Upload
others
View
7
Download
0
Embed Size (px)
Citation preview
Cohabitation vs marriage: Marriage matching with peereffects
Ismael Mourifie and Aloysius Siow
University of Toronto
December 2014, Penn State
Motivation
Trends on US marital behavior since the seventies:The marriage rate has decreased.
The cohabitation rate has increased.
The unmatched rate has increased.
Some evidence on increased positive assortative matching in marriage byeducation (PAM).
Increases in earnings inequality, PAM and married women labor supplyled to increases in family earnings inequality. (Burtless (1999);Greenwood, Jeremy, Nezih Guner, Georgi Kocharkov, and Cezar Santos(2014); Carbone and Cahn (2014)).
Motivation
Figure : Marital Status by Gender and Year.
0.2
.4.6
.81
Fra
ctio
n
1990 2000 2010
Female Male Female Male Female Male
Marital Status by Gender and Year
married cohabsingle
Motivation
Researchers have investigated different causes for these changesincluding
Changes in reproductive technologies.
Changes in access to reproductive technologies.
Changes in family laws.
Changes in earnings inequality and changes in welfare regimes.
E.g. Burtless 1999; Choo Siow 2006a; Fernandez, Guner and Knowles 2005;Fernandez-Villaverde, et. al. 2014; Goldin and Katz 2002; Greenwood, et. al.(2012, 2014) etc...
Most of this research ignored changes in population supplies over time.Often, they also ignore peer effects in marital behavior.
Motivation
Figure : Fraction of individual by gender, education and year.
0.2
.4.6
.8
1990 2000 2010
L M H L M H L M H
Fra
ctio
nFraction of individuals by gender, education and year
male female
Motivation
This change in the sex ratio may have exacerbated the decline in themarriage rate.
It may also potentially changed marriage matching patterns.
Peer effects may have also affected cohabitation and other maritalbehavior. E.g. Waite, et. al. 2000; Fernandez-Villaverde, et. al. (2014))
Adamopoulou (2012) shows that there are positive peer effects on themarital decisions of young adults.
Main objective
The objective of this paper is to first provide an elementary marriagematching function (MMF),
which can be used to parameterize different causes of changes in maritalbehavior.
while allowing for peer effects and changes in population supplies.
Second, we estimate the changes in the parameters of the MMF whichcapture the evolution of marital behavior in the US between 1990 and2010
What we do
We propose a new static empirical marriage matching function (MMF):the Cobb Douglas MMF
The Cobb Douglas MMF encompasses:Choo-Siow frictionless transferable utility MMF (CS)
The CS with frictional transfers.
The Dagsvik Menziel non-transferable utility MMF.
The Chiappori, Salanie and Weiss MMF (CSW).
The marriage matching with peer effects (CSPE).
Properties of this Cobb Douglas MMF are presented.
Existence and uniqueness proof of the marriage distribution areprovided.
Comparative statistics are derived.
What we do(1)
Identification and estimation are presented.
Empirical ApplicationThe CD MMF is estimated on US marriage and cohabitation data by statesfrom 1990 to 2010.
CS with peer effects is not rejected.
There are peer and scale effects in the US marriage markets.
Positive assortative matching in marriage and cohabitation by educationalattainment are relatively stable from 1990 to 2010.
The CS Benchmark
Consider a society with I, i = 1, .., I, types of men and J, j = 1, .. J, typesof women.
Let m and f be the population vectors of men and women respectively.
Let ΞΈ be a vector of parameters.
A marriage matching function (MMF) is an IΓ J matrix valued functionΒ΅(m, f ; ΞΈ) whose typical element is Β΅ij, the number type i men married totype j women.
The CS Benchmark(1)
Building on the seminal papers of Becker (1973, 1974), Choo and Siow(2006; CS) developed a static frictionless transferable utility.
Let the utility a male g of type i will get from marriage with a woman oftype j, j = 0, .., J be:
Uijg = uij β Οij + Ξ΅ijg (1)
uij is the systematic gross return to a male of type i marrying a female oftype j.
Οij is the transfer made by the man to his wife of type j.
uij = uij β Οij a systematic net component common to all (i, j) matches
Ξ΅ijg is an idiosyncratic component which is man specific.
Ξ΅ijg i.i.d. random variable distributed according to the Gumbel distribution.
Similarly, the utility which a woman k of type j who chooses to marry atype i man, i = 0, .., I, is:
Vijk = vij + Οij + Ξ΅ijk. (2)
The CS Benchmark(2)
Now, let extend the CS framework by assuming that we have multipletype of relationships r (ex: Cohabitation vs marriage)
Let the utility of male g of type i who matches a female of type j in arelationship r be:
Urijg = ur
ij β Οrij + Ξ΅r
ijg
Let a man who remains unmatched matched a type 0 woman andΟr
i0 = 0.
Individual g will choose according to:
Uig = maxj,r{Ui0g, Ua
i1g, ..., Uaijg, ..., Ua
iJg, Ubi0g, ..., Ub
ijg, ..., UbiJg}.
Let (Β΅rij)
d be the number of (r, i, j) matches demanded by i,
(Β΅i0)d be the number of unmatched i type men.
