UNIVERSITY OF

ILLINOIS LIBRARY

AT URBANA-CHAMPAIGN

BOOKSTACKS

Digitized by the Internet Archive

in 2011 with funding from

University of Illinois Urbana-Champaign

http://www.archive.org/details/determinantsofho221poir

Faculty Working Papers

THE DETERMINANTS OF HOME BUYING IN THE NEW JERSEY GRADUATED MORK INCENTIVE EXPERIMENT

Dale J. Poirier

#221

College of Commerce and Business Administration

University of Illinois at Urbana-Champaign

FACULTY WORKING PAPERS College of Commerce and Business Administration

University of Illinois at Ur Dana -Champaign December ?, 1974

THE DETERMINANTS OF HOME BUYING IN THE NEW JERSEY GRADUATED ITORK INCENTIVE EXPERIMENT

Dale J. Poirier

#221

The Determinants of Home Buying in the New Jersey Graduated Work Incentive Experiment

Dale J. Poirier September 1974

1« Introduction

The purpose of this study is to investigate home buying during the New Jersey Graduated Work Incentive Experiment. The importance of such an analysis appears to be at least five-fold.

First, Wooldridge [9] in a detailed and thought-provoking analysis of many different housing issues, and Nicholson [5] in a somewhat less detailed analysis, both reached the conclusion that there existed a definite experi- mental effect on homeowner ship. Specifically, Wooldridge fs analysis found that experimentals who were over-breakeven at pre-enrollment bought hoae» at a statistically higher rate (about 5%) than their control counterparts by the end of the experiment. Since the over-breakeven experimentals were eligible to receive payments if their incomes dropped sufficiently, the explanation put forth by Wooldridge was that the guaranteed income gave this group the "financial security" to purchase homes,"... not only in their own eyes, but in- the eyes of potential lending agencieB." While this latter supply side explanation does not seem consistent with finding an experimental affect at the end of the experiment, this supposingly non- payment experimental effect is interesting enough in its own right to re- ceive added attention. Furthermore, since home buying by experimentals

-2-

a costless (for the administrators of the experiment) treatment, its

validity should he rigorously investigated.

Second, the results of the entire experiment are subject to the

criticism that they reflect responses to a temporary three year experiment

which differ from those that would be observed in a permanent national plr.n.

As Metcalf and Nicholson [3, p. 1] point out:

"... many methodologica- approaches "o analyzing the effects of a negative Income tax view houslolds as making coordinated decisions regarding labor-force behavior, consumption behavior, and asset accumulation. Evidence that households view experimental payments as a -transitory income source would therefore be an im- portant indicator that the labor supply erfects of a permanent income maintenance program may not correspond to observed behavior during the experimental."

Similarly, evidence that households view payments as a pemauent in- come source would give additional credibility to the validity of labor supply results for a permanent national plan. There appears to be near

unanimous agreement that hone buying is a function of "normal" or "penaa-

3

nent" income rather then transitory income. Hence, increased hone buying

among experimental families receivi .g payments would tupporf the belief

,4 that experimental families viewed the payments as "permanent.

Third, any strong positive experimental effect among nonwhites compared to whites may help tc reduce limitations on homeownerrhip found among non- whites as a result ce discrimination. As Kain and Quigley [2, p. 273] heve noted, "Homeownership is clearly the most important method of wealth accumu- lation used by low- and middle- income families in the postwar period." Furthermore, they go on to estimate that "... an effective limitation on homeownership can increase Negro housing costs by over 30 per cent, assuming no price appreciation."

-3-

Fourth, for both whites and nonwhites, low homeowner. oh ip rates imply ineligibility from favorable treatm nt accorded horaeoi.nera under federal income tax laws. To the extent that experimental can increase their home buying, they can become eligible for substantial savings in tax breaka.

Fifth, regardless of experimental questions, the panel nature of the data permits a more sophisticated analysis of home buying than is normally possible with cross-sectional studies due to the availability of a "normal1' income variable.

In light of these considerations, the claimed experimental effect on home buying is critically investigated in the remainder of this study. The plan of attack is as follows. Section 2 discusses the selection of the "appropriate" sample (which incidently differs markedly from the one used by Wooldridge [9]. Section 3 then outlines the probit model used in the analysis, leaving to section 4 a description of the independent variables, Empirical results are conteined in section 5 and section 6 attempts to recon- cile them with previous studies.

2. Sample Selection

The selection of the sample in this study is crucially important. Wooldridge [9] used a sample of 776 families, each observed at pre-enroll- ment and quarters four, eight, and twelve. Nicholson's [5] sample consisted of 750 families each observed at quarters four, 3ixt eight, and twelve. The present sample differs from both of these in two important aspects. First, attention is restricted to the "continuous sample" of 693 families which has formed the basis of many of the analyses of the experiment. Thir restriction allows fcr the use of the normal income variable constructed by Watts [8]. Second, and more importantly, attention is restricted to

-4-

only those families who moved (25 7) or those who changed their tenure status wit .out moving (16). Thiy atter group consists of families who apparently bought the house they had been renting. The reasons for restricting attention to only those families who changed their housing status are multi-fcld:

(1) With regard to differences in ethnicity, Kain and Quigley [2, p. 265] remark: "There are some indications that the barriers to Negro occupancy in white neighborhoods are gradually declining. Thus,

it could be argued that current ownership patterns primarily reflect historical discrimination and provide a misleading view of current conditions."

