participants  who  are  singers  
but  misclassifies  3  others  (i.e.,  it  correctly  classifies  97.2%  of  cases).  For  participants  who  
play  guitar,  the  model  correctly  classifies  87  and  misclassifies  4  cases  (i.e.,  it  correctly  
classifies  95.6%  of  cases).  The  overall  accuracy  of  classification  is,  therefore,  the  weighted  
average  of  these  two  values  (96.4%).  So,  when  only  the  constant  was  included,  the  model  
correctly  classified  53.8%  of  children,  but  now,  with  the  inclusion  of  extroversion  as  a  
predictor,  this  has  risen  to  96.4%.  

 
Output  32  

The  next  part  of  the  output  is  crucial  because  it  tells  us  the  estimates  for  the  coefficients  
for  the  predictors  included  in  the  model  (Output  44).  The  interpretation  of  this  coefficient  in  
logistic  regression  is  that  it  represents  the  change  in  the  logit  of  the  outcome  variable  
associated  with  a  one-­‐unit  change  in  the  predictor  variable.  The  logit  of  the  outcome  is  
simply  the  natural  logarithm  of  the  odds  of  Y  occurring.  The  crucial  statistic  is  the  Wald  
statistic,  which  has  a  chi-­‐square  distribution  and  tells  us  whether  the  b  coefficient  for  that  
predictor  is  significantly  different  from  zero.  If  the  coefficient  is  significantly  different  from  
zero  then  we  can  assume  that  the  predictor  is  making  a  significant  contribution  to  the  
prediction  of  the  outcome  (Y).  For  these  data  it  seems  to  indicate  that  both  extroversion  and  
agreeableness  are  significant  predictors  of  instrument  played  (note  the  significance  of  the  
Wald  statistics  are  both  less  than  .05).  

PROFESSOR  ANDY  P  FIELD     39  

DISCOVERING  STATISTICS  USING  SPSS  

 
Output  33  

The  final  thing  we  need  to  look  at  is  exp  b  (Exp(B)  in  the  SPSS  output).  We  can  interpret  
exp  b  in  terms  of  the  change  in  odds.  If  the  value  is  greater  than  1  then  it  indicates  that  as  
the  predictor  increases,  the  odds  of  the  outcome  occurring  increase.  Conversely,  a  value  less  
than  1  indicates  that  as  the  predictor  increases,  the  odds  of  the  outcome  occurring  
decrease.  In  this  example,  the  exp  b  for  extroversion  in  step  2  is  0.238,  which  is  less  than  1  
thus  indicating  that  as  the  predictor  (extroversion)  increases,  the  value  of  the  outcome  
decreases,  that  is,  the  value  of  the  categorical  variable  moves  from  1  (guitarist)  to  0  (singer).  
In  other  words,  more  extroverted  participants  are  more  likely  to  be  singers.  The  exp  b  for  
agreeableness  is  1.43,  which  is  greater  than  1,  thus  indicating  that  as  the  predictor  
(agreeableness)  increases,  the  value  of  the  outcome  also  increases  (i.e.,  the  value  of  the  
categorical  variable  moves  from  0  (singer)  to  1  (guitarist)).  In  other  words,  more  agreeable  
participants  are  more  likely  to  be  guitarists.  Essentially,  singers  are  predicted  from  less  
agreeableness  and  more  extroversion  whereas  guitarists  are  predicted  from  more  
agreeableness  and  less  extroversion.    

Task 11
Which  problem  associated  with  logistic  regression  might  we  have  in  the  analysis  for  
Task  10?  

Looking  at  the  classification  plot,  it  looks  as  though  we  might  have  complete  separation.  The  
model  almost  perfectly  predicts  group  membership.  

