Measuring and Valuing Mobility

Transcription

1 Measuring and Valuing Aggregate Frank London School of Economics January 2011

2 Select bibliography Aggregate, F. A. and E. Flachaire, (2011) Measuring, PEP Paper 8, STICERD, LSE D Agostino, M. and V. Dardanoni (2009). The measurement of rank. Journal of Economic 144, Demuynck, T. and D. Van de gaer (2010). Rank dependent relative measures. Ghent University, Working Paper, 10/628, Fields, G. S. (2007).. Cornell University, I.L.R School Working Paper. Gottschalk, P. and E. Spolaore (2002). On the evaluation of economic. Review of Economic Studies 69, Van Kerm, P. (2004).What lies behind income? Reranking and distributional change in Belgium, Western Germany and the USA. Economica 71,

3 Approaches to Aggregate Model of often depends on application: income or wealth wage educational in terms of social status Measurement addressed from different standpoints in relation to a specific dynamic model as an abstract distributional concept Focus on the measures in the abstract covers income or wealth covers also rank where the underlying data are categorical separates out the fundamental components of the -measurement problem

4 modelling Aggregate Basic information is the temporal pair z i = (x i,t 1, x i,t ) Bivariate distribution distribution function F (x t 1, x t ) marginal distributions F t 1 and F t give income distribution in each period Time-aggregated income derived from (x i,t 1, x i,t ) using weights w t 1, w t x i := w t 1 x i,t 1 + w t x i,t Distribution F w derived from F

5 measures in practice Aggregate Stability indices: 1 I(F w ) w t 1 I(F t 1 )+w t I(F t ) Hart (1976): 1 r(log x t 1, log x t ) where r is the correlation coefficient " R R x t e γr(f,(x t 1,x k t)) df(xt 1,x t ) King (1983): 1 µ k (F t ) k 1, k 6= 0, γ 0 where r(f; (x t 1, x t )) is a rank indicator: µ 1 (F t ) 1 jx t Q(F t ; F t 1 (x t 1 ))j Q(G; q) := inffx : G(x) qg Fields-Ok (1999): c R R jlog x t 1 log x t j df(x t 1, x t ) # 1 k

6 Fundamentals Aggregate How to characterise in terms of individual income? in terms of social position? Ingredients for a theory of measurement: 1 time frame of two or more periods; 2 measure of individual status within society 3 aggregation of changes in status over the time frame.

7 Ingredients of the problem: classes Aggregate Income as a generic term any cardinally measurable, comparable quantity cardinality turns out not to be crucial for our approach Ordered set of K income classes class k is associated with income level x k where x k < x k+1, k = 1, 2,..., K 1 p k 2 R + is the size of class k, k = 1, 2,..., K and K k=1 p k = n, the size of the population k 0 (i) and k 1 (i): income class occupied by person i in periods 0 and 1 respectively characterised by x k0 (1),..., x k0 (n) and x k1 (1),..., x k1 (n)

8 Ingredients of the problem: valuation Aggregate Don t have to use simple aggregation of the x k to compute Could carry out a relabelling of the income classes For example use n 0 (x k ) := k h=1 p h, k = 1,..., K number of persons in, or below, each income class according to the distribution in period 0 Suppose that the class sizes (p 1,..., p K ) in period 0 change to (q 1,..., q K ) in period 1 Relabelling the income classes:n 1 (x k ) := k h=1 q h, k = 1,..., K

9 Ingredients of the problem: status Aggregate u i, v i denote individual i s status in the 0-distribution, the 1-distribution personal history: z i := (u i, v i ) Distribution-independent, static (1). z i = x k0 (i), x k1 (i) Distribution-independent, static (2). z i = ϕ x k0 (i), ϕ x k1 (i) ϕ essentially arbitrary (utility of x?) independent of ϕ? Distribution-dependent, static. z i = n 0 x k0 (i), n 0 x k1 (i) cumulative numbers in class value the class Distribution-dependent, dynamic. z i = n 0 x k0 (i), n 1 x k1 (i).

