An Overview of Longitudinal Research Designs in Social Sciences

CRD and LRD

Based on the number of time periods for which the same variable is measured, the research designs in social sciences are broadly classified into two types: CRD and LRD. In CRD, the researcher collects the data on one or more than one variable for a single time period for each case in the study. The researcher measures the variables only once, and the measurement period can last for weeks, months or years. The data collected using CRD are regarded as ‘contemporaneous’ (Menard, 2002: 2) and infer only about the circumstances of the cases at one particular point in time (Babbie, 2013; Gravlee et al., 2009; Menard, 2008; Ruspini, 2002). For instance, the Police Study 2017 conducted by Common Cause & Lokniti—Centre for the Study of Developing Societies follows a CRD. Here, 15,562 individuals over the age of 18 years were interviewed only once in June–July 2017 across 22 states in India. The data were collected on 42 variables categorized into 6 main themes from these 15,562 individuals. Since these variables were measured only once in this particular survey, it can provide only a snapshot of ‘people’s perception and experience of dealing with the police’ in India (Common Cause & Lokniti—Centre for the Study of Developing Societies, 2018).

As opposed to CRD, in LRD, the researcher repeatedly measures the same variable in precisely the same format for more than one time period (Babbie, 2013; Gravlee et al., 2009; Grotpeter, 2008; Lynn, 2009; Menard, 2008; Ruspini, 2002). The data are collected from similar or comparable cases across these time periods. The Indian Human Development Survey (IHDS) follows LRD (discussed later in detail). Since data are collected on the same variables for two or more time periods in LRD, the researcher can infer about the inter-temporal continuity or change by comparing the values of these variables collected at different points in time. In addition to this, the possible explanations for such change can also be deciphered using LRD (Gravlee et al., 2009; Menard, 2008). Therefore, to ensure the comparability of data collected over different times, it is suggested to use the standardised procedures of data collection at each wave of data collection in LRD (Menard, 2002).

Types of LRDs

Different scholars have classified LRD into different types. LRD is classified into four types based on two criteria: the number of waves, and whether data are collected from similar or different cases in the subsequent waves. These are total population design (TPD), repeated cross-sectional design (RCD), revolving or rotating panel design (RPD) and longitudinal panel design (LPD) (Menard, 2002). Due to the difference in the data collection procedure, each of these LRDs offers a different type of analysis. These different LRDs vary in terms of analysing gross or net change and the intra-individual or inter-individual trajectories of change (Lynn, 2009; Menard, 2008). These four basic types of LRD are discussed in the following sections. It is imperative to note here that the types of LRDs discussed in the present article are only broad categories and do not exhaust the full range of other possible LRDs that exist in the social sciences.

Advantages and Limitations of LRD, Especially Panel Design

The advantages of LRD can be best understood in comparison with the limitations of CRD. The data collected using CRD can infer about the association and not causality between the variables, if the temporal order of the variables is not known. On the contrary, the data collected using LRD can be used to estimate both association and the magnitude and direction of causality between the variables. Hence, in comparison with CRD, the LRD is considered as the most suited design to measure change and infer the possible explanations for such change (Gravlee et al., 2009; Lynn, 2009; Ruspini, 2002).

Besides analysing change at the aggregate level, longitudinal data, especially panel data collected using panel research design, can be used to infer about the intra-individual and inter-individual trajectories of change (Gravlee et al., 2009; Lynn, 2009; Ruspini, 2002). Although most of the LRDs need a substantial investment of money, time and human resource to collect longitudinal data, the researcher can collect this longitudinal data in a time- and cost-effective way using RePD (Babbie, 2013; Grotpeter, 2008). For instance, it is both time- and cost-effective ‘to survey adult respondents’ at one point in time to answer questions about their life experiences over the past 20 years than it is to identify, track, and survey the same people from childhood through adulthood to ask them the same question every year’ (Grotpeter, 2008: 114). Hence, we can argue that the studies based on LRD do not need to last long ¹¹ and all long-term studies need not be longitudinal in nature (Johansen & White, 2002 as mentioned in Gravlee et al., 2009).

While collecting the longitudinal data using LRD, the researcher faces all kinds of limitations in common with CRD plus a few more that are unique to it (Menard, 2008). Similar to CRD, the LRD is at the risk of measurement error (ME), ¹² the issue of management of vast amount of longitudinal data and its analysis ¹³ (Gravlee et al., 2009) and the issue of long-term retrospective recall ¹⁴ (Grotpeter, 2008). Here, instead of focusing on the issues common to both CRD and LRD, the limitations specific to the LRD such as panel conditioning, panel attrition, sample selection bias are discussed in detail in the subsequent sections.