PROFESSOR  ANDY  P  FIELD     40  

DISCOVERING  STATISTICS  USING  SPSS  

Step number: 1

Observed Groups and Predicted Probabilities

80 + +
IS I
IS I
F IS I
R 60 +S G+
E IS GI
Q IS GI
U IS GI
E 40 +S G+
N IS GI
C IS GI
Y IS GI
20 +S G+
IS GI
IS GI
ISSS S S S S S G G G G GGGGI
Predicted ---------+---------+---------+---------+---------+---------+---------+---------+---------+----------
Prob: 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1
Group: SSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGG

Predicted Probability is of Membership for Guitar

The Cut Value is .50
Symbols: S - Singer
G - Guitar
Each Symbol Represents 5 Cases.
 
Output  34  

Task 12
The  musicologist  extended  her  study  by  collecting  data  from  430  musicians.  Again  she  
noted  the  Instrument  a  person  played    (Singer,  Guitar,  Bass,  Drums),  and  the  same  
personality  variables.  However,  she  also  measured  Conscientiousness.  Conduct  a  
multinomial  logistic  regression  to  see  which  of  these  three  variables  (ignore  
interactions)  predicts  which  instrument  a  person  plays  (use  drums  as  the  reference  
category).  The  data  are  in  Band  Personality.sav.  

To  run  multinomial  logistic  regression  in  SPSS,  first  select  the  main  dialog  box  by  selecting    
.  In  this  dialog  box  (Figure  )  there  are  spaces  
to  place  the  outcome  variable  (Instrument),  any  categorical  predictors  (Factor(s))  and  any  
continuous  predictors  (Covariate(s)).  In  this  example,  the  outcome  variable  is  Instrument,  so  
select  this  variable  from  the  list  and  transfer  it  to  the  box  labelled  Dependent  by  dragging  it  
there  or  clicking  on   .  We  also  have  to  tell  SPSS  whether  we  want  to  compare  categories  
against  the  first  category  or  the  last,  and  we  do  this  by  clicking  on   .  By  default  
SPSS  uses  the  last  category,  which  is  perfect  for  us  because  drums  is  the  last  category  and  is  
also  the  category  that  we  want  to  use  as  our  reference  category.  

Next  we  have  to  specify  the  predictor  variables.  We  have  three  continuous  predictors  or  
covariates  (Agreeableness,  Extroversion  and  Conscientiousness).  Drag  all  three  to  the  box  
labelled  Covariate(s)  or  click  on   .  For  a  basic  analysis  in  which  all  of  these  predictors  are  
forced  into  the  model,  this  is  all  we  really  need  to  do.    

PROFESSOR  ANDY  P  FIELD     41  

DISCOVERING  STATISTICS  USING  SPSS  

 
Figure  8  

Output

 
Output  35  

Output  47  displays  the  results  of  the  log-­‐likelihood,  which  is  a  measure  of  how  much  
unexplained  variability  there  is  in  the  data;  therefore,  the  difference  or  change  in  log-­‐
likelihood  indicates  how  much  new  variance  has  been  explained  by  the  model.  The  chi-­‐
square  test  tests  the  decrease  in  unexplained  variance  from  the  baseline  model  (1122.82)  to  
the  final  model  (450.91),  which  is  a  difference  of  1149.53−871  =  672.02.  This  change  is  
significant,  which  means  that  our  final  model  explains  a  significant  amount  of  the  original  
variability  (in  other  words,  it’s  a  better  fit  than  the  original  model).  

PROFESSOR  ANDY  P  FIELD     42  

DISCOVERING  STATISTICS  USING  SPSS  

 
Output  36  

The  next  part  of  the  output  relates  to  the  fit  of  the  model  to  the  data  (Output  48).  We  
know  that  the  model  is  significantly  better  than  no  model,  but  is  it  a  good  fit  of  the  data?  
The  Pearson  and  deviance  statistics  test  the  same  thing,  which  is  whether  the  predicted  
values  from  the  model  differ  significantly  from  the  observed  values.  If  these  statistics  are  not  
significant  then  the  model  is  a  good  fit.  Here  we  have  contrasting  results:  the  deviance  
statistic  says  that  the  model  is  a  good  fit  of  the  data  (p  =  1.00,  which  is  much  higher  than  
.05),  but  the  Pearson  test  indicates  the  opposite,  namely  that  predicted  values  are  
significantly  different  from  the  observed  values  (p  <  .001).  Oh  dear.  Differences  between  
these  statistics  can  be  caused  by  overdispersion.  This  is  a  possibility  that  we  need  to  look  
into.  We  can  compute  the  dispersion  parameters  from  both  statistics:  

