Stochastic Demand and Revealed Preference

Transcription

1 Stochastic Demand and Revealed Preference Richard Blundell Dennis Kristensen Rosa Matzkin UCL & IFS, Columbia and UCLA November 2010 Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

2 Introduction This presentation investigates the role of restrictions from economic theory in the microeconometric estimation of nonparametric models of consumer behaviour. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

3 Introduction This presentation investigates the role of restrictions from economic theory in the microeconometric estimation of nonparametric models of consumer behaviour. Objective is to uncover demand responses from consumer expenditure survey data. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

4 Introduction This presentation investigates the role of restrictions from economic theory in the microeconometric estimation of nonparametric models of consumer behaviour. Objective is to uncover demand responses from consumer expenditure survey data. Inequality restrictions from revealed preference are used to improve the performance of nonparametric estimates of demand responses. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

5 Introduction This presentation investigates the role of restrictions from economic theory in the microeconometric estimation of nonparametric models of consumer behaviour. Objective is to uncover demand responses from consumer expenditure survey data. Inequality restrictions from revealed preference are used to improve the performance of nonparametric estimates of demand responses. Particular attention is given to nonseparable unobserved heterogeneity and endogeneity. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

6 Introduction New insights are provided about the price responsiveness of demand Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

7 Introduction New insights are provided about the price responsiveness of demand especially across di erent income groups and across unobserved heterogeneity. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

8 Introduction New insights are provided about the price responsiveness of demand especially across di erent income groups and across unobserved heterogeneity. Derive welfare costs of relative price and tax changes. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

9 The Problem Assume every consumer is characterised by observed and unobserved heterogeneity (h, ε) and responds to a given budget (p,x), with a unique, positive J-vector of demands q = d(x, p, h, ε), where demand functions d(x, p, h, ε) : R K ++! R J ++ satisfy adding-up: p 0 q = x for all prices and total outlays x 2 R. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

10 The Problem Assume every consumer is characterised by observed and unobserved heterogeneity (h, ε) and responds to a given budget (p,x), with a unique, positive J-vector of demands q = d(x, p, h, ε), where demand functions d(x, p, h, ε) : R K ++! R J ++ satisfy adding-up: p 0 q = x for all prices and total outlays x 2 R. ε 2 R J 1, J 1 vector of (non-separable) unobservable heterogeneity. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

11 The Problem Assume every consumer is characterised by observed and unobserved heterogeneity (h, ε) and responds to a given budget (p,x), with a unique, positive J-vector of demands q = d(x, p, h, ε), where demand functions d(x, p, h, ε) : R K ++! R J ++ satisfy adding-up: p 0 q = x for all prices and total outlays x 2 R. ε 2 R J 1, J 1 vector of (non-separable) unobservable heterogeneity. Assume ε? (x, h) for now. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

12 The Problem Assume every consumer is characterised by observed and unobserved heterogeneity (h, ε) and responds to a given budget (p,x), with a unique, positive J-vector of demands q = d(x, p, h, ε), where demand functions d(x, p, h, ε) : R K ++! R J ++ satisfy adding-up: p 0 q = x for all prices and total outlays x 2 R. ε 2 R J 1, J 1 vector of (non-separable) unobservable heterogeneity. Assume ε? (x, h) for now. The environment is described by a continuous distribution of q, x and ε, for discrete types h. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

13 The Problem Assume every consumer is characterised by observed and unobserved heterogeneity (h, ε) and responds to a given budget (p,x), with a unique, positive J-vector of demands q = d(x, p, h, ε), where demand functions d(x, p, h, ε) : R K ++! R J ++ satisfy adding-up: p 0 q = x for all prices and total outlays x 2 R. ε 2 R J 1, J 1 vector of (non-separable) unobservable heterogeneity. Assume ε? (x, h) for now. The environment is described by a continuous distribution of q, x and ε, for discrete types h. Will typically suppress observable heterogeneity h in what follows. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

14 Non-Separable Demand For demands q = d(x, p, ε): One key drawback has been the (additive) separability of ε assumed in empirical speci cations. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

15 Non-Separable Demand For demands q = d(x, p, ε): One key drawback has been the (additive) separability of ε assumed in empirical speci cations. We will consider the non-separable case and impose conditions on preferences that ensure invertibility in ε (which corresponds to monotonicity for J = 2 which is our leading case). Assume unique inverse structural demand functions exists - Fig 1a. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

16 Non-Separable Demand For demands q = d(x, p, ε): One key drawback has been the (additive) separability of ε assumed in empirical speci cations. We will consider the non-separable case and impose conditions on preferences that ensure invertibility in ε (which corresponds to monotonicity for J = 2 which is our leading case). Assume unique inverse structural demand functions exists - Fig 1a. Here we consider the case of a small number of price regimes and use revealed preference inequalities applied to d(x, p, ε) to improve demand predictions Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

