Improving	
  Evalua/ons	
  and	
  U/liza/on	
  with	
  
Sta/s/cal	
  Edge:	
  Nested	
  Data	
  Designs	
  and	
  
Hierarchical	
  Linear	
  Modeling	
  (HLM)	
  
	
  
CES	
  Conference	
  -­‐	
  June	
  10	
  2013	
  
Marci	
  Pernica	
  -­‐	
  Ministry	
  of	
  Community	
  and	
  Social	
  Services	
  
Judith	
  Godin	
  –	
  J	
  Godin	
  Consul7ng	
  
•  The	
  goal	
  of	
  this	
  presenta0on	
  is	
  to	
  introduce	
  
the	
  concept	
  of	
  HLM	
  and	
  explain	
  how	
  it	
  can	
  be	
  
used	
  in	
  program	
  evalua0on	
  	
  
	
  
Introduc0on	
  
What	
  is	
  Hierarchical	
  Linear	
  Modeling	
  (HLM)	
  
•  HLM	
  is	
  a	
  sta0s0cal	
  technique	
  to	
  analyze	
  data	
  
that	
  is	
  structured	
  in	
  hierarchies	
  (or	
  “nested”)	
  
•  To	
  account	
  for	
  the	
  fact	
  that	
  people	
  that	
  are	
  “clustered”	
  or	
  
“nested”	
  within	
  the	
  same	
  group	
  have	
  more	
  in	
  common	
  than	
  
if	
  they	
  were	
  independent	
  random	
  samples	
  
	
  	
  
Classroom	
  1	
  
Classroom	
  
2	
  
Classroom	
  3	
  
Student	
  4	
  
Student	
  	
  
2	
  
Student	
  5	
  
Student	
  3	
   Student	
  8	
  Student	
  7	
  Student	
  6	
   Student	
  9	
  
Student	
  
10	
  
Student	
  
11	
  
Student	
  
12	
  
Nested	
  
data	
  
designs	
  
Student	
  
Class	
  
School	
  
District	
  
Hierarchical	
  Structure	
  –	
  mul0-­‐level	
  	
  
age	
   I.Q	
  
Measuring	
  
test	
  scores	
  
(dependent	
  variable)	
  
Independent	
  
variables	
  
•  HLM	
  enables	
  a	
  more	
  robust	
  analy0c	
  approach	
  
for	
  nested	
  data	
  (than	
  regression	
  or	
  ANOVA)	
  
•  Data	
  in	
  evalua0on	
  are	
  oZen	
  nested	
  
•  To	
  determine	
  success	
  condi*ons	
  for	
  the	
  
program	
  –	
  e.g.	
  is	
  the	
  program	
  more	
  suitable	
  
for	
  certain	
  sub-­‐popula0ons	
  or	
  more	
  successful	
  
if	
  delivered	
  in	
  a	
  certain	
  way	
  
Why	
  Use	
  HLM	
  in	
  Evalua0on	
  
Program	
  design	
  structure	
  
Data	
  structure	
  
Evalua0on	
  ques0ons	
  
-­‐	
  Which	
  par0cipant	
  or	
  site-­‐level	
  characteris0cs	
  are	
  most	
  influen0al	
  in	
  
explaining	
  the	
  varia0on	
  in	
  test	
  scores	
  among	
  the	
  program	
  
par0cipants?	
  
-­‐	
  What	
  program	
  delivery	
  characteris0cs	
  (site	
  level	
  prac0ces)	
  seem	
  to	
  be	
  
having	
  the	
  most	
  posi0ve	
  impact	
  on	
  the	
  par0cipants’	
  test	
  scores?	
  
-­‐	
  Are	
  some	
  program	
  features	
  more	
  suited	
  to	
  certain	
  sub-­‐popula0ons	
  (e.g.	
  
gender,	
  age	
  group,	
  ethnic	
  or	
  cultural	
  group)	
  
	
  
Applying	
  HLM	
  in	
  Evalua0on	
  
Par0cipant	
  
1	
  
Site	
  1	
   Site	
  2	
   Site	
  3	
  
Par0cipant	
  
4	
  
Par0cipant	
  
2	
  
Par0cipant
5	
  
Par0cipant	
  
3	
  
Par0cipant	
  
8	
  
Par0cipant	
  
7	
  
Par0cipant	
  
6	
  
Par0cipant	
  
9	
  
Par0cipant	
  
10	
  
Par0cipant	
  
11	
  
Par0cipant	
  
12	
  
Example	
  of	
  levels	
  of	
  a	
  hierarchical	
  model	
  
Par0cipants	
  (level	
  1)	
  nested	
  within	
  sites	
  (level	
  2)	
  
