PMT-based social registries

Measuring
income and poverty using
Proxy Means Tests
OVERVIEW

This  document  will  serve  as  an  introduction  to  proxy  means  testing  (PMT).  
The  concept  of  proxy  means  testing  is  usually  understood  as  using  
observable  characteristics  of  the  household  or  its  members  to  estimate  their  
incomes  or  consumption,  when  other  income  data  (salary  slips,  tax  returns)  
are  unavailable  or  unreliable.  

As  the  name  suggests,  the  obtained  income    

estimates  are  in  turn  used  for  
means-­testing,  that  is  to  
determine  whether  a  
household,  or  an  individual,  
is  eligible  for  support  from  
social  safety  net  programs.  If  
their  estimated  income  falls  
below  a  certain  threshold,  
they  are  deemed  eligible  for  
support;;  if  the  income  is  
higher  than  the  threshold,  the  
support  is  withdrawn.  
Throughout  this  guide,  we  
will  look  at  other  possible  
reasons  for  employing  the  
PMT  methodology,  the  
techniques  used  to  create  
and  apply  PMT  models  and  
look  at  some  of  the  
challenges  associated  with  
practical  PMT  application.    
   PMT  questionnaires  are  usually  shorter  and  easier  
What is the Proxy Means Test? to  administer  than  full-­fledged  consumption  surveys,  
Proxy  Means  Test  allows  us  to  estimate  the  income   we  can  also  use  them  to  get  a  quick  snapshot  of  
or  consumption  when  precise  measurements  are   poverty  rates  in  a  given  locality  and  at  a  given  time.  
unavailable  or  difficult  to  obtain.  In  many  situations,    
we  might  not  be  able  to  tell  how  much  a  family   Constructing the PMT model
earns  or  spends  every  month.  Even  the  household   In  the  previous  example,  we  have  decided  that  the  
members  themselves  might  not  be  able  to  tell— expected  consumption  of  families  with  brick-­walled  
they  seldom  maintain  detailed  records.   houses  will  be  higher  compared  to  the  ones  with  
However,  we  can  make  an  informed  guess  based   clad-­walled  dwellings;;  in  other  words,  brick  houses  
on  the  household  characteristics.  We  know  that,  all   are  constructed  by  more  affluent  people.  While  the  
things  remaining  equal,  a  family  living  in  a  brick-­ statement  will  probably  be  correct,  there  are  two  
walled  house  will  likely  be  able  to  afford  more   important  parts  that  we  have  missed:  
compared  to  a  family  living  in  a  house  made  of   •   We  do  not  have  a  proof;;  so  far,  our  statement  has  
clay.  The  type  of  wall  is  our  “proxy”,  because  we   only  been  based  on  our  intuitions;;  
can  use  it  to  approximate  our  value  of  interest—the   •   Even  if  we  know  that  a  family  with  a  brick-­walled  
household  income  or  consumption.     house  will  be  more  affluent  than  the  one  from  a  
Naturally,  using  only  one  proxy  value  will  render   clay  dwelling,  we  do  not  know  by  how  much;;  will  
our  estimations  very  imprecise—there  is  still  great   they  consume  goods  and  services  worth  twice  as  
variability  in  incomes,  even  between  families  living   much?  Or  three  times  as  much?  We  could  try  to  
in  houses  made  of  clay.  But  we  need  not  limit   come  up  with  a  number,  but  it  will  still  be  based  
ourselves  to  only  one  proxy.  If  we  see  that,  on   merely  on  our  own  ideas  and  intuitions.    
average,  richer  families  do  not  only  live  in  brick-­ A  formal  PMT  model  solves  those  two  problems  by  
walled  houses,  but  also  own  livestock,  such  as   being  based  on  actual  quantifiable  data.  The  data,  
cattle,  we  could  add  “livestock  ownership”  as   most  frequently,  comes  from  household  surveys.  
another  proxy  variable.  And  we  need  not  stop   Household  surveys  register  both  household  
there;;  we  can  imagine  adding  layers  of  other  proxy   characteristics  (such  as  the  type  of  dwelling,  the  type  
variables,  until  we  obtain  a  PMT  model—a  set  of   of  assets  that  the  household  owns,  composition  of  
variables  and  weights  associated  with  them  that   the  household,  as  well  as  ages,  genders  and  
accurately  predicts  the  welfare  of  different   occupations  of  its  members).  
household.  In  practice,  most  PMT  models  use  more   Because  the  surveys  tell  us  about  both  household  
than  a  dozen  of  different  variables.     characteristics  and  household  consumption,  they  
  allow  us  to  use  statistical  methods  to  find  out  about  
Why do we need to measure relationships  between  the  two.  Multiple  regression  
income/consumption? allows  us  to  estimate  the  relationship  between  
The  name  proxy  means  test  suggests  that  means   variables  describing  the  household  characteristics  
testing  (making  sure  that  the  social  safety  net   and  the  monetary  value  of  household  consumption.  
benefits  will  only  be  distributed  to  households  that   The  following  table  illustrates  the  example  result  of  
1
are  poor  enough  to  meet  certain  income  criteria)  is   such  an  analysis :  
its  main  purpose.    
This  is  true  and  the  use  of  PMT  for  targeting  is    
something  that  we  will  focus  on  here.  It  should  be    
noted,  though,  that  the  same  model  can  have  a  few  
 
