INTRODUCING BEHAVIOURAL RESPONSES INTO MICROSIMULATION MODELS: SIMULATION OF VAT REFORMS

François GARDES 1 Bertrand LHOMMEAU2 Christophe STARZEC 3

___________________________________ (1).Université Paris 1, e-mail : [email protected] (2) INSEE and TEAM (DEA) when this paper was written; now : DARES, e-mail: [email protected] (3) CNRS-TEAM, INSEE, e-mail: [email protected]; Mail adresse :106-112 Bd. de l’Hôpital, 75647 Paris Cedex 13, France

Résumé

Dans cet article l’impact d’une réforme de la TVA sur la consommation et le revenu disponible des ménages est analysé en prenant en compte les changements possibles dans les comportements de consommation. Les réactions des consommateurs sont intégrées dans un modèle de microsimulation du système redistributif français INES développé à l’INSEE. L’approche adoptée est une modification de la procédure proposée par Baccouche, Laisney (1982, 1990) qui utilise le modèle LES (Linear Expenditure System) comme cadre théorique. Le modèle est estimé et appliqué à une analyse des conséquences redistributives des réformes hypothétiques de la TVA. Les résultats obtenus sont comparés à ceux obtenus dans un cadre statique où les effets des comportements individuels sont ignorés. Mots clés , Microsimulation , Redistribution, Fiscalité, TVA

Abstract In this article we analyse the impact of the VAT reform on consumption and disposable income on the individual level taking into account the possible change in consumer’s behaviour. The behavioural responses are integrated into the classic static microsimulation model developed in INSEE (see appendix 1). Our approach is a modified procedure proposed by Baccouche, Laisney (1982, 1990) based on LES model (Linear Expenditure System).

We apply this model to evaluate the

consequences, in terms of consumption change and income distribution, of the change in VAT. Apart from classic consumption-distribution effects we analyse the sensitivity of the results using the behavioural approach when compared to the static one. Key words: Microsimulation, Redistribution, Taxation, VAT JEL Classification: D12 D31 H24 H31

Introduction Taking into account behavioural responses in a microsimulation model supposes a choice of an appropriate econometric model. In the case of VAT reform analysis a demand system should be chosen. For theoretical reasons the AIDS type models (QAIDS for example) appear as the most satisfactory matching most of the desired theoretical properties . Our choice to start our study with LES model is a very pragmatic solution which copes with the available data constraints but also takes advantage of earlier experience and allows comparisons of our results with those obtained previously. We consider it as a good starting point and hope to progress, if possible, towards more satisfactory solutions in the future. 1. The Linear Expenditure System (L.E.S.) L.E.S, proposed by Richard STONE in 1954 was the first demand system satisfying four principle axioms of the consumer’s behaviour theory (additivity, homogeneity, symmetry, negativity) in the static context. 1.1 Utility function Starting point is a direct, additive utility function: n

(L1)

u(q ) = ∑ β i . Log (qi − γ i ) where γ i and β i are the parameters to be estimated. i =1

Indirect utility can be written in the following way n

(L2)

n

v ( y , p) = ( y − ∑ pi .γ i ) / ( π p j j ) i =1

β

i =1

γ i are interpreted as the subsistence, incompressible minimum consumption. Once the subsistence quantities are bought the consumer can allocate y −

n

∑ p .γ i

i

(supernumerary expenditure or income)

i =1

on different consumption according to the classic choice theory. (L2) can be interpreted as an indicator of household’s welfare. 1.2 The demand system Using matrix notation the optimisation program can be written: according to (L3)

L(q ) = u(q ) + λ .( y − p'. q ) = β '. Log(q − γ ) + λ .( y − p'. q ) or in value terms,

p$ . q = p$ .γ + ( y − p'.γ ). p −1 . β

(L4)

where p$ is a diagonal matrix of price vector p. Then for each good i it can be written: n

pi qi = γ i pi + β i ( y − ∑ p k γ k ) with k =1

n

∑β

i

= 1 (additivity condition),

i =1

Similarly, in terms of quantities : (L5) q = p$ −1 .( β . y + A. p) with A = − ( β .1'− I ).γ$ 1.3 The associated cost function The cost function associated to the utility function takes the following form: (1) 6 47 4 8 n

(2) 6 47 4 8 n

c(v , p) = ∑ pi .γ i + v. π p j j ,

(L6)

i =1

β

j =1

The cost function can be decomposed into two components: (1) constant term, minimal cost or subsistence expenditure, independent from the level of utility v. No substitution effects are possible; n

(2) variable term, which means that the consumer can «buy» utility which price is: π p j j . b

j =1

n

As

∑β

i

= 1 , this price can be interpreted as a marginal cost of living index.

i =1

1.4 Properties of the L.E.S. 1.Additive L.E.S utility function gives a very simple relation between non-compensated price elasticities and income elasticities (Deaton et Muellbauer 1980). Thus, it is very easy (from the point of view of degrees of freedom) to compute all of cross-price elasticities. A L.E.S. of n equations requires the estimation of only 2n parameters (from which 2n-1 are independent). 2.The main drawback of the additive utility function is the variability of budget shares when the number of goods varies. This implies a high level of aggregation. 3. L.E.S does not allow complementary goods and inferior goods. 1.5 The model An other interesting quality of LES is the possibility to introduce easily socio-economic variables which is particularly interesting if one wants to estimate it on cross-section data.

