1 2 Measuring Income Mobility using Pseudo-Panel Data Arturo Martinez Jr., and Mark Western, Michele Haynes, Wojtek Tomaszewski (Institute for Social Science Research, The University of Queensland) Australian Statistics Conference Official Statistics Methodology Session 9th July, 2014 3 Mom: Listen to me, I have to tell you something... Baby: I am listening… (Uh oh, looks like she found out what I did to my nappies) 4 (Mom: We’re poor, we can’t afford to go to Disneyland with your friends.) Baby: I’m sorry, what did you just say?!? 5 20% 20% 20% 20% 20% 7 > 90% Inequality is increasing. 8 Inequality and Income Mobility Q. What does increasing inequality represent? A. It depends on income mobility regime. 9 Baby: Someday, I will see Mickey Mouse too. 10 What makes the PHILIPPINES an interesting case study? The Philippines is a rapidly growing economy. . In 2012, its economy grew by approx. 6%. ECONOMIC GROWTH: approx. 5 % AVE. HHLD INCOME GROWTH: 0.4 % (US$2) POVERTY RATE: 45% to 44% INCOME INEQUALITY: 0.44 to 0.43 Household income distribution๏ stagnant? 13 How do we measure income mobility? Income mobility can be regarded as Y1t Y2t Y3t : : : Ynt Y1t+r Y2t+r Y3t+r : : : Ynt+r a vector transformation from Yit to Yit+r. 14 Panel Data Difference Income Mobility Perspectives Movement Origin independence Equalizer of income 1 ๐ ๐ ๐=1 ๐๐๐ก | ln | ๐๐๐ก−1 1 − ๐ถ๐๐๐๐๐(ln ๐๐๐ก − ln ๐๐๐ก−1 ) ๐ ๐ก=1 ๐๐๐ก ) ๐ผ๐๐๐๐ข๐๐๐๐ก๐ฆ( 1 − ๐ ๐ก=1 ๐ผ๐๐๐๐ข๐๐๐๐ก๐ฆ(๐๐๐ก ) 16 Relationship between various mobility measures 0 .2 .4 0 .5 1 -.2 0 .2 .4 0 .5 1 .6 .4 Field-Ok .2 .4 King .2 0 40 ARJ 20 0 1 Hart .5 0 .6 .4 Shorrocks .2 0 .4 .2 Fields 0 -.2 .1 .05 CDW 0 1 Poverty Persistence .5 0 .1 Poverty Inflow .2 .4 .6 0 20 40 0 .2 .4 .6 0 .05 .1 0 .05 .05 17 0 .1 Repeated Cross-Sectional Data Y1t ? : Y1t+r : : Y2t ? ? Y3t ? : : ? Ynt Y2t+r ? Y3t+r : : Ynt+r ? No one-to-one mapping of individual income. 18 Is there a way out of this problem? 19 Cross-Sectional Data Time t Time t+r Measuring Income Mobility using Pseudo-Panel Data Time t+r Measuring Income Mobility using Pseudo-Panel Data Time t Time t+r Suppose we have two time periods, t-1 and t, and we denote our income mobility measure of interest as M(Yit-1, Yit). Antman-McKenzie (AM) (2005, 2007) Bourguignon, Goh and Kim (BGK) (2004) Dang, Lanjouw, Luoto and McKenzie (DLLM) (2014) 23 AM APPROACH Step 1: For each time period t = 1, 2, group all sampled units into different cohort groups. Step 2: Compute the average income of each cohort {๐ฆ๐๐ก−1 , ๐ฆ๐๐ก }. 24 AM APPROACH Step 3: Estimate the model ๐ฆ๐๐ก = ๐ผ๐ฆ๐๐ก−1 + ๐๐๐ก . Step 4: Compute the variance of the residuals ๐(๐๐๐ก ). Step 5: Compute ๐๐(1)2 = ๐ผ๐๐(1)1 + ๐๐(1)2 where ๐๐(1)2 is a randomly drawn data point from N(0,๐(๐๐๐ก )). Step 6: Estimate the mobility measure M(๐๐(1)1 , ๐๐(1)2 ). Step 7: Repeats Steps 5 and 6 for R times. Step 8: Take the average of M(๐๐11 , ๐๐(1)2 ) across all iterations. 