other  uses.  For  example,  comparing  PMT  scores  
over  time  and  between  regions  can  help  is  in   1
 The  values  are  fictitious  and  do  not  correspond  to  the  
assessing  effectiveness  of  social  policies.  Since   PMT  model  actually  used  by  any  country  
 variable   coefficient   variable   coefficient  

Household  owns  the  house   100   Number  of  household  members  in   50  
employment  
One  child  in  the  household   40  
Number  of  persons  per  room   -­‐20  
Two  children  in  the  household   30  
 
Three  or  more  children  in  the  household   20  
In  this  case,  we  simply  have  to  multiply  the  
Household  owns  cattle   200  
coefficient  listed  in  the  table  by  the  value  applicable  
Household  owns  a  bicycle   300   to  the  household.  For  example,  in  a  household  
Household  owns  a  car   800   where  two  persons  are  employed  and,  on  average,  
three  persons  share  a  room,  we  modify  the  baseline  
Dwelling  walls  made  of  brick   100  
by:  
Dwelling  walls  made  of  tin   0    
Dwelling  walls  made  of  clay   -­‐100   2 ∗ 50 + 3 ∗ −20 =  100 − 60 = 40  
Constant   1000    
In  the  case  of  this  household,  we  increase  their  
     
estimated  consumption  by  40  takas.  
Those  results  allow  us  to  be  more  precise:  let  us  
 