The proposed versions of LES estimation takes largely advantage of these interesting properties and limits the consequences of shortcomings by grouping goods in large aggregates. Starting from the classic, most often used presentation of LES, an equation for a household h, (L4) can be written: (L11)

pi qih = pi γ ih + β i ( y h − pγ h ) i=1,...n , n, number of goods,

with γ ih , subsistence expenditure and β i , marginal budget shares. The same in terms of budget shares can be written: (L12)

wih = β i +

1 ( p q h )1 ( pi γ i − β i pγ ) with wih = i hi y y

Introduction of socio-demographic variables in this model is made using Trognon’s(1981) specification1. In this specification, only the subsistence expenditure is influenced by the sociodemographic variables while the marginal budget shares don’t depend on them. Suppose that all households face the same prices for which in the case of cross -section data, they can be set as 1 : pi = 1 , i=1,...,n. Then, the subsistence expenditure γ depending on the socio-demographic variables Z can be written: K

γ ih = ∑ γ ik Z Kh

for i=1,...,n

k =0

where z h , vector of K socio-demographic variables with

Z h ' ≡ (1, z h ' ) and

1

Other specifications are possible. Following Baccouche and Laisney (1986), this sepecification makes subsistence expenditure independent from socio-demographic characteristics. Only marginal budget shares can vary with them. Equation (L12) can be written: K

wih = ( β i0 + ∑ β ik Z k ) + k =1

with r0 = pγ , total subsistence expenditure, then transforming II-1 we obtain:

wih = β i0 +

K 1 0 ( p γ − ( β β ik . Z k ) pγ ) ∑ i i i h y k =1

K K d i0 − β i0 r0 k k Zk Z r + β − ∑ i k 0∑βi h y yh k =1 k =1

Adding the error term ui , N( σ i ,0), we obtain the stochastic model: h

2

wih = β i0 +

K K d i0 − β i0 r0 k k Zk + β Z − r + uih ∑ 0∑βi i k h h y y k =1 k =1

Another specification would naturally combine the two previous ones: varying both marginal budget shares and subsistence expenditure with respect to the socio-demographic variables. This approach would allow the introduction of cross effects between socio-demographic variables. However this specification will not be tested here for mainly two reasons: - this specification leads to the estimation of a system with a large number of parameters which requires a large number of observations - the potential gain from cross effects is in practice extremely reduced. In their estimation Baccouche and Laisney (1986) obtained for only one variable - number of persons in the household - significant cross effect on both subsistence expenditure and marginal budget shares.

n

rk ≡ ∑ d ik

d ik ≡ pi γ ik

i = 1,..., n

k = 0,..., K

i =1

where d ik et rk are respectively subsistence expenditure for a good i and the total subsistence expenditure corresponding to the one unit of k socio-demographic characteristic with respect to the reference household. Replacing expenditures by budget shares w an introducing error term uih , N( σ i2 ,0), we obtain

π ik Z kh + uih y k =0 h K

wih = β i + ∑

i=1,...,n

with π ik = d ik − β i rk i=1,..,n k=0,..., K For a given k only n-1 equations are independent because of additivity constraint thus K+1 constraints are necessary. Following Baccouche and Laisney (1990), the adopted solution is to set arbitrary the

rk . Under these hypothesis the model can be estimated equation by equation and will automatically satisfy the additivity constraint.

2. Expenditure aggregation As we mentioned in the part describing the LES, the model constraints (additivity of the utility function, no complementary nor primary goods are allowed) require relatively high level of aggregation. Thus an appropriate grouping of goods, giving a reduced number of items, is necessary. The other advantage of aggregation is to reduce measure errors. This is particularly important when the individual, survey observation should be comparable with macro, national accounts data. Table 1 - Composition of aggregated goods i

item

codes 11*** 12*** 13*** 14*** 83311-83317 83511-83512 83411-83419 21*** 22*** 23*** 86211 86232 86221 86222 82111 82113 82115 82211-82213 31*** 91111-91311 93***

1

A1

food at home

2

A2

alcohol and tobacco

3

A3

food away

4

B1

clothing

5

C1

housing

comments included non alcoholic drinks

clothes, shoes, clothes and shoes repairing jewellery, watches and their repairing leather goods rent (but not mortgage) incl.: maintenance charges,

maintenance works

home

insurance,

major

incl. imputed econometrically rent for owners 6 7

C2 D1

heating and lighting equipment and services

8

E1

health & hygiene

9

F1

cars and motorcycles

10

F2

other means of transport

11

F3

telecommunications

12

G1

leisure and culture

13

H1

other goods and services

home

32*** 41111-44218 45115-45215 46211-46214 46111-46115 86223 86111 82112 82114 45111-45114 51*** 52*** 53*** 54*** 91514 91515 61111-61121 61222-62313 91411 91412 63111-63311 65111 64111-64212 92612 71111-71245 71321-71337 61211-61221 91413 91414 96311 71311-71319 71341-74414 86212-86213 86224-86231 91611 91612 91615 91616 92113 92213 92313 82311 82312 83111-83218 91619 other items 86*** and 91***

heating and lighting furniture, household electrical appliances, tableware goods and services for current maintenance domestic services like home child care, external child care, different location costs.

soap, cosmetics medicine drugs and similar medical services: doctors nurses... hospital care and similar private health insurance primes cars and motorcycles:: purchase new and second hand, rent, costs of use and maintenance , insurance primes taxi, public transport ... all goods and services linked with telecommunications audio-video appliances leisure, shows, culture caravan, yacht,... cycles camping, sport books, records,... leisure linked services (repairing, insurance), stationery holidays with insurance

other goods, insurance prime

cf. gs90.l08.bl.scd(bdf1) & gs90.l08.bl.scd3(conso1)

Total expenditure (y) is the sum of 13 expenditure items. The aggregation of expenditures did not eliminate totally the problem of measure errors due to the individual observations. Three items have a very high rate of zero expenditure (25-40%): food away (A3), other transport (F2) and other goods and services (H1). For the two others: alcohol and tobacco (A2) cars and motorcycles (F1) this rate relatively high too (15-16%) (Table 2). Table 2 - Expenditure distribution (9634 households) (not weighted) zero Expenditures in Francs expenditure frequency

Expenditures per CU (Oxford) Maximum

st dev

0.3

27 571

Mean Minimum Maximum 0

188 847

16 284

13 991

0

100 935

7 399

alcohol & tobacco (A2)

16.3

4 567

0

94 910

6 278

2 391

0

93 047

3 562

food away (A3)

30.6

7 825

0

202 807

12 327

3 937

0

118 622

6 358

clothing (B1)

7.3

13 547

0

344 803

16 725

6 901

0

202 825

8 756

housing (C1)

0.0

39 405

700 1 113 504

33 654

22 069

318

655 002

19 894

food at home (A1)

st dev

Mean Minimum

heating and lighting (C2)

1.8

7 208

0

108 865

5 009

3 904

0

52 513

2 944

equipment and services (D1) health, hygiene (E1)