17 Non-Separable Demand For demands q = d(x, p, ε): One key drawback has been the (additive) separability of ε assumed in empirical speci cations. We will consider the non-separable case and impose conditions on preferences that ensure invertibility in ε (which corresponds to monotonicity for J = 2 which is our leading case). Assume unique inverse structural demand functions exists - Fig 1a. Here we consider the case of a small number of price regimes and use revealed preference inequalities applied to d(x, p, ε) to improve demand predictions In other related work Slutsky inequality conditions have been shown to help in smoothing demands for dense or continuously distributed prices Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

18 Related Literature Nonparametric Demand Estimation: Blundell, Browning and Crawford (2003, 2007, 2008); Blundell, Chen and Kristensen (2007); Chen and Pouzo (2009). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

19 Related Literature Nonparametric Demand Estimation: Blundell, Browning and Crawford (2003, 2007, 2008); Blundell, Chen and Kristensen (2007); Chen and Pouzo (2009). Nonseparable models: Chernozhukov, Imbens and Newey (2007); Imbens and Newey (2009); Matzkin (2003, 2007, 2008), Chen and Laio (2010). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

20 Related Literature Nonparametric Demand Estimation: Blundell, Browning and Crawford (2003, 2007, 2008); Blundell, Chen and Kristensen (2007); Chen and Pouzo (2009). Nonseparable models: Chernozhukov, Imbens and Newey (2007); Imbens and Newey (2009); Matzkin (2003, 2007, 2008), Chen and Laio (2010). Restricted Nonparametric Estimation: Blundell, Horowitz and Parey (2010); Haag, Hoderlein and Pendakur (2009), Kiefer (1982), Mammen, Marron, Turlach and Wand (2001); Mammen and Thomas-Agnan (1999); Wright (1981,1984) Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

21 Related Literature Nonparametric Demand Estimation: Blundell, Browning and Crawford (2003, 2007, 2008); Blundell, Chen and Kristensen (2007); Chen and Pouzo (2009). Nonseparable models: Chernozhukov, Imbens and Newey (2007); Imbens and Newey (2009); Matzkin (2003, 2007, 2008), Chen and Laio (2010). Restricted Nonparametric Estimation: Blundell, Horowitz and Parey (2010); Haag, Hoderlein and Pendakur (2009), Kiefer (1982), Mammen, Marron, Turlach and Wand (2001); Mammen and Thomas-Agnan (1999); Wright (1981,1984) Inequality constraints and set identi cation: Andrews (1999, 2001); Andrews and Guggenberger (2007), Andrews and Soares (2009); Bugni (2009); Chernozhukov, Hong and Tamer (2007)... Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

22 Revealed Preference and Expansion Paths Suppose we have a discrete price distribution, fp (1), p (2),...p (T )g. Observe choices of large number of consumers for a small ( nite) set of prices - e.g. limited number of markets/time periods Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

23 Revealed Preference and Expansion Paths Suppose we have a discrete price distribution, fp (1), p (2),...p (T )g. Observe choices of large number of consumers for a small ( nite) set of prices - e.g. limited number of markets/time periods Market de ned by time and/or location. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

24 Revealed Preference and Expansion Paths Suppose we have a discrete price distribution, fp (1), p (2),...p (T )g. Observe choices of large number of consumers for a small ( nite) set of prices - e.g. limited number of markets/time periods Market de ned by time and/or location. Questions to address here: Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

25 Revealed Preference and Expansion Paths Suppose we have a discrete price distribution, fp (1), p (2),...p (T )g. Observe choices of large number of consumers for a small ( nite) set of prices - e.g. limited number of markets/time periods Market de ned by time and/or location. Questions to address here: How do we devise a powerful test of RP conditions in this environment? Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

26 Revealed Preference and Expansion Paths Suppose we have a discrete price distribution, fp (1), p (2),...p (T )g. Observe choices of large number of consumers for a small ( nite) set of prices - e.g. limited number of markets/time periods Market de ned by time and/or location. Questions to address here: How do we devise a powerful test of RP conditions in this environment? How do we estimate demands for some new price point p 0? Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

27 Revealed Preference and Expansion Paths Suppose we have a discrete price distribution, fp (1), p (2),...p (T )g. Observe choices of large number of consumers for a small ( nite) set of prices - e.g. limited number of markets/time periods Market de ned by time and/or location. Questions to address here: How do we devise a powerful test of RP conditions in this environment? How do we estimate demands for some new price point p 0? In this case Revealed Preference conditions, in general, only allow set identi cation of demands. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

28 Revealed Preference and Expansion Paths How do we devise a powerful test of RP? Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

29 Revealed Preference and Expansion Paths How do we devise a powerful test of RP? Afriat s Theorem Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

30 Revealed Preference and Expansion Paths How do we devise a powerful test of RP? Afriat s Theorem Data (p t, q t ) satisfy GARP if q t Rq s implies p s q s p s q t Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

31 Revealed Preference and Expansion Paths How do we devise a powerful test of RP? Afriat s Theorem Data (p t, q t ) satisfy GARP if q t Rq s implies p s q s p s q t if q t is indirectly revealed preferred to q s then q s is not strictly preferred to q t Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