Assessing	
  test	
  scores	
  by	
  age	
  from	
  site	
  to	
  site	
  
Test	
  Score	
  
Age	
  
Four	
  different	
  
program	
  sites	
  
Although	
  the	
  test	
  scores	
  differ	
  from	
  site	
  to	
  site,	
  the	
  
rela0on	
  between	
  age	
  and	
  test	
  score	
  is	
  the	
  same	
  at	
  
different	
  sites	
  
Student	
  1	
   Student	
  2	
   Student	
  3	
  
Baseline	
  
Month	
  7	
  
Month	
  1	
  
Month	
  12	
  
Month	
  6	
   Month	
  9	
  Month	
  8	
  Month	
  6	
   Baseline	
   Month	
  2	
   Month	
  4	
   Month	
  6	
  
Assessing	
  change	
  over	
  0me	
  
Assessments	
  across	
  0me	
  (level	
  1)	
  are	
  nested	
  within	
  individuals	
  
(level	
  2)	
  (i.e.	
  repeated	
  measures	
  design)	
  
Assessing	
  improved	
  performance	
  over	
  0me	
  
Test	
  Score	
  
Time	
  
Four	
  different	
  
study	
  par0cipants	
  
Although	
  some	
  individuals	
  have	
  higher	
  test	
  scores	
  
to	
  start	
  with,	
  the	
  rate	
  of	
  change	
  (improved	
  
performance)	
  is	
  comparable	
  among	
  the	
  
par0cipants	
  
Tradi0onal	
  Methods	
  
Test	
  Score	
  
Age	
  
Rela0on	
  between	
  age	
  and	
  
test	
  score	
  es0mated	
  once	
  for	
  
all	
  sites	
  together	
  
Advantages	
  of	
  HLM	
  
Test	
  Score	
  
Age	
  
Four	
  different	
  
program	
  sites	
  
Here,	
  the	
  rela0on	
  between	
  age	
  and	
  test	
  score	
  varies	
  
across	
  sites.	
  
Are	
  there	
  any	
  site	
  level	
  variables	
  associated	
  with	
  the	
  
strength	
  of	
  this	
  rela0on?	
  
Design	
  Considera0ons	
  for	
  Using	
  HLM	
  
•  Sample	
  size	
  
– Par0cipant	
  level	
  
– Site	
  level	
  
– Repeated	
  measures	
  
•  Missing	
  Data	
  
– Can	
  be	
  easy	
  or	
  difficult	
  to	
  deal	
  with	
  
•  Number	
  of	
  variables	
  
– Comprehensive	
  coverage	
  
– Parsimony	
  
	
  
Final	
  Thoughts	
  	
  
•  Applying	
  HLM	
  in	
  evalua0ons	
  with	
  nested	
  data	
  
enables	
  more	
  robust	
  results	
  and	
  conclusions	
  
•  U0liza0on-­‐focused	
  	
  
– Iden0fy	
  evidence-­‐based	
  success	
  factors	
  or	
  
condi0ons	
  for	
  improving	
  the	
  program	
  delivery	
  
model,	
  to	
  ul0mately	
  achieve	
  beger	
  program	
  
effec0veness	
  	
  
– Promo0ng	
  the	
  value	
  in	
  evalua0on	
  (gathering	
  the	
  
evidence	
  to	
  determine	
  the	
  ‘success	
  factors’	
  for	
  
the	
  interven0on	
  to	
  be	
  effec0ve)	
  	
  
Ques0ons?	
  
	
  
Marci	
  Pernica	
  	
  	
  
Marci.Pernica@Ontario.ca	
  	
  
Ministry	
  of	
  Community	
  and	
  Social	
  Services	
  
	
  
Judith	
  Godin	
  	
  
sta/s/cs@jgodin.com	
  
Independent	
  Consultant	
  

Improving evaluations and utilization with statistical edge nested data designs and hierarchical linear modeling