6.6

12 679

0

616 060

21 327

6 473

0

280 027

10 762

cars, motorcycles (F1)

home

4.7

12 824

0

554 042

22 018

6 692

0

354 331

12 245

14.8

24 874

0

430 261

35 668

12 400

0

253 095

18 591

zero Expenditures in Francs expenditure frequency

Expenditures per CU (Oxford)

40.5

3 161

0

287 316

7 375

1 654

0

169 009

telecommunications (F3)

2.6

3 296

0

53 052

2 993

1 851

0

53 052

1 953

leisure, culture (G1)

2.3

18 026

0

646 813

23 950

9 280

0

380 478

12 587

24.7

2 923

0

158 043

7 457

1 605

0

120 000

4 667

other transport (F2)

other goods and services (H1)

4 010

cf gs90.l08.bl.scd2(mills2)

Thus, expenditure aggregation will be completed by grouping households in cells containing units having similar selection criteria. This grouping should reduce considerably the effects of measure errors at the individual level especially those zero reports which are linked to the short survey duration.

3. Household grouping. Household grouping is made in three steps: - first four socio-demographic characteristics criteria are chosen: age, social category , dwelling’s ownership, family type. Every characteristic is divided in several classes (see table 3) Table 3 -Household grouping criteria: socio-demographic characteristics and their classes. socio-demographic characteristics

variable’s name

classes

code

head’s age

agepr

- less than 30 - 30 - 44 - 45 - 59 - 60 and more

1 2 3 4

head’s social category

cspr

-farmer - self employed - professionals, businessmen, executives

A I C

- middle managers, office workers

P

- workers

O

dwelling’s ownership status

stalog

- owner or free accommodation - mortgage - tenant, lodger

P A L

family type

typmen1

- lone persons or no family link - couple without children - couple with 1 or 2 children - couple with 3 children or more - lone parent families

1 2 3 4 5

in gray characteristics of the reference household. cf. gs90.l08.bl.scd (bdf3)

- in the second step households are classified with respect to the level of per CU total income. Household are grouped in order to form cells of circa 10 observations. This second operation gives 1053 cells of size between 1 and 14. After this first grouping, the zero expenditure frequency is considerably reduced. Only for item « other transport » it is above 2% for all others it is practically 0. However, there are still 53 cells of very small size: less or equal to three. - in the third step small size cells are aggregated using arbitrary proximity criteria. We obtain 1005 cells of size larger than 3. Tableau 4 - Expenditure distribution for 1005 cells (weighted by cell’s size) item

0 in %

food at home(A1) alcohol tobacco(A2) food away (A3)

&

Expenditures in Francs

budget shares ( y)

Mean

Mean

Minimum

Maximum st. dev

Minimum

Maximum st. dev

0.0

27571

5276

73300

10547

0.16

0.04

0.31

0.0477

0.0

4567

43

19931

2501

0.03

0.00

0.10

0.0136

0.7

7825

0

54983

6483

0.04

0.00

0.16

0.0247

clothing (B1)

0.0

13547

389

73584

8998

0.07

0.01

0.23

0.0258

housing(C1)

0.0

39405

9061

224904

17117

0.23

0.11

0.50

0.0612

heating and lighting (C2) . equipment and home services (D1) health, hygiene(E1)

0.0

7208

0

24841

2745

0.04

0.00

0.12

0.0163

0.0

12679.19

270

138011

10328

0.07

0.00

0.26

0.0318

0.3

12824

433

90207

8508

0.07

0.01

0.39

0.0379

cars, motorcycles(F1)

2.1

24874

0

99919

17038

0.13

0.00

0.36

0.0611

other transport (F2)

0.0

3161

0

39550

3294

0.02

0.00

0.14

0.0154

(telecommunications F3) leisure, culture(G1)

0.0

3296

955

14643

1452

0.02

0.01

0.05

0.0073

0.0

18026

666

126276

13516

0.09

0.01

0.26

0.0375

other goods services(H1)

0.2

2923

0

55281

3389

0.02

0.00

0.10

0.0134

177906

39405

639966

79865

and

total expenditure cf. gs90.l08.bl.scd6 (scdd2)

After aggregation we obtained the observation units practically without zero expenditure (cars, motocycles being the only exception). That is why no further specific treatment of zero expenditure is performed. The maximum values of budget shares are not excessive either. They are usually lower than 40% and in one case only exceed 50% (housing)..

4. Model’s specification with aggregated data. Aggregating data modifies the specification of models I and II. A budget share coefficient for one cell composed of G households can be written in the following way:

wi

g

∑ = ∑

h ∈G

y h wih yh

h ∈G

π ik Z kg + ui g g y k =0 K

(I-3) becomes (I-4)

wi g = β i + ∑

where y g and Z kg are cell means and

ui

g

∑ = ∑

h∈G

y h uih

h∈G

yh

The use of weighted regression with weights :

∑ (∑

h ∈G

yh

y 2 ) 1/ 2 h∈G h

reduces the risk of heteroscedasticity

resulting from grouping.

5. Model estimation 5.1 Identification problem. Table 5 - Retained regressors socio-demografic characteristics

variable name

constant term region

rg, cc

location

cc

number of household

working

people

head’s age

head’s marital

education level

in occupr, occupcj

agepr

matripr

etudpr

head’s social category

cspr

dwelling’s ownership status

stalog

car ownership

vehic

household’s size

ucoxf

variable classes

1 : Paris 2 : Ile de France, city of Paris excl. 3 : other region 1 : towns and villages less than 20 000 inhab 2 : towns between 20 and 100 000 inhab 3 :cities over 100 000 inhabitants 1 : head and spouse working 2 : head or spouse working 3 : head and spouse not working (unemployed or inactive 1 : less than 30 2 : 30 - 44 3 : 45 - 59 4 : 60 and more 1 : unmarried 2 : married 3 : widowed f 4 : divorced 1 : primary school and (or) vocational 2 : secondary school (vocational) 3 : secondary school (general) 4 : post-secondary school (2 years) 5 : university level 1 : farmer or farm worker 2 : other 1 :- owner 2 : - mortgage 3 : - tenant, lodger 4 : free accommodation 1 : no car 2 : at least one number of consumer units , Oxford scale