32 Revealed Preference and Expansion Paths How do we devise a powerful test of RP? Afriat s Theorem Data (p t, q t ) satisfy GARP if q t Rq s implies p s q s p s q t if q t is indirectly revealed preferred to q s then q s is not strictly preferred to q t 9 a well behaved concave utility function the data satisfy GARP Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

33 Revealed Preference and Expansion Paths Data: Observational or Experimental - Is there a best design for experimental data? Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

34 Revealed Preference and Expansion Paths Data: Observational or Experimental - Is there a best design for experimental data? Blundell, Browning and Crawford (Ecta, 2003) develop a method for choosing a sequence of total expenditures that maximise the power of tests of RP (GARP). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

35 Revealed Preference and Expansion Paths Data: Observational or Experimental - Is there a best design for experimental data? Blundell, Browning and Crawford (Ecta, 2003) develop a method for choosing a sequence of total expenditures that maximise the power of tests of RP (GARP). De ne sequential maximum power (SMP) path f x s, x t, x u,... x v, x w g = fp 0 sq t ( x t ), p 0 tq u ( x u ), p 0 v q w ( x w ), x w g Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

36 Revealed Preference and Expansion Paths Data: Observational or Experimental - Is there a best design for experimental data? Blundell, Browning and Crawford (Ecta, 2003) develop a method for choosing a sequence of total expenditures that maximise the power of tests of RP (GARP). De ne sequential maximum power (SMP) path f x s, x t, x u,... x v, x w g = fp 0 sq t ( x t ), p 0 tq u ( x u ), p 0 v q w ( x w ), x w g Proposition (BBC, 2003) Suppose that the sequence fq s (x s ), q t (x t ), q u (x u )..., q v (x v ), q w (x w )g rejects RP. Then SMP path also rejects RP. (Also de ne Revealed Worse and Revealed Best sets.) Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

37 Revealed Preference and Expansion Paths - great for experimental design but we have Observational Data continuous micro-data on incomes and expenditures Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

38 Revealed Preference and Expansion Paths - great for experimental design but we have Observational Data continuous micro-data on incomes and expenditures nite set of observed price and/or tax regimes (across time and markets) Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

39 Revealed Preference and Expansion Paths - great for experimental design but we have Observational Data continuous micro-data on incomes and expenditures nite set of observed price and/or tax regimes (across time and markets) discrete demographic di erences across households Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

40 Revealed Preference and Expansion Paths - great for experimental design but we have Observational Data continuous micro-data on incomes and expenditures nite set of observed price and/or tax regimes (across time and markets) discrete demographic di erences across households use this information alone, together with revealed preference theory to assess consumer rationality and to place tight bounds on demand responses and welfare measures. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

41 Revealed Preference and Expansion Paths So, is there a best design for observational data? Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

42 Revealed Preference and Expansion Paths So, is there a best design for observational data? Suppose we have a discrete price distribution, fp (1), p (2),...p (T )g. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

43 Revealed Preference and Expansion Paths So, is there a best design for observational data? Suppose we have a discrete price distribution, fp (1), p (2),...p (T )g. Observe choices of large number of consumers for a small ( nite) set of prices - e.g. limited number of markets/time periods. Market de ned by time and/or location. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

44 Revealed Preference and Expansion Paths So, is there a best design for observational data? Suppose we have a discrete price distribution, fp (1), p (2),...p (T )g. Observe choices of large number of consumers for a small ( nite) set of prices - e.g. limited number of markets/time periods. Market de ned by time and/or location. Given t, q t (x; ε) = d(x, p(t), ε) is the (quantile) expansion path of consumer type ε facing prices p(t). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

45 Revealed Preference and Expansion Paths So, is there a best design for observational data? Suppose we have a discrete price distribution, fp (1), p (2),...p (T )g. Observe choices of large number of consumers for a small ( nite) set of prices - e.g. limited number of markets/time periods. Market de ned by time and/or location. Given t, q t (x; ε) = d(x, p(t), ε) is the (quantile) expansion path of consumer type ε facing prices p(t). Fig 1b Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

46 Support Sets and Bounds on Demand Responses: Suppose we observe a set of demands fq 1, q 2,...q T g which record the choices made by a particular consumer (ε) when faced by the set of prices fp 1, p 2,...p T g. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

47 Support Sets and Bounds on Demand Responses: Suppose we observe a set of demands fq 1, q 2,...q T g which record the choices made by a particular consumer (ε) when faced by the set of prices fp 1, p 2,...p T g. What is the support set for a new price vector p 0 with new total outlay x 0? Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

48 Support Sets and Bounds on Demand Responses: Suppose we observe a set of demands fq 1, q 2,...q T g which record the choices made by a particular consumer (ε) when faced by the set of prices fp 1, p 2,...p T g. What is the support set for a new price vector p 0 with new total outlay x 0? Varian support set for d (p 0, x 0, ε) is given by: S V (p 0, x 0, ε) = q 0 : p0 0 q 0 = x 0, q 0 0 and fp t, q t g t=0...t satis es RP. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