(2) In addition Kain and Quigley [2, pp. 265-6] go on to argue: "Because of past discrimination, Negro movers are less likely than white movers to have been homeowners in the past. This is important because when homeowners change their residence they are more likely to buy than to rent and, conversely, when renters move they are more -ikely to move from ore i.ental property to another."

(3) Along the same lines, current home-buying opportunities may not differ across sites by as much a3 the historically- influenced pre- enrollment differences fo-u.^ by '■/ooldridge [9, pp. 5-7]. Specifi- cally, pre-enroiiment homeovnerhhip rates for Trenton, Paterson- Passaic, Jersey City, and Scranton were 19.7, 5.5, 6.0, and 29.3 per cent, respectively, ror Wooldririge's sample.

(4) Even more pertinent to the analysis of possible experimental effects on home buying, it teeu irrelevant to study experimentais (or for that fact controls; who never moved or changed their housing statue during the experiment. "larher, the appropriate sample in which to

-5-

look lor experimental effects on home buying is one limited only t ' potential home buyers. Dy using samples such as Wooldridge's or Nicholson' v , experimental home purchasing effects could be mistal'.ingly confused vith other experimental effects such as the ability of experimental to retain the homes they already live in at a higher rate than their control counterparts.

For these reasons this study focuses attention solely on the condi- tional probability of buying a home given a move or change in tenure status. This permits uncovering an experimental effect on home buying over an above a simple experimental effect on mobility. Indeed, since an improvement in housing made by moving into a "better" rental unit is a much lower cost action than buying a house, it is a more likely candidate in which to find an experimental effect due to a three year negative income tax experiment.

A detailed description of the 273 observations comprising the sample is provided in the Appendix. Table Al, A3, and A5 b-eak down this sample according to ethnicity, experimental 3tatu3, site, and year. Tables A2, A4, and A6 give the same breakdown for the 80 observations corresponding to home buyers. Because quarterly housing data is not available on a regular basis, the base time period is a year.

Very briefly the importance of these tablej lies in the raw experi- mental-control differentials in probabilities of home buying contained therein. Specifically, experimental advantages of .2687 vs .1765 end .4462 vs .3793 for whites and blacks, respectively, an^ a control advantage of .2500 vs. .1622 for Spanish-speakers. Thus while the positive experi- mental effect found by Wooldridge [9] and Nicholson [5] seems evident

-6-

for whites and blacks, the rxaci opposite seems to be the case for Spanish-speakers. Furthermore, pre .ounced ethnic differences are appar- ent in levels as well as in differentia's. Thin is not surprising In light of the numerous ethnic differences which have been noted elsewhere in analyses of the experiment, however, the studies of both Wooldrtd, i and Nicholson allow for ethnic differences onlv throt gh simple interrept dummies. In contrast the analysis of Section 5 will deal with ethnic differences by a complete depooling of whites, blacks, and Spanish-speakers Analysis will then .proceed to determine whether each of these differences can be explained by factors other than those of the experiment.

3. Probit Model

Let y (J ■ 1, 2,...,n) be a binary variable indicating whether the

jth family purchased a home (y ■ 1) or did not purchase a heme (y ■ 0).

Let x - [x , x ,..., x ] (j - 1, 2,...,n)(k < n) be a row vector of J J-*- J ^ J k

socio-economic-experimental variables pertinent to making this purchase decision, l:t fl - [6 , 6_,..., 3, ]' be the corresponding column vec.or of coefficients, and let I be an index for the jth family which is a linear

function of the regressors, i.e., T4 - x 6 Cj - l,2,...,n). The probit

j J

model used in this stuoy postulates the existence of btandard normal randorr

* variables I (J - 1 , ?,..., n) such that the home purchasing decision can

be described by

U - 1: 2 n).

-7-

In this context the decision of the Jth family to buy a hone is assumed to be a function of the regressor6 (v:- 1 the index I ) nn* of the random variable I which serves as a disturbance term.

Denoting by P(z) the value of the standard ncrnAl cumulative distri- bution evaluated at z, the probability ui the jth family buying a home is Prob (y - 1 1 X > ■ Prob (I < ijl.) - F(I ), and the probability of not

buying Prob {y - 0|l.} ■ Prob fl' > I |l } - l-F(I ). Assuming ind< - J .. j J J J

pendence among family decisions, and ord?ring the sample so

that the first m observations correspond to families who bought, and

the remaining n - r observations to those who did not buy, the log-likclihooc

of the sample is

m n

L - iuL - 2^ InF(I-) + Z- to[i - F(I )]

j-1 j-m-ri

Setting the derivatives of (1) with respect to B, , B2>..., S. equal to zero yields nonlinear normal equations whose solution is the maximum like- lihood (ML) estimator tf - [B. , 30,..., 3.)'. The ML estimator 3 is consist- ent, asymptotically efficient, and has an asymptotic normal distribution »:itn mean B and a variance-covarir.nce matrix which can be approximated by

.2,1-1

6 L

aese*:

j

s-e

4 . Independent VariaMe Selection

The row vector x, of independent variables for the Jth family CA1

J

be conveniently partitioned into socio-economic-demographic variables 'Mich affect the decision of all families in buying a home, md into "treatment"

-8-

variables vtiich affect only the experiment rroup. With regard co the fi^sc set, the following variables (besidtj a constant) form the basis for the subsequent analysis.

