On 9 June 2015 a half day meeting on ‘Making Statistical Sense of the Life Course’ was held at the Royal Statistical Society, jointly organised by the Social Statistics Section and the Society for Longitudinal and Lifecourse Studies (SLLS).
The meeting featured four speakers and a panel discussion and was chaired by Tom King of Newcastle University, who introduced the rationale for the meeting as being born out of the complexity of the life course because of its duration and interrelated facets. Many things change and interact, while small individual differences, characterised as noise in other circumstances, are still important. Patterns within these, such as resilience, can be the target of serious policy and investigation. While longitudinal data for constrained longitudinal problems admit elegant and well-defined solutions, people are more complicated, not least because the world is changing around them.
Andrew Pickles of Kings College London talked about the issue of making sense of children’s language development when the data are collected longitudinally but are different psychometric and parent observations at almost every time. Two approaches to dealing with this were outlined: internal and external calibration, relying on the different measurements on the same people at different times; or on different measures at the same time, rather than simply adopting a mean-centred normalisation at each stage. This was illustrated with examples of growth curves followed by both a clinical, ie impaired, population (the Manchester Language Study) and an at-risk population in a larger sample from the Wirral.
Commonalities emerged: by middle childhood, trajectories seem to be fixed, and therefore stable at a relative level within a population. However, younger children showed the potential for larger shifts in their relative position. Of course, making sense of this is difficult in that potential for change may not mean we have appropriate interventions to achieve change. However, it does suggest that early intervention is a sensible approach to try, especially as late intervention is not so successful. One of the striking features of the data was the persistence of the relative ability in language: the typical finding of a high group, a low group, an improving group and a declining group was not evident in the data.
The second speaker was Marc Scott of New York University who spoke about the affinity of shifting between different classes of jobs within the labour market experiences of people in the US National Longitudinal Study of Youth. Individuals can be observed in a different job in a different sector at each wave of the survey and these transitions (or stability) can relate to individual factors, as well as how long they have been in that current state. His modelling approach highlighted the difficulty of trying to make sense of the whole of the possibilities of the life course within one model, and then allow for changes over time in the sense of a linear drift.
The estimation to operationalise this specific analysis required a sophisticated approach and made use of random effects for things like the job sector effects so as to allow for variation. Within this, some important factors emerge, such as confirming the existence of drift and its apparent relation to other aspects, as well and a stronger random effect for women as opposed to men showing their range of affinities across sectors was greater in some way, compared to other predictive factors. Within specific sectors, there are differences visible in the probability of staying within certain administrative roles for women that are not obvious for men, by considering the pairwise comparison of the sector random effects.
Where the first two talks had been based on inference for the individual, the second pair concentrated on the life course as a whole and how this could be used to represent and describe differences. Jacques-Antoine Gauthier of University Lausanne spoke about the gendered patterns of the life course observed in Swiss data, which captured retrospective life histories of individuals. Using clustering techniques, these showed clear differences, with men predominantly following a full-time pattern of working and women split between this experience and one of full-time motherhood and another of part-time working. Within this, there are caveats that while these are what we might expect historically, there may be some differences in what we can expect in the future; additionally there is no clear explanation of what path men not working full time follow.
While these data relied on retrospective recall to avoid the problem of missing data, prospective panel data face more complicated problems. Using more than one aspect of activity, not just labour force participation, it was possible to see more complex pictures of how these differences are realised, as it would also be with combinations of individuals, ie dyadic data. This is the kind of complexity which drowns the capacity of stochastic models to estimate robustly, leaving more comparative and descriptive methods.
Finally, Brendan Halpin of University Limerick took up the baton to present a solution to the problem of missing data in sequences such as the life course presents, and categorical time series more generally. In repeated observations of the same participant in a longitudinal survey, item 'non-response' is common and typically multiple imputation tools can be used to allow for unbiased estimation. However, Brendan demonstrated that for sequences of the kind arising from a life course, the imputations were poor as they did not draw on the duration of the spells in a state as much as the last one. He presented an approach he developed (available for Stata) to impute based on realigning the data by individual and filling gaps iteratively. He showed an example of using this technique to impute multiple draws of sequences for data around working at the time just before and after mothers gave birth. There was some disappointment from the audience that the routines are only presently available in Stata but tempered by some satisfaction that they do not impose onerous computational requirements.
The event concluded with a panel discussion chaired by John Bynner of UCL Institute of Education featuring the day’s speakers and Amanda Sacker, also of University College London. John extracted key messages about the problems facing the life course research and invited the panel to add their own thoughts. Fundamentally there is a problem with the patterns which emerge from these data that it is hard to move beyond description when so much relies on clustering techniques of some kind and we know that data will yield clusters regardless of whether they are present. Amanda identified ideal types as a promising area for basing inferences rather than the more data-driven approach of clustering by latent classes or other methods, but she also stressed the difficulty in managing and preparing data for these sorts of analysis. Andrew highlighted the challenge in child development of rooting measures of what a child was doing in the context of what was appropriate for the age of the child, that typical behaviour can characterise a problem if it is not typical for the age of the child.
Points from the floor specifically homed in on the challenge of interdisciplinarity faced by life course research, not least in securing academic publications where each discipline can desire satisfaction from the one analysis, setting unrealistic expectations for what was achievable and squeezing out what could still be useful contributions. Speakers recommended alternatives such as submitting to life course-specific journals like the one run by SLLS, or dealing with restrictive word limits by publishing a separate methodology paper with detailed explanations and presenting key results concisely in a relevant substantive journal. Related to this was the difficulty of communicating life course results, which can be rather complex and unfamiliar, to those who may be interested as the life course is of inherently general interest.
It was concluded that the analysis of life course does not have a standard approach that can simply be applied; it takes enormous time for each stage to understand the data, operationalise the question, do the analysis and then make sense of the results.