The CS Benchmark(3)
Following the well known McFadden result, we have:
(Β΅rij)
d
mi= P(Ur
ijg βUrβ²ikg β₯ 0, k = 1, ..., J; rβ² = a, b),
where mi denotes the number of men of type i.From above, we obtain a quasi-demand equation by type i men for(r, i, j) relationships.
ln(Β΅r
ij)d
(Β΅i0)d = urij β ui0 β Οr
ij, (3)
The CS Benchmark(4)
The quasi-supply equation of type j women for (r, i, j) relationships isgiven by:
ln(Β΅r
ij)s
(Β΅0j)s = vrij β v0j + Οr
ij. (4)
The matching market clears when, given equilibrium transfers Οrij, we
have for all (r, i, j):(Β΅r
ij)d = (Β΅r
ij)s = Β΅r
ij. (5)
Substituting (5) into equations (3) and (4) we get the following MMF:
lnΒ΅r
ijβ
Β΅i0Β΅0j= Οr
ij β (r, i, j) (6)
where Οrij = ur
ij β ui0 + vrij β v0j. (called the the gains to marriage)
The CS Benchmark(4)
An equilibrium of the CS matching model can be defined as a vector ofsingle-hood Β΅ β‘ (Β΅10, ..., Β΅I0, Β΅01, ..., Β΅0J)
β² that verifies:The CS MMF.
Β΅rij = Β΅1/2
i0 Β΅1/20j eΟr
ij for r β {a, b}. (7)
The population constraints.
J
βj=1
Β΅aij +
J
βj=1
Β΅bij + Β΅i0 = mi, 1 β€ i β€ I (8)
I
βi=1
Β΅aij +
I
βi=1
Β΅bij + Β΅0j = fj, 1 β€ j β€ J (9)
Β΅0j, Β΅i0 > 0, 1 β€ j β€ J, 1 β€ i β€ I.
The CS Benchmark(5)
In others terms, an equilibrium of the CS matching model is the solutionof the following quadratic system of equations:
J
βj=1
Ξ²iΞ²I+j(eΟa
ij + eΟbij ) + Ξ²2
i = mi, 1 β€ i β€ I
I
βi=1
Ξ²iΞ²I+j(eΟa
ij + eΟbij ) + Ξ²2
I+j = fj, 1 β€ j β€ J
Ξ²i, Ξ²I+j > 0, 1 β€ j β€ J, 1 β€ i β€ I.
where Ξ²i =β
Β΅i0 and Ξ²I+j =β
Β΅0j
The CS Benchmark(6)
Using a variational approach Decker et al (2013) showed that for anyadmissible (m, f ; ΞΈ), Β΅ exists and is unique.
Galichon and Salanie (2013) showed also the existence and theuniqueness using an alternative approach.
Properties of the CS MMF
The CS is just identified.
The CS MMF fits any observed marriage distribution.
The log ratio of the number of (M, i, j) relationships relative to thenumber of (C, i, j) relationships is independent of the sex ratio, mi/fj
lnΒ΅Mij
Β΅Cij=
ΟMij β ΟCij2
Independence is a very strong assumption and unlikely to hold every twotypes of relationships.
Arcidiacono et al (2012, ABM) shows that it does not hold for sexual versusnon-sexual boy girl relationships in high schools.
CS obeys constant returns to scale (CRS) in population vectors (Graham,2013)
Survey on the related literatureWe can distinguish two main directions in the subsequent literature to the CSMMF.
1 One branch relaxes the specification of the idiosyncratic component Ξ΅rijg.
The net systematic gains from matching, urij = ur
ij β Οrij and vr
ij = vrij β Οr
ij,from CS is retained.
The transfer is still used to clear the marriage market.
Chiappori, Salanie and Wiess (2012; CSW) allows the variance of Ξ΅rijg to differ
by gender and type and obtains
lnΒ΅r
ij
(Β΅i0)1βΞ»ij (Β΅0j)
Ξ»ij=
Οrij
Οi + Ξ£jβ (r, i, j) (10)
where Ξ»ij =Ξ£j
Οi+Ξ£j.
Graham (2013) provides a wide set of comparative statistics for a specialcase of CSW i.e. Ξ»ij = Ξ».
Galichon and Salanie (2012, GS) further generalizes the distributions ofidiosyncratic utilities.
Chiappori and Salanie (2014) provides an excellent state of the art survey ofthe above and related models.
2 The second branch studies other behavioral specifications for the netsystematic return from marriage, ur
ij and vrij
Survey on the related literature (1)
We can distinguish two main directions in the subsequent literature to the CSMMF.
1 One branch relaxes the specification of the idiosyncratic component Ξ΅rijg.
2 The second branch studies other behavioral specifications for the netsystematic return from marriage, ur
ij and vrij
Dagsvik (2000) assumes that transfers are not available to clear the marriagemarket i.e. Οr
ij = 0
He assumes that the idiosyncratic payoff of man g of type i marryingwoman k of type j, Ξ΅r
ijgk, is distributed i.i.d. Gumbel.
He uses the deferred acceptance algorithm to solve for a matchingequilibrium and obtain the following non-transferable utility MMF for largemarriage markets:
lnΒ΅r
ij
Β΅i0Β΅0j= Οij β (r, i, j) (11)
Menziel (2012) shows that (11) obtains under less restrictive assumptionsabout the distribution of Ξ΅r
ijgk. Call (11) the DM MMF.
Survey on the related literature (2)
We can distinguish two main directions in the subsequent literature to the CSMMF.
1 One branch relaxes the specification of the idiosyncratic component Ξ΅rijg.
2 The second branch studies other behavioral specifications for the netsystematic return from marriage, ur
ij and vrij
The DM MMF also fits any observed marriage distribution.
Unlike CS, it obeys increasing return to scale in population vectors.
When there is more than one type of relationship, the log odds of thenumbers of different types of relationships is independent of the sexratio, for all the model surveyed i.e. The CS, CSW and DM MMF.