First and foremost, average normal family income (in thousands of dollars) for the year In which the move occurred is included as a regressor. As indicate! in Section 1, economic theory clearly implies that a normal income type of variable should be used instead or a transitory Income variable.

Furthermore, unlike current income, normal income has been purged of any

g experimental effect. Normal income is expected to have a strong positive

effect on the probability of purchasing a home, and because of the ^ay it

was constructed, its effect may "swamp" that of many other non-experimental

variables.

Besides normal income (which includes non-work-conditioned unearned income;, worh-condition unearned income (measured in thousands of dol.'.arn,) is also included as an earnings regressor. Thi3 permits explicit recogni- tion of welfare payments which often have been neglected In other s'rdies. Since welfare status may indeed reflect a treatment effect, especially among experimentals on the least generous plans, an additional regressor is included which interacts welfare payments with an experimental dummy.

Second, the prior tenure of the family is accounted for by the inclusion of dummies for families 1 Lving Ln public housing aad for families who own homes . To the extent that public housing is the least desirable tenure status, a family Jiving in quelle : jusing would 1 cBtt likely to "Jenp" all the wry up to the highest leveJ oi tenure stntus, namelv hopeownershlp than a private renta-i family. By the same token, an urrument along the lines of Kain and Cuigley to he mentioned in section 2, indicates that once a family owns and then moves, it is likely to buy again. However, considering

-9-

the low Income levels of the sample, it. could be that once a family owns a house they i.ave In a =><_r.ae reached he apex of their lifetime housing consump- tion curve, and henv_e they are unlikely to move unless It is a forced move - possibly due to mortgage for closure. ' While this latter hypo the tjir. cannot be testec directly, I . possible explanation for tailing to observe a significantly positive effect or prior ownership such as found by 1'aln and Quigley [2, pp. 266-7].

Third, as in most homeovnersMp tttidieo, family life cycle variables are included. These include the number of kids between the ages 6 and 15, the number of family members other than the head, Epouse, and kids ages 6-15, and the head's age. Kids ages 6-15 are expected to exert a positive influence on the probability of purchase since it is during this school age period that their presence necessitates larger housing space, expecially additional bed- rooms. Pre-school children are much easier to accomodate (e.g., by sharing of bedrooms with many other people), especially infants. An increase in the number of other family members is also likely to have a positive effect on the probability of purchase, albeit, to a lesser degree. The effect of the head '8 age, over and above its influence on income, is expected to be positive, however, its influence is not expected to be as great as that found in other studies because of the inclusion of normal income.

Fourth, in the face of the large differences in current homeovnership patterns acrc3s site-j noted In section 2, durrmy variables are included for Paterson-Passaic , Jersey City, and Scranton. A p r 1 o - i j ;. seems that site along with ethnitity 1 s a legitimate criterion to consider for depooling. Unfortunately, small sample sizes raise problems, ^nd hence only one case can be considered, namely whites in Scranton.

-10-

Fifth, mortgage rate and calendar time are included as repressors. The mortgage rate series used is th< FHLBB effective rite on existing homes which reflects fees and charges as well as contract rates, and assume.", pre- payment at the end of ten years. The data were taken from Federal Reserve Bulletins (December, 1968 thru January, 1973), and the actual rate used wa9 the average of the two middle months in the year in which the family moved. The rationale for the inclusion of the mortgage rate is as an indicator of the cost of buying a home, as well as an indicator of supply side effects. Calendar time is used to capture trends in homebuying and to take into account differences in market conditions facing families moving at different times. The actual values used for mortgage rates and calendar time are given in Tables A7 .and A8.

With regard to treatment variables the following regressors are used: an experimental dummy (equalling one for an experimental family) , yearly experimental payments (measured in thousands of dollars), experimental time (equalling the midyear points .5, 1.5, and 2.5 respectively, for each of the three years), and two experimental interactions, one with the head' 8 age and one with work-conditioned unearned income. The use of experi- mental payments provides a simple parsimonius representation of the treat- ment which differentiates not only between experimentals and controls, but also between experimentals receiving payments and those who are over break- even. Because of the often fruitless results that have been encountered with explicit tax ar.d guarantee representations, as well as their failure to identify experimentals not receiving payments, the payments approach has been adopted.

Experimental time is included in order to determine whether any po ;siMe experiment effect may tend to occur at say the end of the experiment as

-11-

Wooidridge [9] found. Kb ni1:: Cloned earlier the exper imrntal interaction with work-Conditioned unearned inco e lfl Intended to capture, any txperi- mental effects which may arise through welfare status. The exper iniertal interaction with the heal s a^e permits experimental families that are farther along in their life cycle cf consumption (and possibly having additional assets) to react differently than families with younger heads and which may Just be starting out. Indeed the youthfulness of the sample indicates that many families will be Just entering into the home buying age bracket, and hence experimental-age interactions are possible.

5. Empirical Results

In light of the ethnic differences noted earlier, the decision to estimate each ethnicity separately was first tested. A pooled model with the eighteen variables described in section A was estimated and yielded a log-likelihood value of -132.5. Then separate models were esti- mated for each ethnicity yielding the log-liklihood values given in Table 2. The Scranton dummy was omitted f-ora the black and Spanish-speaking models because there were no families in Scranton. Furthermore, the prior owner dummy was omitted from the Spanish-speaking model because neither of the two prior owners "in the sample bought a home — implying that its coefficient cannot be estimated (see Poirier [6]). The explanatory power of the additional thirty-one variables in the black and Spanish-speaking models was then tested by computing -2 times the increase in the log-likeli- hood. This yielded a test siatistic of >. ■ 66.60 which is significant at

the one per cent level. Thus the decision to depool the ethnicities appears

13

to be consistent with the sample information.