25 BGK APPROACH Step 1: For each time period t = 1, 2, group all sampled units into different cohort groups. Step 2: For each cohort c, estimate ๐๐๐2 2 = ๐ฝ๐ก๐ ๐๐๐ 2 2 + ๐๐๐ 2 2 and ๐๐๐3 3 = Step 3: compute ๐ ๐๐ 1 1 = ๐ฝ๐ก๐ ๐๐๐ 3 3 ๐ Retrieve the residuals ๐๐ 1 1 , their respective variances ๐๐2๐1 , ๐ ๐ ๐ฝ๐ก ๐๐ 1 1 + ๐๐๐ 3 3 ๐ ๐๐ 2 2 and ๐๐2๐2 , ๐๐2๐3 . + ๐ ๐๐ 1 1 , ๐ ๐๐ 3 3 and Step 4: For each cohort c, estimate the model ๐ 2๐๐๐ก = ๐ 2 ๐ 2 ๐ ๐ ๐๐๐ก−1 + ๐๐๐๐ก 26 BGK APPROACH Step 5: From the model in Step 4, retrieve the residuals ๐๐2๐๐ก . Step 6: Compute ๐๐๐1 2 = ๐ฝ2๐ ๐๐๐ 1 1 + ๐๐ ๐๐๐ 1 1 + ๐๐๐ก๐ where ๐๐๐ก๐ is 2 a randomly drawn data point from N(0, ๐๐๐๐ก ). Step 7: Estimate the mobility measure M(๐๐ 1 1 , ๐๐ 1 2 ). Step 8: Repeat Steps 6 and 7 for R times. Step 9: Take the average of M(๐๐11 , ๐๐(1)2 ) across all iterations. DLLM APPROACH Step 1: For each time period t, estimate ๐๐(๐ก)๐ก = ๐ฝ๐ก ๐๐(๐ก)๐ก + ๐๐(๐ก)๐ก . Retrieve the parameter estimates ๐ฝ๐ก , residuals ๐๐(๐ก)๐ก , the variance of the residuals, ๐๐2๐ก and the coefficients of determination ๐ ๐ก2 . Step 2: For each j ∈ {Est, LB, UB}, draw n2 pairs of residuals (๐๐(2)1 , ๐๐(2)2 ) from BVN(0, ๐ ) where 2 ๐๐1 ๐๐ ๐๐1 ๐๐2 ๐= 2 ๐๐ ๐๐1 ๐๐2 ๐๐2 DLLM APPROACH ๐๐๐ ๐ก ρyc1yc2 V Yi1 V Yi2 − β1′ V(Xi )β2 = σϑ1 σϑ2 ๐๐ฟ๐ต = ρyc1yc2 , ๐๐๐ต = β′1 V(Xi )β2 V Yi1 V Yi2 DLLM APPROACH Step 3: For each j ∈ {Est, LB, UB}, estimate ๐ ๐๐ 2 1 = ๐ฝ1 ๐๐ ๐ ๐๐ 2 1 . Step 4: Estimate the mobility measure ๐ Mj(๐๐ 2 1 , ๐๐ 2 2 ). Step 5: Repeats Steps 2 to 4 for R times. Step 6: For each j ∈ {Est, LB, UB}, take the average of ๐ Mj(๐๐ 2 1 , ๐๐ 2 2 ) across all iterations. 2 2 + MAIN DATA SOURCE FAMILY INCOME AND EXPENDITURE SURVEY (FIES) Three survey waves: 2003, 2006, 2009 Sub-sample of the data (2003, 2006 and 2009 waves) comprises panel data From the panel sub-sample, I drew independently drew Smaller sub-sample to create cross-sectional data EMPIRICAL APPLICATION • Use pseudo-panel estimation on the cross-sectional data to estimate income mobility. • Compare income mobility estimates derived from actual panel data and pseudo-panel data. 32 EMPIRICAL APPLICATION Table 1. Poverty Dynamics, 2003-2006 33 EMPIRICAL APPLICATION Table 2. Poverty Dynamics, 2006-2009 34 EMPIRICAL APPLICATION Table 3. Poverty Dynamics, 2003-2009 35 EMPIRICAL APPLICATION Table 4. Other Measures of Income Mobility, 2003-2006 36 EMPIRICAL APPLICATION Table 5. Other Measures of Income Mobility, 2006-2009 37 EMPIRICAL APPLICATION Table 6. Other Measures of Income Mobility, 2003-2009 38 SUMMARY Examining income mobility provides a more comprehensive analytical tool for studying income