AN ANALYSIS OF SAMPLE ATTRITION IN PANEL DATA: THE

AN ANALYSIS OF SAMPLE ATTRITION

IN PANEL DATA:

THE MICHIGAN PANEL STUDY OF INCOME DYNAMICS

John Fitzgerald

Bowdoin College

Peter Gottschalk

Boston College

Robert Moffitt

Johns Hopkins University

December, 1996

Revised,

November, 1997

This research was supported by the National Science Foundation through a

grant to the PSID Board of Overseers.

We wish to thank Joseph Altonji,

Greg Duncan,

Guido Imbens,

Charles

Manski,

Gary Solon, Jeffrey

Wooldridge,

and three anonymous referees for comments on various drafts

as well as seminar participants at Berkeley, Michigan State, NYU,

Princeton, Stanford,

and the University of Wisconsin.

Excellent

research assistance was provided by Robert Reville, Lisa

Tichy,

and

Thomas Vanderveen.

Abstract

An Analysis of Sample Attrition in Panel Data:

Michigan Panel Study of Income Dynamics

By 1989 the Michigan Panel Study on Income Dynamics

(PSID)

had

experienced approximately 50 percent sample loss from cumulative

attrition from its initial 1968 membership.

We study the effect of this

attrition on the unconditional distributions of several socioeconomic

variables and on the estimates of several sets of regression

coefficients.

We provide a statistical framework for conducting tests

for attrition bias that draws a sharp distinction between selection on

unobservables and on observables and that shows that weighted least

squares can generate consistent parameter estimates when selection is

based on observables,

even when they are endogenous. Our empirical

analysis shows that attrition is highly selective and is concentrated

among lower socioeconomic status individuals.

We also show that

attrition is concentrated among those with more unstable earnings,

marriage,

and migration histories.

Nevertheless,

we find that these

variables explain very little of the attrition in the sample, and that

the selection that occurs is moderated by regression-to-the-mean effects

from selection on transitory components that fade over time.

Consequently, despite the large amount of attrition, we find no strong

evidence that attrition has seriously distorted the representativeness

of the PSID through 1989,

and considerable evidence that its

cross-

sectional representativeness has remained roughly intact.

The increased availability of panel data from household surveys

has been one of the most important developments in applied social

science research in the last thirty years.

Panel data have permitted

social scientists to examine a wide range of issues that could not be

addressed with cross-sectional data or even repeated cross sections.

Nevertheless, the most potentially damaging and frequently-mentioned

threat to the value of panel data is the presence of biasing attrition--

that is,

attrition that is selectively related to outcome variables of

interest.

In this paper we present the results of a study of attrition and

its potential bias in one of the most well-known panel data sets, the

Michigan Panel Study of Income Dynamics

(PSID).

The PSID has suffered a

large volume of attrition since it began in

1968--almost

50 percent of

initial sample members had attrited by 1989.

We study the effect of

attrition in the PSID on the means and variances of several important

socioeconomic variables

--such as individual earnings, educational level,

marital status, and welfare participation--

and on the coefficients of

variables in regressions for these variables.

We also examine whether

the likelihood of attrition is related to past instability of such

behaviors--

earnings instability, propensities to migrate or to change

marital status, and so on. A companion paper studies the effect of

attrition on estimates of intergenerational relationships (Fitzgerald et

al.,

199733).

An understanding of the statistical issues is important to

understanding our approach. We provide a statistical framework for the

analysis of attrition bias which shows that the common distinction

between selection on unobservables and observables is critical to the

development of tests for attrition bias and adjustments to eliminate it.

However,

we show that selection on observables is not the same as

exogenous selection, for selection can be based on endogenous

observables such as lagged dependent variables which are observed prior

to the point of attrition.

We note that the attrition bias generated by

this type of selection can be eliminated by the use of weighted least

squares,

using weights obtained from estimated equations for the

probability of attrition, and hence without the highly parametric

procedures used in much of the literature.

Many of our tests for

attrition bias are consequently based on whether lagged endogenous

variables affect attrition rates.

However,

we also conduct an implicit

test for selection on unobservables by comparing PSID distributions with

those from an outside data source,

the Current Population Survey (CPS).

We find that while the PSID has been highly selective on many

important variables of interest,

including those ordinarily regarded as

outcome variables,

attrition bias nevertheless remains quite small in

magnitude.

The major reasons for this lack of effect are that the

magnitudes of the attrition effect,

once properly understood, are quite

small (most attrition is random);

and that much attrition is based on

transitory components that fade away from regression-to-the-mean effects

both within and across generations.

We also find that

attrition-

adjusted weights play a small role in reducing attrition bias. We

conclude therefore that the PSID has stayed roughly representative

through

1989.l

A similar conclusion was reached by Becketti, Gould, Lillard,

and Welch (1988) for the PSID using data through 1981 (see also Duncan

and Hill,

1989, for an analysis of representativeness in 1980).

I. The PSID: General Attrition Patterns

The PSID began in 1968 with a sample of approximately 4800

families drawn from the U.S.

noninstitutional population (for a general

description of the PSID see Hill, 1992).

Since 1968 families have

been interviewed annually and a wide variety of socioeconomic

information has been collected. Adults and children in the original

PSID households or who are

descendents

of members of those households

are followed if they form or join new households, thereby providing the

survey the possibility of staying representative of the nonimmigrant

U.S. population.

A consequence of the self-replenishing nature of the

panel is that the sample has grown in size over time. There were

approximately 18,000 individuals in the 1968 families; by 1989,

information on about 26,800 individuals had been collected.'

About three-fifths of the 1968 families were drawn from a

representative sampling frame of the U.S. called the

"SRC"

sample, and

two-fifths were drawn from a set of individuals in low-income families

(mostly in

SMSAs)

known as the

"SEO"

sample. At the time the survey

began,

the PSID staff produced weights that were intended to allow users

to combine the two samples and to calculate statistics representive of

the general population. Those sample weights have been periodically

updated to take into account differential mortality as well as

differential attrition (see Institute for Social Research, 1992,

pp.82-

Institute for Social Research (1992, Table 14). The PSID also

interviews individuals who are not related to a 1968 family but who move

into interviewed households,

most commonly by marrying a PSID member.

Those individuals are termed "nonsample" observations and are assigned a

zero weight. Another 11,600 of these individuals had been interviewed by

1989, on top of the 26,800 mentioned in the text. Generally, such

individuals are no longer interviewed if they leave a PSID household.

However,

all children of a "sample" parent and

"nonsample"

parent are

kept in the survey, which causes the PSID sample size to grow over time;

see below.

98 for a recent discussion of nonresponse and other weighting

adjustments). We shall discuss the effect of this weight adjustment in

our paper.

Table 1 shows response and nonresponse rates of the original 1968

sample

members.3

The first three columns in the table show the number

of individuals remaining in the sample by year---the number in a family

unit,

the portion in

institutions--

whom we treat as respondents, to be

consistent with practice by PSID staff--and their sum, equal to 18,191

individuals in 1968.

As the table indicates in the fourth column, about

88 percent of these individuals remained after the second year, implying

an attrition rate of 12 percent.

The actual number attriting is shown

in the fifth column,

with conditional attrition rates shown in

parentheses below each count.

A smaller proportion left the PSID in

each year after the

first--generally about 2.5 or 3.0 percent annually.

By 1989,

only 49 percent of the original number were still being

interviewed,

corresponding to a cumulative attrition rate of 51 percent.

The table also shows the distribution of the attritors by reason--

either because the entire family became nonresponse ("family unit

nonresponse"),

because of death, or because of a residential move which

could not be successfully

followed.4

The distribution of attrition by

reason has not changed greatly over time,

although there is a slight

increase in the percent attriting because of death and a slight

reduction in the percent attriting because of mobility. Both of these

These attrition rates condition on being interviewed in 1968,

the initial year.

However,

only 76 percent of the families selected to

be interviewed were interviewed (Hill, 1992, p.25).

We return to this

issue below in our comparisons with the CPS.

Some of the "family unit nonresponse" observations may have

attrited because of migration or mortality unknown to the PSID.

trends are no doubt a result of the increasing age of the 1968 sample.

The final column in the table shows the number of individuals who came

back into the survey from nonresponse ("In from nonresponse") each year.

These figures are quite small because,

prior to the early

199Os,

the

PSID did not attempt to locate and reinterview attritors.

Figure 1 illustrates the overall attrition hazards graphically.

The Figure clearly shows the spike in the hazard in the first year. It

is also more

noticable

in the Figure that there has been a slight upward

trend in attrition rates over time, although not large in magnitude.

In a background report (Fitzgerald et al.,

1997a),

we show

cumulative rates of response among 1968 sample members by race, sex, and

age.

Cumulative nonresponse rates have been highest for races other

than black and white,

and next highest for blacks.

Nonresponse rates

are higher among men than among women.

Not surprisingly, nonresponse

rates are highest among the older 1968 sample members and among

respondents initially between 16 and 24. Among the oldest 1968 sample

members,

those 65 and over, only 7 percent were interviewed in 1989.

Nonresponse rates are also higher in the

SE0

subsample than in the SRC

subsample although not by a large amount.

That mortality should have a marked effect on the measured

response rate is not surprising,

but it does imply that the

51-percent

attrition rate in Table 1 overstates sample loss among the living

population.

When individuals who died while in the PSID are excluded,

overall nonresponse rates fall from 51 percent to 45 percent overall and

from 68 percent to 47 percent among those 55-64.

When an additional

adjustment is made for mortality among attritors after the point of

attrition (using national mortality rates by age, race, and sex), the

attrition rate for the older population falls another 12 percentage

points to 35 percent and the overall attrition rate falls to 44 percent

(i.e.,

the estimated percents of still-alive individuals who have left

the

PSID).5

II. Statistical Approach

Although a sample loss as high as 44 percent must necessarily

reduce precision of estimation,

there is no necessary relationship

between the size of sample loss from attrition and the existence or

magnitude of attrition bias.

Even a large amount of attrition causes no

bias if it is "random"

in a sense we will define formally

below.

this section we will outline our approach to addressing this issue by

presenting

statistical model that distinguishes between different

types of bias,

which discusses the different restrictions necessary to

detect and correct for each type, and which outlines which types we will

address in our empirical work.

Selection on Observables and Unobservables.

Attrition bias in the

econometric literature is associated with models of selection bias, and

the applicability of the selection bias model to attrition was

recognized early in the literature (e.g.,

Heckman,

1979).

But

recognition of the problem of nonresponse and the bias it can cause

dates from much earlier in the survey sampling literature (see Madow et

al.,

1983, for a review). Here we will present a model tied more

That is,

individuals who died after the point of attrition

cannot be identified as having died from the PSID data. This implies

that the attrition rates we have calculated,

even netting out those who

died while in the PSID,

overstate the fraction of the living population

that has attrited. We use national mortality rates by age, race, sex,

and year to estimate the number of attritors who have died, and then

recalculate our attrition rates accordingly.

closely to econometric formulations than to those in survey sampling

studies.

Our setup will initially be formulated as a cross-section

model but then will be modified for panel data.

We assume that the object of interest is a conditional population

density

f(ylx)

where y is a scalar dependent variable and x is (for

illustration) a scalar independent variable.

We will work at the

population level and ignore sampling considerations.

Define A as an

attrition dummy equal to 1 if an observation is missing its value of y

because of attrition and 0 if not (we assume for the moment that x is

observed for all, as would be the case if it were a time-invariant or

lagged variable). We therefore observe (or can estimate) only the

density

g(ylx,A=O).

The problem is how to infer f from g.

By necessity

this will require restrictions of some kind.

Although there are many restrictions possible (in fact, an

infinite number), we will focus only on a set of restrictions which can

be imposed directly on the attrition function, which we define as the

probability function

Pr(A=Oly,x,z).

Here z is an auxiliary variable

which is assumed to be observable for all units (e.g., a time-invariant

or lagged variable) but distinct from x,

and whose role will become

clear momentarily.

The variable y is partially unobserved in this

function because it is not observed if A=l.

The key distinction we make is between what we term selection on

observable8 and selection on

unobservables.6 We say that selection on

These terms have not, to our knowledge, been utilized in the

literature on sample selection models (i.e.,

models where a subset of

the population is missing information on y).

However,

the terms have

been used in the treatment-effects literature, most extensively and

explicitly

Heckman

and Hotz (1989) but also by

Heckman

and Robb

(1985,

p.190).

The concept of selection on observables, if not the

exact term,

appears much earlier in the treatment-effects literature.

We should also note that the survey sampling literature often uses the

observables occurs

when

Pr(A=Oly,x,z

) = Pr

(A=Olx,z)

(1)

We say that selection on unobservables occurs simply when (1) fails to

hold; that is,

when the attrition function cannot be reduced from

Pr(A=Oly,x,z).'

These definitions may be more familiar when they are restated

within the textbook parametric model.

Letting

E(~~x)=~~+~,x

and

Pr(A=01x,z)=F(-60-61x-62z),

where F is a proper c.d.f., we can state the

model equivalently with error terms

and v as

p,x

y observed if

A=0

(2)

A* =

alx

622

+ v

(3)

ifA*>

(4)

ifA*<O

where v is the random variable whose c.d.f. is F. In the context

of this model, selection on unobservables occurs when

IElX

but

V-La

(5)

and that selection on observables occurs when

terms

"ignorable" and "missing-at-random" selection to describe what we

are terming selection on observables (Little and

Rubin,

1987).

We could define selection on unobservables to occur when x and z

drop out of the probability function,

and then to define selection on

both observables and unobservables to occur when

y,x,

and z all appear

in the function, but we are not particularly interested in the former

case and hence will not maintain such usage.

but z

-11

(6)

where the symbols

and

denote

"is

independent of" and

"is

not

independent of," respectively.

The selection on observables case is

relatively unfamiliar in the econometrics literature but we will show

that it is relevant for the attrition problem.

However,

we will first

deal with the more familiar case of selection on unobservables.

Selection on Unobservables.

We will discuss this model only

briefly because of its familiarity.

Exclusion restrictions are the

usual method of identifying this model,

and our major goal here is to

discuss the difficulty in finding such restrictions for a nonresponse

model in the PSID.

Working from the parametric form of the model, the conditional

mean of y in the nonattriting sample can be written

E(ylx,z,A=O)

plx

E(e~x,z,v<-60-~lx-62z)

P,x

h(-60-61x-52z)

P,x

h'(F(-50-61x-62z)

(7)

where h and

are functions with unknown parameters.

