Chapter 1

Introduction
• What are longitudinal and panel data?
• Benefits and drawbacks of longitudinal data
• Longitudinal data models
• Historical notes
 1.1 What are longitudinal and panel data?
• With regression data, we collect a cross-section of subjects.
– The interest is comparing characteristics of the subject, that is,
investigating relationships among the variables.
• In contrast, with time series data, we identify one or more
subjects and observe them over time.
– This allows us to study relationships over time, the so-called
dynamic aspect of a problem.
• Longitudinal/panel data represent a marriage of regression
and time series data.
– As with regression, we collect a cross-section of subjects.
– With panel data, we observe each subject over time.
• The descriptor panel data comes from surveys of individuals; a
panel is a group of individuals surveyed repeatedly over time.
 Example 1.1 - Divorce rates
• Figure 1.1 shows the 1965 divorce rates versus AFDC (Aid
to Families with Dependent Children) for the fifty states.
– The correlation is -0.37.
– Counter-intuitive? - we might expect a positive relationship
between welfare payments (AFDC) and divorce rates.

1965 Divorce Rates to AFDC Payments

Divorce
Rates 5

0
20 120 220
AFDC Payments
 Example 1.1 - Divorce rates
• A similar figure shows a negative relationship for 1975 (the
correlation is -0.425)
• Figure 1.2 shows both 1965 and 1975 data, with a line
connecting each state
– The line represents a change over time (dynamic), not a
cross-sectional relationship.
– Each line displays a positive relationship - as welfare
payments increase so do divorce rates.
– This is not to argue for a causal relationship between
welfare payments and divorce rates.
• The data are still observational.
• The dynamic relationship between divorce and AFDC is
different from the cross-sectional relationship.
Figure 1.2 1965 and 1975 Divorce
rates versus AFDC
Comparing 1965 and 1975 Divorce
Rates to AFDC Payments
Divorce 8
Rate
7

0
0 50 100 150 200 250 300 350

AFDC Payments
 Some notation
• Longitudinal/panel data - regression data with “double
subscripts.”
• Let yit be the response for the ith subject during the tth time
period.
• We observe the ith subject over t=1, ..., Ti time periods, for
each of i=1, ..., n subjects.
– First subject - (y11, y12, ... , y1T1 )
– Second subject - (y21, y22, ... , y2T2 )
– ...
– ...
– The nth subject - (yn1, yn2, ... , ynTn )
 Prevalence of panel data analysis
• Importance in the literature
– Panel data are also known as “cross-section time series” data in the social
sciences
– Referred to as “longitudinal data analysis” in the biological sciences
– ABI/INFORM - 326 articles in 2002 and 2003.
– The ISI Web of Science - 879 articles in 2002 and 2003.
• Important panel data bases
– Historically, we have:
• Panel Survey of Income Dyanmics (PSID)
• National Longitudinal Survey of Labor Market Experience (NLS)
– Financial and Accounting
• Compustat, CRSP, NAIC
– Market scanner databases
• See Appendix F
Appendix F. Selected Longitudinal
and Panel Data Sets
• Table F.1 – 20 International Household Panel Studies
• Table F.2 – 5 Studies focused on youth and education
• Table F.3 – 4 Studies focused on the elderly and retirement
• Table F.4 – 7 miscellaneous studies, including
– election data,
– manufacturing data,
– medical expenditure data and
– insurance company data
 1.2 Benefits and drawbacks of
longitudinal data
• Several advantages of longitudinal data compared to
– data that are either purely cross-sectional (regression) or
– purely time series data.
• Having longitudinal data allows us to:
– Study dynamic relationships
– Study heterogeneity
• Reduce omitted variable bias
• With longitudinal data, one can also argue
– Estimators are more efficient
– Addresses the causal nature of relationships
• Main drawback - attrition
 Dynamic relationships
• Static versus dynamic relationships
– Figure 1.1 showed a cross-sectional (static)
relationship.
• We estimate a decrease of 0.95 % in divorce rates for
each $100 increase in AFDC payments.
– Figure 1.2 showed a temporal (dynamic)
relationship.
• We estimate an increase of 2.9% in divorce rates for
each $100 increase in AFDC payments.
• From 1965 to 1975, AFDC payments increased an
average of $59 and divorce rates increased 2.5%.
 Historical approach
• In early panel data studies, pooled cross-sectional data were
analyzed by
– estimating cross-sectional parameters using regression
and
– using time series methods to model the regression
parameter estimates, treating the estimates as known
with certainty.
• Theil and Goldberger (1961) provide an early discussion on
the advantages of estimating these two aspects
simultaneously.
 Dynamic relationships and time
series analysis
• When studying dynamic relationships, univariate time series
methods are the most well-developed.
– However, these methods do not account for relationships
among different subjects.
– Multivariate time series accounts for relationships
among a limited number of different subjects.
– Time series methods requires a fair number (generally, at
least 30) observations to make reliable inferences.
 Panel data as repeated time series
• With panel data, we observe several (repeated) subjects for each time period.
– By taking averages over subjects,
• our statistics are more reliable
• we require fewer time series observations to estimate dynamic patterns.
– For repeated subjects, the model is
yit =  + it, t=1, ..., Ti, i=1, ..., n.
• Here,  is the overall mean and it represents subject-specific dynamic
patterns.
– “Unfortunately,” we don’t get identical repeated looks.
• We hope to control for differences among subjects by introducing explanatory
variables, or covariates.
– A basic model is yit =  + xit´  + it, where xit is the explanatory variable.
• Introducing explanatory variables leaves us with only subject-specific
dynamic patterns, that is, yit - ( + xit´ = it
 Heterogeneity
• Subjects are unique.
– In cross-sectional analysis, we use yit =  + xit´ + it
• ascribe the uniqueness to " it ".
– In panel data, we have an opportunity to model this uniqueness.
– The model yit = i + xit´ + it is
• unidentifiable in cross-sectional regression.
• In panel data, we can estimate  and 1, .., n.
• Subject-specific parameters, such as i, provide an important
mechanism for controlling heterogeneity of individuals.
• Vocabulary:
– When {i} are fixed, unknown parameters to be estimated, we call this a
fixed effects model.
– When {i} are drawn from an unknown population, that is, random
variables, we call this a model with random effects.
 Heterogeneity bias
• Suppose that a data analyst mistakenly uses the model
yit =  + xit´ + it
when yit = i + xit´ + it is the true model.
– This is an example of heterogeneity bias, or a problem
with aggregation with data.
• Similarly, one could have different (heterogeneous) slopes
yit =  + xit´i + it
• or different intercepts and slopes
yit = i + xit´i + it
 Omitted variables
• Panel data serves to reduce the omitted variable bias.
• When omitted variables are time constant, we can still get
reliable estimates.
• Consider the “true” model yit =  + xit´ + zi´ + it.
– Unfortunately, we cannot (or not thought to) measure zi.
– It is “lurking” or “latent.” By considering the changes
yit* = yit - yi,t-1 = ( + xit´ + zi´ + it) - ( + xit-1´ + zi´ + it-1)
= (xit - xit -1 )´ + it - it-1) = xit* ´  + it*
– we do not need to worry about the bias that ordinarily
arises from the latent variable, zi .
• Introducing the subject-specific variable i, accounts for the
presence of many types of latent variables.
 Efficiency of Estimators
• Subject-specific variables i also account for a large portion of
the variability in many data sets
– This reduces the mean square error
– Increases the efficiency (or reduces the standard errors) of
our parameter estimators.
• With panel data, we generally have more observations than
with time series or regression.
• A longitudinal data design may yield more efficient estimators
than estimators based on a comparable amount of data from
alternative designs.
– Suppose that the interest is in assessing the average change in a
response over time, such as the divorce rate.
– A repeated cross-section yields Var y 1  y  2  Var y 1  Var y  2
– Longitudinal data design yields
Var y 1  y  2  Var y 1  Var y  2  2 Covy 1 , y  2 
 Causality and correlation
• Three ingredients necessary for establishing causality, taken
from the sociology literature:
– A statistically significant relationship is required.
– The association between two variables must not be due
to another, omitted, variable.
– The “causal” variable must precede the other variable in
time.
• Longitudinal data are based on measurements taken over
time and thus address the third requirement of a temporal
ordering of events.
• Moreover, longitudinal data models provide additional
strategies for accommodating omitted variables that are not
available in purely cross-sectional data.
 Drawbacks: Sampling Design (attrition)