distribution. Pseudo-panel estimation provides a good alternative approach to genuine panel data-based procedures. Pseudo-panel techniques perform satisfactorily in estimating different mobility indicators that are based on the movement and equalizer of income perspectives. RESULTS Poverty outflow Poverty inflow 15 10 8 10 6 4 5 2 0 01-02 02-03 03-04 04-05 05-06 06-07 07-08 08-09 09-10 0 01-02 02-03 03-04 04-05 05-06 06-07 07-08 08-09 4009-10 RESULTS Poverty persistence Nonpoor 80 10 8 60 6 40 4 20 2 0 01-02 02-03 03-04 04-05 05-06 06-07 07-08 08-09 09-10 0 01-02 02-03 03-04 04-05 05-06 06-07 07-08 08-09 41 09-10 RESULTS ARJ King 80 25 20 60 15 40 10 20 5 0 01-02 02-03 03-04 04-05 05-06 06-07 07-08 08-09 09-10 0 01-02 02-03 03-04 04-05 05-06 06-07 07-08 08-09 09-10 42 RESULTS Fields-Ok 150 Hart 80 60 100 40 50 20 0 01-02 02-03 03-04 04-05 05-06 06-07 07-08 08-09 09-10 0 01-02 02-03 03-04 04-05 05-06 06-07 07-08 08-09 09-10 43 RESULTS CDW Fields 10 20 0 0 -10 -20 -20 -40 -30 -60 01-02 02-03 03-04 04-05 05-06 06-07 07-08 08-09 09-10 01-02 02-03 03-04 04-05 05-06 06-07 07-08 08-09 09-10 44 RESULTS Shorrocks 60 40 20 0 01-02 02-03 03-04 04-05 05-06 06-07 07-08 08-09 09-10 45 THANK YOU! email correspondence: a.martinez2@uq.edu.au Sources of Images http://www.pinoygenius.com/ http://aspanational.wordpress.com/2011/11/22/is-the-american-dream-over-the-disappearing-middle-class/ http://blog.sekiur.com http://blog.shiftspeakertraining.com/lifestyle/is-your-money-mindset-making-you-poor//tag/worm/ http://www.easyvectors.com/gallery/Notes/2 http://www.lovelyphilippines.com/tag/poverty-poverty/ http://www.backtobasicslearning.com/schoolblog/2013/06/want-to-help-set-a-world-record-join-lego-build-days-at-redclay-schools-all-are-welcome/ http://www.businessinsider.com.au/clos-ette-closet-design-wealthy-photos-2011-9 http://archbishop-cranmer.blogspot.com.au/2008/03/poverty-in-uk-blights-1m-rural-homes.html http://www.onyamagazine.com/australian-affairs/the-indigenous-australian-poverty-trap/ http://isiria.wordpress.com/2008/07/18/world-poverty-on-the-increase/ http://www.webdesigncore.com/2010/10/21/faces-of-poverty-33-arresting-photogaphy/ http://noahpinionblog.blogspot.com.au/2014/01/how-will-conservatives-save-poor.html http://nypost.com/2013/10/10/rich-versus-the-filthy-stinking-rich/ http://thepoisedlife.com/967/rich-people-problem-poise/ http://j-walkblog.com/index.php?