Choo-Siow with peer effect (CSPE)
Let the utility of male g of type i who matches a female of type j in arelationship r be:
Urijg = ur
ij + Οri ln Β΅r
ij β Οrij + Ξ΅r
ijg, where (12)
urij + Οr
i ln Β΅rij: Systematic gross return to a male of type i matching to a
female of type j in relationship r.
Οri : Coefficient of peer effect for relationship r, 1 β₯ Οr
i β₯ 0.
Β΅rij: Equilibrium number of (r, i, j) relationships.
ui0 + Ο0i ln Β΅0
i0 is the systematic payoff that type i men get from remainingunmatched, 1 β₯ Ο0
i β₯ 0.
The above empirical model for multinomial choice with peer effects isstandard. See Brock and Durlauf (2001)
What is new is our application to two sided matching.
Choo-Siow with peer effect (CSPE)
Maleβs utilityUr
ijg = urij + Οr
i ln Β΅rij β Οr
ij + Ξ΅rijg
We allow the peer effect to differ by relationship.
There is no apriori reason to rank Ο0i versus Οr
i .
For example, unmarried individuals spend more time with their unmarriedfriends than married individuals with their married friends.
On the other hand, married individuals may want to live in communiteswith couples like themselves so that local firms will provide services to them(E.g. Compton and Pollak (2007); Costa and Kahn (2000)).
CSPE
Similarly, the utility of a woman, is: male g of type i who matches afemale of type j in a relationship r be:
Vrijk = vr
ij + Ξ¦ri ln Β΅r
ij + Οrij + Ξ΅r
ijk,
Using again McFadden result, we can write the men and womenquasi-demand and supply and clears the matching market using thetransfers. We get the following MMF
ln Β΅rij =
1β Ο0i
2β Οri βΞ¦r
iln Β΅i0 +
1β Ο0i
2β Οri βΞ¦r
iln Β΅0j +
Οrij
2β Οri βΞ¦r
i(13)
where Οrij = ur
ij β ui0 + vrij β v0j.
Call (13), the CSPE MMF
Some results on the CSPE
When there is no peer effect or all the peer effect coefficients are thesame (homogeneous peer effects),
Ο0i = Ξ¦0
j = Οri = Ξ¦r
j
we recover the CS MMF
No peer effect, or homogenous peer effects, generates observationallyequivalent MMFs.
If we cannot reject CS using marriage matching data alone, we alsocannot reject homogenous peer effects.
This lack of identification is our version of the reflection problem inManskiβs linear-in-mean peer effects mode.
Some results on the CSPE
When 1βΟ0i
2βΟriβΞ¦r
j6= 1
2 and/or1βΞ¦0
j2βΟr
iβΞ¦rj6= 1
2 , non-homogenous peer effectsare present.
meaning that non-homogenous peer effects are generically detectable.
This is related to identification of linear models with non-homogenouspeer effects. Please see Blume, et. al. (2014) and Djebbari, et. al. (2009).
When1β Ο0
i2β Οr
i βΞ¦rj=
1βΞ¦0j
2β Οri βΞ¦r
j= 1
we recover the DM MMF.E.g. Ο0
i = Ξ¦0j = 0 and Οr
i = Ξ¦rj =
12
Some results on the CSPE
WhenΟ0
i + Ξ¦0j = Οr
i + Ξ¦rj = Οrβ²
i + Ξ¦rβ²j ,
CSW MMF obtains.
The CSPE MMF relaxes the strong independence property imposed byall previous surveyed MMF.
lnΒ΅Mij
Β΅Cij=
(ΟMi + Ξ¦Mj β ΟCi βΞ¦Cj )
(2β ΟMi βΞ¦Mj )(2β ΟCi βΞ¦Cj )
[(1β Ο0
i ) ln Β΅i0 + (1βΞ¦0j ) ln Β΅0j
](14)
+ΟMij
2β ΟMi βΞ¦Mjβ
ΟCij
2β ΟCi βΞ¦Cj
mi/fj ββ ln Β΅0j; ln Β΅i0 ββ lnΒ΅MijΒ΅Cij
However because the coefficients on unmatched men and women havethe same sign, this independence is restricted.
Cobb Douglas MMF
So far all the models (i.e. CS, CSW, DM, CSPE MMF) can be written asfollows:
lnΒ΅r
ij
(Β΅i0)Ξ±r
ij (Β΅0j)Ξ²r
ij= Ξ³r
ij β (r, i, j) (15)
Ξ±rij, Ξ²r
ij β₯ 0
where Ξ³rij is proportional to the mean gross gains to relationship r minus
the sum of the mean gains to them remaining unmatched for tworandomly chosen (i, j) individuals
CD MMF, PAM
Let assume that the exponents on the Cobb Douglas MMF are genderand relationship specific but independent of the types of couples, (i, j):Ξ²r
ij = Ξ²r and Ξ±rij = Ξ±r.
Let the heterogeneity across males (females) be one dimensional andordered. Without loss of generality, let male (female) ability beincreasing in i (j).
l(r, i, j) = lnΒ΅r
ijΒ΅ri+1,j+1
Β΅ri+1,jΒ΅
ri,j+1
= Ξ³rij + Ξ³r
i+1,j+1 β Ξ³ri+1,j β Ξ³r
i,j+1
Whenever Ξ³rij = ur
ij + vrij is supermodular in i and j, then the local log
odds, l(r, i, j), are positive for all (i, j).
Statisticians use l(r, i, j) as a measure of stochastic positive assortativematching.