 
!
𝜒Pearson 1042672.72
𝜙Pearson = = = 914.63  
𝑑𝑓 1140
!
𝜒Deviance 448.032
𝜙Deviance = = = 0.39  
𝑑𝑓 1140

As  we  can  see,  the  dispersion  parameter  based  on  the  Pearson  statistic  is  914.63,  which  is  
ridiculously  high  compared  to  the  value  of  2,  which  I  cited  in  the  chapter  as  being  a  
threshold  for  ‘problematic’.  Conversely,  the  value  based  on  the  deviance  statistic  is  below  1,  
which  we  saw  in  the  chapter  indicated  underdispersion.  Again,  these  values  contradict,  so  all  
we  can  really  be  sure  of  is  that  there’s  something  pretty  weird  going  on.  Large  dispersion  
parameters  can  occur  for  reasons  other  than  overdispersion,  for  example  omitted  variables  
or  interactions  and  predictors  that  violate  the  linearity  of  the  logit  assumption.  In  this  
example  there  were  several  interaction  terms  that  we  could  have  entered  but  chose  not  to,  
which  might  go  some  way  to  explaining  these  strange  results.  The  output  also  shows  us  the  
two  other  measures  of  R2.  The  first  is  Cox  and  Snell’s  measure,  which  SPSS  reports  as  .81,  
and  the  second  is  Nagelkerke’s  adjusted  value,  which  SPSS  reports  as  .86.  As  you  can  see,  
they  are  reasonably  similar  values  and  represent  very  large  effects.  

PROFESSOR  ANDY  P  FIELD     43  

DISCOVERING  STATISTICS  USING  SPSS  

 
Output  37  

The  next  part  of  the  output  shows  the  results  of  the  likelihood  ratio  tests  (Output  49),  
and  these  can  be  used  to  ascertain  the  significance  of  predictors  to  the  model.  This  table  
tells  us  that  extroversion  had  a  significant  main  effect  on  type  of  instrument  played,  χ2(3)  =  
339.73,  p  <  .001,  as  did  agreeableness,  χ2(3)  =  100.16,  p  <  .001,  and  conscientiousness,  χ2(3)  
=  84.26,  p  <  .001.  These  likelihood  statistics  can  be  seen  as  sorts  of  overall  statistics  that  tell  
us  which  predictors  significantly  enable  us  to  predict  the  outcome  category,  but  they  don’t  
really  tell  us  specifically  what  the  effect  is.  To  see  this  we  have  to  look  at  the  individual  
parameter  estimates  (Output  50).  

 
Output  38  

We  specified  the  last  category  (drums)  as  our  reference  category;  therefore,  each  section  
of  Output  50  is  comparing  one  of  the  instrument  categories  against  the  drums  category.  To  
do  this  SPSS  recodes  the  instrument  variable  using  standard  dummy  coding,  so  for  each  
comparison,  drums  will  be  coded  0  and  the  instrument  that  it  is  being  compared  to  (singer,  
guitar,  or  bass)  will  be  coded  as  1.  Let’s  look  at  the  effects  one  by  one;  because  we  are  just  
comparing  two  categories  the  interpretation  is  the  same  as  for  binary  logistic  regression  (so  
if  you  don’t  understand  my  conclusions  reread  the  book  chapter):  

• Extroversion.  How  extroverted  the  participant  was  significantly  predicted  whether  

they  were  a  drummer  or  a  singer,  b  =  1.70,  Wald  χ2(1)  =  54.34,  p  <  .001.  The  odds  
ratio  tells  us  that  as  extroversion  increases  by  one  unit,  the  change  in  the  odds  of  
being  a  singer  (rather  than  being  a  drummer)  is  5.47.  The  odds  ratio  (5.47)  is  greater  
than  1,  therefore  we  can  say  that  as  participants  move  up  the  extroversion  scale,  