r coefficients r0 r1 r2 r3 r4 r5

r6 r7 r8 r9 r10 r11 r12 r13 r14 r15 r16 r17 r18 r19 r20 r21 r22

5.2 Model I estimation As it was mentioned when specifying the model, we need K+1 ( k=0,1,...K) constraints to identify the model. One of possible solutions is to set arbitrary rk (minimal expenditure corresponding to one unit of the characteristic k) and ro (total minimal expenditure of the reference household). However, for every cell it should be checked whether the total expenditure is higher than the subsistence expenditure resulting from the choice of rk . Following Baccouche and Laisney procedure, we obtain::

rk .=0 for k between 1 and 21 ;

r0 and r22 are estimated under constraint of the superiority of the total expenditure with respect to the subsistence one. e.g. : total subsistence expenditure of 28000 francs per CU and the « independent » of the cu total subsistence expenditure of 3000 francs. In order to ensure the concavity of reforms, these amounts are multiplied by 0.9. Finally we obtain:

r0 = 2700  r1 =... = r21 = 0 r = 25200  22 Under these conditions the system can be estimated equation by equation. 5.3 Estimation results (table 6) The goodness of fit the model is acceptable for a majority of expenditures. However, for some items we obtained relatively low adjusted R2 (round .20 and less: alcohol and tobacco, clothing, health, home equipment and services). Most of estimated parameters are not significant at 5% confidence level (*) or 10% confindence level (**). Estimated effects of socio-demographic characteristics. Generally the obtained signs and hierarchies of estimated parameters are in accordance with expectations. For example for the inhabitants of region of Paris, the subsistence expenditure for food at home is higher then elsewhere. Housing expenditure parameters and public transport are higher than for other categories. The correct negative relation is observed between age and housing expenditures. Similarly, expected increase of health expenditure for people of more than 60 was obtained. Marital status seems to be very discriminant as far as leisure and culture expenditures are concerned. The highest values were obtained for singles, divorced and widowed. Education level doesn’t affect very strongly the subsistence expenditure. These parameters increase with the level of education only for item «leisure and culture». For other goods the results are more ambiguous.

This is the household size which turned out to be the most significant regressor. All subsistence expenditures increase when the household size increases. The most significant effect is observed for food at home, car transport, clothing and culture and leisure - the less significant for alcohol and tobacco, telecommunication, heating and lighting.

Table 6 - L.E.S. estimation on aggregated data (1005 cells), Model I: subsistence expenditures variable, marginal budget shares constant with r0=2700, r22=25200

βi γ0 γ1 γ2 γ3 γ4 γ5 γ6 γ7 γ8 γ9 γ 10 γ 11 γ 12 γ 13 γ 14 γ 15 γ 16 γ 17 γ 18 γ 19 γ 20 γ 21 γ 22 2

R ajusté

A1

A2

A3

B1

C1

C2

D1

E1

F1

F2

F3

G1

H1

0.0771

0.0178

0.0542

0.0956

0.1716

0.0187

0.1079

0.0792

0.1878

0.0154

0.0110

0.1380

0.0258

3421

1128**

-577**

-2529**

12144

2086

-4772

-344**

-9254

427**

1884

-534**

-380**

653**

-401**

2512*

2292**

6998

-925**

-815**

-2046**

-11901

2616

1888

-1649**

778**

-2465

-389**

1649*

-1602**

6601

-1182

223**

363**

-5138

1434

83**

1137**

-714**

-1260**

-18**

956**

-1946

-2298**

901

250**

532**

4756

-352**

79**

-1118**

-482**

-796**

321**

716**

-703**

-197**

-759*

295**

211**

227**

540**

209**

-98**

33**

82**

-250**

1001*

1646

-6211

-389**

1314*

-270**

4615

15**

-812

-690**

-51**

-124**

-316**

-1482

-292**

691**

-169**

1555

-558**

1961**

71**

-570

-691**

-76**

-1951

-463

-771

-313**

2229

-40**

733**

-1029*

2298

286**

150**

-877**

-252**

2147

8**

-843

379**

-1403*

276**

-1934

-96**

776**

560

291

-584**

423*

3602

-676*

-1525

374**

-3309

540*

-1981

3635

-643**

-156**

386

16**

-264**

-4888

-143**

1667

1598

-3423

-1192

-1349**

-1863*

5946

1093

-206**

3031

-271**

-5552

-1270

1548

274**

4235

51**

714**

-2873

253**

44**

393

1268

916

-3497

-1241

1334*

631**

681**

304**

-835**

-1300**

1425**

254**

448

1467

330**

2993

1471

-2122

-2234

253**

458**

2828

-513**

5875

-2491

-1051

-5923

457**

2708

1878

-2222

-2376

1306**

695*

2656

-112**

3757*

-3085

-681

-5226

703**

1182**

739**

-3351

-651**

7885

1503

1932**

-39**

-1146**

-1396

-371**

-7046

757**

-787**

-150**

-91**

-1207**

5169

931*

-75**

-3702*

-2376**

933**

574*

176**

605**

-115**

-365*

382**

1409

-611**

-1023

635**

543**

-1323**

25**

-247

240**

451

1400

-611

-1392

-978

4749

1876

-161**

-51**

-3628

-361*

-221

-925*

303**

-312**

-818

-506**

-891

4731

1067

870*

-1053*

-2097

-262**

-338

-256**

-134**

-357**

296**

801**

1173

-7414

-480*

668**

1062**

990**

467**

-348

2831

310**

-1119**

-255**

1792

2815

1455**

-237**

782**

1708*

-9033

810

-47**

1766

-437**

7063

540

1477

3085

2224

887

736

1853

3102

1029

417

2927

-139**

0.74

0.22

0.37

0.20

0.50

0.66

0.17

0.18

0.35

0.46

0.48

0.37

0.13

5.4 Estimated elasticities. Price and income elasticities can be easily computed from estimated L.E.S using following formulas (Deaton et Muellbauer 1980) : Income elasticities: (L13) ei =