Thus even when peer effects or frictionS are present, we can test forsupermodularity of the marital output function, a cornerstone ofBeckerβs theory of PAM in marriage.
CD MMF, Models and Restrictions
Models and restrictions on Ξ±r and Ξ²r
Model Ξ±r Ξ²r Ξ³rij Restrictions
Cobb Douglas MMF Ξ±r Ξ²r Ξ³rij Ξ±r β₯ 0, Ξ²r β₯ 0
CS 12
12 Οr
ij Ξ±r = Ξ²r = 12
DM 1 1 Οrij Ξ±r = Ξ²r = 1
CSW ΟΟ+Ξ£
Ξ£Ο+Ξ£
ΟijΟ+Ξ£ Ξ±, Ξ² > 0; Ξ± + Ξ² = 1
CSPE 1βΟ0
2βΟrβΞ¦1βΞ¦0
2βΟrβΞ¦rΟr
ij2βΟrβΞ¦r Ξ±r, Ξ²r β₯ 0, Ξ±M
Ξ±C=
Ξ²M
Ξ²C
CD MMF, Existence and Uniqueness
Existence and Uniqueness of the Equilibrium matching
For every fixed matrix of relationship gains and coefficients Ξ²rij; Ξ±r
ij β₯ 0, theequilibrium matching of the Cobb Douglas MMF model exists and is unique.
CD MMF, Existence and Uniquenessβs proof
As shown previously, an equilibrium of the Cobb Douglas MMF is thesolution of the following system of equations:
mi = Β΅i0 +J
βj=1
¡αMiji0 ¡
Ξ²Mij0j eΞ³Mij +
J
βj=1
¡αCiji0 ¡
Ξ²Cij0j eΞ³Cij , for 1 β€ i β€ I, (16)
fj = Β΅0j +I
βi=1
¡αMiji0 ¡
Ξ²Mij0j eΞ³Mij +
I
βi=1
¡αCiji0 ¡
Ξ²Cij0j eΞ³Cij , for 1 β€ j β€ J. (17)
Could we write this system as a F.O.C of a variational problem? It seemsnot true in general.
The method of GS cannot be directly applied due to the presence of thepeer effect.
CD MMF, Existence and Uniquenessβs proof
However, let us propose an alternative strategy.
To ease the notation, denoteMβ‘ a and C β‘ b
let consider the following mapping g : (Rβ+)I+J β (Rβ+)
I+J
gi(Β΅; ΞΈ) = Β΅i0 +J
βj=1
¡αa
iji0 Β΅
Ξ²aij
0j eΞ³aij +
J
βj=1
¡αb
iji0 Β΅
Ξ²bij
0j eΞ³bij , for 1 β€ i β€ I, (18)
gj+I(Β΅; ΞΈ) = Β΅0j +I
βi=1
¡αa
iji0 Β΅
Ξ²aij
0j eΞ³aij +
I
βi=1
¡αb
iji0 Β΅
Ξ²bij
0j eΞ³bij , for 1 β€ j β€ J. (19)
CD MMF, Existence and Uniquenessβs proof
Hadamardβs Theorem (Krantz and Park (2003, Theorem 6.2.8 p 126))
Let M1 and M2 be smooth, connected N-dimensional manifolds and letf : M1 β M2 be a C1 function if
(1) f is proper,
(2) the Jacobian of f vanishes nowhere,
(3) M2 is simply connected,
then f is a homeomorphism.
CD MMF, Existence and Uniquenessβs proof
Let see if our mapping g verifies those conditions.
There are many ways to verify that g is proper.
A continuous function between topological spaces is called proper if theinverse images of compact subsets are compact.
An easy way is to use the following lemma
Lemma (Krantz and Park (2003, p 125))
Let U and V be connected open sets in RI+J, g: Uβ V is a proper mappingif and only if whenever {xj} β U satisfies xj β βU then g(xj)β βV.
Does the Jacobian of g vanish nowhere?
CD MMF, Existence and Uniquenessβs proof
Let write
¡αr
iji0 Β΅
Ξ²rij
0j eΞ³rij = eΞ±r
ijlnΒ΅i0+Ξ²rijlnΒ΅0j+Ξ³r
ij ,
β‘ eΞ΄rij .
Let Jg(Β΅) be the Jacobian of g.
After a simple derivation we can show that Jg(Β΅) takes the followingform:
Jg(Β΅) =
((Jg)11(Β΅) (Jg)12(Β΅)(Jg)21(Β΅) (Jg)22(Β΅)
)with
CD MMF, Existence and Uniquenessβs proof
(Jg)11(Β΅) =1 + βJ
j=1
[Ξ±a
1jΒ΅10
eΞ΄a1j +
Ξ±b1j
Β΅10eΞ΄b
1j]Β· Β· Β· 0
.... . .
...
0 Β· Β· Β· 1 + βJj=1
[Ξ±a
IjΒ΅I0
eΞ΄aIj +
Ξ±bIj
Β΅I0eΞ΄b
Ij]
(Jg)21(Β΅) =
Ξ±a
11Β΅10
eΞ΄a11 +
Ξ±b11
Β΅10eΞ΄b
11 Β· Β· Β· Ξ±aI1
Β΅I0eΞ΄a
I1 +Ξ±b
I1Β΅I0
eΞ΄bI1
.... . .