Tabls I Problt C Lclcntl with Standard Erros in Parentheses

Coeff icl«-r.:

Variable

3, B.

8i»

Whites ^L

Black.fi Spanish-apeikers

Cocb:

Normal incone Work-conditioned income Prior owner dtll Public housing dune Nunber of kids n^ei 6-15 Fanlly aire - 2 - kida 6-15 Head's age

PaterBon-pRBflai d duaay Jersey City dusny Scranton dunny Calender tise Mortgage rate

1.420

.1247* (.07002)

-.853i** ;. 3750)

.3974 (.6611)

-.1870 (.4528)

.2375 (.1483)

.2712 (.1935)

-.09260*** (.03370)

-.4319 (.9066)

-1.676 (1.259)

-.5992 (1.079)

-.03378 (.04144)

.2252 (.7609)

-21.76*** (6.873)

.2640***

(.08682)

-.3637 (.26b9)

.4688 (.7269)

-.050?6 .4123

.1784 (.1088)

-.1658 (.1345)

.08018** .03952

-1.569** (.6966)

-2.121*** (.8220)

.1149*** .043.^1

1.930*** (.7396)

-4.611 (15.08)

.5606** (.2341)

-.2412 (.2763)

1.041 (1.223)

.1506 (.2697)

.1372 (.3320)

.03424 (.06632)

-2.568 (2.184)

-1.506 (2.070)

.00743.1 (.06034)

.007213 (1.853)

i||

•A ill

s

i

17

. I

Experimental dusmy Experimental paynents

Expert* ital 1 1

(Experi Vread'i age)

(Experl-'.T.t. ' . .--

-4.867** (2.004)

.1483 (.1873)

.230 b (.4600)

.0986/** (.04*34)

.5121

•(Work-condll tac) I ( . I

4. 1 72** 2.014

-.1034

(.1810)

-1.332*** (.5189)

-.04768 (.04691)

.2873 '.. 3220)

-11.94 (8.216)

-2.161* (1.233)

.1666

(.8636)

.1062 (.1874)

-3.256 (2.541)

*Througbou: this study "**• " "**," and "*" will denote significance at th« one, five, and ten 3<:rc< . lis, respectively.

-13-

Table 2 Probit Statistics

Statis	tic	Description 1 i ______			Whites	Blacks	Spanish-apea'-
n		Sample size			118	94	61
■		Number of buyers			27	40	12
k-1		Degrees of freedom			17	16	15
L		Log-likelihood			-42.41	-42.37	-14.45
X		-2* (log-likelihood	ra	tio)	42.10***	43.48***	31.58***

Table 3 Likelihood Ratio Tests

Test
rer	Description		Whiter	Blacks	Spanish-speakers
1	Work-conditioned income <B3 " £18 = °>		■ 9. 394***	1.935	6.893**
2	Prior tenure (6^ - B$ - 0)		.5770	.4862	- - -
3	Family characteristics (S6 ■= 67 -6,-0)		13.41***	11.56***	2.539
4	Head's age (2e - S:? - 0)		10.33***	5.901*	8.446**
5	Site (8 - B - 6 )		2.840	7.870**	4.233
6	Calender time and mortgage rate (Bla = Sn - 0)	1.783	11.83***	.02180
7	Experimental variable (S ■ fl -...■ fl V U Ml 5 18	0)	9.232	10.85*	9.827*
8	Experimental interactions (3 - 6 - 0) 1 7 Ml 8 '	7.734**	1.69	6.758**

-14- Th* probit coefficients for each of the three ethnic models .ire presented Ln Table 1. Standard Statistics srs given in Table 2, and the liklihooa ratio tests for various groups of independent variables are given in Table 3. Considering these tables the following remarks seem in order.

With regard to e.irnln^s, as expected, normal income is significant and positive in all three models, albeit, to a lesser degree in t he case of whites. Work-condit ioacd unearned income is consistently negative (and significant in the case of whites), most likely reflecting the fact that families with sizeable welfare payments have few assets and are probably unlikely to be able to get a mortgage. Furthermore, there is little evidence that the response for experimentals to work-conditioned unearned income differs from that of controls, although in the case of whites and blacks it tends to lessen the previously mentioned negative response. For Spanish-speakers the experimental response reinforces the negative response and the joint effect is significant (see Test 1 in Table 3).

Prior tenure status has virtually no effect on the home buying decision for any of the ethnicities (see Test 2 in Table 3). This result is some- what surprising since besides the rationale for its inclusion presented in section £, prior tenure was thought to be an excellent proxy for assets. As Kain and Quigley [2, p. 269] note, "For most households , black and white, equity in owner-occupied housing is itself the largest component of net worth."

Family characteristics are significant for whites and blacks, but not for Spanish-speakers (see Test 3 in Table 3) . As expected the number of kids ages 6-15 has a consistent positive effect although it is not

-15-

qulte significant at the 101 level. The same holds for family size except that its effect is negative for blocks. The coefficient of the head's age has the expected sign for blacks and Spanish-speakers (significant for blacks), houevcr, it is negative and significant for whites. The "explanation" for this seems to lie in the youthfulness of the controls who bought (see Table A). Interestingly, the experimental response inter- acts positively (and significantly) with the head's age for whites and in effect wipes out the significant negative effect for controls. For all three models the Joint effect of the head's age is significant (see Test 4 in Table 3).