49 Support Sets and Bounds on Demand Responses: Suppose we observe a set of demands fq 1, q 2,...q T g which record the choices made by a particular consumer (ε) when faced by the set of prices fp 1, p 2,...p T g. What is the support set for a new price vector p 0 with new total outlay x 0? Varian support set for d (p 0, x 0, ε) is given by: S V (p 0, x 0, ε) = q 0 : p0 0 q 0 = x 0, q 0 0 and fp t, q t g t=0...t satis es RP In general, support set will only deliver set identi cation of d(x, p 0, ε).. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

50 Support Sets and Bounds on Demand Responses: Suppose we observe a set of demands fq 1, q 2,...q T g which record the choices made by a particular consumer (ε) when faced by the set of prices fp 1, p 2,...p T g. What is the support set for a new price vector p 0 with new total outlay x 0? Varian support set for d (p 0, x 0, ε) is given by: S V (p 0, x 0, ε) = q 0 : p0 0 q 0 = x 0, q 0 0 and fp t, q t g t=0...t satis es RP In general, support set will only deliver set identi cation of d(x, p 0, ε). Figure 2(a) - generating a support set: S V (p 0, x 0, ε) for consumer of type ε. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

51 e-bounds on Demand Responses Can we improve upon S V (p 0, x 0, ε)? Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

52 e-bounds on Demand Responses Can we improve upon S V (p 0, x 0, ε)? Yes! We can do better if we know the expansion paths fp t, q t (x, ε)g t=1,..t. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

53 e-bounds on Demand Responses Can we improve upon S V (p 0, x 0, ε)? Yes! We can do better if we know the expansion paths fp t, q t (x, ε)g t=1,..t. For consumer ε: De ne intersection demands eq t (ε) = q t ( x t, ε) by p 0 0 q t ( x t, ε) = x 0 Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

54 e-bounds on Demand Responses Can we improve upon S V (p 0, x 0, ε)? Yes! We can do better if we know the expansion paths fp t, q t (x, ε)g t=1,..t. For consumer ε: De ne intersection demands eq t (ε) = q t ( x t, ε) by p0 0 q t ( x t, ε) = x 0 Blundell, Browning and Crawford (2008): The set of points that are consistent with observed expansion paths and revealed preference is given by the support set: q S (p 0, x 0, ε) = q 0 : 0 0, p0 0 q 0 = x 0 fp 0, p t ; q 0, eq t (ε)g t=1,...,t satisfy RP Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

55 e-bounds on Demand Responses Can we improve upon S V (p 0, x 0, ε)? Yes! We can do better if we know the expansion paths fp t, q t (x, ε)g t=1,..t. For consumer ε: De ne intersection demands eq t (ε) = q t ( x t, ε) by p0 0 q t ( x t, ε) = x 0 Blundell, Browning and Crawford (2008): The set of points that are consistent with observed expansion paths and revealed preference is given by the support set: q S (p 0, x 0, ε) = q 0 : 0 0, p0 0 q 0 = x 0 fp 0, p t ; q 0, eq t (ε)g t=1,...,t satisfy RP By utilizing the information in intersection demands, S (p 0, x 0, ε) yields tighter bounds on demands. These are sharp in the case of 2 goods. (BBC, 2003, for RW bounds for the many goods case). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

56 e-bounds on Demand Responses Can we improve upon S V (p 0, x 0, ε)? Yes! We can do better if we know the expansion paths fp t, q t (x, ε)g t=1,..t. For consumer ε: De ne intersection demands eq t (ε) = q t ( x t, ε) by p0 0 q t ( x t, ε) = x 0 Blundell, Browning and Crawford (2008): The set of points that are consistent with observed expansion paths and revealed preference is given by the support set: q S (p 0, x 0, ε) = q 0 : 0 0, p0 0 q 0 = x 0 fp 0, p t ; q 0, eq t (ε)g t=1,...,t satisfy RP By utilizing the information in intersection demands, S (p 0, x 0, ε) yields tighter bounds on demands. These are sharp in the case of 2 goods. (BBC, 2003, for RW bounds for the many goods case). Figure 2b, c - S (p 0, x 0, ε) the identi ed set of demand responses for p 0, x 0, ε given t = 1,..., T. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

57 Unrestricted Demand Estimation Observational setting: At time t (t = 1,...,T ), we observe a random sample of n consumers facing prices p (t). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

58 Unrestricted Demand Estimation Observational setting: At time t (t = 1,...,T ), we observe a random sample of n consumers facing prices p (t). Observed variables (ignoring other observed characteristics of consumers): p (t) = prices that all consumers face, q i (t) = (q 1,i (t), q 2,i (t)) = consumer i s demand, x i (t) = consumer i s income (total budget) Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

59 Unrestricted Demand Estimation We rst wish to recover demands for each of the observed price regimes t, q (t) = d(x(t), t, ε), t = 1,..., T, where d is the demand function in price regime p (t). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

60 Unrestricted Demand Estimation We rst wish to recover demands for each of the observed price regimes t, q (t) = d(x(t), t, ε), t = 1,..., T, where d is the demand function in price regime p (t). We will here only discuss the case of 2 goods with 1-dimensional error: ε 2 R, d(x(t), t, ε) = (d 1 (x(t), t, ε), d 2 (x(t), t, ε)). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