...Ξ±a
1JΒ΅10
eΞ΄a1J +
Ξ±b1J
Β΅10eΞ΄b
1J Β· Β· Β· Ξ±aIJ
Β΅I0eΞ΄a
IJ +Ξ±b
IJΒ΅I0
eΞ΄bIJ
,
CD MMF, Existence and Uniquenessβs proof
(Jg)12(Β΅) =
Ξ²a
11Β΅01
eΞ΄a11 +
Ξ²b11
Β΅01eΞ΄b
11 Β· Β· Β· Ξ²a1J
Β΅0JeΞ΄a
1J +Ξ²b
1JΒ΅0J
eΞ΄b1J
.... . .
...Ξ²a
I1Β΅01
eΞ΄aI1 +
Ξ²bI1
Β΅01eΞ΄b
I1 Β· Β· Β· Ξ²aIJ
Β΅0JeΞ΄a
IJ +Ξ²b
IJΒ΅0J
eΞ΄bIJ
,
(Jg)22(Β΅) =1 + βI
i=1
[Ξ²a
i1Β΅01
eΞ΄ai1 +
Ξ²bi1
Β΅01eΞ΄b
i1
]Β· Β· Β· 0
.... . .
...
0 Β· Β· Β· 1 + βIi=1
[Ξ²a
iJΒ΅0J
eΞ΄aiJ +
Ξ²biJ
Β΅0JeΞ΄b
iJ
].
.
CD MMF, Existence and Uniquenessβs proof
Let us denote every element of Jg(Β΅), Jk,l with 1 β€ k, l β€ I + J.
We can remark that |Jll| > βI+Jk 6=l |Jkl| for l = 1, ..., I + J.
So , Jg(Β΅) is a column diagonally dominant matrix or diagonallydominant in the sense of McKenzie (1960) far all Β΅ > 0.
Thus, Jg(Β΅), for all Β΅ > 0, is a non-singular matrix. (Proof see McKenzie(1960).
Then, our mapping g is an homeomorphism.
Therefore, the system of equation (18) admits a unique solution.
We can easily check that the solution is economically relevant in thesense that 0 < Β΅eq < (mβ², f β²)β².
Cobb Douglas MMF : Comparative statics
Unlike the Variational approach this methodology does not provide usintuitions about some comparative statics.
Let assume Ξ²rij = Ξ²r and Ξ±r
ij = Ξ±r
Following Graham (2013), let us propose a fixed point representation ofthe equilibrium of the Cobb Douglas MMF.
The Cobb Douglas MMF can be written as follows: for all (i, j) pairs:
Β΅rij
Β΅i0= exp
[Ξ³r
ij + (Ξ±r β 1) ln Β΅i0 + Ξ²r ln Β΅0j
]β‘ Ξ·r
ij for r β {a, b}, (20)
Β΅rij
Β΅0j= exp
[Ξ³r
ij + Ξ±r ln Β΅i0 + (Ξ²r β 1) ln Β΅0j
]β‘ ΞΆr
ij for r β {a, b}. (21)
Cobb Douglas MMF : Comparative statics (1)
Manipulating the population constraints (8), (9) we have the following:
Β΅i0 =mi
1 + βJj=1
[Ξ·a
ij + Ξ·bij
] β‘ Bi0, 1 β€ i β€ I (22)
Β΅0j =fj
1 + βIi=1
[ΞΆa
ij + ΞΆbij
] β‘ B0j, 1 β€ j β€ J. (23)
Let B(Β΅; m, f , ΞΈ) β‘ (B10(.), ..., BI0(.), B01(.), ..., B0J(.))β².
For a fixed ΞΈ we have shown that the (I + J) vector Β΅ of unmatched is asolution to (I + J) vector of implicit functions
Β΅β B(Β΅; m, f , ΞΈ) = 0. (24)
Now, let J(Β΅) = II+J β5Β΅B(Β΅; m, f , ΞΈ) with5Β΅B(Β΅; m, f , ΞΈ) =βB(Β΅;m,f ,ΞΈ)
βΒ΅β²
be the (I + J)Γ (I + J) Jacobian matrix associated with (24).
Cobb Douglas MMF : Comparative statics (2)
To derive the different comparative statics, we generalize Graham(2013)βs approach.
1 J(Β΅eq) = U(Β΅eq)H(Β΅eq)U(Β΅eq)β1 where U(Β΅eq) is a diagonal matrix andH(Β΅eq) a matrix with non-negative elements when Ξ²r β€ 1 and Ξ±r β€ 1
H(Β΅eq) is a row stochastic matrix (a matrix with non-negative elements where therows sum to one) whenever Ξ±r + Ξ²r = 1
2 H(Β΅eq) is row diagonally dominant if
1 max(Ξ²C β Ξ±C , Ξ²M β Ξ±M) < miniβI
( 1βΟmi
Οmi
);
2 min(Ξ²C β Ξ±C , Ξ²M β Ξ±M) > βmaxjβJ
( 1βΟfj
Οfj
);
where Οmi is the rate of matched men of type i and Ο
fj is the rate of matched
women of type j.
3 Therefore, we can show that Hβ1 has the following sign structure
Hβ1(Β΅eq) =
+... β
. . . . . .
β... +
.
Cobb Douglas MMF : Comparative statics (3)
the sign structure of Hβ1 is obtained because:1 The Schur complements of the H(Β΅) upper IΓ I and lower JΓ J diagonal
blocks are SH11 = H22 βH21(H11)β1H12 and SH22 = H11 βH12(H22)
β1H21.
2 H row diagonally dominantβ the two Schur complement are alsodiagonally dominant. See Theorem 1 of Carlson and Markham (1979 p 249).
3 SH11 and SH22 are also Z-matrices (i.e., members of the class of real matriceswith nonpositive off-diagonal elements).
4 Then, SH11 and SH22 are M-matrices =β SHβ111 = 0 and SHβ1
22 = 0. seeTheorem 4.3 of Fiedler and Ptak (1962).
