Read on Twitter
My copies have arrived! So time for a quick summary thread of my new edited book, which looks at a variety of approaches to age-period-cohort analysis. And it’s currently 20% off: https://www.routledge.com/Age-Period-and-Cohort-Effects-Statistical-Analysis-and-the-Identification/Bell/p/book/9780367174439 @routledgebooks (1/16)
C1: it’s worth saying that all authors agree that there is no solution to the APC identification problem, and non of the chapters that follow present a ‘solution’. The intro provides a short accessible explanation of APC issues (2/16)
PS – if you want an even more accessible introduction to APC, see this animated video (3/16)
C2: constraining variables. Dangerous to do this atheoretically, but if we have strong theory to justify particular constraints, it can work. We should be cautious that what comes out is sensible, and test other ‘plausible’ constraints to ensure they give the same answer
C3: multilevel models: How they have been used, why they don’t solve the identification problem, and how they can be used as a sensible way of modelling APC and separating long-running and discrete APC effects, and explicating assumptions on the former (5/16)
C4: Visualising APC with Lexis plots @jonminton. We can use Lexis plots (like below) as a map to help us understand, theorise and then model APC, particularly APC interactions. In this case looking at various causes of death. (5/16)
C5: in a similar vein, Lexis plots are particularly valuable to see not long-running changes, but deviations from those changes, in this case considering mortality across 25 countries (7/16)
C6: Making slightly weaker assumptions: often instead of assuming a particular value for one of APC, we can assume a range; or we might assume a causal pathway. In *some* instances, these will be more plausible. @ethanfosse (8/16)
C7: Carefully understanding the causal pathway is the key here, too, with the example of smoking as a key predictor of cancer rates. Understanding the theory/processes by which that happens allows clearer modelling of those processes. (9/16)
(C7 also includes this diagram, which is pleasingly nihilistic out of context) (10/16)
C8 – Bayesian models @ethanfosse. Bayesian models are often touted as a solution to, well, most things, but this chapter shows that it is not a solution to the identification problem. It is, however, a useful way by which assumptions can be explicated and modelled (11/16)
C9 and 10 provide overarching views on the topic. C9 considers the debate around APC over time, and argues that, in many cases, simply assuming flat period trends is a sensible approach (an assumption made in other chapters). (12/16)
Finally, C10 brings the above chapters together around the idea of the ‘line of solutions’ – which is central to the identification problem and what ‘solutions’ to the identification problem mean. (13/16)
Overall, the chapters in the book are in agreement: there can be no solution to the identification problem – only careful stating of the (strong) assumptions we are making, and theory to back up those assumptions (14/16)
Thanks to @britishacademy_ for funding the conference that this book came out of, and to all the authors of the chapters for their brilliant work and willingness to contribute. (15/16)
I should also mention Manfred te Grotenhuis, who was one of the first to suggest the book, but sadly died before he could contribute more. The book is dedicated to him. (16/16)