With regard to LRD specifically, the reliability of the longitudinal data, particularly panel data, is affected by the panel conditioning of the respondents. It occurs due to the respondent’s participation in earlier waves of the study. This panel conditioning of the respondent can take place due to two reasons: the first reason is to avoid ‘response burden’. For instance, if the respondent learned that reporting ‘no’ as a response to certain questions leads to a battery of relevant follow-up questions in wave I, the respondent may report affirmatively in the second wave to make the interview shorter (Gravlee et al., 2009).

The second reason for the panel conditioning could be in an actual change in the behaviour of the respondent. The participation in earlier waves of the panel study might have made the respondent aware of the alternative options to explore. This change in the respondent’s behaviour would not have happened if the respondent participated in the study for the first time. The information gathered by participating in earlier waves of the study may change the natural trajectory that the respondents would have otherwise followed (Cantor, 2008; Gravlee et al., 2009). For instance, Clausen (1968 as mentioned in Cantor, 2008) observed that participating in a pre-election survey motivates some respondents to subsequently vote in 1964 presidential elections in America (for details, see Cantor, 2008: 125). Hence, panel conditioning need not always be negative and it can also be positive. In some instances, it can motivate the respondents to participate in subsequent waves and provide more correct information. While in some other cases, it can de-motivate them from participating in subsequent waves, thus leading to panel attrition.

The loss of respondents at each wave of the panel study is referred to as panel attrition ¹⁵ (Gravlee et al., 2009). Laurie (2008) observed that the probable causes or sources of panel attrition are the death of the respondent, respondent’s refusal to participate in subsequent rounds and the inability of the researcher to trace the sample member due to the respondent’s geographical mobility. The respondent’s refusal to participate in subsequent waves of the panel study can be due to various reasons such as frequent interviews, interview length, complexity of the questionnaire, personal or overly intrusive subject matter, relevance of the topic and mode of data collection.

Panel attrition has severe implications on the sample size of the study and consequently on the range of statistical analysis possible with the data collected. For instance, if the sample size becomes too small in the subsequent waves due to panel attrition, the range of statistical analysis possible will be limited, as would the efficiency of the parameter estimated. Furthermore, if the sample becomes unrepresentative due to non-random attrition, it can produce inconsistent parameter estimates (Gravlee et al., 2009; Laurie, 2008).

Due to the issue of panel conditioning and panel attrition in PPD, it is advised to use RePD to collect longitudinal data. However, the RePD suffers from sample selection biases as it tends to include only the ‘survivors’ in the sample (Grotpeter, 2008). For instance, the eight cohorts interviewed retrospectively in WGLHS (as mentioned earlier) do not include the respondents who were eligible to be included in the sample but could not be included due to their death or emigration.

Gravlee et al. (2009) and Laurie (2008) suggested a few ways to minimise panel attrition which is the main issue in LPD. These are oversampling, building necessary rapport with the respondents and collecting additional contact details at the baseline measurement period. The change in the mode of data collection to fit the respondent’s need, providing incentives and performing ‘keeping-in-touch exercises’ ¹⁶ (Laurie, 2008: 174) in between the measurement periods are some of the other ways to minimise panel attrition.

Field Experience of Using PPD

Despite the limitations discussed earlier with regards to LRD, there are certain issues which can be best understood using LRD. For instance, Grotpeter (2008) argued that PPD could be used to ‘collect prospectively oriented data concerning individual aspirations and expectations and to compare these to actual outcomes at later points in time’ (p. 113). Following this, in my PhD research work, I used PPD to explore the gap (if any) between the post-high-school education aspiration and attainment of grade XII students in Thanesar town of Kurukshetra District, Haryana. The grade XII students from different streams were interviewed about their post-high-school education aspirations when they were in grade XII, i.e. from November 2017 to February 2018 (wave I). These students were re-traced after an academic year, i.e. during October–November 2018 (wave II), when they were supposed to be enrolled in some higher education institute to explore whether they were able to translate their educational aspirations into reality, and the possible explanations for the aspiration–attainment gap. In the following paragraphs, a few field insights of using PPD are discussed.

The PPD was used to collect the panel data. As discussed earlier, the issue of panel attrition is inherent to PPD and to minimize it the original sample size of 112 grade XII students ¹⁷ was doubled to 224 students at the baseline period (wave I). So, if 50 per cent sample mortality occurred, the original sample size of 112 students would be retained in wave II. These 224 grade XII students were selected using a purposive sampling technique. ¹⁸

As mentioned earlier, the LRD faces the limitations of CRD and some other limitations that are unique to it. Despite setting a target of interviewing 224 grade XII students within the school premises in wave I, only 174 students were eventually interviewed. This gap between the intended and the achieved sample is because principals of the private schools (contacted for the present study) who act as gatekeepers of these schools allowed me to interact with the students for a limited period of time. ¹⁹ Furthermore, the measurement period for the wave I of this study, i.e. from October 2017 to February 2018, also acted as the source of this gap between the achieved and the intended sample size. During wave I, these grade XII students had to appear for their mid-exams, pre-board exams and board exams and had other types of tests and practicals. This, coupled with the holidays, substantially reduced the time period for conducting interviews in wave I.