Surprisingly, site is significant only for blacks, (see Test 5 in Table 3) , although Trenton (the omitted site) consistently fares better in all three models. As mentioned in section 2, the historic differences in home- ownership rates across sites do not seem indicative of the housing markets during the experiment.

Calendar time and the mortgage rate are significant determinants of the home buying decision only in the case of blacks (see Test 6 in Table 3) . The consistently positive coefficient for the mortgage rate seems to indi- cate that what is being measured is not effect of a housing "price" but rather other effects which are correlated with movements in interest rates.

Of course the independent variables of primary interest are the experi- mental variables. The Joint teat on all experimental variables (Test 7 in Table 3) indicates that there is a slight experimental effect for all ethnicities (the test for whites Just misses being significant at the 10 per cent level). However, the nature of this response differs markedly across ethnicities. The weakest response is for whites and the experimental- age interaction accounts for a great deal of it. While experimental payments

-16-

Table 4 Head's Average

Group		Whites	Blai	Span 1 sh-speakera
Controls		40.33	35.89		35.47
Buyers		33.05	39.82		39.82
Non-buyers		41.89	33.49		34.02
Experlmentals		35.30	35.02		37.63
Buyers		36.38	36.46		42.48
Non-buyers		34.91	33.87		36.69

Table 5 Suaunary of Predicted Probabilities

Whites Blacks J !

Spanish-speakers

Average for non-buyers

Average for buvers

Using Independent variable ireans

1517 4806

1362

.2638 .6521

.1966

.09451 .6117

.2007

-17-

are not significant, their positive sign, together with the negative and statistically significant dummy coefficient, indicates that experimentals near their breakeven point have a negative response. In fact for a mean- aged white experimental midway through the experiment who is receiving no work-conditioned unearned income, the estimated experimental response does not become positive until payments reach $7000.

In the case of blacks the experimental response is significant and positive for those receiving little or not payments. Specifically, for a mean-aged black experimental midway through the experiment who is not re- ceiving work-conditioned unearned income, the estimated response is positive up to a payments level of $4877. This somewhat strange response — posi- tive for those not- receiving payments — is similar to that found by Wooldridge [9]. However, the negative and experimental coefficient of experimental time indicates that the greatest response occurred at the be- ginning of the experiment. This is consistent with the argument that the experiment provided security for over-breakeven experimentals to buy a home early in the year. As pointed out earlier, Wooldridge [9], somewhat confusingly, found this security effect setting in at the end of the experiment.

The experimental response for Spanish-speakers is different from that of both whites and blacks. The coefficient of the experimental dummy Is negative, and the payments coefficient is not only negative and significant, but its absolute value is much larger than that for either whites or blacks. The implication is that experimentals not receiving payments had a negative response and this response became more negative as payments increased. The evidence clearly indicates that a negative response was present for

-18-

Spanish- speakers.

Looking deeper, a few more inrights into the experimental responses can be found. In the case of blacks and Spanish-speakers (both of which had negative responses to payments), the payments coefficient is signifi- cantly different (at the 5 per cent level) from the normal income coeffi- cient. In the case of whites (for which the payments coefficient is posi- tive), it is virtually the same as the normal income coefficient.

Considering Table 5, the ratio of average predicted probabilities of buyers to non-buyers is greatest for Spanish-speakers (over 6 to 1) and less for whites (about 3 to 1) and blacks (about 2h to 1). Considering Table 6, the average predicted probabilities among buyers are nearly identi- cal for experimentals and controls in all three models. However, among non- buyers the experimental-control averages do differ somewhat. Further break- ing down these averages by site results in rather eradic patterns.

Finally, two subgroups of special interest were further analyzed and they yielded the results presented in Tables A9 - A12. The first subgroup was formed by deleting those black families who bought the house they were renting. This amounted to a comparatively large eleven families (versus four for whites, and one for Spanish-speakers). The major differences in the results from subgroup (mover) model from the model considered earlier are that site, calendar time, and the mortgage rate are no longer signifi- cant. The explanation for this is somewhat elusive, but it may point out a data artifact in these observations. The experimental response is basically the same as noted for the black model in this section except that experi- mental time is no longer significant and the age-experimental interaction now is.

The second subgroup consists of only those whites in Scranton. This is the only subgroup large enough to permit depooling by both site and ethnicity.

-19-

Table 6 Breakdown of Average Predicted Probabilities

Non-buyera

ControlH I Experimental*

Buyers

Coatroln I Hxpcrimentals

Whites Trenton scan nuober

Paterson-Passaic swan

nuober

Jersey City

1122

nuober

Scranton aeir.

nunber

2113 3

1046 39

.1856

.3817 2

.4006 8

.1505 4

,1293 35

.4730

.5529 1

.0461 1

,5226 7

.4844

.6194 2

.5491 6

.4185 10

Black*

,2289

,2812

,6406

.6564

Trenton
moan	.3358	.3627	.5943	.6574
nuober	5	4	5	3
Paterson-Pajsaic
•ean	.2811	.2265	.9085	.6646
number	2	'1	2	8
Jersey City
nean	.1708	.2943	.5646	.6526
nuober	11	21	4	18
Spanish-speakers	.1252	.07670	.6116	.6119
Trenton
mean		.0662		.5659
nuaber	0	6	0	1
Pateraoa-Pasaaic
swan	.0814	.0763	.3140	.6919
nuaber	5	14	1	3 •
Jersey City
■ean	.1421	.0828	.6711	.5149
nuabsr	13	11	5	2

-20-

A* noted earlier, historical housing patterns might suggest strong differ- ences across sites. However, as Tt ;t 5 in Table 3 suggests, there is little difference between the Scranton-only model and the white model considered earlier. The main difference is that the experimental response is slightly stronger and the payment coefficient is now significant at the 10 per cent level. The nature of this response remains the same.