By applying the implicit function theorem to the equation (24) we have:1 βΒ΅
βmi= J(Β΅)β1 βB
βmifor 1 β€ i β€ I
2 βΒ΅βfj
= J(Β΅)β1 βBβfj
for all 1 β€ j β€ J
Cobb Douglas MMF : Comparative statics (4)
Comparative Statics (1)
Let Β΅ be the equilibrium matching distribution of the Cobb Douglas MMFmodel. If the coefficients Ξ²r and Ξ±r respect the restrictions
1 0 < Ξ²r; Ξ±r β€ 1 for r β {M, C};2 max(Ξ²C β Ξ±C , Ξ²M β Ξ±M) < miniβI
(1βΟm
iΟm
i
);
3 min(Ξ²C β Ξ±C , Ξ²M β Ξ±M) > βmaxjβJ
( 1βΟfj
Οfj
);
Type-specific elasticities of unmatched.
miΒ΅k0
βΒ΅k0βmiβ₯
1
mβimkmβk
βJj=1
[Ξ±MΒ΅Mkj +Ξ±CΒ΅Ckj][Ξ²MΒ΅Mkj +Ξ²CΒ΅Ckj]
f βj> 0 if k 6= i
mimβi
[1 + 1mβi
βJj=1
[Ξ±MΒ΅Mij +Ξ±CΒ΅Cij ][Ξ²MΒ΅Mij +Ξ²CΒ΅Cij ]
f βj] > 1 if k = i,
1 β€ k β€ I.
Cobb Douglas MMF : Comparative statics (5)
Comparative Statics (2)
fjΒ΅0k
βΒ΅0kβfjβ₯
1f βj
fkf βk
βIi=1
[Ξ±MΒ΅Mik +Ξ±CΒ΅Cik ][Ξ²MΒ΅Mik +Ξ²CΒ΅Cik ]
mβi> 0 if k 6= j
fjf βj[1 + 1
f βjβI
i=1[Ξ±MΒ΅Mij +Ξ±CΒ΅Cij ][Ξ²
MΒ΅Mij +Ξ²CΒ΅Cij ]
mβi] > 1 if k = j,
miΒ΅0j
βΒ΅0j
βmiβ€ β
[Ξ±MΒ΅Mij + Ξ±CΒ΅Cij ]
mβi f βjmi < 0, for 1 β€ i β€ I and 1 β€ j β€ J,
fjΒ΅i0
βΒ΅i0βfjβ€ β
[Ξ²MΒ΅Mij + Ξ²CΒ΅Cij ]
mβi f βjfj < 0, for 1 β€ i β€ I and 1 β€ j β€ J,
Cobb Douglas MMF : Return to scale
CRSThe equilibrium matching distribution of the Cobb Douglas MMF modelsatisfies the Constant return to scale property if Ξ²r + Ξ±r = 1 i.e.
Ξ²r + Ξ±r = 1 for r β {M, C} βI
βi=1
βΒ΅
βmimi +
J
βj=1
βΒ΅
βfjfj = Β΅.
Cobb Douglas MMF : Return to scale
One important question is to know under which conditions on Ξ±r and Ξ²r
the Cobb Douglas MMF admit constant (increasing or decreasing) returnto scale?
In other terms, holding the type distributions of men and women fixed,does increasing market size has an effect on the probability of matching?
We can show
I
βi=1
U(Β΅)β1 βΒ΅
βmimi +
J
βj=1
U(Β΅)β1 βΒ΅
βfjfj =
I
βi=1
[H(Β΅)β1]Β·i +J
βj=1
[H(Β΅)β1]Β·(I+j).
Whenever Ξ±r + Ξ²r = 1, H(Β΅) is a row stochastic matrixβ the row of theH(Β΅)β1 matrix sum to oneβ CRS.
Our intuition is whenever Ξ±r + Ξ²r > 1 ( Ξ±r + Ξ²r < 1) we have increasing(decreasing) return to scale.
Identification and estimation
Consider the general Cobb Douglas MMF in presence of multi market data:
ln Β΅rstij = Ξ±r
ij ln Β΅sti0 + Ξ²r
ij ln Β΅st0j + Ξ³rst
ij .
Even with multimarket data, the most general Cobb Douglas MMF isnot identified.
There are 2Γ IΓ JΓ SΓ T elements in the observed matchingdistribution (i.e. Β΅rst
ij ) and there are 2Γ IΓ JΓ SΓ T + 4Γ IΓ Jparameters i.e. (Ξ³rst
ij , Ξ±rij, and Ξ²r
ij).
Identification and estimation
Assumptions
1 (Additive separability of the gain). Ξ³rstij = Οr
ij + Ξ·rsij + ΞΆrt
ij + Ξ΅rstij where Οr
ijrepresents the type fixed effect, Ξ·rs
ij the state fixed effect, ΞΆrtij the time fixed
effect, and Ξ΅rstij the residual terms.
2 (Instrumental Variable (IV)). E[Ξ΅rstij |z
11ij , ..., zST
ij ] = 0, where
zstij = (mst
i , f stj )β².
Identification and estimation
1 (First mean difference)
ln Β΅rstij β lnΒ΅rs
ij = Ξ±rij[ln Β΅st
i0β lnΒ΅si0]+ Ξ²r
ij[ln Β΅st0jβ lnΒ΅s
0j]+ ΞΆrtij β ΞΆ
rij + Ξ΅rst
ij β Ξ΅rsij .