Moving from the

first to the second line of the equation requires that the joint

distribution of a and v be independent of x and z, so that the

conditional expectation depends on x and z only through the index.

Moving from the second to the third line simply replaces the index by

its probability,

which is permissible since they have a one-to-one

correspondence.

Early implementations of this model assumed a specific bivariate

distribution for

and v,

leading to specific forms of the expectation

function (e.g., the inverse Mills ratio for bivariate normality), while

more recent implementations have relaxed some of the distributional

assumptions in the model by estimating functions h or

whose arguments

are either the attrition index or the attrition probability,

respectively (see Maddala,

1983, for a textbook treatment of the early

approach and Powell, 1994,

pp.2509-2510,

for discussions of the more

recent approach).

Armed with estimates of the parameters of the

attrition index or of the predicted attrition probability, equation (7)

becomes a function whose parameters can be consistently estimated.*

However,

aside from nonlinearities in the h, h', and F functions,

identification of

requires an exclusion restriction, namely, that a z

exist satisfying the independence property from

and for which

nonzero.

Such a variable is often loosely termed an "instrument,"

although most estimation methods proposed for eqn (7) do not take a

textbook instrumental-variables form.

Finding a suitable instrument for

unobservable selection is more difficult for the case of nonresponse

than in some other applications because there are few variables that

affect nonresponse that can be credibly excluded from the main equation

for y.

While this depends on the specific model under consideration, on

If nonparametric methods are used to estimate h and h', not all

of the parameters in

(e.g.,

the intercept) may be identifiable. We

should also note at this point that if x is time-varying then it is

necessarily missing for attritors and hence the attrition propensity

equation cannot be estimated as we have written it.

Additional

assumptions are then required to estimate the model.

For example,

adding time subscripts,

one could assume

x(t)=ao+a

x(t-1)ta

z+u(t), thus

letting x be a function of lagged x and z (some

ferent

could be

specified,

alternatively). Substituting this equation for x(t) into the

attrition equation would permit estimation provided

x(t-1)

is available

for all observations. This procedure, however, introduces another

potential source of selection bias from non-independence of u(t) and

e(t)

a priori grounds personal characteristics such as those generally

included in x are unlikely to be promising sources of instruments

because most such characteristics are related to behavior in general and

hence to y.

More promising are variables external to the individual and not

under his control,

such as characteristics of the interviewer or the

interviewing process, or even interview payments.

Although we have

proposed no explicit behavioral model of attrition, a natural theory

would be a simple benefit-cost model in which an individual compares the

value of participating in the survey to the value of not participating.

Good interviewers or interviewing conditions lower the cost of

participation and interview payments directly increase the value of

participation.

However,

a suitable instrument must vary across

respondents,

and must vary in a manner independent of

The staff at

the Institute for Survey Research who have administered the PSID have

assigned interviewers on the basis of respondent characteristics, and

have also varied interviewing conditions (length of interview, in-person

vs.

telephone,

number of callbacks, etc.) entirely and only on the basis

of respondent characteristics; consequently there is no exogenous

component to the variation intreatment.

This rules these variables out

as instruments.

Moreover,

there have also been no exogenous variations in

interview payments over the course of the PSID, for payments have been

adjusted only for inflation over time and vary within year only on the

basis of interview mode.

Based on these and other considerations we

discuss in our background report (Fitzgerald et al.,

1997a),

we conclude

that there are no instruments for nonresponse in the PSID which are

credibly exogenous to behavior in general.'

-I-.

Although we will therefore not test for selection on unobservables

directly,

or correct for such selection,

indirect tests for selection on

unobservables can be conducted whenever an outside data set is available

containing validation information. Administrative data on some variables

(e.g.,

earnings) are occasionally available but this is the exception

rather than the rule,

and they are not available for the

PS1D.l'

However,

the Current Population Survey (CPS) is a heavily-used outside

data set which is a repeated cross section and hence not subject to the

same type of attrition bias as the PSID.

The CPS is subject to

nonresponse itself, but not of the same order of magnitude as the 50

percent nonresponse rate in the

PSID.ll

Hence we will use the CPS as a

comparison data set and compare the marginal distributions of variables

in the CPS and PSID to one another as well as regression coefficients.

If selection on unobservables is present and it biases the coefficients,

for example (see eqn.

(7)),

estimates from the two data sets will be

different.

Unfortunately,

this method of comparison is useful only for

cross-sectionally-defined variables and not for variables which make use

of the panel nature of the PSID, and hence does not offer a general

Exclusion restrictions are only one form of information. For an

example of the use of other types of information, see

Manski

(1994).

Fitzgerald et al.

(1997a) provide some simple bounds calculations of one

type proposed by

Manski.

See Hill (1992, p.29) and Bound et al.

(1994) for a discussion

of validation studies using the PSID.

While the magnitude of nonresponse does not map directly into

the amount of bias, as we noted earlier,

it would be unlikely for the

CPS to be more biased than the PSID given these differences in the

amounts of attrition.

solution to the

prob1em.l'

Selection on Observables. As we noted previously, the case of

selection on observables is relatively unfamiliar in the econometrics

literature.

Because of this unfamiliarity, and because, unlike

selection on unobservables,

it is something we can actually address, we

will discuss it at slightly greater length than we did the previous

case.

The critical variable in the selection on observables case is z, a

variable which affects attrition propensities but is presumed also to be

related to the density of y conditional on x (i.e., z is endogenous to

Such a variable can exist only if the investigator is interested in

"structural" y function which we interpret as a function of a variable

x that plays a causal role in a theoretical sense; other variables

(i.e.,

z) do not "belong" in the function.

More generally, this

situation will arise whenever the investigator is interested in (say)

the expectation of y conditional on x and simply does not wish to

condition on z. In cross-sectional data, for example, the standard

Mincerian theory of human capital proposes that earnings are a function

of education and experience;

other variables which are jointly

determined with earnings, like occupation and industry, should not be

conditioned on to obtain the "correct" estimates.

Yet use of any sample

that is selected on the basis of occupation and industry (e.g., only

certain occupations and industries are included) will clearly bias the

estimates of the earnings equation. The variable z is thus an

Imbens and Hellerstein (1996) show that such outside data sets,

if taken as 'truth,'

can be imposed on the data set of interest (e.g.,

the PSID) and can be used to formally test whether the data

distributions in the two data sets are the same. See related work by

Imbens and Lancaster (1994) and Hirano et al. (1996) along these lines.

"auxiliary"

endogenous variable.

As we will discuss below, in the panel

data case,

a lagged value of y can play the role of z if it is not in

the

"structural"

model and if it is related to attrition.

In the presence of selection on such an endogenous variable, it is

easy to show that least squares estimation of (2) on the nonattriting

sample will generate inconsistent estimates of

and, more generally,

that the estimable density

g(ylx,A=O)

will not correspond to the

complete-population density

f(ylx)

since the event

A=0

is related to y

through z.

Apart from this selection on observables bias, using as much

of the lagged information in the panel as possible helps reduce the

amount of residual,

unexplained attrition variation left over in the

data,

and this will reduce the scope for selection on unobservables.

Formally,

in the Appendix, we show that,

under the selection on

observables restriction given in equation

(l),

the complete-population

density

f(ylx)

can be computed from the conditional joint density of y

and z,

which we denote by g:

f(YlX)

g(y,zlx,A=O)

w(z,x)

where

w(z,x)

Pr(A=Olz,x)

Pr(A=Olx)

(8)

are normalized weights. The numerator of (9) inside the brackets is the

probability of retention in the sample and is,

in the parametric model

described above,

F(-60-51~-62~).

Because both the weights and the

conditional density g are identifiable and estimable functions, the

complete-population density

f(ylx)

is estimable,

as are its moments such

as its expected value

(@,+p,x

in the parametric

model).13

Eqn(8) shows

that the complete-population density can be derived by weighting the

conditional density by the (normalized) inverse selection probabilities;

in the parametric model, it can be shown that this implies that weighted

least squares (WLS) can be applied to eqn(2) using the weights in (9).

We should emphasize that the application of WLS in this case is

unrelated to the heteroskedasticity rationale appearing in most

econometrics texts.

It is also not in conflict with the conventional

view among many applied economists that survey weights can be ignored

because they do not affect the consistency of OLS coefficients, for

survey weights are often intended only to adjust for sample designs

which have stratified the population or differentially sampled it by

variables that are exogenous.

Here, however, selection is indirectly on

the dependent variable, and not adjusting for attrition results in loss

of consistency.

If z is not a determinant of attrition, the weights in (9) equal

one and hence all conditional densities equal unconditional ones and no

attrition bias is present.

Alternatively,

if y and z are independent

conditional on x and A=O,

the density g in (8) factors and it can again

be shown that the unconditional density

f(ylx)

equals the conditional

density,

and there is no attrition bias.

While these results are relatively unfamiliar in the econometric

literature,

they are pervasive in the survey sampling literature, where

they form the intellectual justification for the construction and use of

As we noted in n.8,

if contemporaneous x is unobserved and

hence the attrition probability equation cannot be estimated, lagged x

or additional z variables are required.

attrition-based survey weights (Rao,

1963,197s;

Little and

Rubin,1987,pp.55-60).14'15

In the econometrics literature, while

weighting formulations are sometimes used as a framework for discussing

selection models (e.g.,

Heckman,

1987),

the main point of contact with

the models discussed here is the choice-based sampling literature (for

discrete

see Manski and Lerman,

1977, for an early treatment and

Amemiya,

1985, for a textbook treatment; for continuous y, see Hausman

and Wise,

1981, Cosslett, 1993, and Imbens and Lancaster, 1996).

That

literature generally considers estimation and identification in samples

which are selected directly on the dependent variable, y; weighted

maximum likelihood or least squares procedures are often proposed to

'undo'

the disproportionate endogenous sampling. The difference in the

attrition case is that selection is on an auxiliary variable (z) and not

on y itself; but otherwise the solutions are closely

related.16

For an exception, see Cosslett (1993,

pp.31-32).

In addition,

after the first draft of this paper we discovered an independent

treatment of the selection on observables case by Horowitz and

Manski

(forthcoming),

who show that the mean of a function of y can be

consistently estimated with weights of the type we have discussed under

the same restrictions.

We should note that the weights discussed in the survey

sampling literature sometimes differ from the weights in our model in

two respects.

First,

many survey weights--

including those in the

PSID--

are also intended to capture non-random-sampling at the initial stage

(e.g.,

from stratified designs).

That is not the purpose of the weights

we have discussed and requires a slightly different formulation to

justify.

Second,

the weights in our model are not the type of

"universal"

weights generally computed for many survey data sets;

"universal"

weights are designed to be all-purpose and usable for any

variable or model, whereas our weights are model-specific because one

can easily imagine using different attrition-equations (e.g., with

different lagged y's) depending on the model being estimated and its

definition of y.

We wish to emphasize that WLS is not the only estimation method-

-there are many (imputation,

GMM, various forms of maximum

likelihood)--

nor is it efficient; in addition, there are many issues connected with

the use of weights which we do not discuss here.

The major advantage of

WLS is that it produces consistent estimates and is relatively easy to

It should also be noted that simply conditioning on z does not

solve the problem.

This can be seen most simply by observing that the

object of interest in most models is

E(ylx),

not

E(ylx,z).

Including z

in the regressor set will generate "biased"

coefficients on x in a

linear-regression model, for example, in the sense that it will not

estimate the effect of x on y unconditional on z.

Because z is an

endogenous variable, it distorts the conditional distribution of y on x.

Hence correcting for selection on observables is to be sharply

distinguished from the corrections for unobservable selection shown in

eqn

(7),

which involve conditioning on functions of x and z; those

methods are not appropriate for this case.

Testing.

The application of the selection on observables model to

attrition in panel data is straightforward if a lagged value of y (e.g.,

y at the initial wave of the panel,

when all observations are present)

plays the role of z,

assuming that attrition is affected by such a

lagged value.

Lagged values of y will,

assuming serial correlation in

the y process,

be related to current values of y conditional on x.

The

use of lagged values of y in this role requires the same distinction we

noted earlier between structural and auxiliary determinants of

contemporaneous y, for the use of lagged y as a z makes sense only if

the investigator is interested,

for theoretical or other purposes, in

functions of y not conditioned on those lagged

va1ues.l'

implement.

An investigator who posits a theoretical (i.e., structural)

model that includes all lags of y will necessarily have much reduced

scope for selection on observables.

Taking this point to its extreme, if

there are no observables in the data set that are excluded from the

structural y function, there is no role for for using observables to

adjust for selection. Selection on observables is a data-set-defined and

model-defined category, and what is an observable variable in one data

set or model may be an unobservable in another.

As noted previously, two sufficient conditions for the absence of

attrition bias on observables are either that the weights equal one

(i.e.,

z does not affect A) or that z is independent of y conditional on

Specification tests for selection on observables can be based on

either of these two conditions. Thus one test is simply to determine

whether candidate variables for z (e.g., lagged values of y)

significantly affect A.

We will conduct these tests extensively in our

empirical work.

A second test would be to conduct specification tests

for whether OLS and WLS estimates of eqn (2) are significantly

different,

which is an indirect test for whether the identifying

variables used in the weights are endogenous (see Dumouchel and Duncan,

1983, for an example of such a test).

We will not conduct such tests

in our paper but instead leave them for future research. However, we

will determine whether using the universal weights provided by the PSID

staff affect the estimated coefficients of several models, even though

the "model-based"

weights we have been discussing are not necessarily

the same as the PSID universal weights (see n.15).

Another test for selection on observables which we will perform is

based on an exercise performed by Becketti et al. (1988) and which we

term the BGLW test. In the BGLW test,

the value of y at the initial

wave of the survey,

which we denote by

yo,

is regressed on x and on

future A (i.e.,

whether the individual later

attrites).

The test for

attrition selection is based upon the significance of A in that

equati0n.l'

This test must necessarily

closely related to the test we

We assume x to be time-invariant.

If it is not, this method

requires that only the values of x at the initial wave be included in

the equation.

have already described of regressing A on x and

(which is z in this

case)

;

in fact,

the two equations are simply inverses of one another.