61 Unrestricted Demand Estimation We rst wish to recover demands for each of the observed price regimes t, q (t) = d(x(t), t, ε), t = 1,..., T, where d is the demand function in price regime p (t). We will here only discuss the case of 2 goods with 1-dimensional error: ε 2 R, d(x(t), t, ε) = (d 1 (x(t), t, ε), d 2 (x(t), t, ε)). Given t, d 1 (x(t), t, ε) is exactly the quantile expansion path (Engel curve) for good 1 at prices p (t). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

62 Unrestricted Demand Estimation Assumption A.1: The variable x (t) has bounded support, x (t) 2 X = [a, b] for < a < b < +, and is independent of ε U [0, 1]. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

63 Unrestricted Demand Estimation Assumption A.1: The variable x (t) has bounded support, x (t) 2 X = [a, b] for < a < b < +, and is independent of ε U [0, 1]. Assumption A.2: The demand function d 1 (x, t, ε) is invertible in ε and is continuously di erentiable in (x, ε). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

64 Unrestricted Demand Estimation Assumption A.1: The variable x (t) has bounded support, x (t) 2 X = [a, b] for < a < b < +, and is independent of ε U [0, 1]. Assumption A.2: The demand function d 1 (x, t, ε) is invertible in ε and is continuously di erentiable in (x, ε). Identi cation Result: d 1 (x, t, τ) is identi ed as the τth quantile of q 1 jx(t): d 1 (x, t, τ) = F 1 q 1 (t)jx (t) (τjx). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

65 Unrestricted Demand Estimation Assumption A.1: The variable x (t) has bounded support, x (t) 2 X = [a, b] for < a < b < +, and is independent of ε U [0, 1]. Assumption A.2: The demand function d 1 (x, t, ε) is invertible in ε and is continuously di erentiable in (x, ε). Identi cation Result: d 1 (x, t, τ) is identi ed as the τth quantile of q 1 jx(t): d 1 (x, t, τ) = F 1 q 1 (t)jx (t) (τjx). Thus, we can employ standard nonparametric quantile regression techniques to estimate d 1. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

66 Unrestricted Demand Estimation We propose to estimate d using sieve methods. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

67 Unrestricted Demand Estimation We propose to estimate d using sieve methods. Let ρ τ (y) = (I fy < 0g τ) y, τ 2 [0, 1], be the check function used in quantile estimation. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

68 Unrestricted Demand Estimation We propose to estimate d using sieve methods. Let ρ τ (y) = (I fy < 0g τ) y, τ 2 [0, 1], be the check function used in quantile estimation. The budget constraint de nes the path for d 2. We let D be the set of feasible demand functions, D = d 0 : d 1 2 D 1, d 2 (x, t, τ) = x p 1 (t) d 1 (x, t, ε (t)). p 2 (t) Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

69 Unrestricted Demand Estimation Let (q i (t), x i (t)), i = 1,..., n, t = 1,..., T, be i.i.d. observations from a demand system, q i (t) = (q 1i (t), q 2i (t)) 0. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

70 Unrestricted Demand Estimation Let (q i (t), x i (t)), i = 1,..., n, t = 1,..., T, be i.i.d. observations from a demand system, q i (t) = (q 1i (t), q 2i (t)) 0. We then estimate d (t,, τ) by n 1 ^d (, t, τ) = arg min d n 2D n n ρ τ (q 1i (t) d 1n (x i (t))), t = 1,..., T, i=1 where D n is a sieve space (D n! D as n! ). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

71 Unrestricted Demand Estimation Let (q i (t), x i (t)), i = 1,..., n, t = 1,..., T, be i.i.d. observations from a demand system, q i (t) = (q 1i (t), q 2i (t)) 0. We then estimate d (t,, τ) by n 1 ^d (, t, τ) = arg min d n 2D n n ρ τ (q 1i (t) d 1n (x i (t))), t = 1,..., T, i=1 where D n is a sieve space (D n! D as n! ). Let B i (t) = (B k (x i (t)) : k 2 K n ) 2 R jk nj denote basis functions spanning the sieve D n. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

72 Unrestricted Demand Estimation Let (q i (t), x i (t)), i = 1,..., n, t = 1,..., T, be i.i.d. observations from a demand system, q i (t) = (q 1i (t), q 2i (t)) 0. We then estimate d (t,, τ) by n 1 ^d (, t, τ) = arg min d n 2D n n ρ τ (q 1i (t) d 1n (x i (t))), t = 1,..., T, i=1 where D n is a sieve space (D n! D as n! ). Let B i (t) = (B k (x i (t)) : k 2 K n ) 2 R jk nj denote basis functions spanning the sieve D n. Then ˆd 1 (x, t, τ) = k2kn ˆπ k (t, τ) B k (x), where ˆπ k (t, τ) is a standard linear quantile regression estimator: n 1 ˆπ (t, τ) = arg min π2r jkn j n ρ τ q 1i (t) π 0 B i (t), t = 1,..., T. i=1 Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