10 Example Aggregate Consider the following example: period 0 period 1 x 1 A _ x 2 B A x 3 C B x 4 _ C x 5 Different definitions of status will produce different evaluations of such a change.

11 Basic axioms Aggregate Let m be individual, increasing in ju i Continuity is continuous on Z n Monotonicity. If z, z 0 2 Z n differ only in their ith component then m (u i, v i ) > m u 0 i, v0 i () z z 0. Independence. For z, z 0 2 Z n such that: z s z 0 and z i = z 0 i for some i then z (ζ, i) s z0 (ζ, i) for all ζ 2 [z i 1, z i+1 ] \ z 0 i 1, z0 i+1. Local im. Let z, z 0 2 Z n be such that, for some i and j, u i = v i, u j = v j, u 0 i = u i + δ, v 0 i = v i + δ, u 0 j = u j δ, v 0 j = v j δ and, for all h 6= i, j, u 0 h = u h, v 0 h = v h. Then z s z 0. v i j

12 Representation results (1) Aggregate Theorem. Given basic axioms: is representable by the continuous function given by n i=1 φ i (z i), 8z 2 Z n φ i : Z! R is a continuous function that is strictly decreasing in ju i v i j φ i (u, u) = a i + b i u. Corollary. is also representable by φ ( n i=1 φ i (z i)) φ : R! R is continuous and strictly monotonic increasing.

13 Representation results (2) Aggregate Status scale irrelevance. For any z, z 0 2 Z n such that z s z 0, tz s tz 0 for all t > 0.: Theorem. Given Basic axioms and scale irrelevance: is representable by φ n i=1 u ui ih i v i where H i is a real-valued function.

14 Representation results (3) Aggregate This suggests we can compare the (u, v) vectors in different parts of the distribution in terms of proportional differences m (z i ) = max ui v, v i i u i scale irrelevance. Suppose there are z 0, z0 0 2 Zn such that z 0 s z0 0. Then for all t > 0 and z, z0 such that m (z) = tm (z 0 ) and m (z 0 ) = tm (z0 0 ): z s z0. Theorem. Given our axioms is representable by Φ (z) = φ n i=1 uα i v1 α where α 6= 1 is a constant. i

15 Generating an aggregate Aggregate Consider a subset of Z : Z (ū, v) := fz 2Zj n i=1 z i = (ū, v)g. From theorem 3, that the must take the form: Φ (z) = φ α ; ū, v. n i=1 uα i v1 i Φ (z) should be zero when there is no using the standard interpretation of φ ( n i=1 u i; ū, ū) = 0, i.e. φ (ū; ū, ū) = 0.

16 Using a broader interpretation of zero Aggregate Scaling up everyone s status should not matter v i = λu i, i = 1,.., n (where λ = v/ū) φ λ 1 α n i=1 u i; ū, v = 0 φ ū α v 1 α ; ū, v = 0. This requires φ and φ are equivalent to: h i ψ n ui α h i vi 1 α i=1 µ u µ. v A suitable cardinalisation of ψ(.) gives M. h i M α := 1 ui α h i vi 1 α α[α 1]n n i=1 µ u µ 1. v

17 Limiting cases Aggregate Two limiting cases α = 0: α = 1 M 0 = 1 n n i=1 M 1 = 1 n n i=1. v i µ log ui vi v µ u µ, v. u i µ log ui vi u µ u µ. v We have a class of aggregate measures high α > 0: M sensitive to downward movements α < 0: M sensitive to upward movements

18 1 Aggregate Concerned with ranks not income levels? Then make status an ordinal concept (Chakravarty 1984) Variety of ways to define status ordinally: tables or transition matrices. However, these approaches are sensitive to the adjustment of class boundaries: Consider the case where in the original set of classes p k = 0 and p k+1 > 0; if the is sensitive to small values of p and the income boundary between classes k and k + 1 is adjusted there could be a big jump in the. This will not happen if the is defined in terms of u i and v i