PROFESSOR  ANDY  P  FIELD     44  

DISCOVERING  STATISTICS  USING  SPSS  

they  were  more  likely  to  be  a  singer  (coded  1)  than  they  were  to  be  a  drummer  
(coded  0).  Similarly,  how  extroverted  the  participant  was  also  significantly  predicted  
whether  they  were  a  drummer  or  a  bass  player,  b  =  0.25,  Wald  χ2(1)  =  18.28,  p  <  
.001.  The  odds  ratio  tells  us  that  as  extroversion  increases  by  one  unit,  the  change  in  
the  odds  of  being  a  bass  player  (rather  than  being  a  drummer)  is  1.29,  so  the  more  
extroverted  the  participant  was,  the  more  likely  they  were  to  be  a  bass  player  than  
they  were  to  be  a  drummer.  However,  how  extroverted  the  participant  was  did  not  
significantly  predict  whether  they  were  a  drummer  or  a  guitarist,  b  =  .06,  Wald  
χ2(1)  =  3.58,  p  =  .06.    
• Agreeableness.  How  agreeable  the  participant  was  significantly  predicted  whether  
they  were  a  drummer  or  a  singer,  b  =  −0.40,  Wald  χ2(1)  =  35.49,  p  <  .001.  The  odds  
ratio  tells  us  that  as  agreeableness  increases  by  one  unit,  the  change  in  the  odds  of  
being  a  singer  (rather  than  being  a  drummer)  is  0.67,  so  the  more  agreeable  the  
participant  was,  the  more  likely  they  were  to  be  a  drummer  than  they  were  to  be  a  
singer.  Similarly,  how  agreeable  the  participant  was  also  significantly  predicted  
whether  they  were  a  drummer  or  a  bass  player,  b  =  −0.40,  Wald  χ2(1)  =  41.55,  p  <  
.001.  The  odds  ratio  tells  us  that  as  agreeableness  increases  by  one  unit,  the  change  
in  the  odds  of  being  a  bass  player  (rather  than  being  a  drummer)  is  0.67,  so,  the  
more  agreeable  the  participant  was,  the  more  likely  they  were  to  be  a  drummer  
than  they  were  to  be  a  bass  player.  However,  how  agreeable  the  participant  was  did  
not  significantly  predict  whether  they  were  a  drummer  or  a  guitarist,  b  =  .02,  Wald  
χ2(1)  =  0.51,  p  =  .48.    
• Conscientiousness.  How  conscientious  the  participant  was  significantly  predicted  
whether  they  were  a  drummer  or  a  singer,  b  =  −0.35,  Wald  χ2(1)  =  21.27,  p  <  .001.  
The  odds  ratio  tells  us  that  as  conscientiousness  increases  by  one  unit,  the  change  in  
the  odds  of  being  a  singer  (rather  than  being  a  drummer)  is  0.71,  so  the  more  
conscientious  the  participant  was,  the  more  likely  they  were  to  be  a  drummer  than  
they  were  to  be  a  singer.  Similarly,  how  conscientious  the  participant  was  also  
significantly  predicted  whether  they  were  a  drummer  or  a  bass  player,  b  =  −0.36,  
Wald  χ2(1)  =  40.93,  p  <  .001.  The  odds  ratio  tells  us  that  as  conscientiousness  
increases  by  one  unit,  the  change  in  the  odds  of  being  a  bass  player  (rather  than  
being  a  drummer)  is  0.70,  so  the  more  conscientious  the  participant  was,  the  more  
likely  they  were  to  be  a  drummer  than  they  were  to  be  a  bass  player.  However,  how  
conscientious  the  participant  was  did  not  significantly  predict  whether  they  were  a  
drummer  or  a  guitarist,  b  =  0.00,  Wald  χ2(1)  =  0.00,  p  =  1.00.    
 

PROFESSOR  ANDY  P  FIELD     45