• Selection bias
– may occur when a rule other than simple random
sampling is used to select observational units
– Example – “endogeneous” decisions by agents to join a
labor pool or participate in a social program.
• Missing data
– Because we follow the same subjects over time,
nonresponse typically increases through time.
– Example: US Panel Study of Income Dynamics (PSID):
• In the first year (1968), the nonresponse rate was 24%.
• By 1985, the nonresponse rate was about 50%.
 1.3 Longitudinal data models
• Types of inference
– Primary. We are interested in the effect that an (exogenous)
explanatory variable has on a response, controlling for other
variables (including omitted variables).
– Forecasting. We would like to predict future values of the response
from a specific subject.
– Conditional means.
• We would like to predict the expected value of a future
response from a specific subject.
• Here, the conditioning is on latent (unobserved)
characteristics associated with the subject.
• Types of applications - many
 Social science statistical modeling
• A model based on data characteristics is known as a
sampling based model. The model arises from a data
generating process.
• In contrast, a structural model is a statistical model that
represents causal relationships, as opposed to relationships
that simply capture statistical associations.
• Why bother with an extra layer of theory when considering
statistical models? Manski (1992) offers :
– Interpretation - the primary purpose of many statistical analyses is
to assess relationships generated by theory from a scientific field.
– Structural models utilize additional information from an underlying
functional field. If this information is utilized correctly, then in
some sense the structural model should provide a better
representation than a model without this information. (explanation)
– Particularly for public policy analysis, the goal of a statistical
analysis is to infer the likely behavior of data outside of those
realized (extrapolation).
 Modeling issues
• With subject-specific parameters, there can be many
parameters that describe the model
– “Fixed” versus “random” effects models
• Incorporating dynamic structure is important
– Econometric “dynamic” models (lagged endogenous)
versus serial correlation approach
• Linear versus nonlinear (generalized linear) models
– Marginal versus hierarchical estimation approaches
• Parametric versus semiparametric models
• We wish to separate the effects of:
– the mean
– the cross-sectional variance and
– serial correlation structure
 1.4 Historical notes
• The term ‘panel study’ was coined in a marketing context when
Lazarsfeld and Fiske (1938)
– Considered the effect of radio advertising on product sales.
– People buy a product would be more likely to hear the advertisement, or vice
versa.
– They proposed repeatedly interviewing a set of people (the ‘panel’) to clarify
the issue.
• Econometrics
– Early economics applications include Kuh (1959), Johnson (1960), Mundlak
(1961) and Hoch (1962).
• Biostatistics
– Wishart (1938), Rao (1959, 1965), Potthoff and Roy (1964) – used
multivariate analysis to consider the problem of polynomial growth curves of
serial measurements from a single group of subjects.
– Grizzle and Allen (1969) – introduced covariates