Out of the 174 respondents interviewed in wave I, 171 respondents agreed to participate in a future wave II. Their mobile numbers were collected and were solely used to trace these respondents in wave II. ²⁰ Out of these 171 respondents, 121 respondents were traced back in wave II. The rest of the 50 respondents could not be traced as the mobile numbers provided were wrong or switched off. Out of the 121 respondents, only 46 respondents agreed to participate in wave II, while the rest of the 75 respondents refused to participate further. These respondents gave different reasons for their refusal, such as busy, the geographical mobility of the respondent, health issues, etc. A few respondents who gave the time and place for an interview did not show up or pick up the phone when contacted on the day of the interview. A reason for the respondent’s refusal to participate in wave II might be the lack of necessary rapport building in wave I due to the limited interaction with these students. The ‘refusal conversion exercises’ ²¹ (Laurie, 2008) were also performed with those respondents who did not want to give their interview in wave II by asking them to have a telephonic conversation instead of a face-to-face interaction. Using this technique, nine respondents agreed to participate in wave II.

People participate in a survey after calculating the cost and the benefit of taking part in it from their own perspective (Lynn et al., 2009, Groves & Couper, 1998 as cited in Laurie, 2008: 170). As one respondent enquired: ‘you took my interview last time as well, but why have I not gotten a job yet?’, hence refused further participation in the study. It is possible that participating in the present study is of no benefit to some respondents in terms of any economic gain or specifically in getting a job. This cost–benefit analysis of participating in the present study may be one of the reasons that some of the wave I participants refused to participate in wave II.

The reasons mentioned earlier are the explicit reasons given by the respondents for not participating in wave II. However, there are other implicit factors which may have influenced the further participation of a few respondents. The sensitivity of the research topic can also be one of the reasons for panel attrition in the present study. It was observed that some of the participants who refused to participate in wave II are also dropouts, i.e. not enrolled in any higher educational institute. It can be inferred that the opportunity cost of participating in the study for a few respondents was very high in terms of emotions. As some of the respondents were not able to translate their educational aspirations into reality, they did not want to discuss it, especially with a stranger (researcher), and therefore they preferred not to participate in wave II.

Due to panel attrition, the original composition of the sample gets affected (Laurie, 2008). In the present case, the purposive sampling technique was used to select the respondents from different educational streams to understand their experiences. However, due to panel attrition, none of the girls from the medical stream and not a single boy from the non-medical stream of private schools were present in the final sample in wave II. Therefore, while discussing the findings of this study, it is imperative to mention the composition of the final sample after attrition in a longitudinal panel study.

Besides the above-mentioned issues faced while collecting the panel data, the richness of the information collected using PPD needs special mention. I would like to share an instance where, I believe, I might not have succeeded in gathering such detailed information that I had used some other research design. Kapil, ²² a grade XII student of a private school, had shared his post-high-school educational plans with me in wave I. With so much of enthusiasm, he informed me that he wants to follow his passion for photography and hence wants to pursue a photography course from Chandigarh after grade XII. But, in wave II, it was observed that he was preparing for the International English Language Testing System (IELTS) and was enrolled in a coaching centre in Chandigarh. In between the two waves, his aspiration shifted towards pursuing overseas education with the hidden and a prominent motive of working there. ²³ Hence, the use of PPD allowed me to capture a few more such instances of change in aspirations due to the change in circumstances. Also, this particular research design allowed me to examine which student had a much stronger sense of aspiration, i.e. have the necessary knowledge to translate the aspiration into attainment, which served as one of the probable explanation for the aspiration–attainment gap in case of a few students.

A Final Note

The aim of this study directs the research design, i.e. whether the data are collected for one (CRD) or for more than one time period (LRD). The LRD is considered to be best suited for measuring change and inferring about the possible explanations for such change. This change can be analysed at the aggregate or individual level. The different types of LRDs offer different types of analysis; for instance, RCD analyses change only at the aggregate level, while PPD and RePD offer both aggregate and individual trajectories of change. Based on my experience with the PPD and the richness of the data collected using this particular LRD in terms of individual trajectories of change, there should be more studies in India based on LRD. In this way, change can be analysed at both individual and aggregate levels, and reliable causal interpretations can be inferred about the phenomenon in which temporal order of variables is not known.