73 Unrestricted Demand Estimation Adapt results in Belloni, Chen, Chernozhukov and Liao (2010) for rates and asymptotic distribution of the linear sieve estimator: jj^d (, t, τ) d (, t, τ) jj 2 = O P n m/(2m+1), p nσ 1/2 n (x, τ) ˆd 1 (x, t, τ) d 1 (x, t, τ)! d N (0, 1), where Σ n (x, τ)! is an appropriate chosen weighting matrix. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

74 RP-restricted Demand Estimation No reason why estimated expansion paths for a sequence of prices t = 1,..., T should satisfy RP. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

75 RP-restricted Demand Estimation No reason why estimated expansion paths for a sequence of prices t = 1,..., T should satisfy RP. In order to impose the RP restrictions, we simply de ne the constrained sieve as: D T C,n = DT n \ fd n (,, τ) satis es RPg. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

76 RP-restricted Demand Estimation No reason why estimated expansion paths for a sequence of prices t = 1,..., T should satisfy RP. In order to impose the RP restrictions, we simply de ne the constrained sieve as: D T C,n = DT n \ fd n (,, τ) satis es RPg. We de ne the constrained estimator by: ^d C (,, τ) = arg min d n (,,τ)2d T C,n 1 n T t=1 n i=1 ρ τ (q 1,i (t) d 1,n (t, x i (t))). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

77 RP-restricted Demand Estimation No reason why estimated expansion paths for a sequence of prices t = 1,..., T should satisfy RP. In order to impose the RP restrictions, we simply de ne the constrained sieve as: D T C,n = DT n \ fd n (,, τ) satis es RPg. We de ne the constrained estimator by: ^d C (,, τ) = arg min d n (,,τ)2d T C,n 1 n T t=1 n i=1 ρ τ (q 1,i (t) d 1,n (t, x i (t))). Since RP imposes restrictions across t, the above estimation problem can no longer be split up into T individual sub problems as the unconstrained case. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

78 RP-restricted Demand Estimation Theoretical properties of restricted estimator: In general, the RP restrictions will be binding. This means that ^d C will be on the boundary of DC T,n. So the estimator will in general have non-standard distribution (estimation when parameter is on the boundary). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

79 RP-restricted Demand Estimation Theoretical properties of restricted estimator: In general, the RP restrictions will be binding. This means that ^d C will be on the boundary of DC T,n. So the estimator will in general have non-standard distribution (estimation when parameter is on the boundary). Too hard a problem for us... Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

80 RP-restricted Demand Estimation Theoretical properties of restricted estimator: In general, the RP restrictions will be binding. This means that ^d C will be on the boundary of DC T,n. So the estimator will in general have non-standard distribution (estimation when parameter is on the boundary). Too hard a problem for us... Instead: We introduce DC T,n (ε) as the set of demand functions satisfying x (t) p (t) 0 d (x (s), s, τ) + ɛ, s < t, t = 2,..., T, for some ("small") ɛ 0. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

81 RP-restricted Demand Estimation Theoretical properties of restricted estimator: In general, the RP restrictions will be binding. This means that ^d C will be on the boundary of DC T,n. So the estimator will in general have non-standard distribution (estimation when parameter is on the boundary). Too hard a problem for us... Instead: We introduce DC T,n (ε) as the set of demand functions satisfying x (t) p (t) 0 d (x (s), s, τ) + ɛ, s < t, t = 2,..., T, for some ("small") ɛ 0. Rede ne the constrained estimator to be the optimizer over D T C,n (ε) DT C,n. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

82 Under assumptions A1-A3 and that d 0 2 DC T, then for any ɛ > 0: jj^d C ɛ (, t, τ) d 0 (, t, τ) jj = O P (k n / p n) + O P kn m, for t = 1,..., T. Moreover, the restricted estimator has the same asymptotic distribution as the unrestricted estimator. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

83 Under assumptions A1-A3 and that d 0 2 DC T, then for any ɛ > 0: jj^d C ɛ (, t, τ) d 0 (, t, τ) jj = O P (k n / p n) + O P kn m, for t = 1,..., T. Moreover, the restricted estimator has the same asymptotic distribution as the unrestricted estimator. Also derive convergence rates and valid con dence sets for the support sets. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

84 Under assumptions A1-A3 and that d 0 2 DC T, then for any ɛ > 0: jj^d C ɛ (, t, τ) d 0 (, t, τ) jj = O P (k n / p n) + O P kn m, for t = 1,..., T. Moreover, the restricted estimator has the same asymptotic distribution as the unrestricted estimator. Also derive convergence rates and valid con dence sets for the support sets. In practice, use simulation methods or the modi ed bootstrap procedures developed in Bugni (2009, 2010) and Andrews and Soares (2010); alternatively, the subsampling procedure of CHT. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

85 demand, food Demand Bounds Estimation Simulation Study: Cobb-Douglas demand function. Demand bounds,τ = true 95% conf. band price, food Figure: Performance of demand bound estimator. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