/weblog/posts/picture_of_the_day/ http://www.backtobasicslearning.com/schoolblog/2013/06/want-to-help-set-a-world-record-join-lego-build-days-at-redclay-schools-all-are-welcome/ http://2politicaljunkies.blogspot.com.au/2009_11_01_archive.html http://jasonshofner.wordpress.com/ Sources of Images http://2politicaljunkies.blogspot.com.au/2009_11_01_archive.html http://jasonshofner.wordpress.com/ http://www.indonesia.hu/news.php?id=169&news=achieving_indonesia%E2%80%B2s_golden_moment_of_economic_gro wth&l=en http://scriptshadow.blogspot.com.au/2009/08/malcom-mccree-and-money-tree.html http://www.123rf.com/photo_10566708_several-people-out-of-work-compete-for-a-single-available-job-in-a-crowdedlabor-market-symbolizing-.html http://www.123rf.com/photo_15206271_hidden-risk-and-false-advertising-concept-with-a-beautiful-tropical-island-onthe-sea-as-a-natural-g.html http://www.accountancyage.com/aa/opinion/2180874/accountancys-taking-steps-social-mobility http://www.genome.duke.edu/genomelife/2011/03/take-pause/ http://www.prx.org/pieces/27431-what-if-counterfactuals-examine-what-might-h http://www.adelaidenow.com.au/news/gap-between-rich-and-poor-widening/story-e6frea6u-1226063787085 http://www.bbc.com/news/magazine-20255904 http://www.theepochtimes.com/n2/australia/gap-between-rich-and-poor-growing-welfare-group-5528.html http://www.macrobusiness.com.au/2013/01/australian-income-inequality-worsens/ AM Approach yc(t )t ๏ฝ ๏กyc(t ๏ญ1)t ๏ญ1 ๏ซ ๏ขxc(t )t ๏ซ f c(t )t ๏ซ ๏ฌc(t )t λ๐ ๐ก ๐ก = ๐ผ[๐๐ ๐ก ๐ก−1 − ๐๐ ๐ก−1 ๐ก−1 ] BGK Approach c i (t )t ๏ฝ๏ข X ๏ฅ ๏ฝ๏ฒ ๏ฅ Y c i (t )t c t c c i (t )t ๏ญ1 ๏ณ ๏ฝ (๏ฒ ) V (๏ฅ 2 ๏ฅct c 2 ๏ซ๏ฅ c i (t )t c i (t )t ๏ซe c i (t )t ๏ญ1 c i (t )t ) ๏ซ๏ณ 2 ect ๏ฉc ๏ฉc c c ˆ ˆi (t )t z ๏ญ ๏ข x ๏ญ ๏ฒ ๏ฅ ๏ฉ t ๏ซ 1 i ( t ) t ๏ซ 1 c c c c 2 P(Yi (t )t ๏ซ1 ๏ผ z | xi (t )t , xˆi (t )t ๏ซ1 , ๏ขˆt ๏ซ1 , ๏ณ ect ๏ซ1 ) ๏ฝ ๏( ) 2 ๏ณˆ ect ๏ซ1 DLLM Approach Yi (1)1 ๏ฝ ๏ข1 X i (1)1 ๏ซ ๏ฎ i (1)1 Yi ( 2 ) 2 ๏ฝ ๏ข 2 X i ( 2 ) 2 ๏ซ ๏ฎ i ( 2 ) 2 ๏ฉ ~ ˆ Yi ( 2)1 ๏ฝ ๏ข1 X i ( 2)1 ๏ซ vi ( 2)1 ๏ฝ ๏ขˆ1 X i ( 2) 2 ๏ซ v~i ( 2)1 2 ๐๐1 ∑ϑ = ๐๐๐1 ๐๐2 ๐๐๐1 ๐๐2 2 ๐๐2 DLLM Approach ∅( ∅( ๐ง−๐ฝ1 ๐๐(2)2 ๐๐1 ๐ง−๐ฝ1 ๐๐(2)2 ๐๐1 ∅( − ∅( − , ๐ง−๐ฝ1 ๐๐(2)2 ๐๐1 ๐๐2 ,− ๐ง−๐ฝ1 ๐๐(2)2 ๐๐1 ๐ง−๐ฝ2 ๐๐(2)2 , ,− | ๐ = 0) ≤ P(๐i(2)1 < z, Yi(2)2 < z) ≤ ∅( ๐ง−๐ฝ2 ๐๐(2)2 ๐๐2 ๐ง−๐ฝ2 ๐๐(2)2 ๐๐2 ๐ง−๐ฝ2 ๐๐(2)2 ๐๐2 ๐ง−๐ฝ1 ๐๐(2)2 | ๐ = 1) ≤ P(๐i(2)1 < z, Yi(2)2 > z) ≤ ∅( ๐๐1 , ๐ง−๐ฝ1 ๐๐(2)2 ๐๐1 | ๐ = 1) ≤ P(๐i(2)1 > z, Yi(2)2 < z) ≤ ∅( − | ๐ = 0) ≤ P(๐i(2)1 > z, Yi(2)2 > z) ≤ ∅( − ๐ง−๐ฝ2 ๐๐(2)2 ๐๐2 ๐๐1 ๐ง−๐ฝ1 ๐๐(2)2 ๐๐1 ๐ง−๐ฝ2 ๐๐(2)2 ,− ๐ง−๐ฝ1 ๐๐(2)2 | ๐ = 1) ๐๐2 , ๐ง−๐ฝ2 ๐๐(2)2 ,− ๐๐2 | ๐ = 0) | ๐ = 0) ๐ง−๐ฝ2 ๐๐(2)2 ๐๐2 |๐ =1 What makes AUSTRALIA an interesting case study for examining income mobility? THE WORLD AGREED ON 8 GOALS TO BE ACHIEVED BY 2015 54 OECD’s Better Life Index Housing Governance Income Health Jobs Life Satisfaction Community Safety Education 55 Classical Pseudo-Panel Estimation • Pioneered by Deaton (1985) • Creates synthetic panels by aggregating analytical units into cohorts which are repeatedly observed in RCS • Applications in sociology, economics, finance, biology, etc. • Useful for estimating origin independence based concept of income mobility, e.g., income elasticity in a regression framework, ๐๐๐ก = ๐ผ๐๐๐ก−1 + ๐๐๐ก ๐๐๐ก = ๐ผ๐๐๐ก−1 + ๐๐๐ก 56 Classical Pseudo-Panel Approach: General Idea Cross-sectional surveyt Age-cohort averaget Cross-sectional surveyt+r Age-cohort averaget+r 1950s 1950s 1960s 1960s 1970s 1970s 1980s 1980s 1990s 1990s Creates one-to-one mapping of cohort average income. 