Formally,

suppose that the attrition function is taken as the

latent index in the parametric model, i.e.,

alx

a22

+ v

(10)

Inverting this equation, taking expectations, and applying

Bayes'

Rule,

it can be shown that

(yOIArx)

(yolx)

w(A,Y~~x)

dye

(11)

where

w(A,yo,x)

(AIyO,x)

(12)

Pr (Alx)

which are essentially the same as the weights appearing in (9) but

including the probabilities of A=1 as well as A=O.

Eqn (11) shows that

if the weights all equal one,

the conditional mean of

is independent

of A and hence A will be insignificant in a regression of y on x and A

(the conditional mean of

in the absence of attrition bias is

B,+B,x,

so a regression of

on x will yield estimates of this equation). As

noted previously,

the weights will equal one only if

is not a

determinant of A conditional on x. Thus the BGLW method is an indirect

test of the same restriction as the direct method of estimating the

attrition function

itself.lg

However,

if the weights do not equal one,

it would be difficult

derive an explicit solution for

equation(l1)

from the estimates of (10)

that we will obtain in our attrition propensity models.

To do so would

require conducting directly the integration shown in (11).

It would be

simpler to just estimate a linear approximation to (11) by OLS, as did

Becketti et al.,

to determine the magnitude of the effect of A on the

intercept and coefficients of the equation for

as a function of x.

We shall therefore also estimate such equations in our empirical work.

However,

it should be kept in mind that this is not an independent test

of attrition bias separate from that embodied in our estimates of

eqn(l0);

it is only a shorthand means of deriving the implications of

our estimates of eqn(lO) for the magnitudes of differences in 1968 y

conditional on X between attritors and nonattritors.

Panel Data and Permanent-Transitory Effects.

Finally,

we wish to

relate the selection on observables model we have been discussing to

more traditional models of attrition in panel data, and to point out a

connection with permanent-transitory distinctions which we will also

apply in our empirical work below.

The most well-known model of

attrition in the econometrics literature is the model of Hausman and

Wise (1979); that model has been generalized and extended by Ridder

(1990,1992),

Nijman and Verbeek

(1992),

Van den Berg et al.

(1994),

and

others (see Verbeek and Nijman, 1996, for a review).

These models

generally assume a components structure to the error term, sometimes

In general, of course, if

v=o+(3u+e,

regressing u on v instead

of v on u results in a "biased" coefficient on v (i.e., it is not a

consistent estimate of the inverse of

8).

Nothing here contravenes

that.

The

"coefficient"

on x in a regression of y on x and A bears no

simple relationship to 61 or

eqn(lO),

as can be from

eqn(l1).

including individual-specific time-invariant effects and sometimes

serially-correlated transitory effects, for example, and impose

restrictions on how attrition relates to the components of the

structure.

A common assumption in some studies in the literature, for

example,

is that the unobserved components of attrition propensities are

independent of the transitory effect but not the individual effect; in

that case,

simple first-differencing (among other methods) can eliminate

the bias.

Our approach differs from this past work because of our sharp

distinction between identifiability under selection on observables and

on unobservables, a distinction not made in these past studies.

Many

error components models which allow attrition propensities to covary

with individual components of the process can be treated within the

selection on observables framework because lagged values of y can be

mapped into those components.

If we let

in our model stand for a

vector of lagged values of y instead of a scalar, we have

Pr(A=Olx,yt-1,Yt-2,Yt-3r..

y,)

as our attrition function.

Assume full

observability of those lagged values. Then any model in which the error

components of the y process which covary with attrition can be uniquely

mapped into the set of t values of lagged y can be captured by our

selection on observables model.

An example is the autoregressive model:

=P,

P,x

t-1

=,-$j

PTY

(13)

(14)

(15)

t-1

A* = 6

alx

+ c 6

r=O

2T%

Estimation of (13) on the non-attriting sample results in bias because

is serially correlated and A*

is a function of the lagged values of

that error.

But solving

eqn(l3)

for

in lagged periods, and

substituting into

eqn(l5)

for the lagged errors, leads to an equation

for A* where only lagged y appear.

This example also illustrates a case in which controlling for

lagged observables in the A*

equation is not sufficient to avoid

attrition bias,

for it is necessary that the contemporaneous shock at

(i.e.,

that which is not forecastable from lagged

be independent of

conditional on the observables.

For example,

shocks to earnings

which occur simultaneously with, not prior to, attrition from the

sample, cannot be captured by lagged values of y; attrition bias from

this source falls under the selection on unobservables rubric we

discussed earlier.

However,

a full conditioning on the available data

on the history of y reduces the scope of possible unobservable

selection,

as we noted earlier, because it isolates the only remaining

source of such bias to contemporaneous, non-forecastable shocks.

The general form of our attrition probability

Pr(A=Olx,yt-l,

Yt-2'Yt-3'*

.'I

y,)

is capable of capturing a large variety of

alternative forms of attrition dependence on lagged y other than the

simple linear form portrayed in the autoregressive case.

For example,

the mean of a set of lagged values of y,

is a consistent estimator

(as

T-m)

for the individual effect,

after conditioning on observables

x and assuming mean-zero transitory disturbances.

The deviations of

each value of

from

represent transitory disturbances in each

period

By estimating flexible forms of the attrition function which

contain both

and the deviations of lagged y from

in different

periods,

we can determine whether attrition probabilities

covary

with

"permanent" levels of y and with transitory shocks one period, two

periods,

and more periods back in time.

The variance of

over any

specified length of past periods is yet another transform of lagged y

values which may

covary

with attrition; this would occur if it is

variability per se,

not the mean or value of any set of individual

disturbances, that affects whether individuals stay in or out of the

sample.20

We will test these and other transforms of lagged y in our

models.

Summary of Analyses to be conducted.

To summarize, in the

following analysis of the PSID we will (i) conduct tests for the

presence of attrition on unobservables by comparing cross-sectional

marginals and regression coefficients in the CPS and the PSID; (ii)

conduct tests for the presence of selection on observables by estimating

attrition equations as a function of lagged y values as well as by

regressing first-period y on future attrition;

and (iii) we will conduct

tests for "dynamic" attrition effects by estimating attrition equations

as a function of lagged permanent, transitory,

and other moments of the

lagged y distribution.

We should note at this point that a problem with implementing

procedures using lagged values of y is that those measures are available

for the full sample only at the initial year of the PSID, 1968.

Conditioning on values of y after 1968 necessarily opens the door to

bias because some attrition has already occurred and estimation must be

restricted to observations for whom all data on all lagged variables in

the equation are available. Consequently, for the most part, we will

It is clear that formal modeling of the error process of y

could be conducted here but we will leave that for future research, and

will only test various transforms of lagged y in a reduced-form context.

restrict our tests of lags to only those available in the first year,

1968.

While this approach necessarily ignores much of the information

in the PSID on attritors prior to the point of attrition, it yields

results least subject to the post-1968 attrition bias problem.

Our

dynamic attrition analysis will be an exception, for there we will

estimate attrition hazards--that is, probabilities of exit conditional

on being in the sample

t-he

previous period--as a function of all the

lags available up to each decision point.

That analysis will be

conducted ignoring the potential bias induced by this sample restriction

(usually called "unobserved heterogeneity" in duration analyses);

consequently, no "structural"

interpretation will be given to the

estimated coefficients in those attrition

equations.*l

III. Observable Correlates of Attrition in the PSID

Rather than begin our analysis with the comparison of the PSID to

the CPS, we will first examine the observable correlates of attrition in

the PSID, primarily focusing on characteristics, any one of which could

be a

"y"

or a

"x",

in 1968.

We will also estimate attrition probability

equations as a function of 1968 characteristics for selected

"y"

variables and will conduct BGLW tests in this section.

The last year of the PSID available at the time our data files

were created is 1989. We focus on the seemingly simple question of

whether 1968 characteristics differ between those who were present in

Note,

however,

that a bias in the structural coefficients of

attrition hazards does not affect the consistency of the WLS estimator

using the predicted probabilities from those equations as weights.

The

selection on observables model does not require independence of z and v

in eqn(3).

1989 and those who were not (hence the distributions of x and y

conditional on A,

in a tabular

form)." For our analysis sample, we take

every individual who was present in a PSID household in 1968, or about

eighteen thousand individuals, as noted previously. We disaggregate the

sample by sex and 1968 household

headship

status, and focus on five

population subgroups: male heads, wives, female heads, male nonheads,

and female nonheads.

The asymmetric treatment of men and women is

required by the gender-specific definitions of

headship

in the PSID, and

the division of groups by

headship

in the first place is required

because sharply differential amounts of information were collected on

heads and

nonheads

(many variables are not available for the latter

group).23

We also exclude subfamily heads from the PSID because they were

defined inconsistently over time and also differently than in the CPS,

whose comparisons to the PSID are an important part of our analysis.

For the bulk of our work, we include the

SE0

oversample together

with the SRC representative sample.

We therefore use PSID-constructed

1968 sample weights whenever

appropriate.24

However, we also provide

In our background report,

we also conduct analyses of the

middle year,

1981, because that was the latest year analyzed by BGLW.

The issue that analysis addresses is whether any attrition bias we find

has arisen since the BGLW study was conducted.

The PSID makes no distinction between male heads similar to

that made between wives and female heads, for all married women are

automatically classified as wives.

The PSID also incorporates

cohabitation to a degree:

any couple living together in a "partner"

status for more than one interview is then and thereafter treated as

"married"

--the male is classified as a "head" and the female is

classified as a "wife".

We include them in our sample.

These weights reflect only the sample design of the PSID (and

initial nonresponse) and contain no adjustments for attrition.

Hence

they are not the types of weights we were discussing in Section II.

However,

they must be utilized because the

SE0

observations were sampled

on variables that are correlated with income,

which is closely related

to many of our dependent variables.

estimates on the SRC sample alone and show that attrition effects are

sometimes worse for that sample than for the combined SEO-SRC sample.

Distributions of 1968 Characteristics. Table 2 shows the mean

values of 1968 characteristics of men who were 25-64 and household heads

in 1968, by their attrition status as of

1989--"always

in" versus 'ever

out" by that

year.25

As the first two columns indicate, attritors and

non-attritors have many significant differences in characteristics.

Attritors are more likely to be on welfare,

less likely to be married,

and are older and more likely nonwhite.

In addition,

attritors have

lower levels of education, fewer hours of work, less labor income, and

are less likely to own a home and more likely to

rent.*'j

The clear

implication of this pattern is that attritors are concentrated in the

lower portion of the socioeconomic distribution.

The second moments for

labor income in the table indicate that the variance of labor income is

greater among attritors than among nonattritors, and, interestingly,

that the attritor labor income distribution is more dispersed at the

upper tail than the nonattritor distribution.

This suggests that, to

some degree,

some high labor-income families may be more likely to

attrite

than middle-income families."

The last two columns in the table provide an assessment of the

Because only a tiny fraction of attritors ever return--see

Table 1 above--

those individuals who were "always in' between 1968 and

1989 are almost identical to the set of individuals present in 1989, and

the set of individuals who were 'ever out' between 1968 and 1989 is

almost identical to those who were nonresponse in 1989.

All monetary figures in the paper are in real 1982 dollars using

the personal consumption expenditure deflator.

We should also note

that the top and bottom 1 percent of the labor income variable is

excluded to circumvent top-coding problems and to avoid distortion from

outliers.

A similar finding was reported by BGLW.

effect of mortality. The third and fourth columns disaggregate the "ever

out"

subsample into those "not dead" and those "dead" according to

whether individuals died while in the PSID (as noted previously,

some

individuals die after attriting,

of which we have no knowledge).

Comparing the third column (not dead) with the first two shows that the

gap between the Always In and Ever Out is sometimes narrowed by

excluding the dead from the attritors,

but rarely

very much; indeed,

in some circumstances,

the gap even increases.

The latter occurs when

mortality is related to a variable in opposite sign to its relation to

attrition conditional on being alive: consequently, ignoring mortality

actually makes the selectiveness of attrition

seem

milder than it

actually is.

Tables 3 and 4 show the corresponding tables for wives and female

heads.28

The general findings are the same as for male heads:

attritors and nonattritors frequently differ in their characteristics,

and the differences cannot be explained by mortality.

A few of the

details do differ across demographic groups, however.

Female heads have

much larger differences in welfare participation, for example (female

heads also have higher participation rates in the U.S. welfare system

than other groups).

Interestingly, the variance of labor income is

smaller among attritors than nonattritors among female heads, although

the differences among women are not significant. We conclude that the

many significant differences in attritors and non-attritors in the PSID

appear broadly across all

headship

and gender groups.

Attrition

Probits.

The first multivariate analysis we present

consists of estimates of binary-choice models for the determinants of

In our background report,

we also provide tabulations for

nonheads.

attrition,

using the same data in the tables we have been presenting

(i.e.,

whether having ever been nonresponse by 1989 as a function of

1968 characteristics).

We therefore estimate

probit

equations for the

probability of having ever been nonresponse by 1989." As in Tables 2-4,

the sample consists of all 1968 respondents 25-64 and all regressors are

measured in 1968.

We shall also make a distinction between

"x"

and

"y"

in this

analysis by focusing on three

"y"

variables: labor income, marital

status,

and welfare participation (female heads only).

We select these

three because they are some of the more common dependent variables used

by economists and sociologists,

and therefore their relations to

attrition are of particular interest.

Our tabular analysis in Tables 2-

4 showed some evidence of significant attrition effects for these key

variables,

which should generate some cause for concern for analysts who

study these

outcomes.30

One issue that can be addressed in a

multivariate analysis is whether these effects are attenuated when a set

of other socioeconomic variables is controlled for in a regression

framework.

Table 5 shows a set of expanding specifications of attrition

probits

which focus on the effect of our first

"y,"

labor income, on the

attrition of male heads.

The first two columns of the table 5 show the

effect of labor income on attrition without conditioning on any other

Although we do not estimate a dynamic model of year-by-year

attrition,

these estimates can be viewed as a model of cumulative

attrition that reflects the working-out of a year-by-year model.

Since

all the regressors are held at their 1968 values, our equation can be

viewed as a approximation to the reduced-form model.

To repeat a point in Section II,

the concern arises because the

1968 values of these variables are likely to

covary

with their later

values.

regressors ("No Labor Income"

is a dummy equal to 1 if the individual

has no labor income).

The results show that the 1968 labor income levels

of male heads have a very strong correlation with future nonresponse.

Attrition probabilities are quadratic in labor income--lowest at middle

income levels and greatest at high and low income levels, a pattern

also found by BGLW, as noted earlier.

