Selection of articles
Materials for this review were primarily identified through searches of three electronic databases for peer-reviewed published papers: the MEDLINE/Pubmed, PsychInfo, and Scopus. Search terms were selected based on initial review of relevant subject headings across databases that were likely to yield relevant results. Using MEDLINE/Pubmed, initial search terms included: “age-period-cohort” or “birth cohorts” and “alcohol” (N=47), “liver” (N=47), “cirrhosis” (N=21), “substance use” (N=10). We then added a search term for “gender” (N=2) and “women” (N=18). Using PsychInfo, initial search terms included: “cohort analysis”, “time trends” and “alcohol drinking patterns” (N=27), “liver” (N=26), “cirrhosis” (N=24), and “substance abuse” (N=15), and a search terms for “gender” (N=27) and “women” (N=13). Using Scopus, initial search terms included: “age-period-cohort” and “alcohol” (N=20) “liver” (N=18) “cirrhosis” (N=6), “substance abuse” (N=1) and additional search terms for “gender” (N=1) and “women” (N=1). A search of related articles in these databases, references in relevant papers, and consultation with recognized experts yielded additional 8 papers for review.
details the process of article refinement for inclusion in the review. After duplicate articles were removed, an initial set of 151 papers was reviewed. Inclusion criteria were as follows: (1) published in the last 20 years (i.e., after 1988, in order to focus the review on the most recent information regarding birth cohort effects); (2) written in English; (3) peer-reviewed; (4) original research (i.e., no review articles); (5) population- or community-based sample; (5) directly assessed possible cohort effects by comparing at least two birth cohorts; (6) included alcohol use, alcohol use disorder, or alcohol-related mortality as an outcome.
Search strategy and method of article selection for review
Measures and definitions
Lifetime prevalence of alcohol use was defined as either the lifetime proportion of alcohol use (percentage of population who have ever used alcohol) or as the incidence density of alcohol use (percentage of new alcohol users per unit of person-time at risk of alcohol).
Prevalence of current alcohol use was defined as the proportion of alcohol users in a specified time frame (e.g., past week, month, or year).
Alcohol disorders/problems included any prevalent DSM (substance abuse or dependence) or ICD (harmful substance use or substance dependence) defined alcohol disorder, chart review for alcohol problems, or scales capturing alcohol-related impairment in social, occupational, or physical domains.
Alcohol-attributable morbidity/mortality was included as an outcome for papers that examined morbidity and mortality that is solely caused by alcohol (e.g., ICD-10 codes F10 [mental and behavioral disorders due to use of alcohol], X45 [accidental poisoning by and exposure to alcohol], and Y15 [poisoning by alcohol, undetermined intent]), or for which alcohol is a primary contributing case (e.g., ICD-10 code K70 [alcoholic liver disease]).
Identification of cohort effects: quality of evidence
We separated the studies into three quality levels of evidence, hierarchically organized by the rigor of the statistical modeling and the appropriateness of the dataset to answer age-period-cohort research questions. For convenience, we labeled these as Levels 1, 2 and 3.
Studies given a (1) provided the most rigorous evidence. These include cross-sectional studies repeated over time as well as panel designs (longitudinal follow-ups of sequential cohorts), and formal statistical modeling to decompose variance in trends over time into components that can be attributed to age, period, and cohort. We note that modeling of APC effects remain an ongoing statistical debate (Keyes, Utz et al. 2010
), full elaboration of which is outside the scope of this review. All articles included in this review utilize common and well-documented methods for APC analysis.
Those studies given a (2) provided an intermediate level of evidence in terms of methodological rigor. These include longitudinal cohort studies with a limited number of birth cohorts (as opposed to panel designs with a greater number of birth cohorts), with and without statistical age-period-cohort models. While longitudinal studies offer more rigorous evidence for cohort effects than single time point cross-sectional studies, period and age effects are difficult to unconfound with cohort effects in the longitudinal design due to the often restricted variation in birth cohorts, and the fact that the cohorts age simultaneously with time so age and period effects cannot be disentangled.
Those studies given a (3) are those providing the least rigorous evidence, and for which the most caution should therefore be taken during synthesis and inference. Cross-sectional studies make up the majority of this category, since they offer the advantage of readily-available data. However, one-time cross-sectional studies rely heavily on retrospectively reported information, raising methodological issues in the validity of retrospective recall. Further, age and cohort effects cannot be rigorously disentangled in the cross-sectional design, as respondents’ current age at the time of the survey determines birth cohort. However, cross-sectional studies often control for a collection of important confounders, thus, we include these types of APC analyses in this review, but use caution in interpreting the results. We included only those cross-sectional studies that report on birth cohorts, rather than cross-sectional studies that report differences across age alone. While age can be used as a proxy for cohort in the cross-sectional design, we focused on studies that explicitly aimed to estimate parameters for birth cohorts.