19 2 Aggregate The derivation is value free. Can we introduce a social valuation of? Could construct an explicit welfare approach something analagous to Atkinson inequality? (Gottschalk-Spolaore 2002) but you must go beyond simplistic welfare models (Markandya 1982, 1983) Can also introduce normative elements in the above framework definition of status value range of α

20 Aggregate Simplest case: status before and after i.e. distribution-independent, static Movements of incomes: u i = x 0i and v i = x 1i. Define moment µ g(u,v) = n 1 n i=1 g(u i, v i ) g(.) is a specific function consider three cases: M α, M 0 and M 1

21 General case Aggregate Rewrite the as M α = 1 α(1 α) In terms of moments: M α = 1 α(α 1) n 1 ui αv1 α i µ u α µ1 v α h µu α v 1 α µ u αµ1 v α i Central Limit Theorem implies asymptotic normality under standard regularity conditions. h dvar(m α ) = D ˆΣD > with D = Mα M µ ; α u µ ; v D in terms of sample moments: D = h µu α v 1 α µ u α 1 µ v α 1 ; µ u α v 1 α µ α (α 1) α u µα v 2 ; µ u αµα v 1 α(α 1) i M α µ u α v 1 α i.

22 Covariance matrix Aggregate ˆΣ is the estimator of the covariance matrix of µ u, µ v and µ u α v 1 α We have:

23 Limiting case (1) Aggregate M 0 as a function of four moments: M 0 = µ v log v µ v log u µu µ + log v µ. v dvar(m 0 ) = D 0 ˆΣ 0 D 0 > h M0 M D 0 = µ ; 0 u µ ; v ˆΣ 0 : M 0 µ v log v ; i M 0 µ v log u

24 Limiting case (2) Aggregate M 1 as a function of four moments: M 1 = µ u log u µ u log v µv µ + log u µ u dvar(m 1 ) = D 1 ˆΣ 1 D > 1 with D 1 = D 1 = ˆΣ 1 : h M1 M µ ; 1 u µ ; v h µu log u +µ u log v µ 2 u M 1 µ u log u ; µ u i M 1 µ u log v i ; µ ; v µ ; u µ u

25 Aggregate we now consider the distribution-dependent, dynamic status. Scale independence means we can define status using proportions u i = ˆF 0 (x 0i ) and v i = ˆF 1 (x 1i ) ˆF 0 (.) and ˆF 1 (.) are the empirical distribution functions ˆF k (x) = 1 n n j=1 I(x kj x) Method of moments does not apply as the values in u and v are non i.i.d.

26 Establishing the asymptotic distribution Aggregate Ruymgaart and van Zuijlen (1978): asymptotic normality for the multivariate rank statistic T n = 1 n n i=1 c inφ 1 (u i )φ 2 (v i ). c in are given real constants, φ 1 and φ 2 are (scores) functions defined on (0,1), these are allowed to tend to infinity near 0 and 1 but not too quickly. need to assume the existence of K 1, a 1 and a 2, s.t. φ 1 (t) K 1 [t(1 t)] a 1 and φ 2 (t) K 1 [t(1 t)] a 2 with a 1 + a 2 < 1 2 for t 2 (0, 1). Then, φ 1 (t) and φ 2 (t) tend to infinity near 0 at a rate slower than the functions t a 1 and t a 2. Variance of T n is finite, even if not analytically tractable.

27 Applying the results - general case Aggregate M α can be written as a function of T n Note that µ u = µ v = n 1 n i=1 n i = n+1 2n. M α as a function of one moment: M α = 1 2n α(α 1) n+1 µ u α v 1. 1 α Hence, M α = 1 α(α 1) [T n 1].