Individuals with no labor income

at all have higher attrition rates as well. The third column in the

table shows that when "standard"

earnings-determining variables are

added--race, age, and education

--labor income remains a significant

determinant of attrition. Implicitly, therefore, the residual in a labor

income equation containing these regressors is correlated with

attrition.

When a large number of other variables--income/needs,

home

ownership, SE0

status, and others--are added, the labor income effects

remain.

Table 6 shows the coefficients on the earnings variables in these

models (except for the first) for wives and female heads, and also the

coefficients for other 1968

"y"

variables.'l

For female heads and wives,

labor income effects are much weaker.

For neither group is there much

of an effect of labor income on nonresponse except for the effects of

having no labor income at all,

which continues to have a positive effect

on nonresponse.

For wives,

even this effect is relatively weak when the

larger set of covariates is included in the equation.

When the earnings

variables are replaced by our other two

"y"

variables--l968 marital

status and welfare participation

--rather similar patterns are found.

Again, there are some significant coefficients on these variables when

nothing else is controlled for,

but in all cases those effects fall to

The full set of regression coefficients on all models is

available in our background report.

insignificance at conventional levels in the most expanded

specification.

Table 7 shows the coefficients in attrition

probits

when all three

types of y variables are included.

Although including the variables

singly gives the best specification for comparison with the BGLW

specification (which inverts the attrition

probit

to solve for a single

there is no reason not to include all available data in an attrition

probit

intended for weight construction, or for general

interest.32

The

results in the table indicate that very little is changed when multiple

y variables are included;

most effects are insignificant, with the

absence of labor income continuing to be the one variable with

often-

significant effects even after controlling for other regressors.

We should also note that the R-squareds from these

probits

are

extremely

small.33

In Table 5 they never exceed

.069

and in the models

in Tables 6 and 7 they range from

.028

.071,

and even lower in Models

1,2,

and 3 when fewer other regressors are conditioned on.

Thus, even

in those cases where significant correlates of attrition are found, they

explain very little of the variation in attrition probabilities in the

data.

One implication of this result is that weights based on these

equations would,

in all likelihood, have little effect on estimated

As we stressed in Section II,

all these y variables are

potentially "endogenous"

in the sense that they might be related to a

contemporaneous y of interest,

and adding more lagged y variables to the

attrition equations increases the chances of capturing such endogeneity.

But it is only through the existence of such endogeneity that weights

can reduce attrition bias.

The R-squared measure we use is defined in the footnote to the

Table and is a common measure of fit in binary-choice models.

This

measure has recently been shown to have desirable properties relative to

other measures (Cameron and Windmeijer,

1997) and can be interpreted as

the proportionate reduction in uncertainty from the fitted model, where

uncertainty is defined by an entropy measure.

outcome

equations.34

We conclude from these results that the unconditional effects of

labor income,

welfare participation,

and marital status significantly

covary with attrition probabilities,

consistent with our conclusions

from the tabular analysis in Tables 2-4 (although the BGLW form of the

test,

reported next,

corresponds more closely to Tables 2-4).

However,

we also find that, in a majority of the cases,

these effects fall to

insignificance at conventional levels when a sufficiently broad set of

covariates are conditioned on.

The main exceptions to this occur for

various specifications of labor income models, particularly for male

heads but occasionally as well for female heads and for women in general

and for the occasional other model.

Thus these results provide support

for some concern for cross-sectional attrition bias in the PSID for

unconditional distributions, and for conditional distributions for

earnings,

especially of male heads.

BGLW Tests.

we noted in Section II, the inversion of our

attrition

probits-- the effect of future attrition on 1968 outcome

variables,

rather than the other way around--is also of interest.

Such

regressions were estimated by Becketti et al.

(1988) and used as a test

for attrition bias.

As we noted previously,

apart from nonlinearities

and some differences in the stochastic assumptions, the results should

have the same general tenor as the attrition

probits

but will show more

directly the degree to which regression coefficients in typical outcome

equations are affected.

This statement must be qualified because even weights with very

small variance could have a large impact if they are sufficiently highly

correlated with the error term and the regressors.

Table 8 shows 1968 log labor income regressions for male

heads.35

Separate regressions are estimated for individuals who were always in

the sample through our final year, 1989, and for the total sample in

1968.

We compare the total sample and the nonattriting sample--not

attritors and nonattritors

--because the issue is how different parameter

estimates would be from those in the total sample if only the

nonattriting sample is

used.36

We show results separately when the

SE0

sample is included and excluded. For male heads, none of the

coefficients on the variables of most past research interest--Black,

Ed<12,

College Degree, Age and Age-Squared

--are

significantly different

between the total and nonattriting samples in estimates including the

SEO, and the magnitudes of the differences in the coefficients are

seldom large from a substantive research point-of-view. Significant

differences do appear for the "Other Race" and "Some College" variables

(and one of the region variables), for reasons we have not been able to

determine. More significant difference appear for the estimates when

the

SE0

is excluded, but these are again not large in magnitude. In

summary,

at least for SRC-SE0 combined sample, we find very few

important effects of attrition on the

coefficients.37'3e

Individuals with zero labor income are excluded.

While this

introduces some noncomparability with our attrition

probits

as well as

raising well-known selection issues,

we wish to maintain correspondence

with the bulk of the earnings function literature, which also generally

conditions on positive income.

The two sets of differences are transforms of one another, but

they have different standard errors.

Under the null of equality of the

true coefficient vectors,

the variance of the difference in the

coefficients is the difference in the separate variances (the variance

in the smaller sample must be larger, necessarily, under the null).

Similar findings were reported by BGLW.

However, their

analysis only went through 1981 and, in addition, they tested the

difference in coefficients between attritors and nonattritors whereas we

properly test between the total sample and nonattritors.

In our background report (Fitzgerald et al.,

1997a),

we show

estimates of labor income equations for wives and female heads; marital

status

probits

for men and women;

and welfare-status

probits

for female

heads,

all estimated in 1968 separately for the total and nonattriting

samples.

For wives,

the labor income results are essentially similar to

those for men although some significant differences in the magnitude

(though not the sign) appear for the education coefficients. For female

heads,

the only significant labor-income differences are for the

coefficients on age, but the separate coefficients for the total and

nonattritor samples are each insignificant (a sign that female heads

have very flat age-earnings profiles),

so it is not clear how important

this result is.

In the marital-status

probits,

some significant

differences appear for men (Black coefficient) and women (education

coefficient), generating some what more concern for these outcome

variables than for labor income.

The welfare

probits

show no

significant differences in any of the coefficients.

Wald tests for the joint significance of the differences in all

slope coefficients and intercepts generally reject the hypothesis of

equality between the vectors. However,

when test are conducted for the

equality of the slope coefficients allowing the intercepts to differ,

most fail to reject equality.

The estimated intercept differences

(i.e.,

constraining all coefficients on the other regressors to be the

same for the two groups) are shown in Table 9.

Thus we conclude that,

while the coefficients on

"standard" variables in labor income and

welfare-participation equations and, to a lesser extent, marital-status

We calculated White standard errors for the coefficients but

found them to be only 5 percent higher, at most, than those shown. We

therefore do not calculate them for the remainder of the analysis.

equations,

are unaffected by attrition, there are still be differences

in the levels of these outcome variables conditional on the regressors.

IV. Cross-Sectional Comparisons to Census Data

The second piece of our analysis is to compare cross-sectional

distributions and regression coefficients between the PSID and the CPS,

allowing us to conduct a more direct analysis of the existence of

attrition bias for these types of variables.

Comparing the PSID and the

CPS has some difficulties, however.

The

most

important is that the

sampling frames are not identical,

for the CPS includes individuals and

families who have immigrated to the U.S. since 1968, while the PSID

excludes those

families.3g

We will find this issue to be of some

importance and, consequently,

we will present some tabulations on the

characteristics of immigrants since 1968 taken from the Decennial Census

in 1990. Second,

many of the variables are defined differently in the

two data sets (headship, for example,

as well as labor income) and hence

this will generate some noncomparability.

Tables 10 and 11 show PSID-CPS comparisons for male heads 25-64 in

1968 and 1989, respectively.

Table 10 compares the two data sets in

1968, and is thus relevant to the issue of whether the approximate

25-

percent nonresponse in the drawing of the PSID sample systematically

biased the first wave of the

data.

The table indicates that the

distributions of age, race, education, marital status, and regional

location in the CPS and PSID were roughly in line in 1968, both for the

SRC sample and the combined (weighted) SRC-SE0

sample.40

A few

The PSID

Latin0

supplemental sample,

which includes a few

immigrants, was not begun until 1990.

miscellaneous divergences appear (e.g., in the educational distribution)

which may be a result of different questionnaire wording.

As for labor

force and earnings,

neither the CPS nor the PSID have unbracketed

variables for weeks worked or hours in 1968, so only the fraction of

those with positive weeks worked can be compared, and in this dimension

the PSID again lines up with the CPS. In addition, the PSID

unfortunately did not obtain an unbracketed earnings variable in 1968 so

we must rely on a measure of labor income, which includes some earned

income other than wages and

salaries.41

The means of the two earnings

measures are about $1,000 apart in the two data sets, and a bit farther

apart if the SRC sample is used. Whether this is a result of the

difference in the measures cannot be ascertained.

The table also shows

measures of dispersion in the two data sets,

although these are also

contaminated by the differences in measures. The log variance of

earnings is considerably smaller in the PSID than in the CPS, but the

measures of percentile points are not far apart, suggesting that

differences at the very lowest percentiles are driving the

difference.42

Statistical tests for the differences in the distributions almost

always reject equality of the distributions because the standard errors

The PSID weights in 1968 were not obtained from direct

post-

stratification against Census or CPS distributions, but were derived

from combining the weights from the University of Michigan's SRC

sampling frame and the Census Bureau's

SE0

sampling weights.

The

weights for the combined SRC-SE0 sample were set to make the combined

SRC-SE0 sample representative.

The PSID procedure for creating labor income is described in

Institute for Social Research (1972,

pp.307+).

We exclude from our

calculations those with zero wage and salary income and those who said

on a separate question that they were self-employed.

Our CPS wage and

salary measure therefore also excludes individuals with self-employment

income.

The log variance is sensitive to changes in the lower tail of

the distribution.

from the CPS, with its very large sample sizes, are extremely small.

However,

the magnitudes of the differences in most of the variables are

small from a substantive research point of view, so we shall continue to

make comparisons along this dimension rather than through formal

statistical

tests.43

Table 11 shows the comparable distributions in 1989. In this table

we show two columns for the combined SEO-SRC PSID sample, one using 1968

weights and one using the 1989 weights calculated by the PSID staff and

including an attrition

adjustment.44

Some differences between the PSID

and CPS appear but they are not large,

and are often narrowed slightly

by the weights.

For example,

the higher attrition rate for blacks can

be seen from the slightly lower percent black for the

1968-weight

PSID

(.07)

versus the current-weight PSID

t.09).

The

SAC-only

sample is the

worst

(.06),

no doubt because no attrition-adjusted weights have been

calculated for that sample.

Nevertheless, both for race and for age,

education,

marital status, and region,

the differences between the CPS

and the PSID, and among the different PSID samples, is quite small and

gives an overall impression of fairly strongly continued

representativeness of the PSID for male heads, even through 1989.

In addition, the PSID has a wage and salary earnings variable in

1989 which can be compared to that in the CPS, allowing a better

comparison between the data sets on this score than was the case for

1968.

In 1989 the two are within $500 of each other, only half of the

However,

on the more important issue of differences in

regression coefficients,

we will rely more heavily on tests of

differences.

See below.

The construction of these attrition-adjusted weights is

described in Institute for Social Research (1992,

pp.82-98).

The

variables included in the attrition equation are age, gender, race,

education,

number of children, region, lagged family income, and others.

$1000 difference in 1968.

The continued difference with the labor

income variable suggests that much of the 1968 difference was indeed a

result of noncomparability of variables.

For earnings itself, the

current-weight PSID is the closest to the CPS, followed by the

1968-

weight PSID and followed by the SK-only,

which is the farthest from the

CPS.

As for dispersion,

the log variance measures in the PSID are still

smaller in 1989 when comparable measures are used (the SRC-only sample

continues to be the farthest from the CPS). Again, however, the

percentile point measures are reasonably close in the different data

sets, perhaps suggesting that the log variance measures are affected by

outliers at the bottom of the distribution.

It might also be noted that

the percentile measures show strong increases in dispersion over time

(compare Tables 10 and

111,

consistent with the evidence now recognized

of increasing earnings inequality among men in the U.S.

This

comparability was also noted previously by Gottschalk and Moffitt

(1992).

It is necessary to reconcile these findings, which indicate that

the PSID has roughly maintained representativeness through 1989 for the

unconditional means and distributions of major sociodemographic lines,

with those from the previous analysis indicating significant differences

between attritor and nonattritor unconditional characteristics in 1968

(Tables

2-4).45

Taking both results at face value, they necessarily

imply that the differences in the value of the variables for the two

Actually,

the differences are a bit exaggerated because Tables

2-4 compare attritors to nonattritors instead of the total sample to

nonattritors,

which is the implicit comparison in the CPS analysis. At

an approximate attrition rate of

SO%,

the differences shown in Tables 2-

4 should be halved for comparison with the CPS. This by itself reduces

the perceived seriousness of the discrepancy somewhat.

samples in 1968 must have converged over time.

Further investigation of

this possibility reveals it to indeed be the case, as we demonstrate in

Tables 12 and 13. Table 12 shows the characteristics of PSID males who

were 25-40 in 1968 and therefore were 46-61 in 1989, but including in

the 1968 sample only those men who responded in 1989; consequently, the

sample is composed of the same individuals in both years (unlike Tables

10 and 11, the former of which include some men who have attrited or

died by 1989 and the latter of which includes a second generation).

The

table also shows CPS tabulations of men in these same age groups in the

same years. It is clear that,

while time-invariant characteristics such

as race must necessarily remain as far apart between the data sets in

1989 as they were in 1968,

this is not the case for time-varying

characteristics.

Indeed, the distributions of education and marital

status change over time for the PSID men in a way that reduces the

initial selection and moves the distributions closer to the CPS.

The

initial selection on relatively high-educated men in the PSID is offset

by a slower rate of growth of education over the life cycle among

nonattriting individuals in the PSID than in the CPS; and the initial

selection on married men is partly offset by a more rapid decline in

marriage rates in the PSID than in the CPS.

