longitudinal research 1

8/6/2019 Longitudinal Research 1

1/34


2/34

IntroductionLongitudinal Research is touted as:

a panacea for establishing temporal ordermeasuring change

making

stronger

causal

interpretations

Pure cross sectional research :

For each case, the measurement of each variable occurs at a single time period, with contemporaneous


3/34

DefinitionLongitudinal research must be defined in terms of both the

data and the methods of analysis used in the research.

Data are collected for each item or variable for one or

more time periodsThe subjects or cases analyzed are the same or at least comparable from one period to the next

The analysis involves some comparison of data between or among periods


4/34

Types of Longitudinal ResearchProspective panel design: data may be collected at two or more distinct periods, for those distinct periods, on the same set of cases and variables in each period.Retrospective panel research : data may be collected at a

single period for several periods, usually including the period that ends with the time at which the data are collected.Repeated cross sectional design: data collected for the same set of variables for/at two or more periods, includes nonidentical but comparable cases for each period.


5/34

Purposes of Longitudinal Research

To describe patterns of change To establish the direction and

magnitude of causal relationships


6/34

Age , Period, and Cohort effects

Cross sectionally, age differences are between birth cohorts or age groups at a particular time.

Longitudinally, age differences may be interpreted as developmental differences within a single cohort or age group over time.


7/34

Age , Period, and Cohort effectsNeither a cross sectional nor a single cohort longitudinal study can eliminate both cohort membership and period effects as rival hypothesis, it would seem logical to combine

the two approaches and use a multiple year, multiple cohort design.The problem with this is that if we assume that the effects of age, period, and cohort membership are linear, controlling for any two variables controls for the third as well.


8/34

Age , Period, and Cohort effectsCohort (year of birth) =Period (calendar year) Age (years since birth)

Dummy variable regression analysis is an attempt to overcome the problem of linear dependence among age, period, and birth cohort.

Cohorts : aggregates of individuals, are units of analysis , cases for study.

Possible conceptualizations of cohort:Error or disturbanceDimension of generalization (use of cohorts as units of analysis rather than as theoretical variables)Theoretical and process variable


9/34

Age , Period, and Cohort effects

We do not measure aggregate characteristics of ages or periods as such, but we may measure aggregate

characteristics

of

cases

during

a

particular

age

or

period.

Ages and periods are aggregates of time, variables rather than units of analysis.


10/34

Age , Period, and Cohort effectsAge provides a developmental explanation for behavior.

Period provides explanation or at least a weak proxy that is historical in nature and that may help identify those historical events that most plausible as explanations for given behavior.

The effect of birth cohort measured as year of birth, may serve as proxy for the interaction of effects of age and period.


11/34

Cohort EffectsImpact of cohort size on a variety of social problems, including unemployment, divorce and crime.Age period are more appropriate as explanatory variables, age more so than period.One would eliminate period and cohort and replace them with the variables for which they act as a proxies in any causal analysis.Developmental, historical, and cohort membership effects cannot be clearly separated without longitudinal data.


12/34

Changes Over TimeSocial indicators of progress or decline in meeting basic social needs or achieving desirable social goals.

Age specific comparisons :only age specific comparisons, only those cases of a certain age in one year are compared with cases of the same age in some subsequent year.


13/34

Historical Trends in Relationships Between Variables

Examining changes in the strength or patterns of relationships over time, one important concern is the replication of previous result with new data .

Real changes in relationships : 1. Instability or unreliability of measurement

2. Misspecification of a causal model


14/34

5 Objectives of Longitudinal Research(Baltes & Nesselroade, 1979)1. Direct identification of intraindividual change (change from

one period to another)2. Direct identification of intraindividual similarities, or

differences in intraindividual change (whether individuals

change in the same or different ways)3. Analysis of interrelationships in behavior change (whether

certain changes are correlated with each other)

4. Analysis

of

causes

or

determinants

of

intraindividual change

(why individuals change from one period to another)5. Analysis of causes or determinants of interindividual

similarities or differences in intraindividual change (why different individuals change in different ways from one period to another)


15/34

Life Cycle and Developmental ChangesAbsence of pretest or baseline data has the effect of rendering uncertain whether differences after some treatment or intervention may be wholly attributable to the treatment or intervention, or to preexisting differences between the group that did and the group that did not receive the treatment or intervention.


16/34

Development Trends

in

Relationship

Between Variables

Cross sectional (intercohort)Longitudinal (intracohort)

With cross sectional data, such differences may be attributed to age or to intercohort differences; the extent to which the differences are developmental, as

opposed to period or intercohort differences, only can be examined longitudinally.


