Cohort analysis is a set of statistical techniques that can be used to disentangle age,
cohort, and period effects. To conduct such an analysis, one needs comparable
repeated cross-sectional data gathered numerous times at regular intervals. Age
effects are changes observed in specific groups as they age; period effects (indexed
by date of measurement) are changes due to events that influence the behavior of all
age groups; cohort effects (indexed by year of birth) refer to the long-term impact of
events on people born in different periods and do not change with one’s age or stage
of life (Glenn 2005).
However, there arises a major statistical problem, as there is a perfect colinearity
between the three variables: by definition age = period − cohort. In a regression
approach, the matrix of explanatory variables does not have full rank, and the
ordinary least-squares estimate cannot be computed because the same predicted
values (and errors) may be obtained by different combinations of the age, period,
and cohort coefficients.
One solution omits one explanatory variable: age, period, or cohort. However,
such an omission creates a specification error so that the estimates of the coefficients
of the remaining variables are biased and not convergent. Another solution imposes
a priori constraints on specific coefficients, e.g., that consumers belonging to two
neighboring cohorts have identical coefficients. The problem here is that the
estimated results may vary widely depending on which constraints are imposed. A
third approach substitutes additional information, but it is often difficult to obtain
good a priori information. If, for example, cohort is replaced with a measured
variable such as cohort education, it can be unclear whether the right variable for
cohort has been substituted. Finally, current research is exploring the possibility of
combining a priori information (e.g., information indicating that period effects
should be negligible) and a partial least squares (PLS) estimation approach to obtain
precise estimates that are not entirely data-driven. An interesting aspect of PLS is the
possibility of analyzing at the same time multiple dependent variables, such as repeat
purchase and type of brand purchased.
5 Future
Cohort analysis is a set of statistical techniques that can be used to disentangle age,cohort, and period effects. To conduct such an analysis, one needs comparablerepeated cross-sectional data gathered numerous times at regular intervals. Ageeffects are changes observed in specific groups as they age; period effects (indexedby date of measurement) are changes due to events that influence the behavior of allage groups; cohort effects (indexed by year of birth) refer to the long-term impact ofevents on people born in different periods and do not change with one’s age or stageof life (Glenn 2005).However, there arises a major statistical problem, as there is a perfect colinearitybetween the three variables: by definition age = period − cohort. In a regressionapproach, the matrix of explanatory variables does not have full rank, and theordinary least-squares estimate cannot be computed because the same predictedvalues (and errors) may be obtained by different combinations of the age, period,and cohort coefficients.One solution omits one explanatory variable: age, period, or cohort. However,such an omission creates a specification error so that the estimates of the coefficientsof the remaining variables are biased and not convergent. Another solution imposesa priori constraints on specific coefficients, e.g., that consumers belonging to twoneighboring cohorts have identical coefficients. The problem here is that theestimated results may vary widely depending on which constraints are imposed. Athird approach substitutes additional information, but it is often difficult to obtaingood a priori information. If, for example, cohort is replaced with a measuredvariable such as cohort education, it can be unclear whether the right variable forcohort has been substituted. Finally, current research is exploring the possibility ofcombining a priori information (e.g., information indicating that period effectsshould be negligible) and a partial least squares (PLS) estimation approach to obtainprecise estimates that are not entirely data-driven. An interesting aspect of PLS is thepossibility of analyzing at the same time multiple dependent variables, such as repeatpurchase and type of brand purchased.5 Future
การแปล กรุณารอสักครู่..