2 (second mean difference)
yrstij = xst
ijβ²Ξ»r
ij + Ξ΅rstij
where yrstij β‘ ln Β΅rst
ij β lnΒ΅rsij β lnΒ΅rt
ij + lnΒ΅i0,
xstij β‘ (ln Β΅st
i0 β lnΒ΅si0 β lnΒ΅t
i0 + lnΒ΅i0, ln Β΅st0j β lnΒ΅s
0j β lnΒ΅t0j + lnΒ΅0j)
β²,
Ξ»rij β‘ (Ξ±r
ij, Ξ²rij)β², and Ξ΅rst
ij β‘ [Ξ΅rstij β Ξ΅rs
ij ]β [Ξ΅rtij β Ξ΅
rij]
Identification and estimation
Since Β΅sti0 and Β΅st
0j are potentially correlated with the residual terms Ξ΅rstij a
simple ordinary least square (OLS) will not be able to identify Ξ»rij
We will instrument Β΅sti0 and Β΅st
0j respectively with msti and f st
j .
IdentificationUnder the latter Assumption, the general Cobb Douglas MMF is identified if
E[zstij xst
ijβ²] is of full column rank. The identification equation is given by
Ξ»rij = {E[zst
ij xstijβ²]}β1E[zst
ij yrstij ].
Empirical Application
We study the marriage matching behavior of 26-30 years old women and28-32 years old men with each other in the US for 1990, 2000 and 2010.
The 1990 and 2000 data is from the 5% US census.
The 2010 data is from aggregating three years of the 1% AmericanCommunity Survey from 2008-2010.
A state year is considered as an isolated marriage market.
There were 51 states which includes DC.
Individuals are distinguished by their schooling level: less than highschool (L), high school graduate (M) and university graduate (H).
A cohabitating couple is one where a respondent answered that they arethe βunmarried partnerβ of the head of the household.
Empirical Application
Table : Summary Staticsβ.
Variable Obs Mean Std. Dev. Min Max N cohabitations 1113 710.7491 1367.211 3 15362 N marriages 1283 5023.924 9981.822 6 118867 N males 1377 49097.82 67662.33 174 568449 N females 1377 48319.55 67411 143 580493 N unmatched males 1377 20361.38 30701.33 165 262267 N umatched females 1377 18757.32 28271.38 76 236391 Year 1377 2000 8.167932 1990 2010
*An observation is a state/year. There are 51 states which includes DC. Observations with 0 cohabitation or marriages are excluded.
Empirical Application
Table : Ordinary Least Square (OLS).
1a 1b 2a 2b 3a 3b
Dep. Var. LCOH LMAR LCOH LMAR LCOH LMAR LU_M (πΌ) 0.562 0.536 0.320 0.244 0.557 0.626 (0.041)** (0.048)** (0.075)** (0.061)** (0.084)** (0.056)** LU_F (Ξ²) 0.531 0.665 0.630 0.609 0.827 0.939 (0.040)** (0.047)** (0.076)** (0.058)** (0.078)** (0.051)**
πΏπ»π» βπππ»π βππ» 2.31
(0.078)**
2.44 (0.069)**
2.290 (0.071)**
2.412 (0.048)**
πΏππ β πΏπΏπΏπ βππΏ 1.78
(0.087)**
2.52 (0.082)**
1.781 (0.080)**
2.464 (0.063)**
πΏπ»π βππΏππ β π»πΏ 0.784
(0.147)**
1.43 (0.092)**
0.796 (0.145)**
1.426 (0.084)**
πΏππ» β πΏπππ β πΏπ» 1.33
(0.141)** 1.37
(0.101)** 1.344
(0.135)** 1.379
(0.083)** Y2000 0.289 -0.287 0.313 -0.277 (0.042)** (0.032)** (0.037)** (0.022)** Y2010 0.627 -0.604 0.615 -0.667 (0.041)** (0.037)** (0.040)** (0.030)** STATE Y Y _cons -4.788 -3.981 -2.842 1.172 -7.359 -5.015 (0.383)** (0.441)** (0.196)** (0.158)** (0.772)** (0.543)** R2 0.51 0.45 0.90 0.95 0.92 0.97 N 1,113 1,283 1,113 1,283 1,113 1,283
πΌπ½ 1.058
(0.137) 0.806
(0.115) 0.508
(0.178) 0.400
(0.138) 0.673
(0.146) 0.667
(0.082) πΌ + π½ 1.093
(0.039) 1.202
(0.045) 0.950
(0.018) 0.853
(0.016) 1.384
(0.079) 1.57
(0.057) πΌβ³
π½β³π½π
πΌπ
0.762 (0.147)
0.788 (0.387)
0.991 (0.246)
ππππ πΌβ³ = πΌπ
π½β³ = π½π
0.091 0.001 0.117
* p<0.05; ** p<0.01
Empirical Application
Table : Instrumental Variable (IV).
