Has the attitude of US citizens towards 
redistribution changed over time? 
By Maria Grazia Pillau and Roberto Zelli 
Andrew E. Clark (Paris School of Economics and IZA) 
http://www.parisschoolofeconomics.com/clark-andrew/ 
IARIW, Rotterdam, August 29th 2014
This is an interesting paper for a number of different 
reasons 
1) What people say predicts how they behave, and the 
authors here consider preferences for redistribution 
(PFR). 
2) The US is sometimes considered as a poster child 
for bad behaviour. 
3) We almost always consider that one size fits all, so 
that one regression coefficient tells the whole story 
of the relationship between Y and X. This paper does 
not. 
4) And some of the results are going to turn your 
prejudices on their head.
US inequality has grown sharply...
Was this growth in inequality a result of some kind of 
structural change in the economy (the death of 
manufacturing?)... 
Or was it in some sense “demand-driven”, as Americans 
now wanted less redistribution (and thus were 
willing to accept greater inequality)? 
The authors use US GSS data from 1972 to 2010 to find 
out.
Their three questions. 
⁻ Have PFR fallen? 
⁻ Have the determinants of PFR changed? 
⁻ Are these changes period effects or cohort effects 
(e.g. those born during the Great Depression will 
always be pro-redistribution). In other words, are 
these between changes or within changes?
They create five-year birth cohorts, which appear in 
repeated cross-section waves of the GSS. 
Not all cohorts appear in all waves, although most do. 
The 1900 cohort only appears up to the 1993 wave… 
and the 1990 cohort doesn’t appear until the 2008 
wave.
Support for redistribution from question EQWLTH: 
“Some people think that the government in Washington 
ought to reduce the income differences between the 
rich and the poor, perhaps by raising the taxes of 
wealthy families or by giving income assistance to 
the poor. Others think that the government should 
not concern itself with reducing this income 
ditterence between the rich and the poor” 
PFR are on a 1 to 7 scale (from Should to Should Not, 
which they recode to binary (1-3 vs. 4-7). 
N = 23,765
What’s happened to PFR in the US? 
- Not much. 
Different symbols for different cohorts
There is a lot of variability, but the trend line is almost 
flat. 
[Would have been nice to see this figure overlaid with 
red and blue to denote Democratic and Republican 
Presidencies] 
Model PFR with the following right-hand side variables: 
household equivalent income, age, gender, marital 
status, children at home, race, years of education, 
labour-force status, past experience of 
unemployment in the last ten years, religious 
denomination, religious attendance, and political 
views
The usual way of running a regression is something like 
this: 
PFR = β’X + γ’Wave + ε 
We have all done this, for our sins. 
But this estimates only one β per right-hand side 
variable. 
And we might think that the effect of a given X variable 
may have changed over time, as in the Oaxaca- 
Blinder decomposition
We could address this coefficient heterogeneity by 
estimating separate equations for each GSS year. 
But the coefficients will move around probably due to 
sample variability (the individual N’s in each year 
are not that large). 
The authors use multi-level models (partial pooling). 
The multilevel estimates are a weighted average of 
the specific regression estimates in each year and of 
the overall regression coefficient estimated pooling 
together all the years. They are also known as 
shrinkage estimates.
Individual observations are nested within survey years 
and cohorts. 
They use a method inspired by Yang and Land to carry 
out Age-Cohort-Period (ACP) analyses. 
Model PFR as 
The coefficients differ by time (t) and cohort (k).
Decompose intercepts and slopes as:
T time periods 
K birth cohorts 
P individual-level predictors whose coefficients vary 
over time 
R individual-level predictors whose coefficients are 
unmodeled (think this means non time-varying). 
Three kinds of variation: the model is complex and can 
fall over. 
ML estimation with random effects that are integrated 
out.
The authors say they use:
My 
understanding 
of this method
Which variables are P (time-varying coefficients) and 
which R? 
Estimate separately by year and see which of the 
estimated β’s are small and don’t change 
[But isn’t this the procedure that you criticised in the 
first place?] 
R = marital status, gender, religion, religious practice, 
labour-force status, previous unemployment.
PFR is higher for women, previous unemployment, and 
lower for the married, the self-employed, and the 
religious: I guess that these results are standard. 
The main time-varying effects are... 
Age 
• The young have higher PFR, and this hasn’t 
changed over time; 
• The older have lower PFR and this has changed 
over time.
In the late 1970s there was no difference between the old 
and the middle-aged in terms of PFR; by 2010 there was 
a 10%-point difference. No cohort effect.