assume  that  our  household:   Choosing data source
•   Is  composed  of  two  adults  and  three  children;;  
As  we  mentioned,  the  essence  of  the  proxy  means  
•   Owns  its  house;;   test  method  is  using  household  characteristics  to  
•   Owns  a  bicycle,  but  not  a  car;;   predict  household  welfare.  Therefore,  we  need  a  
•   Doesn’t  have  any  cattle;;   data  source  that  would  contain  information  about  
•   And  that  the  walls  of  the  house  are  made  of  clay.   both  household  characteristics  (household  
  composition,  dwelling  qualities,  assets)  and  
In  order  to  calculate  the  estimated  consumption   household  welfare  (normally  measured  by  
expenditure  of  the  household,  we  should:   consumption  volume),  and  that  would  allow  us  to  
draw  correspondence  between  the  two.  
Start  with  the  value  listed  in  the  “constant”  row.  
This  will  be  our  baseline.  In  our  case,  this  amounts   This  means  that  we  normally  need  results  of  a  
to  1000  takas  (Bangladeshi  currency)   household  survey.  Household  Income  and  
Expenditure  Surveys  (sometimes  called  household  
•   Find  all  rows  that  apply  to  the  household  
economic  surveys)  or  Living  Standards  
situation  and  add  or  subtract  the  corresponding  
Measurement  Study  surveys  are  examples  of  
coefficients  to  or  from  the  “baseline”  value.  
surveys  that  contain  data  that  could  be  used  in  
•   Consequently,  the  estimated  value  of  
developing  a  PMT  model.  
consumption  for  the  example  household  will  be  
Regardless  of  the  exact  type  of  the  survey  that  we  
equal  to  1000  (baseline)  +  100  (house  
want  to  use,  there  are  two  important  criteria  that  our  
ownership)  +  20  (three  or  more  children)  +  300  
data  source  has  to  meet:  
(ownership  of  bicycle)  –  300  (clay  walls)  =  1320  
takas.   •   The  data  should  be  as  recent  as  possible—as  the  
  living  conditions  of  the  population  change,  the  
relationship  between  particular  household  
Not  all  of  the  variables  will  refer  to  simple  yes/no  
characteristics  and  its  welfare  is  also  prone  to  
questions.  It  could  be  the  case  that  an  entry  in  the  
change.  Ideally,  the  survey  should  have  been  
table  will  read:  
conducted  in  the  last  five  years.  
 
•   The  survey  results  should  be  representative  for  
the  area.  If  we  want  to  use  the  PMT  scores  for  
 targeting  social  safety  nets  beneficiaries   model.  The  simplest  way  of  validating  the  model  
nationwide,  we  should  use  a  survey  that  is   involves  using  the  data  from  the  same  survey  that  
nationally  representative.   has  been  described  above.  Next  to  the  survey,  we  
  will  also  need  to  know  the  poverty  line.  In  its  
Choosing the variables absence,  we  can  select  a  percentage  and  assume  
Results  of  a  typical  LSMS  survey  will  contain   that  the  poor  comprise  the  bottom  30  or  40%  of  
hundreds  of  variables.  Among  them,  we  need  to   households  with  lowest  consumption.  
choose  the  ones  that  will  be  used  in  the  PMT    
model.  While  doing  that,  we  should  take  into   Following  that,  we  conduct  the  validation  by:  
account  two  basic  criteria:   •   using  the  survey  data,  calculate  PMT  scores  for  
  every  household  in  the  survey  dataset;;  
•   “Verifiability”  of  the  variable   •   listing  the  households  below  the  poverty  line,  
•   and  correlation  between  the  variable  and   according  to  their  consumption  (not  the  PMT  
household  welfare  (consumption  levels).   score);;  
  •   listing  the  households  below  the  poverty  line,  
Good  PMT  variables  will  be  easy  to  verify  by   according  to  their  PMT  scores;;  
enumerators  or  social  workers.  For  example,  the   •   comparing  the  two  lists.  
type  of  walls  in  the  dwelling  can  be  easily    
discerned  by  the  enumerator  visiting  the   Since  the  actual  consumption  data  is  more  reliable  
household—brick  is  visibly  different  from  wood  or   than  the  PMT  data,  if  a  household  was  included  in  
tin.  This  is  generally  the  case  with  other  variables   the  consumption-­based  list,  but  not  in  the  PMT-­
describing  housing  quality.  Although  household   based  list,  we  speak  of  exclusion  error.  The  
characteristics  (number  of  members,  their  ages,   household  was  poor,  but  the  PMT  model  has  not  
education  or  occupation)  could  be  misrepresented,   identified  it  as  such.  
this  does  not  happen  often,  in  particular  when  the    
enumerators  or  social  workers  belong  to  the  same  
On  the  other  hand,  if  a  household  was  not  identified  
local  community.  Caution  should  be  exercised  with  
as  poor  according  to  its  consumption  measured  by  
including  ownership  of  small  assets  in  the  model—
the  household  survey,  but  the  PMT  model  has  
these  could  be  removed  or  concealed.  
identified  is  as  poor,  we  speak  of  inclusion  error.  
  The  household  was  not  poor,  but  the  PMT  model  
The  relationship  between  the  selected  variables   considered  it  poor.  
can  be  determined  through  regression  analysis;;    
while  comparing  results  of  regressions  with  
Ideally,  we  would  like  to  minimize  both  inclusion  and  
different  variables  included  or  omitted,  the  authors  
2
exclusion  errors.  In  practice,  however,  there  is  often  
of  the  PMT  model  can  compare  R  and  statistical  
trade-­off  between  the  two,  with  typical  PMT  models  
significance  of  the  coefficients  for  different  models.  
featuring  inherent  30-­40%  inclusion  and  exclusion  
Once  one  model  has  been  chosen,  the  regression   errors.  
coefficients  will  become  the  coefficients  for  the  
 