The analysis of earnings is

complicated by the noncomparability of measures, but the growth of labor

income in the PSID was much smaller than the growth of earnings in the

CPS,

thus partly offsetting the initial selection on relatively

high-

income men in the PSID.

The simplest explanation for this pattern is that the time series

processes for education, marital status,

and earnings contain a serially

correlated component which at least partly regresses to the mean, and

that selection is at least partly based on that component.

The

existence of

ARMA

errors,

after a time-invariant or even unit root

component has been controlled for, has been amply demonstrated in the

literature on earnings dynamics

(MaCurdy,

1982; Abowd and Card, 1985;

Moffitt and Gottschalk, 1995);

the transitory components in these models

do not fade out very quickly over time, at least in levels.

In our next section, where we more directly examine attrition dynamics,

we will show explicitly that attrition is based upon lagged shocks which

are deviations from average levels, although contemporaneous shocks

cannot be directly examined.

A similar regression-to-the-mean effect appears to be at work in

the PSID across generations,

although milder in magnitude (see

Fitzgerald et al.,

1997b,

for a fuller examination of intergenerational

attrition issues). Table 13 shows the original Table 11 for 1989 split

out between those 25-45 and those 46-64; the former were mostly children

in 1968 and hence constitute the 'second generation' that was implicitly

contained in Table 11. The CPS-PSID differences are often slightly

narrower for the younger generation than for the old, as can been seen

from the percent with less than 12 years of education, the percent

married and the percent owning a home. The pattern is not uniform

across all categories, however.

Nevertheless,

for many categories the

data are consistent with an intergenerational model with similar

serially-correlated mean-regressing components.

Returning to Table 11,

it can be seen that a second explanation

for the comparability with CPS is a small role played by the updating of

the PSID weights for attrition on observables.

The PSID staff readjusts

its weights over time to take into account both differential mortality

by age, race,

and sex but also differential nonresponse (Institute for

Social

Research,1992,pp.82-98).

The latter adjustment is based on an

estimated nonresponse model in which nonresponse probabilities for

different time intervals since 1968 are made a function of past

socioeconomic characteristics such as age, race, sex, income, family

structure,

urban-rural location, and regional location.

The predicted

nonresponse probabilities from the model are used to adjust the weights

for each member of the sample on the basis of his or her

characteristics.

This procedure is capable, in principle, of adjusting

for attrition on observables, as discussed above in Section II, even

though these are "universal"

weights rather than model-specific

weights.46

Comparison of the columns for current-weight and

1968-weight

estimates in Table 11 shows that this adjustment has an effect on the

PSID means for only a few variables. The adjustments are generally

(though not always) in the "right" direction--that is, to move the PSID

means closer to those in the CPS.

This is particularly the case for the

race distribution, where the percent white is improved by this

adjustment. The labor force and income variables are likewise moved

slightly toward the CPS by the weight

adjustment.47

Nevertheless, the

We state

"in principle" because it is necessary that the

nonresponse model be properly specified for the adjustment to restore

representativeness. It is worth emphasizing that no outside benchmarks

from the CPS or other data set are used for these nonresponse

adjustments. The adjustments are all "internal," and result only in a

multiplication factor being applied to the prior year's weights to

obtain current weights.

See n.44.

However,

the table also suggests a problem with the PSID weight

because time-invariant characteristics, such as race, are capable of

perfect attrition adjustment because the true population means of those

variables must be the same as they were in 1968; hence it is easy to

calculate a weight that perfectly restores the 1968 mean. But if the

weights are based on nonresponse models which are parametric functions

of several variables (like race), and hence smooth over them, the

resulting weights will never fully adjust any single variable, even

time-invariant ones.

This is a problem with all universal weights.

magnitude of the changes resulting from the weight adjustment are

generally quite small.

The major reason for this result is that,

despite the correlation of observables with attrition propensities,

attrition remains mostly noise.

This was clear from the low R-squared

values reported in our attrition

probits.

The variances of the predicted

attrition rates from those

probits

are small,

which necessarily implies

that the variance of attrition-adjusted weights is small; weighting may

have little effect in this case (subject to the caveat mentioned

previously).

Although we have now provided explanations for the closeness of

the CPS and PSID cross-sectional distributions, we note that there are

some remaining differences.

These can be further narrowed once

immigration into the U.S. since 1968 is accounted for. The importance of

immigration is illustrated in Table 11,

which shows means for male heads

in 1989 taken from the 1990 Decennial Census Public Use Microdata sample

(PUMS).

Although the CPS did not, as of 1989, ask date-of-immigration

questions,

the Decennial Census did so. The PUMS figures in the table

introduce some additional complications because the PUMS means without

immigrants are not always equal to those of the CPS, in part because of

sampling error in the CPS and in part because the 1989 CPS sampling

frame is based on that of the 1980, not the 1990, Census.

Nevertheless,

in several instances the PUMS tabulations indicate that

immigrant/non-

immigrant differences in characteristics are in the direction that would

explain some of the CPS-PSID differences.

Immigrants are

disproportionately nonwhite,

for example, possibly explaining the

remaining gap between the CPS and PSID; and immigrants have lower labor

force activity and earnings,

consistent with the direction of the

PSID-

CPS gap (i.e.,

higher labor force activity and earnings levels in the

PSID).

Thus,

while the evidence is not conclusive, it does suggest that

immigration is part of the explanation for the remaining PSID-CPS

difference for some variables.

CPS-PSID comparisons for other demographic groups--wives, female

heads,

male non-heads, and female non-heads (see our background report)

indicate that the results for wives are quite similar to those for male

heads and,

if anything, the CPS-PSID differences are even smaller.

The

results for female heads show again small CPS-PSID differences, with a

few exceptions.

We conclude from this examination, therefore, that, despite the

seemingly large differences in characteristics of attritors and

non-

attritors in the PSID,

it nevertheless remains cross-sectionally

representative of the non-immigrant U.S. population.

CPS-PSID Regression Comparisons.

Table 14 shows estimates of

cross-sectional log earnings equations for male heads in the PSID and

CPS in 1968,

1981, and 1989, using current-year values for the

independent variables as well as dependent variable.

In general, the

differences in parameter estimates are larger than might be expected on

the basis of the unconditional means which,

as we just demonstrated, are

quite close to one another. The regression coefficients in the three

years show generally similar signs but a number of differences are

sizable in magnitude.

Two of these--the "other race" and "some

college

"--are probably due to differences in definitions of other race

and of post-high-school

education.4*

The same type of differences appear

In the PSID, "Hispanic"

was coded as a racial category prior to

1985 whereas in the CPS, "Hispanic" comes from a separate ethnicity

question.

For our regressions, we recoded "Hispanic" to "white" in the

PSID in years prior to 1985.

For the

"Some College" variable, the

treatment of junior colleges and vocational schools is different in the

two data sets. On the other hand, these coefficients are also those for

for earnings regressions of wives and female heads (see background

report).

Table 15 shows F and chi-squared statistics for the significance

of the differences between PSID and CPS earnings regressions as well as

probit

equations for marital status and welfare participation in each

year.

For the log earnings regressions for male heads, both the full

set of coefficients, those excluding the constant, and those excluding

the constant and the regional coefficients are significantly different

in the two data sets in 1968.

However,

interestingly, the size and

significance of the test statistics tends to fall over time, in general.

Indeed, by 1989, the coefficients other than the constant and region are

insignificantly different in the two data sets. This finding suggests

that attrition is not the cause of these differences in coefficient

vectors. We speculate that the initial selectivity of who consented to

be a part of the PSID (a 25-percent nonresponse rate) could have

generated the 1968 differences we observe.

That the dissimilarity then

tends to fade out over the length of the PSID may be the result of the

regression-to-mean phenomenon we demonstrated earlier for the

unconditional means. This is an area for future research."

The test statistics shown in Table 15 generally show somewhat

similar patterns in the test statistics for other demographic groups and

which differences appeared in Table 8.

Becketti et al.

(1988) found the same result: through 1981, the

F-statistics for the difference in earnings regression coefficients

(they did not examine other dependent variables) tended to fall over

time.

They speculated that the cause might be a result of their

inclusion of nonsample individuals after 1968.

However,

we exclude

nonsample individuals and find the same result, so we conclude that the

pattern is a result of something else. We should also note that the

patterns in Table 15 are unaltered by either the exclusion of the

SE0

sample or estimation without weights.

for other dependent variables although the size of the statistics is

sometimes smaller and sometimes larger. For the earnings equations for

both wives and female heads, the coefficients in the two data sets are

insignificantly different from one another when the constant is excluded

(and when both the constant and the region coefficients are excluded) in

all three years. For the other dependent variables, the test statistics

are larger than for earnings but, like the male head earnings

statistics,

generally fall over time.

In addition,

in 1989 not a single

test statistic for any group or any dependent variable is significant

when coefficients other than the constant and region are

compared.50

In any case,

the major finding of our analysis is that, while the

PSID-CPS differences in regression coefficients are larger than would be

expected after our examination of the unconditional means, these

differences go back to 1968. Further investigation, particularly of the

causes of the initial, 1968 difference, would be warranted in future

research.

V. Dynamic Attrition Models

In the final piece of our analysis, we explore the dynamic

attrition issues we discussed in Section II concerning the effects of

permanent and transitory components of lagged

"y"

variables and make

use,

in general,

of the full y-history by estimating year-by-year

attrition hazards through 1989.

This exercise has interest for two

This general pattern of falling test statistics might be

thought to be partly the result of declining sample sizes, but in fact

the combined CPS-PSID sample size increases over time because the CPS

has been gradually expanded over time,

and more than enough to outweigh

PSID attrition.

reasons.

First, for the development of weights based on estimated

attrition functions, these equations may be superior to those based only

on the levels of the 1968 variables.

However,

given the results of our

analysis thus far,

attrition bias in the PSID does not appear to be very

severe for cross-sectionally defined variables.

The second reason is

therefore more important,

for these equations have implications for

attrition bias in equations used in past and future PSID studies which

use dynamic, or panel-defined,

outcome variables rather than

cross-

sectionally-defined ones (earnings and employment dynamics, welfare and

marital status transition models, etc.).

"y"

in our models in

Section II is reinterpreted as such a dynamic outcome variable, then

that analysis implies that if lags of those variables are significant

determinants of attrition then analyses which attempt to model the

contemporaneous values of those variables on the nonattriting sample may

produce inconsistent parameter estimates (namely, if the lagged values

of those variables covary with the contemporaneous values). Because

there is no counterpart to the CPS for panel-defined variables in the

PSID, this can be our only (indirect) test of attrition bias for PSID

dynamic analyses.

Although we have not developed a formal model of the causes of

attrition,

it is plausible to hypothesize that not only are

low-

socioeconomic-status individuals likely to

attrite

(as our results on

levels of the relevant variables have demonstrated thus far) but also

that individuals with a recent change in earnings, marital status, and

other variables are more likely to

attrite.

Taking this notion one step

further,

we hypothesize that individuals observed over their full past

history to have had above-average rates of fluctuations in earnings,

above-average numbers of transitions in marital status, or above-average

rates of geographic migration--to take the three which we will examine--

are more likely to

attrite.

We conjecture that it is plausible to

suppose that disruption in general may be related to attrition because

it may make individuals either more difficult to locate by the PSID

field staff, or less receptive to participation in the panel, or both.

To investigate this issue, we estimate attrition functions with a

latent index of the form:

Ait

f(Y

jJt-ltYi,t-2'

'*sf

YiO)

Xi00

Vit

(16)

where the outcome variable,

Ait,

equals 1 if the individual attrites at

time t,

conditional on still being a respondent at t-l.

The vector

Xi0

consists of time-invariant

"x"

variables, with coefficient vector

Eqn(13) allows the lagged dependent variables to affect current

attrition propensities in a general way (function f) but, in our

empirical work,

we test functions which transform the lagged y into only

four different summary variables: (a) the individual-specific mean of

the variable over all years since 1968; (b) the individual-specific

variance of the variable over all years since 1968; (c) deviations of

lagged variables from the individual-specific means; and (d) durations

of time spent in various states defined by the variables in question.

The first of these measures tests whether attrition is affected by

individual-specific mean levels of earnings, marital status, and other

variables (we include family structure and geographic mobility as well).

This analysis should yield broadly similar findings to those in Section

III above,

for they only replace the 1968 values of these variables with

their means over a period of years.

The second of the statistics

measures individual heterogeneity in turnover (labor market, marital,

geographic location, etc.).

As we noted previously, if attrition

covaries

with lagged values for these variables, then it follows that

models estimated on nonattritors but using the contemporaneous

counterparts to these measures as dependent variables (turnover,

durations,

transition rates,

etc.) will be biased provided that the

contemporaneous and lagged measures

covary

as well.

The third of the

measures tests whether lagged changes ("shocks") to these variables

affect attrition.

This is logically separate from the question of

individual heterogeneity in turnover. It relates closely to the issue

of whether transitory events affect later attrition, although we cannot

be sure of that interpretation because we cannot, by definition,

determine whether recent events will persist in the future or not if the

individual attrites (and hence whether the events will, in retrospect,

be seen to be permanent or transitory shocks).

This analysis has

implications for bias in the estimation of transition rate models for

contemporaneous variables on the nonattriting sample. The fourth

measure is more familiar and tests whether durations in a state

(marriage,

migration) affect attrition propensities; these equations

have implications for the estimation of contemporaneous models for the

length of spells.

For our models we pool all observations on individuals 25-64 in

original 1968 sample families for all years 1970-1989 for which they are

observed.51

We estimate

logits

for whether the individual attrites in

We omit 1968 and 1969 so that we can construct at least two

lagged variables for individuals last observed in 1970.

We also make

no adjustment to the standard errors for the pooled nature of the data

(relatedly,

as we noted earlier, there are no adjustments for unobserved

heterogeneity).

However,

year-by-year estimation of the models reveal

qualitatively similar results;

hence the standard error issue does not

the next period as a function of the four summary measures discussed

above defined as of the current period.

We also include 1968 variables

for education, age,

and other socioeconomic characteristics.

In some

runs we include year dummies, which fully capture duration dependence.

Table 16 shows a series of estimated attrition equations focusing

on lagged earnings.

Column (1) shows that attrition propensities for

men are significantly negatively affected both by lagged mean earnings

as well as earnings in the prior period.