28 Applying the results - limiting cases Aggregate M 0 as a function of T n M 0 = 2n n+1 (k µ v log u ) = l T n. M 1 as a function of T n M 1 = 2n n+1 (k µ u log v ) = l T n It can be shown that the relevant conditions are met for 0.5 < α < 1.5. M α is asymptotically normal asymptotic justification for the bootstrap

29 Aggregate coverage error rate of a confidence interval probability that CI does not include the true value of a parameter should be close to the nominal rate e.g. 5% for a 95% CI use Monte-Carlo simulations to approximate coverage error rates for different methods: asymptotic percentile bootstrap and studentized bootstrap

30 Asymptotic confidence intervals Aggregate CI asym = [M α c dvar(m α ) 1/2 ; M α + c dvar(m α ) 1/2 ] c is a critical value from the Student distribution T(n 1). finite sample often poor bootstrap confidence intervals can be expected to perform better

31 Percentile bootstrap method Aggregate does not require the (asymptotic) standard error Method: generate B bootstrap samples by resampling in the original data for each resample, we compute the. obtain B bootstrap statistics, M b α, b = 1,..., B. The percentile bootstrap confidence interval is equal to CI perc = [c b ; cb ] c b and cb are the 2.5 and 97.5 percentiles of the EDF of the bootstrap statistics.

32 Studentized bootstrap method Aggregate uses the asymptotic standard error Method: generate B bootstrap samples by resampling in the original data for each resample, compute a t-statistic. obtain B bootstrap t-statistics t b α = (M b α M α )/dvar(m b α) 1/2, b = 1,..., B, where M α is the original CI stud = [M α c d Var(M α ) 1/2 ; M α c d Var(M α ) 1/2 ] c0.025 and c are percentiles of the EDF of the bootstrap t-statistics.

33 Comparison Aggregate studentized bootstrap based on asymptotically pivotal statistic t-statistics follow (asymptotically) a known distribution superior of the bootstrap over asymptotic confidence intervals both bootstrap intervals asymmetric, asymptotic confidence interval is symmetric bootstrap CIs more accurate if the underlying distribution is asymmetric.

34 Experiments Aggregate Bivariate Lognormal distribution: (x 0, x 1 ) LN(µ, Σ) 1 ρ with µ = (0, 0) and Σ = ρ 1 increases as ρ decreases we try different indices, different sample sizes and different levels for each combination: draw 10,000 samples (from the bivariate lognormal distribution) compute M α compute confidence interval at 95% How often does the interval does not include the true parameter?

35 Asymptotic confidence intervals Aggregate α n = 100, ρ = n = 100, ρ = n = 100, ρ = n = 100, ρ = n = 100, ρ = n = 100, ρ = n = 100, ρ = n = 200, ρ = n = 500, ρ = n = 1000, ρ = n = 5000, ρ = n = 10000, ρ = Table: Coverage error rate of asymptotic confidence intervals at 95% of income measures. The nominal error rate is 0.05, replications

36 Results Aggregate Distribution-independent, static. Recall: coverage error rate should be close to 5% asymptotic intervals perform poorly for α = 1, 2 coverage error rate is stable as ρ varies (for α = 0, 0.5, 1 and n = 100) coverage error rate decreases as n increases coverage error rate close to 0.05 for n and α = 0, 0.5, 1. asymptotic confidence intervals perform well in very large sample, with α 2 [0, 1]. dismal of asymptotic confidence intervals for small and moderate samples poorest results for ρ = 0.8 try other two methods for this value of ρ

37 Other methods Aggregate α n = 100, ρ = 0.8 Asymptotic Boot-perc Boot-stud n = 200, ρ = 0.8 Asymptotic Boot-perc Boot-stud n = 500, ρ = 0.8 Asymptotic Boot-perc Boot-stud n = 1.000, ρ = 0.8 Asymptotic Boot-perc Boot-stud Table: Coverage error rate of asymptotic and bootstrap confidence intervals at 95% of income measures replications, 199 bootstraps

38 Results Aggregate percentile bootstrap and asymptotic confidence intervals perform similarly studentized bootstrap confidence intervals outperform other methods siginificant improvement over asymptotic confidence intervals

39 Aggregate distribution-dependent, dynamic status variance of M α is not analytically tractable cannot use asymptotic and studentized bootstrap confidence intervals use the percentile bootstrap method

40 Percentile bootstrap method (n=100) Aggregate α ρ = ρ = ρ = ρ = ρ = ρ = ρ = Table: Coverage error rate of percentile bootstrap confidence intervals at 95% of rank- measures replications, 199 bootstraps and 100 observations.