86 demand, food Demand Bounds Estimation Simulation Study: Cobb-Douglas demand function. 95% con dence bands of demand bounds. Demand bounds,τ = true 95% conf. band price, food Figure: Performance of demand bound estimator. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

87 Testing for Rationality Constrained demand and bounds estimators rely on the fundamental assumption that consumers are rational. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

88 Testing for Rationality Constrained demand and bounds estimators rely on the fundamental assumption that consumers are rational. We wish to test the null of consumer rationality. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

89 Testing for Rationality Constrained demand and bounds estimators rely on the fundamental assumption that consumers are rational. We wish to test the null of consumer rationality. Let S p0,x 0 denote the set of demand sequences that are rational given prices and income: S p0,x 0 = q 2 Bp T 0,x 0 : 9V > 0, λ 1 : V (t) V (s) λ (t) p (t) 0 (q (s) q (t)). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

90 Testing for Rationality Test statistic: Given the vector of unrestricted estimated intersection demands, bq, we compute its distance from S p0,x 0 : ρ n (bq,s p0,x 0 ) := inf kbq q2s p0,x 0 qk 2 Ŵ test n, where kkŵ test n is a weighted Euclidean norm, kbq qk 2 Ŵ test n = T (bq (t) q (t)) 0 Ŵn test (t) (bq (t) q (t)). t=1 Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

91 Testing for Rationality Test statistic: Given the vector of unrestricted estimated intersection demands, bq, we compute its distance from S p0,x 0 : ρ n (bq,s p0,x 0 ) := inf kbq q2s p0,x 0 qk 2 Ŵ test n, where kkŵ test n is a weighted Euclidean norm, kbq qk 2 Ŵ test n = T (bq (t) q (t)) 0 Ŵn test (t) (bq (t) q (t)). t=1 Distribution under null: Using Andrews (1999,2001), ρ n (bq,s p0,x 0 )! d ρ (Z,Λ p0,x 0 ) := inf kλ Z k 2, λ2λ p0,x 0 where Λ p0,x 0 is a cone that locally approximates S p0,x 0 and Z N (0, I T ). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

92 Estimating e-bounds on Local Consumer Responses For each household de ned by (x, ε), the parameter of interest is the consumer response at some new relative price p 0 and income x or at some sequence of relative prices. The later de nes the demand curve for (x, ε). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

93 Estimating e-bounds on Local Consumer Responses For each household de ned by (x, ε), the parameter of interest is the consumer response at some new relative price p 0 and income x or at some sequence of relative prices. The later de nes the demand curve for (x, ε). A typical sequence of relative prices in the UK: Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

94 Estimating e-bounds on Local Consumer Responses For each household de ned by (x, ε), the parameter of interest is the consumer response at some new relative price p 0 and income x or at some sequence of relative prices. The later de nes the demand curve for (x, ε). A typical sequence of relative prices in the UK: Figure 4: Relative prices in the UK and a typical relative price path p 0. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

95 Estimating e-bounds on Local Consumer Responses For each household de ned by (x, ε), the parameter of interest is the consumer response at some new relative price p 0 and income x or at some sequence of relative prices. The later de nes the demand curve for (x, ε). A typical sequence of relative prices in the UK: Figure 4: Relative prices in the UK and a typical relative price path p 0. Figure 5: Engel Curve Share Distribution Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

96 Estimating e-bounds on Local Consumer Responses For each household de ned by (x, ε), the parameter of interest is the consumer response at some new relative price p 0 and income x or at some sequence of relative prices. The later de nes the demand curve for (x, ε). A typical sequence of relative prices in the UK: Figure 4: Relative prices in the UK and a typical relative price path p 0. Figure 5: Engel Curve Share Distribution Figure 6: Density of Log Expenditure. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

97 Estimation In the estimation, we use log-transforms and polynomial splines log d 1,n (log x, t, τ) = q n π j (t, τ) (log x) j j=0 + r n k=1 π qn +k (t, τ) (log x ν k (t)) q n +, where q n 1 is the order of the polynomial and ν k, k = 1,..., r n, are the knots. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

98 Estimation In the estimation, we use log-transforms and polynomial splines log d 1,n (log x, t, τ) = q n π j (t, τ) (log x) j j=0 + r n k=1 π qn +k (t, τ) (log x ν k (t)) q n +, where q n 1 is the order of the polynomial and ν k, k = 1,..., r n, are the knots. In the implementation of the quantile sieve estimator with a small penalization term was added to the objective function, as in BCK (2007). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

99 Unrestricted Engel Curves τ = 0.1 τ = 0.5 τ = % CIs Figure: Unconstrained demand function estimates, t = Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

100 RP Restricted Engel Curves τ = 0.1 τ = 0.5 τ = % CIs Figure: Constrained demand function estimates, t = Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

101 Estimation conditional quantile splines - 3rd order pol. spline with 5 knots Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

102 Estimation conditional quantile splines - 3rd order pol. spline with 5 knots RP restrictions imposed at 100 x-points over the empirical support x. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