βi wi

Non-compensated price elasticities: (L10) eij = Kij φei − ei w j − ei β j φ with φ = − ( y −

n

∑p γ k

k

)/ y

k =1

The hierarchy of obtained elasticity estimates is for both models correct: food at home, alcohol and tobacco, energy home equipment, public transport, telecommunication appear as necessity goods. Health, car expenditures, leisure and culture and more surprisingly clothing look like luxury goods (tables 9 and 10). Table 7 Model I: Total expenditure elasticities 2 average elasticity A1 A2 A3 B1 C1 C2 D1 E1 F1 F2 F3 G1 H1

0.497 0.693 1.231 1.256 0.775 0.462 0.263 1.099 1.343 0.869 0.593 1.362 1.571

cell mean 0.522 0.938 3.147 1.529 0.784 0.486 0.370 1.378 2.698 2.383 0.619 1.750 4.932

minimum

maximum

st. deviation

0.246 0.170 0.339 0.411 0.344 0.162 0.072 0.205 0.526 0.113 0.214 0.538 0.247

1.912 31.937 121.087 15.912 1.569 1.474 6.721 12.091 328.605 197.427 1.675 11.617 199.331

0.184 1.441 7.994 0.843 0.207 0.186 0.307 0.851 11.365 7.371 0.206 0.954 12.333

minimum

maximum

st. deviation

-1.667 -23.537 -65.879 -8.290 -1.332 -1.191 -3.961 -6.214 -119.582 -128.098 -1.523 -5.575 -99.886

-0.102 -0.050 -0.197 -0.212 -0.213 -0.051 -0.054 -0.208 -0.368 -0.064 -0.050 -0.331 -0.159

0.162 0.986 4.550 0.472 0.147 0.162 0.174 0.558 4.272 4.705 0.170 0.442 6.253

Tableau 8 - Model I: own, non compensated price elasticities average elasticity -0.398 A1 -0.494 A2 -0.869 A3 -0.890 B1 -0.620 C1 -0.336 C2 -0.183 D1 -0.787 E1 -0.951 F1 -0.614 F2 -0.421 F3 -0.959 G1 -1.096 H1 cf. gs90.l08.scd6(scdd2)

cell mean -0.410 -0.649 -1.947 -1.004 -0.608 -0.346 -0.232 -0.932 -1.524 -1.534 -0.425 -1.107 -2.868

5.5 Sensitivity of results on the initial values of r parameters. The identifying parameters of Model I (r0 to r22) was chosen arbitrary. In fact the total subsistence expenditure depends only on one characteristic: household size (r22). Other household characteristics are supposed to be neutral with respect total subsistence expenditure and influence only the distribution among different goods. To what extent this choice could influence the estimation results? What could have been the estimates in case of different choices? In the first variant (V1) we applied the following values for r parameters, where the subsistence expenditures are multiplied by .9: r0 = 2430  r1 =... = r21 = 0 r = 22680  22

The second variant (V2) takes into account the fact that some characteristics (location in Paris location in region of Paris influence significantly so important expenditures like food at home, housing and car-motorcycle expenditures (see table 7). Thus the alternative choice is proposed for r1, r2 et r21 following probable real values (taking into account difference between Parisian and non-Parisian residents, car and public transport expenditures) instead of zero values: r0 = 2700 r = 6000 1 r2 = 3000  r3 =... = r20 = 0 r21 = −6000  r22 = 25200

Tableau 9 - Comparison of variants V1 and V2 for Model I elasticities for average cell A1 A2 A3 B1 C1 C2 D1 E1 F1 F2 F3 G1 H1

2

model I -0.398 -0.494 -0.869 -0.890 -0.620 -0.336 -0.183 -0.787 -0.951 -0.614 -0.421 -0.959 -1.096

V1 -0.412 -0.514 -0.904 -0.924 -0.640 -0.349 -0.190 -0.817 -0.984 -0.639 -0.439 -0.994 -1.142

average V2 -0.399 -0.495 -0.871 -0.892 -0.622 -0.337 -0.183 -0.789 -0.953 -0.615 -0.423 -0.962 -1.100

model I -0.410 -0.649 -1.947 -1.004 -0.608 -0.346 -0.232 -0.932 -1.524 -1.534 -0.425 -1.107 -2.868

Cells whose budget share is zero are excluded for the calcul of elasticities.

V1 -0.425 -0.678 -2.056 -1.051 -0.629 -0.361 -0.244 -0.974 -1.606 -1.617 -0.444 -1.161 -3.064

V2 -0.413 -0.658 -1.996 -1.017 -0.611 -0.348 -0.236 -0.940 -1.597 -1.562 -0.427 -1.124 -2.949

Decreasing by 10% the total subsistence expenditure (V1) leads to an increase of about 4% of all own price elasticities. On the other hand alternative variant V2 gives very similar results when compared to the reference situation. In the light of these results we decided to use for simulation the reference definition rather than any of variants V1 and V2. However, it would be interesting in the future to estimate r parameters by an approach based on declared in the survey information reporting the minimal needed income rather than arbitrary chosen values.

6. VAT reforms simulations based on the Family Budget Survey. 6.1 Hypothesis 6.1.1 Prices In principle the VAT modifications should change automatically the consumer’s price structure. However several substitution effects can take place in production and distribution as well which makes a global, precise evaluation can be very difficult. We suppose that the consumer’s prices entirely reflect the change in taxation. All costs (or gains) of the reform are borne only by the purchaser. Consumer’s prices are supposed to be composed of production price plus VAT:

pi = qi (1 + t i ) where t i is VAT rate Modifying the rate of VAT from t i0 to t i1 , the price of the good j can be written: (L14) pi1 = pi0 .

(1 + t i1 ) (1 + t i0 )

6.1.2 Homogeneity of tax rates Using aggregated LES implies strong hypothesis about VAT rates within the grouped items. The aggregation by goods homogenous from the point of view of consumer’s behaviour does not always mean that it is coherent with taxation rates. In reality heterogeneity among goods and rate of taxation composing a group can be very strong (see table 10).