The latter implies that

negative deviations of current earnings from mean earnings raise the

likelihood of attrition. Column (2) shows that the effect of deviations

does not extend back beyond the current period. Column (3) tests the

effect of the individual-specific variance and finds that attrition

rates are positively affected by variances, even conditioning on current

period and lagged mean earnings. Column (4) shows that this result is

robust to the inclusion of age and year dummies, for it might be the

case that if attrition rates vary with calendar year or age, this might

create spurious estimates since earnings vary with year and

age.52

However,

column (5) shows that the inclusion of several standard

socioeconomic variables (education, race, etc.) is sufficient to render

insignificant the effect of lagged mean earnings on attrition rates, a

result not surprising inasmuch as permanent earnings are likely to be

more predictable by such regressors than are earnings deviations or

earnings variances.

The latter two remain significant even after

inclusion of the additional regressors.

The last column shows, in

addition,

that there are no significant effects of this kind for women.

affect our conclusions.

The year dummies show no significant duration dependence in the

hazard after 1970.

We speculate that earnings are not as good a predictor of instability of

other behaviors for women as for men because there are considerably more

planned fluctuations in earnings for

women.53

These results, therefore, are consistent, at least for men, with

attrition being selective on stability.

Therefore it should be expected

that measures of second moments,

of turnover and hazard rates, and of

related variables should be smaller in the nonattriting PSID sample than

in the population as a whole.

Tables 17 and 18 show that this result extends to marital, family

structure,

and migration behavior.

Table 17 demonstrates that men

recently experiencing a transition out of marriage (due to divorce,

separation,

or widowhood) are more likely to

attrite

than those not

experiencing such a transition.

In addition,

men who have experienced

larger numbers of marital transitions in the past are more likely to

attrite.

Interestingly, however,

no effects of this kind appear for

females.

Table 18 shows that men who have split off from other families

are more likely to attrite--

although the effects are insignificant when

other characteristics are controlled

--and that men who have moved

recently or who show a high average propensity to move are more likely

attrite.

Again, however,

no significant effects appear for women.

Although these results clearly demonstrate a tendency for men with

more unstable histories to

attrite,

the seriousness of the problem for

the PSID is difficult to judge. The R-squared values in these attrition

equations are uniformly very small, as shown in the tables, which

implies that attrition along these dimensions may not have a large

We thank a referee for suggesting as well that the female

results may reflect the existence of married-couple households in which

the husband's earnings is the dominant factor affecting the family's

attrition.

effect on the comparable contemporaneous measures on the nonattriting

sample from selection on these observables. This cannot be known for

certain because the size of the bias depends not only on the R-squared

values,

but also on the size of relation of these lagged instability

measures with both the regressors in the main outcome equation of

interest and with the error term in that equation (recall the model of

Section II).

However,

weights based on these equations could be

developed which would capture dynamic effects more adequately than do

the current,

universal PSID weights,

and these could be used in

specification tests to see the importance of their effect on estimates

of outcome equations.

Nevertheless,

this approach cannot capture any

bias from selection on unobservables in such equations (unfortunately,

as previously noted, there is no equivalent to the CPS for these

variables with which to gauge the presence of such selection).

VI. Conclusions

Our study of attrition in the PSID has yielded several findings:

The observed baseline characteristics of those who later do and

do not

attrite

from the PSID are quite different; these differences are

often statistically significant.

Attritors tend to have lower earnings,

lower education levels,

lower marriage propensities, and appear to be

generally drawn from the lower tail of the socioeconomic distribution.

These unadjusted differences fall in magnitude and are usually

rendered statistically insignificant as determinants of attrition

propensities after conditioning on a number of other socioeconomic

characteristics. In one leading case, however--

earnings

for male heads-

-a significant relationship continues to exist even after such

conditioning.

In a regression context, attrition appears to primarily affect

intercepts rather than slopes of regressions for earnings and welfare

participation, but also some slopes for marital-status regressions.

Cross-sectional comparisons of unconditional moments between

the PSID and the CPS show a close correspondence all the way through

1989.

We reconcile the seemingly inconsistent findings of, on the one

hand,

significant measured correlates of attrition and, on the other

hand, continued cross-sectional representativeness by showing that

regression-to-the-mean effects are present that cause initial

differences in characteristics to fade away over time both within and

across generations.

A small role is also played by PSID weights used to

adjust for attrition related to observables, although, because attrition

is mostly noise,

the weights do not alter PSID means by a very large

amount.

We also find that some portion of the remaining CPS-PSID

difference is a result of the exclusion of individuals who have

immigrated to the U.S.

since 1968 from the PSID sampling frame.

Regression coefficients in models for earnings, marital status,

and welfare participation in the CPS and the PSID are usually quite

similar in sign and magnitude but not always so,

and the differences in

coefficient vectors as a whole are usually significant in the baseline

year (1968).

However,

the test statistics for the difference in

coefficient vectors fall over time and imply that, by 1989, the CPS and

PSID coefficients are insignificantly different as a whole.

We find evidence that attrition propensities are correlated

with individual-specific levels of turnover and instability in earnings,

in marital status,

and in geographic mobility.

We also find that recent

unfavorable events along these dimensions--a drop in earnings,

a marital

dissolution, or a geographic move--induce more attrition.

The

magnitudes of the effects of these variables on attrition,

as measured

by R-squareds, are not large, which suggests that they are unlikely to

induce significant bias in studies which have such dynamic measures as

outcome variables. As noted earlier, however, this conclusion depends

on model specific correlations, and we recommend that authors of these

types of studies be aware of possible attrition biases and check the

sensitivity of their results accordingly.

APPENDIX

Let

f(y,z(x)

be the complete-population joint density of y and

and let

g(y,zlx,A=O)

be the conditional joint density.

Then

(Y,

z,A=Olx)

(y,

zlx,A=O)

Pr(A=Olx)

(A=Oly,

2,~)

(y,

zlx)

Pr(A=OJx)

Pr(A=Olz,x)

f(y,zlx)

Pr(A=Olx)

(Y,

zlx)

w(z,x)

where

w(z,x)

is given in eqn (9) in the text.

Hence

(Y,

z/x)

w(z,x)

g(y,zlx,A=O).

Integrating both sides over z gives eqn (8) in the text.

REFERENCES

Abowd, J. and D. Card.

Hours Changes."

"On the Covariance Structure of Earnings and

Econometrica 57 (March 1989): 411-445.

Amemiya, T. Advanced Econometrics. Cambridge: Harvard University

Press, 1985.

Becketti, S.; W. Gould; L. Lillard; and F. Welch. "The Panel Study of

Income Dynamics after Fourteen Years: An Evaluation."

Journal of

Labor Economics 6 (October 1988): 472-492.

Bound, J.; C. Brown; G. Duncan; and W. Rogers. "Evidence on the

Validity of Cross-Sectional and Longitudinal Labor Market Data."

Journal of Labor Economics 12 (July 1994): 345-68.

Cameron,

A.C. and F.A.G. Windmeijer.

"An R-squared Measure of Goodness

of Fit for Some Common Nonlinear Regression Models." Journal of

Econometrics 77 (April 1997): 329-342.

Cosslett, S.

"Estimation from Endogenously Stratified Samples." In

Handbook of Statistics, Vol. 11, eds. G.S. Maddala, C.R. Rao, and

H.D. Vinod, eds.

Elsevier, 1993.

DuMouchel,

W. and G. Duncan.

"Using Sample Survey Weights in Multiple

Regression Analysis of Stratified Samples."

Journal of the

American Statistical Association 78 (September 1983): 535-543.

Duncan,

G. and D. Hill.

"Assessing the Quality of Household Panel

Data:

The Case of the Panel Study of Income Dynamics."

Journal

of Business and Economic Statistics 7 (October 1989): 441-452.

Fitzgerald, J.; P. Gottschalk; and R. Moffitt. A Study of Sample

Attrition in the Michigan Panel Study of Income Dynamics.

Mimeographed,

Johns Hopkins University, 1997a.

Fitzgerald, J.;

P. Gottschalk; and R. Moffitt.

The Impact of Attrition

on the Estimation of Intergenerational Relationships in the PSID.

Mimeographed, Johns Hopkins University, 1997b.

Gottschalk, P. and R. Moffitt.

"Earnings and Wage Distributions in the

NLS, CPS, and

PSID."

Part I of Final Report to the U.S.

Department of Labor. Providence: Brown University, 1992.

Hausman,

J. and D. Wise. "Attrition Bias in Experimental and Panel

Data: The Gary Income Maintenance Experiment."

Econometrica 47

(March 1979): 455-474.

"Stratification on Endogenous Variables and

Estimation:

The Gary Income Maintenance Experiment." In

Structural Analysis of Discrete Data with Econometric

Applications, eds.

Manski

and D. McFadden.

Cambridge: MIT

Press, 1981.

Heckman,

"Sample Selection Bias as a Specification Error."

Econometrica 47 (January 1979): 153-162.

Heckman,

"Selection Bias and Self-Selection." In The New Palgrave:

A Dictionary of Economics, eds.

Eatwell,

M. Milgate, and P.

Newman,

Vol.IV.

London: Macmillan, 1987.

Heckman,

J. and V. J. Hotz.

"Choosing Among Alternative Nonexperimental

Methods for Estimating the Impact of Social Programs.'* Journal of

the American Statistical Association 84 (December 1989): 862-874.

Heckman,

J. and R. Robb. "Alternative Methods for Evaluating the

Effects of Interventions."

In Longitudinal Analysis of Labor

Market Data, eds. J.

Heckman

and B. Singer. Cambridge: Cambridge

University Press, 1985.

Hill, M. The Panel Study of Income Dynamics: A User's Guide.

Park, Ca.:

Sage Publications, 1992.

Newbury

Hirano,

K; G. Imbens; G. Ridder; and D.

Rubin.

"Combining Panel Data

Sets with Attrition and Refreshment Samples." Mimeographed,

Harvard University, September 1996.

Holt,D.;

T.M.F. Smith; and P.D. Winter.

"Regression Analysis from

Complex Surveys."

Journal of the Royal Statistical Society

Series A 143 (1980): 474-487.

Horowitz,

J. and C.

Manski.

"Censoring of Outcomes and Regressors Due

to Survey Nonresponse:

Identification and Estimation Using

Weights and Imputations." Journal of Econometrics, forthcoming.

Imbens,

G. and T.

Lancaster.

"Combining Micro and Macro Data in

Microeconometric Models."

Review of Economic Studies 61 (October

1994):

655-680.

Imbens,

G. and T. Lancaster. "Efficient Estimation and Stratified

Sampling."

Journal of Econometrics 74 (October 1996): 289-318.

Imbens,

G. and J. Hellerstein.

"Imposing Moment Restrictions from

Auxiliary Data by Weighting."

Mimeo,

1996.

Harvard University, July

Institute for Social Research.

Design, Procedures,

A Panel Study of Income Dynamics: Study

Years,

Volume I.

and Available Data, 1968-1972 Interviewing

Ann Arbor, Michigan, 1972.

Institute for Social Research.

A Panel Study of Income Dynamics:

Procedures and Tape Codes, 1989 Interviewing Year, Vol. I,

Procedures and Tape Codes. Ann Arbor, Michigan, 1992.

Little,

R. and D.

Rubin.

Statistical Analysis with Missing Data. New

York: Wiley, 1987.

MaCurdy,

"The Use of Time Series Processes to Model the Error

Structure of Earnings in a Longitudinal Data Analysis."

Journal

of Econometrics 18 (January 1982): 83-114.

Maddala,

G.S.

Limited-Dependent and Qualitative Variables in

Econometrics. Cambridge:

Cambridge University Press, 1983.

Madow,

W.; I. Olkin; and D.

Rubin,

eds.

Incomplete Data in Sample

Surveys,

3 volumes.

New York: Academic Press, 1983.

Manski,

"The Selection Problem."

In Advances in Econometrics:

Sixth World Congress, Vol.

University Press, 1994.

I, ed. C. Sims. Cambridge: Cambridge

Manski,

C. and S. Lerman. "The Estimation of Choice Probabilities from

Choice-Based Samples."

Econometrica 45 (November, 1977): 1977-

1988.

Moffitt, R. and

Gottschalk.

"Trends in the Autocovariance Structure

of Earnings in the U.S.: 1969-1987."

Working Paper 355, Johns

Hopkins University, 1995.

Nijman,

T. and M. Verbeek.

"Nonresponse in Panel Data:

Estimates of a Life Cycle Consumption Function."

The Impact on

Journal of

Applied Econometrics 7 (July-September 1992): 243-257.

Powell, J.

"Estimation of Semi-Parametric Models." In Handbook of

Econometrics, Vol. IV, eds. R. Engle and D. McFadden.

Amsterdam

and New York:

North-Holland, 1994.

Rao,

C.R.

"On Discrete Distributions Arising Out of Methods of

Ascertainment."

In Classical and Contagious Discrete

Distributions, ed. G.P. Patil.

Calcutta:

Society, 1965.

Statistical Publishing

C.R.

"Weighted Distributions Arising Out of Methods of

Ascertainment:

What Population Does a Sample Represent?"

In A

Celebration of Statistics, eds.

A.Atkinson and S.Fienberg. New

York:

Springer-Verlag, 1985.

Ridder, G.

"Attrition in Multi-Wave Panel Data." In Panel Data and

Labor Market Studies, eds. J. Hartog, G. Ridder, and J. Theeuwes.

Amsterdam:

North-Holland, 1990.

"An Empirical Evaluation of Some Models for Non-Random

Attrition in Panel Data." Mimeographed,

University of Groningen,

1992.

Van den Berg, G.; M. Lindeboom; and G. Ridder. "Attrition in

Longitudinal Panel Data and the Empirical Analysis of Dynamic

Labour

Market Behavior." Journal of Applied Econometrics 9

(October-December 1994): 421-435.

Verbeek, M. and T. Nijman.

"Incomplete Panels and Selection Bias." In

The

Econometics

of Panel Data, eds. L. Matyas and P. Sevestre.

Dordrecht: Kluwer, 1996.

Figure

Attrition Hazards: Sample With No New Entrants

0.12

0.1

0.08

0.06

0.02

Table 1

Response and Nonresponse Rates in the PSID

Year

Remaining in Sample

Attritors=

from

In a

In an

Total

As a

Total

Fam.

Died

Moved

Non-

Family Insti-

Pet

Unit

Resp.

Unit tution

1968

Non-

Total

Resp.

