Longitudinal Surveys
From ICE Primer: A Tobacco Control Research Methodology Primer
Longitudinal surveys are surveys wherein data are obtained from most or all participants at several points in time, called "waves" or "cycles". Longitudinal health surveys typically follow subjects over several years, so that medium and longer term changes can be observed.
Statistics Canada surveys which follow participants over several years include:
The National Population Health Survey (NPHS) The National Longitudinal Survey of Children and Youth (NLSCY) The Youth in Transition Survey (YITS)
The General Social Survey (GSS), the Canadian Community Health Survey (CCHS) and the Canadian Tobacco Use Monitoring Survey (CTUMS) are not longitudinal. They are examples of periodic cross-sectional surveys, where the sample is newly recruited each time.
Advantages of longitudinal survey data
Longitudinal surveys allow us to observe change for individual sample members, and thus allow us to associate one kind of change with another.
For example, the following is a quote from a summary report of a study on the impact of longer working hours on lifestyle (Shields, Statistics Canada Health Reports, 1999):
"Unlike weight gain, there was no relationship between working long hours and daily smoking in 1994/95, when other factors such as age and education that might affect smoking were taken into account. But for both sexes, changing from standard to long hours between 1994/95 and 1996/97 was significantly associated with an increase in daily smoking during the period."
Longitudinal surveys facilitate modeling the effects of interventions by comparing responses of participants over time, effectively matching each participant with self at earlier times.
Longitudinal surveys may also allow us to observe order of events, which assists in causal interpretation. If event A always or nearly always precedes event B, event B is not likely to be a cause of event A. (At the same time, causation of event B by event A is not established, since both might result from event C. Longitudinal surveys can establish causation only in a context of treatments randomly allocated to participants.)
Recruitment costs for longitudinal survey interviews may be lower than for the same number of cross-sectional survey interviews, since most of the recruitment work will be done at the first cycle in the longitudinal case.
Imputation of item non-response may be made easier by earlier or later responses to the same item.
Limitations of longitudinal survey data
Successive interviews of a participant may take place one or two years apart: what happens between cycles (and when) may not be observed with sufficient accuracy to be useful. For example, in examining data from an all-purpose health survey, it may be possible to observe that a participant has quit smoking between cycles. If the time of smoking cessation has not been asked, its observation is said to be interval censored.
Longitudinal sample participant responses may be influenced by participation in the survey. Possible mechanisms are respondent fatigue, conditioning, or heightened awareness in recalling events. Such phenomena are known as time-in-sample effects. The sample after a few cycles may represent mainly the hypothetical population of people like the original recruits, had these people also been in the survey.
Surveys like the U.S. Survey of Income and Program Participation (SIPP) show what is called the seam effect. The SIPP conducts interviews every four months and respondents are asked about month-to-month changes since the previous interview. There is a "tendency for month-to-month changes in the data to concentrate suspiciously in adjacent months that were covered in different interviews". It is thought that "seam effects may be due in part to respondents' tendency to minimize change within the reference period". Tourangeau, R., Rips, L. J. and Rasinski, K. (2000) The Psychology of Survey Response. Cambridge University Press.
Last but most important is the limitation that the sample decreases in size over time, through explained and unexplained attrition.
Analysis of longitudinal survey data
The most basic analyses are those which model the variables from an interview in terms of background variables and variables from a previous interview. More causally oriented are those which model changes in one variable in terms of the changes of others, or those which simply look at associations between the changes in important variables.
Some analyses model growth or development of a variable over time (repeated measures analysis, growth curve analysis).
There are survival analyses which model times in various states, such as durations of quit attempts.
There are event history analyses which model transitions among states, such as Never Smoker, Smoker, and Quitter.