57 Classical Pseudo-Panel Approach Cons Pros • Not prone to bias caused by attrition in panel surveys • Useful in measuring income mobility at the macro-level even in the presence of measurement error • Facilitates analysis of trends for longer periods • Arbitrary choice of cohorts • Induces bias in the presence of timevarying cohort-level measurement error Loss of information ๏ not useful in examining income mobility at the microlevel ; only useful for originindependence perspective 58 Addressing Loss of Information {๐๐1 , ๐๐2 } {๐๐1 , ๐๐2 } • Dang, Elbers, et al. (2011) proposed estimating income regression models of the form: ๐๐1 = ๐ฝ1 ๐๐ + ๐ฃ๐1 ๐๐2 = ๐ฝ2 ๐๐ + ๐ฃ๐2 where Z’s are time-invariant explanatory variables of income and v’s are the error terms (can be assumed to follow ~ BVN(0, ∑ϑ )). WHAT WE KNOW: ๐ฝ1 , ๐ฝ2 , ๐๐ , ๐๐ , {๐ฃโ1 }, {๐ฃ๐2 } WHAT WE WANT: ๐๐1 = ๐ฝ1 ๐๐ + ๐ฃ๐1 59 Quick Facts about Australia’s Income Distribution The wealthiest quintile account for 61% of total household net worth while the poorest quintile account for 1%. Ave net worth (richest 20%): $2.2 million per hhld Ave net worth (poorest 20%): $31,205 per hhld Income inequality is slightly higher in Australia than OECD’s average (OECD 2014). The Gini coefficient in Australia is 0.33 vs. OECD’s 0.31, Undesirable? A Necessary Feature of Rapid Economic Growth? 60 DLLM APPROACH Step 1: For each time period t, estimate ๐๐(๐ก)๐ก = ๐ฝ๐ก ๐๐(๐ก)๐ก + ๐๐(๐ก)๐ก . Retrieve the parameter estimates ๐ฝ๐ก , and the residuals ๐๐(๐ก)๐ก . Step 2: Compute the mean and the variance of the residuals, ๐๐๐ก and ๐๐2๐ก .. Step 3: Step 3: Set the residual correlation ๐๐ , j ∈ {LB, UB}, such that ๐๐ฟ๐ต = 0 and ๐๐๐ต = 1. DLLM APPROACH Step 4: Sort the residuals ๐๐(2)2 from lowest to highest. Step 5: For each j ∈ {LB, UB}, draw n2 pairs of residuals (๐๐(2)1 , ๐๐(2)2 ) from BVN(0, ๐ ) where 2 ๐๐1 ๐๐ ๐๐1 ๐๐2 ๐= 2 ๐๐ ๐๐1 ๐๐2 ๐๐2 Rank the residual pairs (๐๐(2)1 , ๐๐(2)2 ) in ascending order according to the values of ๐๐(2)2 . DLLM APPROACH Step 6: Pair the first element ๐๐(2)1 of each sorted residual pair (๐๐(2)1 , ๐๐(2)2 ) with the sorted ๐ ๐๐ 2 1 . Step 7: For each j ∈ {Est, LB, UB}, estimate ๐ฝ1 ๐๐ 2 2 + ๐ ๐๐ 2 1 = ๐ ๐๐ 2 1 . Step 8: Estimate the mobility measure ๐ Mj(๐๐ 2 1 , ๐๐ 2 2 ). DLLM APPROACH Step 9: Repeats Steps 5 to 8 for R times. Step 10: For each j ∈ {LB, UB}, take the average of across all iterations. ๐ Mj(๐๐ 2 1 , ๐๐ 2 2 )