1a 1b 2a 2b 3a 3b
Dep. Var. LCOH LMAR LCOH LMAR LCOH LMAR LU_M (πΌ) 0.574 0.599 0.267 0.164 0.556 0.728 (0.044)** (0.053)** (0.086)** (0.074)* (0.095)** (0.070)** LU_F (Ξ²) 0.527 0.717 0.682 0.731 0.915 1.186 (0.042)** (0.052)** (0.089)** (0.070)** (0.090)** (0.058)**
πΏπ»π» βπππ»π βππ» 2.315
(0.079)** 2.436 (0.073)**
2.282 (0.071)**
2.393 (0.051)**
πΏππ β πΏπΏπΏπ βππΏ 1.783
(0.087)** 2.514
(0.081)** 1.773
(0.080)** 2.440
(0.051)**
πΏπ»π βππΏππ β π»πΏ 0.784
(0.149) 1.430
(0.092)** 0.802
(0.147)** 2.440
(0.064)**
πΏππ» β πΏπππ β πΏπ» 1.340
(0.141) 1.377
(0.104)** 1.350
(0.135)** 1.428
(0.086)** Y2000 0.293 -0.277 0.320 -0.258 (0.042)** (0.035)** (0.038)** (0.025)** Y2010 0.624 -0.612 0.607 -0.694 (0.041)** (0.038)** (0.039)** (0.030)** STATE Y Y _cons -4.806 -5.215 -2.806 0.775 -8.045 -8.216 (0.384)** (0.452)** (0.196)** (0.180)** (0.798)** (0.663)** R2 0.51 0.45 0.90 0.95 0.92 0.97 N 1,113 1,283 1,113 1,283 1,113 1,283
πΌπ½ 1.045
(0.132) 0.808
(0.107) 0.346
(0.165) 0.201
(0.103) 0.605
(0.135) 0.606
(0.065) πΌ + π½ 1.102 1.316 0.950 0.892 1.501 1.908
(0.038) (0.045) (0.018) (0.016) (0.082) (0.064) πΌβ³
π½β³π½π
πΌπ
0.771 (0.140)
0.521 (0.347)
1.00 (0.249)
ππππ πΌβ³ = πΌπ
π½β³ = π½π
0.000 0.090 0.000
* p<0.05; ** p<0.01
Empirical Application
Table : IV with time varying match effects.
1a 1b 2a 2b
Dependent variable LCOH LMAR LCOH LMAR LU_M (πΌ) 0.415 0.357 0.576 0.754 (0.071)** (0.064)** (0.077)** (0.057)** LU_F (Ξ²) 0.528 0.524 0.688 0.885 (0.071)** (0.063)** (0.073)** (0.050)**
πΏπ»π» βπππ»π βππ»
2.288
(0.148)** 2.458
(0.087)** 2.278
(0.145)** 2.440
(0.045)**
πΏππ β πΏπΏπΏπ βππΏ
1.504
(0.116)** 2.255
(0. .121)** 1.514
(0.103)** 2.198
(0.067)**
πΏπ»π βππΏππ β π»πΏ
1.169
(0.283)** 1.698
(0.136)** 0.787
(0.211)** 1.702
(0.110)**
πΏππ» β πΏπππ β πΏπ»
0.693
(0.260)** 1.305
(0.134)** 1.177
(0.266)** 1.313
(0.119)
πΏπ»π» βπππ»π βππ»
β π2000 0.259
(0.181) -0.011 (0.191)
0.663 (0.230)**
-0.097 (0.119)
πΏππ β πΏπΏπΏπ βππΏ
β π2000 0.346
(0.323) 0.346
(0.186) 0.257
(0.153) 0.337
(0.124)**
πΏπ»π βππΏππ β π»πΏ
β π2000 0.032
(0.330) -0.279 (0.187)
0.332 (0.317)
-0.274 (0.151)
πΏππ» β πΏπππ β πΏπ»
β π2000 1.133
(0.252) -0.042 (0.193)
0.040 (0.311)
-0.052 (0.152)
πΏπ»π» βπππ»π βππ»
β π2010 1.133
(0.252)** -0.208 (0.199)
1.062 (0.232)**
-0.396 (0.128)
πΏππ β πΏπΏπΏπ βππΏ
β π2010 0.692
(0.190)** 0.534
(0.200)** 0.679
(0.181)** 0.540
(0.162)**
πΏπ»π βππΏππ β π»πΏ
β π2010 -0.384 (0.261)
-0.609 (0.268)**
-0.323 (0.261)
-0.655 (0.246)**
πΏππ» β πΏπππ β πΏπ»
β π2010 0.394
(0.392) 0.264
(0.234) 0.391
(0.376) 0.237
(0.203) Y2000 0.693 0.014 0.669 -0.071 (0.104)** (0.075) (0.091)** (0.054) Y2010 1.122 -0.097 1.061 -0.274 (0.102)** (0.079) (0.092)** (0.057)** STATE Y Y _cons -3.075 0.618 -6.512 -5.900 (0.185)** (0.161)** (0.735)** (0.506)** R2 0.91 0.96 0.93 0.98 N 1,113 1,283 1,113 1,283
πΌπ½
0.786 (0.239)
0.681 (0.203)
0.836 (0.174)
0.852 (0.096)
πΌ + π½ 0.943 0.881 1.264 1.640 (0.017) (0.016) (0.076) (0.055)
πΌβ³
π½β³π½π
πΌπ
0.866 (0.369)
1.019 (0.241)
ππππ πΌβ³ = πΌπ
π½β³ = π½π
0.051 0.000
Conclusion
We propose a new static empirical marriage matching function (MMF):the Cobb Douglas MMF
The Cobb Douglas MMF encompasses CS, CSW, CSFT, CSPE, DM MMF
Properties of this Cobb Douglas MMF are presented.
Existence and uniqueness proof of the marriage distribution areprovided.
Comparative statistics are derived.
Conclusion
The CD MMF is estimated on US marriage and cohabitation data bystates from 1990 to 2010.
CS with peer effects is not rejected.
There are peer and scale effects in the US marriage markets.
We find evidence against all existing MMFs present so far in theliterature.
Positive assortative matching in marriage and cohabitation byeducational attainment are relatively stable from 1990 to 2010.