Tableau 10 - Tax rate heterogeneity of the expenditure groups.

food at home) alcohol & tobacco (A2) food away (A3) clothing (B1) housing (C1) heating & lighting (C2) equip and home . (D1) health,, hygiene (E1) car, motorcycle (F1) other trans. (F2) telecommunications (F3) leisure and culture (G1) other goods and services (H1)

share of taxed item at 0 rate at special rate 0.0 0.0 0.0 0.0 17.4 0.0 0.0 0.0 70.5 0.0 0.0 0.0 10.7 0.0 77.4 16.7 32.9 0.0 7.7 0.0 10.7 0.0 7.7 6.4 74.8 0.0

at reduced rate 97.1 100.0 12.4 0.0 2.8 0.0 0.0 0.0 0.0 91.1 0.0 42.0 4.0

at normal rate 2.9 0.0 70.2 1.0 26.7 100.0 89.3 5.9 67.1 1.2 89.3 43.9 21.2

The adopted remedy to the problem is a desegregation of each item into homogenous subgroups with respect to the VAT rate like in Nichèle and Robin (1993). 6.1.3 Computation of indirect tax revenues. If a group i contains ki sub-groups homogenous in terms of VAT price variation formula is modified: (L15)

∆pi = log pi1 − log pi0 becomes pi0

(L16)

∆pi = ∑ wik (log pik1 − log pik0 ) 0 pi k

(L17)

∆pi t ik1 − t ik0 = ( w ∑k ik 1 + t 0 ) pi0 ik

or

and

t ik1 − t ik0 ) + pi0 (L18) p = p . ∑ wik ( 0 1 + t ik k 1 i

0 i

where t ik0 and t ik1 , are VAT rates before and after the reform for the sub-group k in the group i. The indirect tax revenue R can be written: (L19) R 0 =

G

t i0 wig0 0 1 + t i =1 i n

∑ π g y g0 ∑ g =1

Decomposing into homogenous VAT rate sub-groups:

ki G n G t ik0 t ik1 0 1 1 w ). w = π y ( w ). w − l πg . (L20) R = ∑ π g y ∑ ( ∑ ∑ ∑ ∑ ∑ ik ig ik ig g g 0 1 g =1 i =1 k =1 1 + t ik g =1 i =1 k =1 1 + t ik g =1 0

G

0 g

n

ki

where:

π g ,- weights for extrapolation to the total population, l - a flat rate payment to all household (if need to be). Budget shares before reform are those observed in the survey. The coefficients after reform ( wig1 ) are estimated with the equation system (I-3) with prices p 1 , total expenditure y 1g = y g0 + l and residuals

u$ g0 = w g0 − w$ g0 , (i.e. residuals obtained after model estimation). This is stochastic version (Baccouche et Laisney (1989)), the deterministic one ( u$ g0 et u$ 1g ,) would be obtained by putting 0 for residuals.

The actual VAT revenues reported in national accounts are 645 billion Francs (1994). From this amount, about 60% (390 billion Francs) are paied by households. The estimation from the family budget data gives about 344 billion Francs of tax revenue..

6.2 Simulated VAT changes We will try to evaluate the effects of different changes in VAT using our simulation procedure. Some additional, simplifying hypothesis are necessary: - production prices are not affected by the change in VAT rates. Thus, the VAT changes pass on totally in the selling price - household’s total expenditure is supposed unchanged: households don’t modify their saving behaviour or indebtedness - all a household can do is to reallocate its unchanged total expenditure among different goods. 6.2.1 Simulation 1: increase by 2 points in normal VAT rate. This is a real reform decided in August ‘95. The normal VAT rate was then increased from 18.6% to 20.6%. Table 11 - Simulation 1: VAT rates modifications rate 0 special rate reduced rate normal rate VAT revenues (billion of francs)

before reform 0.0 % 2.1 % 5.5 % 18.6 % 338.75

after reform 0.0 % 2.1 % 5.5 % 20.6 % 365.48

6.2.2 Simulation 2 : unification of the reduced and normal VAT rates The normal (18.6) and reduced (5.5) rates are replaced by one averaged rate giving the same tax revenues (13.9). Zero and special rates are maintained as they concern very specific consumption: - medicines, certain necessity goods, - the press; - second hand goods. Table 12 - Simulation 2: VAT rates modifications rate 0 special rate reduced rate normal rate VAT revenues (billion of francs)

before reform 0.0 % 2.1 % 5.5 % 18.6 % 338.75

after reform 0.0 % 2.1 % 13.9 % 13.9 % 338.99

6.1.3 Simulation 3: Compensated total tax exemption for food at home and energy The third simulation supposes the exemption from VAT item A1 and C2 (food at home, heating and lighting).). This tax reduction is compensated by an increase of the normal VAT rate. Obtained by iterations the new normal rate would be 23.7% Table 13 - Simulation 3: VAT rates modifications total exemption of food at home and energy rate 0 special rate reduced rate normal rate VAT revenues (billion of francs)

before no

after yes

0.0 % 2.1 % 5.5 % 18.6 % 338.75

0.0 % 2.1 % 5.5% 23.7 % 338.18

7. Integration of indirect taxes into the model of microsimulation 7.1 Static module In the current version of INSEE microsimulation model the indirect taxes module is based on the Family Budget Survey ‘95 and allows simple static evaluations of indirect taxes reforms. This static module can be used to evaluate very detailed reforms because of possibility to use the full nomenclature of goods present Family Budget Survey.

However, like in every static case no behavioural reaction is taken into account and large scale reforms would certainly underestimate long term substitution effects. 7.2 Behavioural effects This module is being developed and is not entirely operational. Like in the static case evaluations are based on both Family Budget survey and Fiscal Survey matched using four classes of head’s age 6 classes of family type deciles of total per CU family income (transfers included but wealth income excluded for lack of precision in both sources). The cross section of these 3 variables gives about 224 cells which serve to match two surveys . Table 14 - Comparison of totals of VAT revenues computed on Family Budget Fiscal Surveys for proposed simulations (billion of francs). reference simulation 1 simulation 2 simulation 3