41 Results Aggregate the coverage error rate can be very different for different values of ρ and α, it decreases as ρ increases, except for the case of nearly zero (ρ = 0.99). the coverage error rate is close to 0.05 for ρ = 0.8, 0.9 and α = 0, 0.5, 1. What happens as the sample size increases?

42 Percentile bootstrap method (variable n) Aggregate α n = 100, ρ = n = n = n = n = 100, ρ = n = n = n = n = 100, ρ = n = n = n = Table: Coverage error rate of percentile bootstrap confidence intervals at 95% of rank- measures replications, 199 bootstraps.

43 Results Aggregate the coverage error rate gets closer to 0.05 as the sample size increases, the coverage error rate is smaller when α = 0, 0.5, 1. better statistical properties as the sample size increases

44 Asking about Aggregate A questionnaire study use same type of methods as for inequality? focus on whether people value contrast with preferences for equality? Study using 356 students from three countries: Italy (120) UK (89) Israel (147) problem obviously more complex is a from-to concept not a snapshot graphics representation is tricky Method: bus queue pictures combine equality and intergenerational get personal characteristics

45 Q1 Full Mixing v Rigidity Aggregate

46 Q2 Full Mixing and Widening Aggregate

47 Q3 Rigidity v Full Mixing+Widening Aggregate

48 Q4 Partial mixing v Rigidity Aggregate

49 Q5 Partial Mixing and Widening Aggregate

50 Q6 Rigidity v Partial Mixing+Widening Aggregate

51 Q7 Full v Partial Mixing Aggregate

52 Q8 Rigidity v Simple Widening Aggregate

53 Asking about Aggregate Yes if A chosen more often than B in Q1 (Full mixing v rigidity) Q4 (Partial mixing v rigidity) Q7 (Full v partial mixing) A responses B responses Q1 Q4 Q7 Q1 Q4 Q7 Italy UK Israel All

54 Asking about equality Aggregate Yes if A chosen more often than B in Q2 (Full mixing and widening) Q5 (Partial mixing and widening) Q8 (Rigidity v Simple widening) A responses B responses Q2 Q5 Q8 Q2 Q5 Q8 Italy UK Israel All

55 Categorical variables Aggregate Check for each person the answers to Q1,Q4,Q7 () Q2,Q5,Q8 (equality) Percentages in each category 0A 1A 2A 3A Italy UK Israel TOTAL Equality Italy UK Israel TOTAL

56 Regression model Aggregate Seek to explain 1 attitudes to 2 attitudes to equality Use categorical variables 1 preferences 0A, 1A, 2A, 3A 2 equality preferences 0A, 1A, 2A, 3A Variety of personal characteristics Standard ordered probit

57 Personal characteristics 1 Aggregate

58 Personal characteristics 2 Aggregate

59 Personal characteristics 3 Aggregate

60 Baseline model Aggregate equality age gender familyincome livingstandards prospects perspectivn independendencea independendenceb governmentsrolee

61 Model with country dummies Aggregate equality age gender familyincome livingstandards prospects perspectivn independendencea independendenceb governmentsrole italy uk

62 Model with nationality dummy Aggregate equality age gender nationality familyincome livingstandards prospects perspectivn independendencea independendenceb governmentsrole

63 : Measurement Aggregate Key step involves a logical separation of fundamental concepts measure of individual status aggregation of changes in status Status concept derived directly from the information in the marginals Apply standard principles to movements in status get a superclass of measures generally applicable to wide variety of status concepts parameter α that determines type of measure Principal status types yield statistically tractable indices

64 : Values Aggregate key factors: The more independent are children s and parents economic positions in a society... I am from around here Italy country dummy (Italians don t value....!) Equality key factors: family income role of government prospective social position