6. Conclusion

The results of this study are markedly different from those found in the studies of Wooldridge [9] and Nicholson [5], The experimental response appears fairly weak, and its sign varies substantially across ethnicities. Only in the case of blacks does there appear to be evidence of a possible experimental effect among experimentals receiving small or zero amounts of payments (e.g., those over-breakeven). Unlike black experimental responses that have been found in other areas of analysis, this result does not appear to reflect a poor performance on the part of black controls. Even after excluding the five black controls v'io bought their own homes, the home buying rate among black controls was .2500, which is substantially larger than the rates for white or Spanish-speaking controls. While the exact explanation for these divergent results between studies is not certain, a number of possible candidates exist.

First, it seems that part of the explanation must lie in differences in the various samples used. Both Wooldridge' s and Nicholson's results most likely contain at least some experimental mobility effect. For the sample used here, a mobility effect appears present only for blacks. The proportion of experimentals for whites, blacks, and Spanish-speakers are .5678, .6915, and .6066, respectively, compared to the corresponding

-21-

proportions .5839, .6453, and .6242 in the parent continuous 693 sample. It seems that Nicholson in some sei.je captures the spirit of the sample selection procedure used here, by using a pre-enrollment homeownership dummy. Since homeowners are less inclined to move, this dummy helps to distinguish the effects of the mover part of the sample. Also along the lines of sample selection, the ethnic pooling used by both Wooldridge and Nicholson in the face of the repeatedly large ethnic differences found in other analyses, notably in the area of consumption (see Metcalf and Nicholson [3]), must be an important factor.

Second the estimation techniques vary across studies Nicholson uses the ordinary least squares linear probability model exclusively, and Wooldridge uses both probit analysis and the former. This author's choice to use probit analysis exclusively is based on the well-known inappropriate- ness of the linear probability model for a model with a binary dependent variable. One important distinction between these two models is that the probit model is interactive, whereas the linear probability model is addi-

A Si

tive. For the probit model, ,]_ f, a\n » which clearly depends on the

dxji j * levels of all Independent variables.

Third, all three studies suffer from some methodological problems.

Wooldridge and Nicholson analyze their inherently panel data problems

by running separate cross sections at different points In time. While

this procedure is inefficient, it must be admitted that the use of probit

analysis on panel data is a difficult question. However, if one does not

object to using the linear probability model, then all the standard panel

data techniques are available. As noted earlier, the sample selection

utilized here eliminates (for the most part) this panel data problem.

-22-

On the other hand, tnr ethnic depooiing used here has of course re- sulted in some rather small sample sizes. Whether these sample sizes are large enough to justify the use of large sample maximum likelihood proper- ties is unclear.

Fourth and finally, some of the poorest data in the experiment i3 the housing and housing debt drta. This has also been noted by both Wooldridge [9] and Nicholson [5]. The problems that all authors necessarily face in piecing together often contradicting information must be expected to intro- duce unintentional- data differences — the effects of which are unknown.

-23-

6 . Appendix

Tat e Al White Sample

			Paterson-	Jersey
		Trenton	Paasaic	City	Scranton	Total
First year		2	9	2	27	40
Experimentals	(I)	2(1.0)	8(.89)	K.50)	18(.67)	29(.73)
Second year		1	3	2	23	29
Experimentals	(X)	1(1.0)	3(1.0)	0(0.0)	12 (.5 2)	16(.55)
Third year		1	3	4	41	49
Experimentals	(Z)	1(1.0)	3(1.0)	3(.75)	15(.37)	22 (.45)
Total		4	15	8	91	118
Experimentals	(%)	4(1.0)	14(.93)	4(.50)	45 (.49)	67(.57)

Table A2 Hone-Buying Among Whites

		F terson-	Jersey
	Trenton	Passaic	City	Scranton	Total
First year	0	5	1	6	12
Experimentals ("/.)	0(0.0)	4(.80)	0(0.0)	6(1.0)	8(.89)
Second year	1	1	0	6	8
Experimentas (%)	1(1.0)	1(1.0)	0(0.0)	3(.15)	4(.67)
Third year	1	1	0	5	7
Experimentals (%)	1(1.0)	1(1.0)	0(0.0)	K.17)	3(.38)
Total	2	7	1	17	27
Experimentals (Z)	2(1.0)	6 (.86)	0(0.0)	10 (.59)	18 (.67)

-24-

Table A7 Mortgage Rates
Paterson- Passaic	Jersey City

Year

Trenton

Scranton

1	7.49	8.05
2	8.385	8.415
3	7.60	7.67

8.265

8.08

7.64

8.43 7.70 7.44

Table A8 Calender Time

		Paterson-	Jersey
Year	Trenton	Passaic	City	Scranton
1	7	12	16.5	19
2	19.5	24.5	29	30.5
3	31	35.5	40	43