Family Budget 338.75 365.48 338.99 338.14

Fiscal Survey 342.07 369.09 342.09 341.45

The difference of about 1% is due to the slightly higher weighted number of households in Fiscal Survey that in Family Budget Survey. 7.3 Results Simulation 1 The increase of 2 points in normal VAT rate yields the extra revenues of 27 billion of francs, 1152 francs per year and per household. The disposable income effect of this increase is almost uniformly distributed among all type of households (table 17). The share of VAT is increasing between from 0.6 to 0.8 depending on the family type. Table 15 - Simulation 1 : results by family type household type

disposable income VAT share (per CU,in francs), income(%) normal rate 18.6% single,other 101732 8.9 couple 0 children 119491 7.6 couple 1 child 112391 9.0 couple 2 children 104924 7.9 couple3 children + 85461 8.5 lone parents 77773 10.0 total 105542 8.4

in

disposable variation

normal 20.6% 9.6 8.2 9.7 8.5 9.2 10.8 9.0

rate +0.7 +0.6 +0.7 +0.6 +0.7 +0.8 +0.6

The same differences can be observed when analysing per CU disposable income (graph 1). This change in VAT has practically no redistributive effect but a proportional uniform loss in purchasing power among all family and income categories. Graph 1: Simulation 1: 2 point increase in normal rate of VAT all households classed by centile of disposable income (per CU) 2 point increase in VAT normal rate

VAT, in % of disposable income

14 13 12 11 10 9 8 7 40000

60000

80000

100000 120000 140000 160000 180000 200000 disposable income (perCU)

vat(normal rate18.6%)

vat (normal rate 20.6%)

Simulation 2 The second simulation consists in unification of reduced and normal rates. This is done respecting constant VAT revenues principle (this was not the case of the first simulation).Thus the obtained average between normal and reduced rates of 13.9% should give the same total revenues. Table 16 - Change in VAT paid by household resulting from the unification of reduced and normal rates: by family type and income. in francs centiles of disposable income (per CU)

disposable income (per CU)

single, other

couple children

5 10 20 30 40 50 60 70 80 90 95 100 All

24213 44120 53297 63226 72897 82393 92986 106262 123715 152387 196847 351461 105542

159 214 176 134 52 26 -24 -156 -245 -384 -492 -443 -53

267 323 429 407 400 326 172 79 5 -206 -412 -544 88

no

couple 1 child

couple children

493 497 281 203 55 63 -27 -236 -384 -620 -1172 -1134 -203

653 615 442 269 94 25 -30 -147 -264 -456 -1236 -1336 -79

2

couple children

1050 1058 812 694 585 412 308 192 -258 -461 -367 -122 505

3

lone parent family

All

230 213 198 40 -103 -154 -314 -413 -333 -266 -493 -579 -47

389 392 364 273 174 123 25 -102 -203 -381 -639 -667 1

The consequence of the change in VAT rates are, as it could be expected, relatively high losses observed among poor people and significant gains for rich part of population (see table 18).

The main reason of this result is the effect of increase of food VAT rate. High budget share for food among poor and small among rich as well as a low price elasticity of this necessity good explain almost entirely observed differences. This fact explains also the case of families with 3 or more children (and probably those of couples with no children) for which the losses are the largest for poor and gains the lowest for the rich. It can be interesting to compare the effect to use a behavioural model (L.E.S) instead of static one (with no consumers’ reactions and unchanged budget shares): behavioural model predicts 1.2 % lower total VAT revenues then the static one (338.93 and 342.09 billion respectively). More detailed comparison confirm relatively small difference between two approaches. This is surprising because the change in prices is strong (minus 5 points for good taxed at the normal rate and +8 for reduced rate. These changes do not seem strong enough to create substantial substitution effects for different expenditure items. In terms of disposable income evolution the difference between static an behavioural versions is usually less than 0.1%. (graph 2). Graph 2 : Simulation 2 : comparison of static and behavioural approaches. Simulation 2 : comparing static and behavioural approches 1 0.8

couple without children

couple with 1 child

couple with 2 children

1* 2 3 4 5 6 7 8 9 10*

1* 2 3 4 5 6 7 8 9 10*

1* 2 3 4 5 6 7 8 9 10*

couple with 3 children & +

% disposable income

0.6 0.4 0.2 0 -0.2 -0.4 -0.6

1* 2 3 4 5 6 7 8 9 10*

deciles of per CU disposable income * 1stpercentileand lastpercentile are excluded

L.E.S.

static

Simulation 3 This simulation evaluates the effects of

total exemption from VAT (rate 0) of two necessity

expenditure groups: food and energy compensated by an appropriate increase in normal rate (in order to maintain the similar amount of total tax revenues. As expected , the results are almost opposite to

those obtained in simulation 2: poor households are the main beneficiaries of this change and rich one main contributors. The influence of family composition is very limited Table 17 - Simulation 3: Change in VAT paid by household resulting from exemption from VAT of food and energy compensated by an increase in normal rate: by family type and income. in % of disposable income 5 10 20 30 40 50 60 70 80 90 95 100 all

single, other -0.7 -1.4 -1.0 -0.7 -0.4 -0.4 -0.2 0.1 0.2 0.3 0.3 0.1 -0.1

couple no couple children child -1.6 -1.4 -1.0 -0.8 -1.1 -0.4 -0.9 -0.2 -0.8 -0.1 -0.6 -0.0 -0.3 0.1 -0.2 0.2 -0.1 0.3 0.1 0.4 0.2 0.5 0.1 0.3 -0.2 0.2

1 couple children -1.7 -0.8 -0.5 -0.3 -0.1 -0.0 0.0 0.2 0.3 0.3 0.5 0.3 0.1

2 couple children -0.7 -0.6 -0.4 -0.2 -0.2 -0.0 0.1 0.2 0.4 0.4 0.4 0.2 -0.0

3 lone parent family -0.7 -0.5 -0.3 0.0 0.2 0.3 0.3 0.4 0.2 0.2 0.2 0.1 0.1

all -1.0 -0.9 -0.7 -0.5 -0.3 -0.2 -0.1 0.1 0.2 0.3 0.3 0.2 0.0

Households from the first decile increase their disposable income by 1%, those from 6th and 7th don’t change and the upper deciles loose 0.3% of their disposable income. The differences between different family types are very small; this simulation seems to be relatively neutral with respect to the family type. When compared with the simulation 2 this one is much more sensitive to the substitution effects controlled in the frame of LES model. Static model gives 3.5% higher tax revenues then behavioural (LES) model. The Static model overestimates taxes paid by all households (.3%) more or less to the same extent whatever the point of the income distribution (see graph 3). Graph 3 - Simulation 3: comparison between static and LES results. Exoneration of food and energy 13

% of disposable income

12

11

10

9

8

7

1*

2

3

4 5 6 7 decile of disposable income per cu

*1st and last percentile are excluded

8

reference

9

L.E.S.