17/34

Causal RelationshipsCriteria essential to establish the existence of causal

relationship:1. The phenomena of variables in question must covary (by

differences between experimental and control groups, or by a nonzero

correlation between two variables)2. The relationship must not be attributable to any other

variable or set of variables (must not be spurious, but must persist even when variables are controlled)

3. The supposed cause must precede or be simultaneous with the supposed effect in time (as indicated by the change in the cause occurring no later than the associated change in the effect)


18/34

Causal RelationshipsEvidence for the first two criteria may be obtained from purely cross sectional or time ordered cross sectional data.

There is a possibility of a non recursive causal relationship.


19/34

Stage State

Analysis

of

Temporal

and

Causal Order for Qualitative Variables

One cannot necessarily infer that the first variable to change causes the second (the criteria of covariation and nonspuriouseness must still be met), but such a test would provide evidence that the second variable to change did not

cause the first.

Stage State analysis might be used; changes measured as

dichotomies (yes change has occurred, no change has not occurred)


20/34

Temporal Order of Measurement In linear panel analysis, any variables between which we are trying to

ascertain causal order are treated as dependent (endogenous) variables and are measured for at least two periods (waves).

The time between the periods for which the variables are measured is

called the measurement interval (not to be confused with interval scales or interval measures)

At least one prior value of each endogenous variable (a lagged endogenous variable) can be included as a possible predictor of the most recent value of that same variable.

Linear panel model may include either or both of lagged and instantaneous effects


21/34

Temporal Order of MeasurementHypothetical Causal relationship Between X and Y:

X1

Y1 Y2

X2 X3

Y3


22/34

Temporal Order of MeasurementRecursive no simultaneous reciprocal effects Nonrecursive including simultaneous reciprocal effects

Stage

state

and

linear

analyses

are

possible

When competing theories posit different and conflicting causal orderings, analysis of temporal or causal order may provide a strong test of competing theories.


23/34

Temporal Order of MeasurementGranger causality

Y causes X if some b j does not equal zero (implicitly, b j must be statistically significantly different from zero).

X causes Y if some c jis not equal zero. In effect , the question

posed by the test for the Granger causality becomes, Is there variation in one variable that cannot be explained by past values of that variable but that can be explained by past values of another variable?


24/34

Temporal Order of MeasurementIt may well be that a lag of 1 is sufficient both to explain the variance in the dependent variable and to reject the hypotheses of Granger causality. ( for very long series it should be possible to test for stationarity)Longitudinal data and analysis in conjunction with causal modeling may be used to examine the distinction between long and short term effects on behavior.


25/34

Serendipity

and Intentionality

in Longitudinal Data

The earliest data in social sciences is the national census data.Until recently the fact that census data may be used to measure change and perhaps to infer nature of causal relationships has been more serendipitous than intentional.


26/34

Designs for Longitudinal Data Collection1. Not Quite Longitudinal Designs2. Total Population Designs3. Repeated Cross sectional Designs4. Revolving Panel Designs5. Longitudinal Panel Designs6. Other variations


27/34

Designs for Longitudinal Data CollectionNot Quite Longitudinal designsTime ordered cross sectional design consists of time

ordered data and cross sectional analysis.

Time

ordered

cross

sectional

data

is

describable

once

temporal order has been established, but it is insufficient to ensure that one does not predict a

cause

from

its

effect.


28/34

Designs for Longitudinal Data CollectionNot Quite Longitudinal designs With cross sectional data, even time ordered cross

sectional data, we run the risk of undetectable misspecification, because of incorrect causal ordering

in the model being estimated. With longitudinal data, incorrect causal ordering is

more likely to be detected (Ex: Stage state analysis,

linear panel analysis of temporal or causal order)


29/34


30/34

Designs for Longitudinal Data Collection


31/34

Designs for Longitudinal Data Collection


32/34

Designs for Longitudinal Data CollectionRepeated Cross Sectional Designs

Draw independent probability samples at each measurement period

Principal

limitations

are

inappropriateness for studying developmental patterns within cohorts

inability to resolve issues of causal ordermany argue not really a longitudinal design


33/34

Designs for Longitudinal Data CollectionRevolving Panel Designs

Collect data on a sample of cases retrospectively or prospectively for some sequence of measurement periods, then drop some and replace some casesHelps rule out the possibility the results are a consequence of bias from repeated interviewing


34/34

Designs for Longitudinal Data CollectionLongitudinal Panel Designs

Retrospective or prospectiveMissing data result from failure of the respondent to

Remember past events, behaviors, or attitudesFeel comfortable divulging sensitive informationCooperate or stay in touch with the study

longitudinal research 1

Documents