1968 17807

384 18191 100.0

1969

15561

367

16028

88.1

1970

15126 333

15459

85.0

1971

14767

322 15089

82.9

1972

14400 293 14693

80.8

1973 13969

307

14276 78.5

1974

13581

307

13888

76.3

1975

13226

302

13528

74.4

1976

12785 291 13076 71.9

1977

12377 310

12687

69.7

1978

12078 320

12398

68.2

1979

11718

316 12034

66.2

1980

11357

305

11662

64.1

1981

11022 340

11362

62.5

1982

10780

326 11106

61.1

1983

10487

322

10809 59.4

2163

1797

(.119)

(-099)

600 351

(.037)

(-022)

404

208

(.026)

(-013)

429 190

(.028) (-013)

449

247

(.031)

(-017)

410

229

(.029) (-016)

386

200

(.028) (-014)

487

310

(.036)

(-023

411

234

(.031)

(.018)

330

210

(.026)

(.017)

387

224

(.031)

(.018)

405

233

82 33

(.034)

(.019) (.007) (-007)

337

208

(.029)

(.O

-8)

285

135

(.025)

(.012)

336 194

(.030)

(.017)

(.OOS)

282

t-016)

175

(.005)

(.Oll)

101

(.006)

(.007)

115

124

(.008)

100

(.007)

(.006)

(.007)

(.006)

(.007)

(.005)

(-004)

(.006)

(-007)

(.007

(.008)

102

(.007)

92 22

(.006)

(-004)

(.005)

(.007) (.005)

Year

In a

In an

Total

As a

Total

Fam Died

Moved In

Fam.

Inst.

Pot

Att.

Non-

from

Unit

of 68

Resp

Non

Resp

1984

10178

319

10497

1985

9891 275 10166

1986

9517

292

9809

1987

9230

257

9487

1988

9002

206

9208

1989

8743

170

8913

57.7

348

225

(.032)

(.021)

(.009) (.003)

55.9

371

(-035)

53.9 390

(.038)

52.2 357

(.036)

50.6

310

(.033)

49.0

323

(.035)

229

(-022)

275

(-027)

215

(.022)

178

(.019)

212

(.023)

(.009)

(-008)

(.OlO)

(.009)

(-004)

(.003)

(.005)

(.004)

(-003)

Notes:

Excludes new births and nonsample entrants.

Figures in parentheses show attrition rates

as a percent of the total sample

remaining in the prior year (column four).

Table 2

1968 Characteristics by Attrition Status:

Male

Heads,

Age 25-64

Always

Ever Out Ever Out/

Ever

Out/

Not Dead

Dead

Welfare Participation

(%)

Marital Status(%):

married

never married

widowed

divorced/separated

Percent with Annual

Hours Worked

Annual Labor Income

Annual Labor Income for those

w/income

Annual Hours Worked for those

w/hours

Variance of log annual labor

income for those w/income

Labor income quintile ratios for

those w/labor income

Quintile

20/median

Quintile

40/median

Quintile

60/median

Quintile

8O/median

Education (%):

12-15

16+

0.8

1.3

1.4

1.2

95.8 90.1*

87.1

98.1

2.4

3.7*

4.9

0.4

0.3

1.5*

2.0

0.1

1.2 4.6*

5.9

1.3

98.7 94.lf

95.7

89.8

21345

21631

17011 17277

16298

18152

18106

18281

2378 2246

2268

2182

.248 .529

.481

.667

-658

886

1.101

1.392

31.5

32.8

15.8

19.9

611

615

-905

923

1.139

1.123

1.498

1.462

52.5*

50.8

25.6*

27.3

11.5*

11.5

10.4*

10.4

.558

.865

1.164

1.493

57.2

21.0

11.5

10.4

Race

(%):

White

Black

92.7

6.6

88.3*

87.4

90.7

10.7*

11.5

8.0

Region

(%):

Northeast

24.7 25.8 26.9

22.3

North Central

32.2

27.5*

26.5

30.1

South

26.7

30.1*

29.6

31.2

West

16.4 16.7

17.0

15.7

Table 2 continued

Always In Ever Out Ever Out/

Ever Out/

Not Dead

Dead

Age

40.7

45.6*

Tenure(%): Own home

74.9 62.9*

Rent

21.5

33.8*

Number of Children in Family

2.0 1.5

Sample Size 1238 1533

Notes:

Sample weights used.

*: Significantly different from "Always In" at 10% level.

43.1 52.1

58.0 75.9

38.9 20.2

1.6

1.3

1116

417

Table 3

1968 Characteristics by Attrition Status: Wives,

Age 25-64

Always In Ever Out Ever Out/

Ever Out/

Not Dead

Dead

Welfare Participation

(%)

Percent With Annual Hours

Worked

Annual Labor Income

Annual Labor Income for those

w/income

Annual Hours Worked for those

w/hours

Variance of log annual labor

income for those w/income

Labor income quintile ratios for

those w/labor income

Quintile

20/median

Quintile

40/median

Quintile

60/median

Quintile

80/median

Education

(%):

12-15

16+

Race

(2):

White

Black

Region

(%)

: Northeast

North Central

South

West

Age

Tenure

(%):

Own home

Rent

Number of Children in Family

Sample Size

1.1

47.7

36308

7653

1311 1315

1342

1173

1.546 1.624

1.548

2.014

.240

-800

1.205

2.000

30.5

49.1

10.7

9.8

92.0

7.4

23.9

31.7

28.0

16.5

40.9

77.8

18.8

2.0

1377

1.6

1.4

2.2

44.0

44.4

42.3

3299 3366

2960

7509 7580

7128

.218

.222

611

622

1.164 1.667

1.637 2.078

45.6*

44.7

38.7*

39.9

10.2

9.6

5.5*

5.8

89.5*

90.0

9.4*

8.7

27.3*

28.3

26.4*

25.1

31.2* 31.8

15.1

14.8

44.5*

43.5

69.1*

67.9

28.5*

29.6

1.5

1.6

1043

847

.216

.557

1.195

1.670

50.0

32.8

12.7

4.5

86.6

12.5

22.4

32.8

27.9

16.9

49.6

75.5

22.6

1.4

196

Notes:

Sample weights used.

. Significantly different from “Always In" at 10% level.

Table 4

1968 Characteristics by Attrition Status: Female Heads,

Age 25-64

Always In Ever Out

Ever

Out/

Ever

Out/

Not Dead

dead

Welfare Participation

(%)

Marital Status

(%):

married

never married

widowed

divorced/separated

Percent With Annual Hours

Worked

Annual Labor Income

Annual Labor Income for those

w/income

Annual Hours Worked for those

w/hours

Variance of log annual labor

income for those w/income

Labor income quintile ratios for

those w/labor income

Quintile

20/median

Quintile

40/median

Quintile

60/median

Quintile

80/median

Education

(%):

12-15

16+

Race (8):

White

Black

Region

(%):

Northeast

North Central

South

West

Age

Tenure

(%):

Own home

Rent

Number of Children in Family

Sample Size

4.3 10.5*

10.0

17.9

1.4

1.8 1.7

2.5

21.2 14.6*

14.7

13.1

38.7 39.1

39.1

39.0

36.7 40.8

40.6

43.8

80.4 67.4*

67.0

73.7

8199

10214

6950 7167

3482

10296 10679

4723

1593

1645 1676

1203

1.426

1.185 1.045

1.739

316

.737

1.163

1.553

45.1

28.3

13.8

12.8

-424

800

1.178

1.468

49.2

32.4

9.6*

8.8*

.471

.438

.838

.653

1.178

2.483

1.440

5.724

46.8

88.4

33.7

11.6

10.2

0.00

9.3

0.00

80.3 76.0* 77.3

55.4

18.8 23.2*

21.9

44.6

25.2

26.2 26.3

24.8

30.0 24.6* 25.6

9.3

25.8 27.7

25.9

57.5

19.0 21.4

22.2

8.4

44.9 47.4*

47.2

50.4

45.0

40.3 40.3

40.7

50.3 55.9* 55.8

58.2

1.3 1.0 1.0

1.8

502

526

475

Notes:

Sample weights used.

. Significantly different from "Always

In"

at 10% level.

Table

Ever-Out Attrition

Probits

Male Heads Age 25-64,

Focus on Labor Income

Model 1

Model 2

Model 3 Model 4

Coeff.

aP/ax

Coeff.

ap/ax

Coeff.

ap/ax

Coeff.

ap/ax

Intercept

.334’

.128

1.059)

.360'

.139

1.770*

t.096)

(.454)

.671

1.130*

t-518)

Labor

Incomea)

-.0239*

-.0092

(.0030)

0272*

0105

(.0103)

.254

.lOO

(.177)

.009

.003

-.0192

(.0108

-.0237*

(.0120]

No Labor

Income

La;;:a:zpT=-

Black

.284*

-110

t.1601

.291

l.180)

-.0073

.110

.181

(.186)

.018

t.025)

f.026)

.074

f.066)

.006

.022

t.026)

.028

.037

(.081)

Other Race

-356

t.248)

.134

.198

t.251)

Age

-.088f

(.022)

-.033 -.039

c.024)

Age Squared

Education < 12

Years

Some College

.107*

t.025)

.041

054f

[:028,

.076

.208*

i.071)

-.114

t.096)

-.043

-.195*

i.097)

College Degree

-.305*

(.107)

-.116

-.384'

.109)

Northeast

-.051

t.9391

North

-.139

Central

(.091)

South

-.120

t.088)

In SE0 Sample -.070

t.080)

Lives in Rural

Area

(SMSA

1000)

-.271*

t.072)

Number of

Children in

Family

Presence of

Child

-.033*

(.017)

095

(:061)

Cwns

House

-.310*

f.068)

Might Move

in Future

-.015

t.072)

Income/Needs

Ratio

031

(:033)

.417

-.0088

.067

.008

.014

.073

-.014

.020

.077

-.072

-.142

-.019

-.051

-.044

-.025

-.lOO

-.012

.035

-.114

-.006

.012

Table 5,

continued

Model 1 Model 2 Model 3

Model 4

Coeff.

.028

ap/ax

Coeff.

ap/ax

Coeff.

ap/ax

Coeff.

ap/ax

.028 .044 .068

Sample

Size

2253

2253 2253

Number

1074

Ever Out

Loq Like.

-1516.05

-1515.99 -1490.27 -1453.02

Notes:

Excludes known dead. Characteristics measured in 1968.

Significant at 10% level.

aP/dX

signifies the effect of a unit change in the variable on the probability of attrition evaluated at

the mean.

R‘

equals one minus the ratio of the log likelihood of the fitted function to the log likelihood of a

function with only an intercept.

Coefficients multiplied by

103.

Coefficients multiplied by

108.

Coefficients multiplied by

102.

Table 6

Ever-Out Attrition

Probits:

Other Results

Model 2

Model 3

Coeff.

se/ax

Coeff.

se/ax

Model 4

Coeff.

de/ax

Wives, 25-64,

Focus on Labor Income

Labor

Income

No Labor

Income

Labor

Incgye

Squared

Female Heads, 25-64,

Focus on Labor Income

Labor

Income

No Labor

Income

Labor

Incgye

Squared

Men,

25-64,

Focus on Marital Status

Married

Widowed

Divorced/

Separated

i.0166)

0010

.0004

.0056

.0021 .0016 .0006

(.0168)

t.0172)

.133

.051

.128

.048

-135 .049

l.083) l.085) (.086)

011

i.073)

.004 .021 .008

030

.Oll

(.074) i.075)

-.OOlO

-.0004

-.0018

-.0007

-.0035 -.0013

1.0195) (.0201) (.0214)

438f

i.125)

.171

424*

.162

i.128)

424*

.160

i.133)

009

i.073)

004

0186

i.074)

.007

033

i.078)

.012

-.436+

-.165

-.192

-.0710

-.156

-.058

l.134) (.140)

C.142)

-.130

-.049

054

i.238)

020

026

009

t.234)

i-239)

I:1911

259

-.098

.255

.094

,288

.106

(.193)

C.194)

Women, 25-64,

Focus on Marital Status

Married

-.182f

-.069 -.036

-.014

-.039 -.015

t.1011

C.104)

C.106)

Widowed

-.024

-.009

0425

i.125)

.0160

.065

024

f.123) t.126)

Divorced/

Separated

090

(:112)

.034

114

(:114)

.043

131

(:115)

.049

Table 6 continued

Model 2

Model 3

Model 4

Coeff.

se/ax

Coeff.

Be/ax

Coeff.

de/ax

Female Heads, 18-54,

Focus on Welfare

Welfare

.270* .106

.214

.083 .0704

.027

Receipt

1.139)

(.143) (.149)

Notes:

Excludes known dead. Characteristics measured in 1968.

Significant at 10% level.

dP/aX

signifies the effect of a unit change in the variable on the probability of attrition evaluated at

the mean.

Coefficients multiplied by

103.

Coefficients multiplied by

10'.

Table 7

Ever-Out Attrition

Probits

Multiple Focus Variables

Labor Income

Female Heads

Men

Women

18-54 25-64

18-54

Coeff.

apm

Coeff.

apm

Coeff.

awax

Labor

Incomea)

No Labor Income

Marital

Statusc)

Married

Widowed

Divorced/

Separated

Welfare

Welfare Receipt

-.0350

-.0130

l.0022)

,::z;

162

-.003 -.OOl

(.008)

-- --

;1;:4,

053

.249*

-094

l.121)

070

;.149,

027

-.0199*

-.0073

f.0120)

.203

(.179)

071

006

-.156

-.060

(.142)

iz9,

009

.288

(.194)

106

-.239

-.088

(.213)

0000

.221*

-082

(-071)

000

i.001)

000

-.039

-.015

l-106)

i.126)

065

024

i.115)

131

049

;";;9,

031

Notes:

Excludes known dead. Characteristics measured in 1968.

Significant at 10% level.

aP/aX

signifies the effect of a unit change in the variable on

the probability of attrition evaluated at the mean.

Other variables included are those in Model (4) in Table 5.

a)Coefficients

multiplied by

103.

b)Coefficients

multiplied by

108.

Omitted category for female heads is never-married.