  • 1.
    Improving  Evalua/ons  and  U/liza/on  with   Sta/s/cal  Edge:  Nested  Data  Designs  and   Hierarchical  Linear  Modeling  (HLM)     CES  Conference  -­‐  June  10  2013   Marci  Pernica  -­‐  Ministry  of  Community  and  Social  Services   Judith  Godin  –  J  Godin  Consul7ng  
  • 2.
    •  The  goal  of  this  presenta0on  is  to  introduce   the  concept  of  HLM  and  explain  how  it  can  be   used  in  program  evalua0on       Introduc0on  
  • 3.
    What  is  Hierarchical  Linear  Modeling  (HLM)   •  HLM  is  a  sta0s0cal  technique  to  analyze  data   that  is  structured  in  hierarchies  (or  “nested”)   •  To  account  for  the  fact  that  people  that  are  “clustered”  or   “nested”  within  the  same  group  have  more  in  common  than   if  they  were  independent  random  samples       Classroom  1   Classroom   2   Classroom  3   Student  4   Student     2   Student  5   Student  3   Student  8  Student  7  Student  6   Student  9   Student   10   Student   11   Student   12   Nested   data   designs  
  • 4.
    Student   Class   School   District   Hierarchical  Structure  –  mul0-­‐level     age   I.Q   Measuring   test  scores   (dependent  variable)   Independent   variables  
  • 5.
    •  HLM  enables  a  more  robust  analy0c  approach   for  nested  data  (than  regression  or  ANOVA)   •  Data  in  evalua0on  are  oZen  nested   •  To  determine  success  condi*ons  for  the   program  –  e.g.  is  the  program  more  suitable   for  certain  sub-­‐popula0ons  or  more  successful   if  delivered  in  a  certain  way   Why  Use  HLM  in  Evalua0on  
  • 6.
    Program  design  structure   Data  structure   Evalua0on  ques0ons   -­‐  Which  par0cipant  or  site-­‐level  characteris0cs  are  most  influen0al  in   explaining  the  varia0on  in  test  scores  among  the  program   par0cipants?   -­‐  What  program  delivery  characteris0cs  (site  level  prac0ces)  seem  to  be   having  the  most  posi0ve  impact  on  the  par0cipants’  test  scores?   -­‐  Are  some  program  features  more  suited  to  certain  sub-­‐popula0ons  (e.g.   gender,  age  group,  ethnic  or  cultural  group)     Applying  HLM  in  Evalua0on  
  • 7.
    Par0cipant   1   Site  1   Site  2   Site  3   Par0cipant   4   Par0cipant   2   Par0cipant 5   Par0cipant   3   Par0cipant   8   Par0cipant   7   Par0cipant   6   Par0cipant   9   Par0cipant   10   Par0cipant   11   Par0cipant   12   Example  of  levels  of  a  hierarchical  model   Par0cipants  (level  1)  nested  within  sites  (level  2)  
  • 8.
    Assessing  test  scores  by  age  from  site  to  site   Test  Score   Age   Four  different   program  sites   Although  the  test  scores  differ  from  site  to  site,  the   rela0on  between  age  and  test  score  is  the  same  at   different  sites  
  • 9.
    Student  1  Student  2   Student  3   Baseline   Month  7   Month  1   Month  12   Month  6   Month  9  Month  8  Month  6   Baseline   Month  2   Month  4   Month  6   Assessing  change  over  0me   Assessments  across  0me  (level  1)  are  nested  within  individuals   (level  2)  (i.e.  repeated  measures  design)  
  • 10.
    Assessing  improved  performance  over  0me   Test  Score   Time   Four  different   study  par0cipants   Although  some  individuals  have  higher  test  scores   to  start  with,  the  rate  of  change  (improved   performance)  is  comparable  among  the   par0cipants  
  • 11.
    Tradi0onal  Methods   Test  Score   Age   Rela0on  between  age  and   test  score  es0mated  once  for   all  sites  together  
  • 12.
    Advantages  of  HLM   Test  Score   Age   Four  different   program  sites   Here,  the  rela0on  between  age  and  test  score  varies   across  sites.   Are  there  any  site  level  variables  associated  with  the   strength  of  this  rela0on?  
  • 13.
    Design  Considera0ons  for  Using  HLM   •  Sample  size   – Par0cipant  level   – Site  level   – Repeated  measures   •  Missing  Data   – Can  be  easy  or  difficult  to  deal  with   •  Number  of  variables   – Comprehensive  coverage   – Parsimony    
  • 14.
    Final  Thoughts     •  Applying  HLM  in  evalua0ons  with  nested  data   enables  more  robust  results  and  conclusions   •  U0liza0on-­‐focused     – Iden0fy  evidence-­‐based  success  factors  or   condi0ons  for  improving  the  program  delivery   model,  to  ul0mately  achieve  beger  program   effec0veness     – Promo0ng  the  value  in  evalua0on  (gathering  the   evidence  to  determine  the  ‘success  factors’  for   the  interven0on  to  be  effec0ve)    
  • 15.
    Ques0ons?     Marci  Pernica       Marci.Pernica@Ontario.ca     Ministry  of  Community  and  Social  Services     Judith  Godin     sta/s/cs@jgodin.com   Independent  Consultant