10*

Conclusion Including behavioural responses into a microsimulation model is not an easy task. Even in the frame of consumption models, relatively rich of theoretical and practical examples a number of problems questions have to be solved by more or less arbitrary hypothesis. This is particularly the case of goods’ and households’ aggregation, identifying assumptions for estimation models and inadequacy of VAT tax rates and expenditure aggregates. Despite of all these difficulties the obtained estimation and simulation results are acceptable, and close to expectations. Simulations based on the estimated model give reasonable results. However, the comparisons between results of static and behavioural models show generally only relatively little differences although the static model overestimates systematically tax revenues. Other methods of estimations of the L.E.S. (pseudo-panel data in particular) are planed to be tested so as to improve the estimation of elasticities and limit the number of arbitrary asumptions adopted in the presented model.

References Baccouche, R. et Laisney, F. « Analyse microéconomique de la réforme de la TVA de juillet 1982 en France », Annales d’économie et de statistiques, n°2, avril/juin 1986, p. 37-74. Baccouche, R. et Laisney, F. « Simulation of value-added tax reforms for France using cross-section data », Microeconometrics : surveys applications, Basil Blackwell 1990, p.265-301. Barten, A. et Böhm, V. (1982) « The consumer theory », Handbook of Mathematical Economics, vol II, pp382-429. Blundell R. (1988) « Consumer behaviour : theory and empirical evidence - a survey », The Economic Journal, mars 1988, pp.16-65. Deaton A. (1983) « Demand Analysis », Handbook of Econometrics, vol III, pp1768-1837. Deaton A. et Muellbauer J. (1980) « Economics and consumer behavior », Cambridge University Press chapitres 1, 2, 3, 5 et 6. Decoster, A. et Van Camp, G. « Evaluation of simultaneous reforms in direct and indirect taxes, Belgium 19881993 », presentation paper for the 5th Nordic seminar on microsimulation models, Stockholm, june 1997. Gauyacq D. « Les systèmes interdépendants de fonctions de demande » (1985), Prévisions et analyse économique, cahiers du GAMA vol.6, n°2, juin 1985, chapitres 1,2 et 8. Nichèle, V. et Robin, J.M. « Evaluation des effets budgétaires et redistributifs de réformes de la fiscalité indirecte française », Economie et prévision 1993 n°4-5, pp.105-128. Pollak, R.A. et Walles, T.J. « Estimation of complete demand systems from household budgetdata : the Linear and Quadratic Expenditure Systems », The American Economic Review, vol. 68 n°3, june 1978, pp.348-359. Trognon, A. « Composition des ménages et système linéaire de dépenses », Annales de l’INSEE, vol. 81, jan/mars 1981, pp.4-42.

Appendix 1 The Microsimulation Model

The microsimulation model used in this studyis is a set of data bases and procedures conceived to enable the evaluation of redistributive effects of changes in tax benefit system at the individual level (household, tax unit, individual). It is a simple static model with a possibility to include the individual behaviours and to update the demographic structure. Description: 1. Tax-benefit system All taxes contributions and transfers are included (with only small exceptions): 1.1 Taxes and contributions income tax social contributions VAT and other indirect taxes local taxes 1.2 monetary transfers family allocations back-to-school benefit young child benefit (APJE) housing benefit (AL, APL) lone parent benefit (API) family supplement Minimum Income (RMI) minimum pension 2. Behavioural responses The model is static. However it is possible to introduce consumer’s reactions in VAT part of the model. 3. Coverage The whole population of ordinary households. Persons in institutions are excluded. Corporate taxes are not included. 4. 4.Model’s structure The model is divided in several units. The most important one is the large data basis on the individual level, computation procedures (taxes, means tested benefits...), simulation procedures, report module to present results under a standard format. 4.1. Data basis. Tax file records 1990 (28000 households, 34 000 tax units, 80 000 individuals) is a central source of information. Other sources (surveys and administrative files) to complete lacking information or update: Surveys - Family Budget Survey (BDF 1995), Housing Survey, Financial Assets Survey, Employment Survey Administrative files - Annual Wage Register (DADS),Pensions register (SESI). 4.2 Computing procedures Income Tax (IR) - exhaustive computing procedure adapted from Ministry of Finance module. Housing benefit - exhaustive computation procedure for means testing and the level of benefit computation Social contributions module - computes contributions for wage earners and self employed, general social contributions (CSG, RDS, CSGbis). 4.3 Updating procedures : reweighting module (CALMAR), updating incomes,(multipliers from DADS, SESI, DGI files SESI records) 4.4 Imputation procedures for some transfers and benefits. 4.5 Simulation programs: income tax modification change in indirect taxes introducing new direct taxes (CSG, RDS) substitution between different taxes and contributions 5. Reports The redistributive effects of the change in the tax benefit system can be presented at the individual level by typical sociodemographic characteristics : demographic structure, family type, percentile of income, social category, location... Standard indicators to measure inequality and change in income distribution are available.

Appendix 2 Matching criteria for Family Budget Survey and Fiscal Survey Results are weighted to obtain the total population of ordinary households (institutions excluded) Head’s age, 4 classes Family Budget

Fiscal Survey

less than 30

11.5

11.9

30 - 44

30.1

30.1

45 - 59

24.8

24.7

60 and more

33.6

33.3

all

100.0

100.0

Family Budget

Fiscal Survey

single, other

32.7

31.8

couple no children

26.0

26.5

couple 1 child

14.2

14.1

couple 2 children

13.7

13.5

couple 3 children

7.2

7.3

lone parents

6.1

6.8

all

100.0

100.0

Family type

« Common » income per CU decile 1 2 3 Family Budgets Fiscal Survey

4

5

6

7

8

9

10

35 932

45 667

54 191

62 669

72 000

82 165

94 781

111 200

143 970

2 295 000

38 511

47 429

55 599

64 158

73 558

83 873

96 946

115238

148 817

30 003 850