Table 8

1968 Log Labor Income Regresssions

Male Heads

SRC and SE0 Combined

SRC Only

Total

Always In

Difference

Total

Always In

Difference

Intercept

Black

Other Race

Some College

College Degree

Age

Age Squared

Northeast

North Central

South

Sample Size

F-statistic

Variance of

Error

8.24*

(.197)

-.249*

(-044)

-.221

(.136)

-.293*

(.034)

(2;;

.271*

(.043)

,:::i;

948*

(.108)

,x:

-.076*

(.039)

8.38*

(.232)

-.272*

(.056)

-.246

(.173)

-.271*

l.039)

068*

(:039)

.283*

(.045)

.074*

(-011)

-.856*

(-132)

(

ia:;

006

(.;)43)

-.105*

(.045)

.14

t.121

-.022

(.035)

.196*

(.106)

.023

(.019)

-.033*

(.014)

012

(.;)ll)

-.059

(.061)

(

,:3;

-.039*

(.020)

-.028

(.023)

8.28*

t.23)

-.173*

(.055)

-.393*

(0.164)

-.291*

(.040)

(:E;

,:::i;

080*

(:oll)

-.947*

(.125)

(

(.il:i;

-.111*

(.045)

8.35*

(0.26)

-.195*

(0.070)

-.193

(-184)

-.244*

(.045)

,::i:;

,:Z;

(

-.922*

(.149)

(

-.056

(.048)

-.147*

(.051)

2182 1159

1406

788

.19

.24

-22

.26

50.5

35.7

38.8

27.8

.326

.220

.285

.194

.08

(.13)

-.022

(.043)

.200*

(.0830)

(

-.005*

(.OOl)

(

-.OOl

(.007)

-.022

(-023)

-.069*

(.021)

-.036

(.025)

votes:

Standard errors in parentheses.

Sample excludes known dead.

SRC+SEO sample are weighted.

Significant at 10% level.

E;Coefficients

multiplied by

103.

F-statistic for hypothesis that all coefficients except the intercept are

equal to zero.

Table 9

1968 Income,

Marital Status, and Welfare Equations:

Difference in Total and Always-In Samples, Intercept-Only Model

SRC+SEO SRC Only

Labor Income Regressions:

Male Heads

Wives

Female Heads

-.059*

-.053*

(.012)

(.013)

.016

(.028)

;":',4,

Marital-Status

Probits:

Men

-.232*

Women

Welfare-Status

Probits:

(.037)

(.044

-.063*

-.078

(.022)

(.028

Female Heads

-.264*

-.383*

(.087) (.186)

Notes:

Models include all variables shown in Table 8 but allow the intercept

to differ for the Total and Always-In Samples.

Coefficient equals

Total-

Sample intercept minus Always-In Sample intercept.

Standard errors in parentheses

Sample exludes known dead

SRC+SEO is weighted

*Significant at the 10% level

Table 10

Characteristics of Male Heads 25-64: 1968

PSID and CPS

e!F

Race

White

CPS

43.7

. 91

PSID

Weighted

Unweighted

(SRC and SEO)

(SRC

only)

43.3

43.6

.90

-91

Black

.09

-08

Hispanic

Education

Less than

.43

.29

13-15

.ll

.14

16+

. 15

-15

.41

.30

.14

-15

Marital Status

Never married

03 .03

Married

-92

Divorced/separated

.03 .03

Widowed

.Ol

Region

Northeast

.25

-25

North Central

.28

.30

South

.29

-28

West

.18

Own Home

.69

.03

.94

.03

-01

.22

.31

.30

.17

.71

Table 10 continued

Labor Force

CPS

PSID

Weighted

Unweighted

(SRC and SEO)

(SRC only)

Positive weeks

worked

-96 -96

.96

Weeks

workeda)

Annual hours

workeda)

Earningsa)

Real wage and salary

$19478

Real labor income

$20460

$20709

Wage and Salary

Distribution"'

Log variance

.452

.389

.354

20th

Percentile

671

. 667

.667

40th

Percentile

:886

.893

.907

60th

Percentile

1.114

1.087

1.107

80th

Percentile

1.429

1.373

1.400

Welfare Participation

Notes:

.Ol

a) Workers only.

b) PSID figures use labor income rather than wage and salary income.

Table 11

Characteristics of Male Heads 25-64: 1989

PSID, CPS, and PUMS

PUMS CPS

PSID

with

without

Current Wgts.

1968 Wgts.

Unweighted

immignts

(SRC and SEO)

(SRC

and SEO)

(SRC

Only)

Race

White

Black

Hispanic

Education

Less than 12

13-15

16+

Marital Status

Never married

Married

Divorced/separated

Widowed

Region

Northeast

North Central

South

West

Own Home

42.4

42.7

42.0

42.0 42.0

42.2

-08

-93

-09

.07

-06

03 .02

.Ol

-17

-16

.17

.28

.29 .36

.29

.27

.28

.19

.23

.27

.27 .28

.29

.18

.17

.29

.23

.30

.31

.lO

.79

.lO

.Ol

.lO

.79

.Ol

.08

.Ol

.81

.09

.Ol

.09

.08

.81

.82

.09

.Ol

-20

.25

.34

-21

-72

.19

.26

-35

.20

.22

.23

.25

.28 .28

.34

.31

-21

-18 -18

-71

.73

.74

.20

.32

.17

.75

Table 11 continued

PUMS CPS PSID

with without

Current Wgts.

1968 Wgts.

Unweighted

irrunignts

immignts (SRC and

SEO)

(SRC and SEO)

(SRC Only)

Labor Force

-92

.94

Positive weeks

worked

Weeks

workeda) 48.1

Annual

ours

worked

2156

48.3

2164

49.0 46.6

46.6

46.7

2165 2172

2176

2199

Earnings

Real wage and

salary

$24239 $24582

$22970 $23481

$23645

$23905

Real labor income

$24090

$24273

$24537

Wage and Salary

Distribution"'

Log Variance

.63 -61

.624

-501

.491

-452

Ratios of Percentile

Points to Median

20th Percentile

.557 .571

.566

40th Percentile

.857

.886

.868

60th Percentile

1.117 1.143 1.132

80th Percentile

1.500 1.525 1.509

Welfare

Participation

.02

-02 -02

.582 .571 .589

.873 -873

.875

1.163 1.143 1.143

1.519 1.500 1.500

.Ol -01 .Ol

Notes:

Workers only.

Table

Characteristics

of Males 25-40

1968 and 46-61

in 1989

PSID and CPS

CPS

PSID

25-40 46-61

25-40

in 1968

46-61

in 1968

in 1989

%!?

Race

White

Black

Education

Less than 12

13-15

16+

Marital Status

Never married

Married

Divorced/separated

Widowed

Region

Northeast

North Central

South

32.4

53.1

32.8

53.8

-09

.31

-38

.25

.24

-12

.04

-80

West

.24

.21

.28

-25

-30 -35

-18

-19

Own Home

.25

-34

.22

.95

-01

.26

.25

.30

.28

.29

.31

-86

-92

.27

.18

.26

-02

-86

-10

-02

Table 12 continued

CPS

25-40

46-61

in 1968

in 1989

PSID

46-61

in 1989

Earningsb)

Real wage and

salary

$18429

$24694

$25464

Real labor income

$21265 $24638

Notes:

PSID sample includes SE0 and SRC and both years use 1968 weights.

Sample includes only those responding in 1989.

Workers only.

Table 13

Characteristics of Male Heads 25-45 and 46-64 in 1989

PSID and CPS

Age 25-45

Age 46-64

CPS

PSID

CPS

PSID

Race

White

Black

Education

Less than 12

13-15

16+

Marital Status

Never married

Married

Divorced/separated

Widowed

Region

Northeast

North Central

South

West

Own Home

34.9

34.8

54.6

55.3

-08

-36

-30

-22 .26

.30

.32

.25

-25

.28

-29

-17

.26

.76

-04

.84

-00

.78

.09

-00

.02

-01

.88

.02

.20

-25

.34

.22

-64

.21

.28

.31

-21

-25

-66

.83

-25

-28

.29

.17

Table 13 continued

Earningsa)

Age 25-45

Age 46-64

CPS PSID

Real wage and

salary

$22096

$23162

$24878 $25262

Real labor income

$23622

$25890

Notes:

PSID sample uses SRC-SE0 and 1968 weights.

Workers only.

Table 14

PSID and CPS Log Earnings Regressions: Male Heads

1968

1981

1989

PSID

CPS

PSID CPS

PSID

CPS

Intercept

Black

Other

Low Ed

Some

College

Grad

Age

Squared'

Northeast

North

Central

South

8.642*

(.015)

-.229*

(.032)

-.102

l.099)

-.288*

(.026)

,:E,

.247*

(.032)

( :

z:;

-.007*

(.OOl)

054*

(:029)

092*

(.028)

102*

(:029)

8.456*

(.065)

-.393*

(.014)

-.264*

(-040)

-.271*

(.OlO)

.119

(.014)

.248*

(.013)

.070*

(-003)

-.008*

(-004)

035*

(:012)

(:Z,

-.177*

(.012)

8.478*

(-086)

-.159*

(.043)

-.244*

(-037)

.016*

(.033)

.293*

(.036)

063*

(:009)

-.006*

(.OOl)

-.016

(.035)

-.067*

(.034)

7.545*

(.071)

-.283*

(.016)

-.210*

(-030)

-.313*

(.012)

101*

(:013)

.263*

(-012)

105f

(:003)

-.011*

(.OOl)

060*

(:014)

046*

(:013)

-.039*

(.013)

8.066*

(.067)

-.278*

(-048)

046

(:125)

-.140*

(.046)

167*

039)

.442*

(-040)

078*

c:o11,

-.008*

(.OOl)

(

z;:;

-.067

(.041)

-.099*

(.041)

7.560*

(.080)

-.241*

(.017)

-.210*

(-028)

-.366*

(.015)

(

.390*

(.012)

101*

(:004)

-.011*

(.OOl)

,:z;

057*

(:015)

-.013

(.014)

Standard errors in parentheses

significant at 5% level

Combined SRC-SE0 sample (weighted) is used for PSID

Omitted categories for dummies are white, 12 years of education, and West.

'Coefficients multiplied by 10

Table 15

Significance Tests for CPS-PSID Differences

1968

1981

1989

Earnings:

Male Heads

All Coeffs

All Coeffs but Const.

Region

Earnings:

Female Heads

All Coeffs

All Coeffs but Const.

Region

Earnings: Wives

All Coeffs

All Coeffs but Const.

Region

Marital Status: Males

11.3*

3.7*

4.1*

8.9*

5.6*

2.5

3.4*

3.0

4.0

2.8*

2.6*

3.9*

1.2

1.3

1.6

1.5

2.2

1.5

8.1*

4.8*

1.5

0.9

2.4

1.5 0.9

2.3

All Coeffs

All Coeffs but Const.

Region

Marital Status: Females

124.6* 96.4*

96-l*

23-O* 23.5*

18.3*

14.7* 22.0*

13.6

All Coeffs

All Coeffs but Const.

Region

Welfare Part: Female Heads

21.1*

16.2

27.1*

20.5*

9.1

22.1*

7.5*

8.7

13.5

All Coeffs

107.7* 25.8"

28.7*

All Coeffs but Const.

42.0*

23.9*

18.4

All Coeffs but Const.

Region

33.2*

17.2*

14.2

Notes:

Earnings statistics are F-statistics;

are likelihood ratio statistics.

marital status and welfare participation

*: significant at the 5 percent level

Table 16

Dynamic Attrition Models With Focus On Lagged Earnings

(Logit

Coefficients)

Males

Females

(1)

(2)

(3)

(4)

(5)

(6)

-.20*

(.07)

-.24*

t.08)

-.22* -.17*

t.06) t.08)

-.09

(

09)

-.28*

t.08)

-.18*

t-06)

-.26*

t.08)

-.20*

t-06)

-.07

09)

-.15*

(.07)

.23

(.14)

-.11

(.ll)

yt-1

yt-2

Var

(y)

;“Z,

;?Z,

38*

i.09,

-.04

t.23)

Time Dummies

and age

Other

Characts.

n n

-018

.017 .020

.025

.043

.018

Notes: Dep. var.

is 1 if individual attrites in next period, 0 if not.

the mean earnings from 1968 to current period;

ytwl

and

yt-2

are earnings in

the current period and one period back;

and

var(y)

is the variance of earnings

from 1968 to the current period.

The coefficients on the first three

variables are multiplied by

lo4

and the coefficient on the fourth is

multiplied by

108.

Standard errors in parentheses.

For R-squared definitions, see Table 5.

significant at the 10 percent level.

a) Education, race, region, age of youngest child, rural residence, homeowner.

Table 17

Dynamic Attrition Models With Focus On Lagged Marital Status

(Logit

Coefficients)

Males

Females

(1)

(2)

(3)

(4)

(5)

-.24

t.19,

yt-1

-.81*

I.151

"tr

Duration

Other

Characts.

Pseudo

.022

-.22

I.201

-.31

(

19)

-.72* -.67*

t.151 t.151

.20*

t.051

.024

-.04*

t.011

023

-.21

t.201

-.72*

t.16)

.21*

l.09)

i.02)

-043

-.14

f.19)

-.15

l.17)

39,

-.Ol

(-02)

009

Notes:

Dependent variable is the same as in Table 16.

is the average

probability of being married from 1968 to the current period;

ytsl

is a

married dummy for the current period;

ntr

is the number of marital transitions

from 1968 to the current period; and, 'duration' is the number of years since

the last marital transition. All equations contain age and year dummies.

Standard errors in parentheses. For R-squared definitions, see Table 5.

*: significant at the 10 percent level.

See Table 16.

Table 18

Dynamic Attrition Models With Focus On Splitoff and Migration

(Logit

Coefficients)

(1)

Splitoff Migration

Male

Female

Male

Female

(2)

(3)

(4)

(5)

(6)

(7)

Splitoff

Split in t-l

Ever Split Off

.73*

.74*

t.371 t.371

.28*

t-151

Migration

yt-1

Duration

Other Characts.

n n

Pseudo

006

.007

i.37)

-.04

(

.16)

.036

-.05

t.591

.oo

t.18)

017

;‘,“*,,

ix,

-.02*

t.011

015

.77*

t.301

.28*

(-12)

-.Ol

t-021

.040

-.02

t.36)

i.13)

-.oo

t.01,

.017

Notes :

Dependent variable is the same as in Table 16.

is the average

number of moves from 1968 to the current period;

ytml

is a

dun-my

for having

moved in the current period; and 'duration'

is the number of years since the

last move.

All equations include age and year dummies.

Standard errors in parentheses. For R-squared definitions, see Table 5.

*: significant at the 10% level.

See Table 16.