Income: PFR falls with income, as is standard; absolute 
effect larger over time
There is no cohort effect here either. 
The difference in PFR between the poor (-1SD) and rich 
(+2.5SD) was 17% points in the late 1970s; now 
estimated to be 27% points. 
So that there is increasing polarisation in terms of 
redistribution as a function of income. 
Is this due to the inexorable rise in income inequality? 
Or something else?
Education: Correlated with PFR (because it predicts 
income?)
There is no cohort effect here either. 
Here’s the first thing you didn’t know. 
The above time trend is positive for the high-educated, 
negative for the lower-educated. 
So the “standard” education gap in PFR is now reversed. 
The high-educated are now more PFR than the low-educated. 
This is an astonishing result. The legacy of Reagan 
Democrats?
Political Views: Self-declared L-R scale. The gap has 
opened up
Ethnicity: Non-white have higher PFR, but this reverses 
over time
Here there is, for the only time, a cohort effect. 
But it looks pretty noisy
Questions questions… 
1) “Traits” are not traits, but rather seem to change, 
often quite drastically, over time. Link to the work 
on the systematic changes in personality? See 
Boyce, C., Wood, A., and Powdthavee, N. (2013). 
"Is Personality Fixed? Personality Changes as Much 
as “Variable” Economic Factors and More Strongly 
Predicts Changes to Life Satisfaction". Social 
Indicators Research, 111, 287-305. 
2) What is missing here (I think) is an over-arching 
story which is consistent with all of these 
(somewhat surprising) changes.
3) A lot of the time effects look very similar and very 
linear. Is this mechanical? Can you show me a 
variable for which this does not hold? 
4) As the time effects look so linear, could we have put 
a time trend in on the estimated coefficient, and 
saved ourselves quite a lot of bother? 
5) Is the US different? Would other countries give you 
different results? What do you expect? 
6) Selection. Education has expanded and has become 
differently selected over time. Do cohort effects 
pick this up adequately?
7) Cohort effects largely don’t matter: can we just 
ignore them? If we had panel data, does the 
insignificance of cohort effects mean that individual 
fixed effects wouldn’t matter? 
8) The heterogeneity explored here is ex post: we 
decide which groups are going to be different (high 
vs. low education, for example). 
9) Why not let the data decide this, and estimate a 
finite mixture model, which determines any 
heterogeneity ex ante (don’t know how this deals 
with ACP though).

Session 8 a clark discussionzellipittau

  • 1.
    Has the attitudeof US citizens towards redistribution changed over time? By Maria Grazia Pillau and Roberto Zelli Andrew E. Clark (Paris School of Economics and IZA) http://www.parisschoolofeconomics.com/clark-andrew/ IARIW, Rotterdam, August 29th 2014
  • 2.
    This is aninteresting paper for a number of different reasons 1) What people say predicts how they behave, and the authors here consider preferences for redistribution (PFR). 2) The US is sometimes considered as a poster child for bad behaviour. 3) We almost always consider that one size fits all, so that one regression coefficient tells the whole story of the relationship between Y and X. This paper does not. 4) And some of the results are going to turn your prejudices on their head.
  • 3.
    US inequality hasgrown sharply...
  • 4.
    Was this growthin inequality a result of some kind of structural change in the economy (the death of manufacturing?)... Or was it in some sense “demand-driven”, as Americans now wanted less redistribution (and thus were willing to accept greater inequality)? The authors use US GSS data from 1972 to 2010 to find out.
  • 5.
    Their three questions. ⁻ Have PFR fallen? ⁻ Have the determinants of PFR changed? ⁻ Are these changes period effects or cohort effects (e.g. those born during the Great Depression will always be pro-redistribution). In other words, are these between changes or within changes?
  • 6.
    They create five-yearbirth cohorts, which appear in repeated cross-section waves of the GSS. Not all cohorts appear in all waves, although most do. The 1900 cohort only appears up to the 1993 wave… and the 1990 cohort doesn’t appear until the 2008 wave.
  • 7.
    Support for redistributionfrom question EQWLTH: “Some people think that the government in Washington ought to reduce the income differences between the rich and the poor, perhaps by raising the taxes of wealthy families or by giving income assistance to the poor. Others think that the government should not concern itself with reducing this income ditterence between the rich and the poor” PFR are on a 1 to 7 scale (from Should to Should Not, which they recode to binary (1-3 vs. 4-7). N = 23,765
  • 8.
    What’s happened toPFR in the US? - Not much. Different symbols for different cohorts
  • 9.