103 Estimation conditional quantile splines - 3rd order pol. spline with 5 knots RP restrictions imposed at 100 x-points over the empirical support x (T=8). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

104 Estimation conditional quantile splines - 3rd order pol. spline with 5 knots RP restrictions imposed at 100 x-points over the empirical support x (T=8). Figures 9-11: Estimated e-bounds on Demand Curve Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

105 Estimation conditional quantile splines - 3rd order pol. spline with 5 knots RP restrictions imposed at 100 x-points over the empirical support x (T=8). Figures 9-11: Estimated e-bounds on Demand Curve Demand (e-)bounds (support sets) are de ned at the quantiles of x and ε Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

106 Estimation conditional quantile splines - 3rd order pol. spline with 5 knots RP restrictions imposed at 100 x-points over the empirical support x (T=8). Figures 9-11: Estimated e-bounds on Demand Curve Demand (e-)bounds (support sets) are de ned at the quantiles of x and ε tightest bounds given information and RP. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

107 Estimation conditional quantile splines - 3rd order pol. spline with 5 knots RP restrictions imposed at 100 x-points over the empirical support x (T=8). Figures 9-11: Estimated e-bounds on Demand Curve Demand (e-)bounds (support sets) are de ned at the quantiles of x and ε tightest bounds given information and RP. varies with income and heterogeneity Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

108 Demand Bounds Estimation T = 4 T = 6 T = Figure: Demand bounds at median income, τ = 0.1. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

109 Demand Bounds Estimation T = 4 T = 6 T = Figure: Demand bounds at median income, τ = 0.5. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

110 Demand Bounds Estimation T = 4 T = 6 T = Figure: Demand bounds at median income, τ = 0.9. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

111 Demand Bounds Estimation 250 T = 4 T = 6 T = Figure: Demand bounds at 25th percentile income, τ = 0.5. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

112 Demand Bounds Estimation 250 T = 4 T = 6 T = Figure: Demand bounds at 75th percentile income, τ = 0.5. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

113 Endogenous Income To account for the endogeneity of x we can utilize IV quantile estimators developed in Chen and Pouzo (2009) and Chernozhukov, Imbens and Newey (2007). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

114 Endogenous Income To account for the endogeneity of x we can utilize IV quantile estimators developed in Chen and Pouzo (2009) and Chernozhukov, Imbens and Newey (2007). Chen and Pouzo (2009) apply to exactly this data using the same instrument as in BBC (2008). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

115 Endogenous Income To account for the endogeneity of x we can utilize IV quantile estimators developed in Chen and Pouzo (2009) and Chernozhukov, Imbens and Newey (2007). Chen and Pouzo (2009) apply to exactly this data using the same instrument as in BBC (2008). Our basic results remain valid except that the convergence rate stated there has to be replaced by that obtained in Chen and Pouzo (2009) or Chernozhukov, Imbens and Newey (2007). Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

116 Endogenous Income To account for the endogeneity of x we can utilize IV quantile estimators developed in Chen and Pouzo (2009) and Chernozhukov, Imbens and Newey (2007). Chen and Pouzo (2009) apply to exactly this data using the same instrument as in BBC (2008). Our basic results remain valid except that the convergence rate stated there has to be replaced by that obtained in Chen and Pouzo (2009) or Chernozhukov, Imbens and Newey (2007). Alternatively, the control function approach taken in Imbens and Newey (2009) can be used. Again they estimate using the exact same data and instrument. Specify ln x = π(z, v) where π is monotonic in v, z are a set of instrumental variables. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

117 Summary Objective to elicit demand responses from consumer expenditure survey data. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

118 Summary Objective to elicit demand responses from consumer expenditure survey data. Inequality restrictions from revealed preference used to produce tight bounds on demand responses. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

119 Summary Objective to elicit demand responses from consumer expenditure survey data. Inequality restrictions from revealed preference used to produce tight bounds on demand responses. Derive a powerful test of RP conditions. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

120 Summary Objective to elicit demand responses from consumer expenditure survey data. Inequality restrictions from revealed preference used to produce tight bounds on demand responses. Derive a powerful test of RP conditions. Particular attention given to nonseparable unobserved heterogeneity and endogeneity. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

121 Summary Objective to elicit demand responses from consumer expenditure survey data. Inequality restrictions from revealed preference used to produce tight bounds on demand responses. Derive a powerful test of RP conditions. Particular attention given to nonseparable unobserved heterogeneity and endogeneity. New (empirical) insights provided about the price responsiveness of demand, especially across di erent income groups. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37

122 Summary Objective to elicit demand responses from consumer expenditure survey data. Inequality restrictions from revealed preference used to produce tight bounds on demand responses. Derive a powerful test of RP conditions. Particular attention given to nonseparable unobserved heterogeneity and endogeneity. New (empirical) insights provided about the price responsiveness of demand, especially across di erent income groups. Derive welfare costs of relative price and tax changes across the distribution of demands by income and taste heterogeneity. Blundell, Kristensen and Matzkin ( ) Stochastic Demand November / 37