-25-

		Tabli	2 A3
		Black j	) itr.ple
			Paters on-	Jersey
		Trenton	Passalc	City	Total
First year		8	8	28	44
Experimentals	(X)	5(.63)	7(.88)	19(.68)	31 (.70)
Second year		4	6	11	22
Experimental	C-)	2 (.50)	4(.67)	8(.73)	14(.64)
Third year	.	2	9	15	28
Experimentals		0(0.0)	8(.89)	12 (.80)	20(.71)
Total		17	23	54	94
Experimentals	(30	7(.41)	19 (.83)	39 (.72)	65(.69)

Table A4 Home-Buying Among Blacks

Paterson- Jersey Trenton Passaic City

Total

First year

Experimentals ("0

Second year

Experimentals (2)

Third year

Experimentals (X)

3 3

2(.67) 3(1.0)

2 5 K.50) 3(.60)

3 2 0(0.0) 2(1.0)

10	16
9(.90)	14(.88)
10	17
7(.70)	1K.65)
2	7
2(1.0)	4(.57)
22	40
18(.82)	29(. 73)

Total

Experimental* (7.)

8 10

3(.38) 8(.80)

-26-

Tablc A5 Spanish-Spt. diking Sample

			Paterson-	Jersey
		Trenton	Passai c	City	Total
First year		2	10	10	22
Experimentals	cz)	2(1.0)	6(.60)	3(.30)	1K.50)
Second year		3	4	7	14
Experiment i] s	(Z)	3(1.0)	4(1. 0)	2(.28)	9(.64)
Third year	-	2	9	14	25
Experimentals	(Z)	2(1.0)	7(.78)	8(.57)	17 (.6 8)
Total		7	23	31	61
Experimentals	(Z)	7(1.0)	17(.74)	13(.42)	37(.61)

Table A6 Home-Buying Among Spanish-Speaking

			Paterson-	Jersey
		Trenton	Passaic	City	Total
First year		1	1	2	4
Experimentals	{%)	1(1.0)	1(1.0)	0(0.0)	2(.50)
Second year		0	0	3	3
Experiment .	(Z)	0(0.0)	0(0.0)	K.33)	K.33)
Thi rd year		0	3	2	5
Experimentals		0(0.0)	2C.67)	K.50)	3(.60)
Total		1	4	7	12
Experiment.	(Z)	.0)	3(.75)	2(.28)	6(.50)

-27-

Table A 7 Mortgage Rates

	r '	r— erson-	~i -	•
Year	Trout	Lc		•	Scranton
1	7.49	8.05	-	8.265	8.43
2	8. -	8.415		8.08	7.70
3	7.60	7.67		7.64	7.44

Table A8 Calender Time

Year

Trenton

racerson- Passaic

Jersey City

Scranton

1	7	12	16.5	19
2	19.5	24.5	29	30.5
3	31	35.5	40	43

-2 -

Table A9 Problt Coefficients With Standard Errors In Parcnth«e«a

Coefficient

Variable

Black

(movers only)

Whites (Scranton onlv)

A

J,

s.

Constant

Normal Income

Work-conditioned Income

Prior ovr.ei dummy

Public housing dummy

Number of kld9 ages 6-15

Family size - 2 - kids 6-15

Head's age

Paterson-Paaeaic dummy

Jersey City dummy

Calender time

Mortgage rate

Experimental dummy

Experimental payments

Experimental time

(Experimental dummy) ♦(Head's ai;e)

(Experimental d'inny)

*(Work-concii t ionod income;

■18.28** (7.174)

-.6425 (.«05?)

-1.245 (.9359)

.07206 (.05310)

1.311 (.7987)

5.944** (2.439)

-.2941 (.2462)

-.9187 (.6553)

-.09630* (.05539)

.2041 (.3833)

30.93 (23.53)

.3768***	.2335*
(.1095)	(.1213)
-.2600	-.8594
(.3356)	(.6591)
.6489	.1091
(.7293)	(.7964)
-.006084	.3324
(.4689)	(.6359)
.1906	-.05647
(.1318)	(.2260)
-.1334	.08571
(.1614)	(.3228)
.08875**	-.1759**
(.04518)	(.06977)

-.07112 (.09586)

-3.246 (2.777)

-6.121* (3.538)

.6292* (.3406)

-1.680 (1.353)

.2263*** (.08233)

-.4888 (1.285)

-29-

Table A10 Probit Statistics

Statistic

Descripi Lon

blacks Whites

(movers only) (Scranton only)

n

m

k-1

L

Sample size

Number o: rs

Degrees of freec

Log-likelihood

-2* (log-likellhooci ratio)

83	91
29	17
16	14
32.26	-20.95
42.90***	45.74***

Table All Likelihood Ratio Tests

Number

Description

Blacks Whites

(movers only) (Scranton only)

Work-conditioned income

ce, ■ 320 ■ o)

Prior tenure

(B. - S - 0)

Family characteristics (S6 ■ 37 - fc3 - 0)

Head's age

8 1 9

Site

Calender time and mortgage rate (B:2 ■ Bn ■ 0,

Exper; i variables

..- I., ■ 0)

«» ■

BxpexinentaJ Lnt tractions

.6446 .8536

9.273**

4.618*

2.112

3.754

11.58** 3.410

5.591*

3.494

15.80***

16.97***

1.677

17.20***

11.90***

-30-

Footnote9

*The author is an Assistant Professor of Economics at the University of Illinois at Urb ana-Champaign. Part of the research that went into this study was performed while the author was a Visiting Assistant Professor at the Institute for Research on Foverty in Madison, Wisconsin during the summer of 1974. He wishes to express his gratitude to Helen Lowry of the University of Illinois for her help in implementing the probit computer package used in this study. Thanks are also owed to Robert Avery, Joseph Hotz, and Harold Watts of the University of Wisconsin at Madison and Douglas Bendt and Judith Wooldridge of Mathematics Inc. for their thoughtful comments.