    There is alot of variability, but the trend line is almost flat. [Would have been nice to see this figure overlaid with red and blue to denote Democratic and Republican Presidencies] Model PFR with the following right-hand side variables: household equivalent income, age, gender, marital status, children at home, race, years of education, labour-force status, past experience of unemployment in the last ten years, religious denomination, religious attendance, and political views
  • 10.
    The usual wayof running a regression is something like this: PFR = β’X + γ’Wave + ε We have all done this, for our sins. But this estimates only one β per right-hand side variable. And we might think that the effect of a given X variable may have changed over time, as in the Oaxaca- Blinder decomposition
  • 11.
    We could addressthis coefficient heterogeneity by estimating separate equations for each GSS year. But the coefficients will move around probably due to sample variability (the individual N’s in each year are not that large). The authors use multi-level models (partial pooling). The multilevel estimates are a weighted average of the specific regression estimates in each year and of the overall regression coefficient estimated pooling together all the years. They are also known as shrinkage estimates.
  • 12.
    Individual observations arenested within survey years and cohorts. They use a method inspired by Yang and Land to carry out Age-Cohort-Period (ACP) analyses. Model PFR as The coefficients differ by time (t) and cohort (k).
  • 13.
  • 14.
    T time periods K birth cohorts P individual-level predictors whose coefficients vary over time R individual-level predictors whose coefficients are unmodeled (think this means non time-varying). Three kinds of variation: the model is complex and can fall over. ML estimation with random effects that are integrated out.
  • 15.
    The authors saythey use:
  • 16.
  • 17.
    Which variables areP (time-varying coefficients) and which R? Estimate separately by year and see which of the estimated β’s are small and don’t change [But isn’t this the procedure that you criticised in the first place?] R = marital status, gender, religion, religious practice, labour-force status, previous unemployment.
  • 18.
    PFR is higherfor women, previous unemployment, and lower for the married, the self-employed, and the religious: I guess that these results are standard. The main time-varying effects are... Age • The young have higher PFR, and this hasn’t changed over time; • The older have lower PFR and this has changed over time.
  • 19.
    In the late1970s there was no difference between the old and the middle-aged in terms of PFR; by 2010 there was a 10%-point difference. No cohort effect.
  • 20.
    Income: PFR fallswith income, as is standard; absolute effect larger over time
  • 21.
    There is nocohort effect here either. The difference in PFR between the poor (-1SD) and rich (+2.5SD) was 17% points in the late 1970s; now estimated to be 27% points. So that there is increasing polarisation in terms of redistribution as a function of income. Is this due to the inexorable rise in income inequality? Or something else?
  • 22.
    Education: Correlated withPFR (because it predicts income?)
  • 23.
    There is nocohort effect here either. Here’s the first thing you didn’t know. The above time trend is positive for the high-educated, negative for the lower-educated. So the “standard” education gap in PFR is now reversed. The high-educated are now more PFR than the low-educated. This is an astonishing result. The legacy of Reagan Democrats?
  • 24.
    Political Views: Self-declaredL-R scale. The gap has opened up
  • 25.
    Ethnicity: Non-white havehigher PFR, but this reverses over time
  • 26.
    Here there is,for the only time, a cohort effect. But it looks pretty noisy
  • 27.
    Questions questions… 1)“Traits” are not traits, but rather seem to change, often quite drastically, over time. Link to the work on the systematic changes in personality? See Boyce, C., Wood, A., and Powdthavee, N. (2013). "Is Personality Fixed? Personality Changes as Much as “Variable” Economic Factors and More Strongly Predicts Changes to Life Satisfaction". Social Indicators Research, 111, 287-305. 2) What is missing here (I think) is an over-arching story which is consistent with all of these (somewhat surprising) changes.
  • 28.
    3) A lotof the time effects look very similar and very linear. Is this mechanical? Can you show me a variable for which this does not hold? 4) As the time effects look so linear, could we have put a time trend in on the estimated coefficient, and saved ourselves quite a lot of bother? 5) Is the US different? Would other countries give you different results? What do you expect? 6) Selection. Education has expanded and has become differently selected over time. Do cohort effects pick this up adequately?
  • 29.
    7) Cohort effectslargely don’t matter: can we just ignore them? If we had panel data, does the insignificance of cohort effects mean that individual fixed effects wouldn’t matter? 8) The heterogeneity explored here is ex post: we decide which groups are going to be different (high vs. low education, for example). 9) Why not let the data decide this, and estimate a finite mixture model, which determines any heterogeneity ex ante (don’t know how this deals with ACP though).