PMT  model—they  will  determine  the  extent  to   Different models
which  each  variable  affects  the  household  
It  is  not  required  that  a  single  model  is  used  for  
consumption.    
entire  country.  Sometimes  it  might  improve  accuracy  
 
of  prediction  if  separate  models  for  different  regions  
Validating the PMT model
are  created.  Often  a  separate  model  for  urban  and  
After  choosing  the  model,  it  is  important  to  see  how   rural  poor  is  developed,  as  welfare  indicators  differ  
it  performs  in  identifying  the  poorest  households.   between  cities  and  the  countryside  (for  instance,  
This  is  often  referred  to  as  ‘validating’  the  PMT  
livestock  ownership,  which  often  accurately  predict   good  data  quality  and  the  fact  that  quite  details  
welfare  in  the  countryside,  is  irrelevant  for  the  city   household  characteristics  variables  were  included.  In  
dwellers).   addition  to  that,  the  survey  was  used  to  compute  
More  advanced  PMT  models  can  use  multi-­staged   official  poverty  figures  and  poverty  maps  for  the  
approach,  where  the  results  of  one  model’s   country;;  consequently,  its  use  meant  that  a  
prediction  are  ‘plugged  in’  to  another  model,  in   consistent  definition  of  poverty  could  be  applied  
order  to  minimize  inclusion  and  exclusion  errors.   across  the  board.  
   
Updating the model The  country  uses  a  single  model,  although  the  final  
As  societies  change,  PMT  models  may  become   score  is  adjusted  based  on  the  region.  That  
outdated.  Intuitively,  the  models  can  and  should  be   approach  simplifies  the  model,  which  is  important  
updated  when  more  recent  data  become  available.   when  it  comes  to  promoting  its  use  across  different  
The  household  surveys  used  to  produce  the   government  agencies.  Most  of  the  variables  selected  
models  (HEIS,  LSMS)  are  often  released  in  5-­  to   for  the  models  can  be  directly  observed  by  the  
10-­year  cycles,  which  suggests  a  similar  pattern  for   enumerator  visiting  the  household,  minimizing  the  
PMT  model  updates.   risk  of  fraud  or  misrepresentation.    
   
Bangladesh experience The  variables  used  by  the  Bangladesh  model  and  
To  finish  this  note,  it  may  be  useful  to  present  the   the  corresponding  scores  are  included  in  the  
experience  of  building  the  PMT  model  for   following  table.  Undercoverage,  for  the  cut-­off  point  
Bangladesh  to  see  how  the  principles  described   of  40%,  is  equal  to  36%.  While  the  value  could  still  
above  can  be  applied  in  practice.   be  considered  a  sizeable  exclusion  error,  it  has  been  
  proven  through  pilots  that  targeting  benefits  based  
In  terms  of  the  data  source,  the  mode  has  been   on  the  model  is  still  more  accurate  compared  to  the  
based  on  the  results  of  the  most  recent  2010   methods  previously  used  for  social  safety  net  
Household  Expenditure  and  Income  Survey.   programs  in  Bangladesh.  
The  main  rationale  behind  the  choice  has  been      

Prepared by the Social Protection & Labor team

Dhaka, Bangladesh
Photo credits: The World Bank, Karol Karpinski