^Wooldridge [9, p. 38].

2

Further doubt is cast on any type of supply side explanation by the re- cent findings of Robert Avery. Avery conducted personal interviews with lend- ing institutions in all sites* and found little, if any, awareness on their

part of the experiment.

3 See for example Carliner [1], Kain and Quigby [2],Reid [7], and Morgan

[4].

As Metcalf and Nicholson [3, p. 5] and Reid [7, p. 11] have noted, the main proponent of the permanent income concept, Milton Friedman, has used three year income averages as proxies for permanent income implying that individuals may only have <\ three year time horizon.

See Kain and Quigley [2, p. 273].

In the models considered in this study this independence assumption may be slightly violated since a small protion of the samples appear more than once. Specifically, 16 white, 10 black, and 8 Spanish-speaking families appear more than once.

-31-

In Watts [8] the natural logarithm of normal Income is estimated.

Here normal income itself is used. Since In y - N(p,a2) implies E(y) -

1 I

exp(p+ X0 )i the "blowing-up" procedure took into account the estimated

standard deviation of a family's income from their normal income.

8See Watts [8].

9 Normal income is also serving as a proxy for assets. Unfortunately,

the financial asset series constructed by Metcalf and Nicholson [3] is only available at pre-enrollment and quarters two, six, and ten. Hence, for ex- ample, for a family itt Trenton who bought in the second year, it is not possible to determine whether their asset figure refers to before or after they bought. This is not viewed as a major shortcoming since as Nicholson [5, pp. 13-14] notes, the average family stock holdings of stocks, bonds, and savings accounts was only $140, and that of cash was $31. The major asset for these families is their house if they own one, and prior ownership is included as an inde- pendent variable.

Wooldridge [9] states that for her sample "18.8 and 24.8 per cent of moves were made because of poor conditions or condemned housing in respec- tively the last and penultimate moves of families (from the tenth quarterly interview) ."

Since the time period involved is a year, it is not even clear how to construct an 'over-breakeven* dummy except for families who were over- breakeven the entire year. Since even experimentals over-breakeven received fees for reporting their incomes (this -amounted to $260 a year for the family over-breakeven the entire year) over breakeven experimentals are also dif- ferentiated from controls.

-32-

12

Actually there were tVH bleck families in Scronton who moved. How- ever, i Lee in other studies, they have been omitted from the . La.

The three possible ethnic : i ie conb inat ions (whites and blacks, whites and Spanish-speakers . and blacks and Spanish-speakers) were also tested separately for depoolir.g. This yielded test statistics of 38.63, 33.75, and 28.30, respectively, which are significant at the .5, 1.0, and 2.5 percent levels. Furthermore, the Bonferroni joint testing procedure indicates that the simultaneous significance level for these three tests is at most 4.0

percent .

14

Furthermore the ethnic proportions of .4322, .3443, and .2234 for

whites, blacks, and Spanish-speakers, respectively, are close to the corres- ponding ethnic proportions of .4473, .3377, and .2150 in the parent continuous

-33-

Re ■ es

[1] Car liner, ' ■ tS of Home Ownership." Madison:

Institute for Reseat ;. on Poverty D. Lon Paper No. 169-73, 1973.

[2] Rain, tahn P., - . gley. "':. et ijiscr initiation, Bomeovnership , and • Economic Review.

Vol. l.Xil (Jun< , 21 3-77.

[3] Netcalf, ^ E.. -icholson. "Low-Income House-

holds and t . -icome Hypothes Ls: : up 11 cat ions of Che

Urban Experinec . published manuscript presented at a joint

meeting of I ri □ Economic Association and the Economic Society, New York, 1973.

[4] Morgan, James N. "Housing and Ability to Pay." Econometrica . Vol. XXXIII (April, 1965), 289-306.

[5 J Nicholson, Walter. "Expenditure Patterns in the Graduated Work

Incentive Experiment: A Descriptive Survey." Final Report of the New Jersey Graduated kork Incentive Experiment- Madison: Institute for Research en Poverty, 1973, DIIIb-1 to DIIIb-51.

[6] Poirier, Dale J. "A Note onf Perfect Classifications' in Binary Dependent Variable Models." Unpublished manuscript, 1974.

[7] Reid, Margaret 6. mousing and income. Chicago: University of Chicago Press, 1962.

[8] Watts, Harold W. "The Estimation of Normal Family Income." Final

Report of the New Jersey Graduated Work Incentive Experiement . Madison: Institute for Research or. Poverty, 1973, 31-61 to BI-93.

[9] Wooldridge, Judith. "Housing Consumption in the New Jersey- Pennsylvania Experiment." Pinal Rdpo-t of the New Jersey Graduate Work Incentive Expc-.' ant. Madison: Institute for Research on Poverty ,