Background

Longitudinal cohort studies play a central role in advancing understanding of the onset and progression of physical and mental health problems. Cohort studies assess, and often compare, the incidence of a condition within a group of people who share common characteristics (e.g., being born in the same year) [1]. A key advantage of longitudinal cohort studies over other research designs is that repeated measures data temporally orders exposures and outcomes to facilitate causal inference [2]. However, significant and systematic attrition can reduce the generalisability of outcomes and the statistical power to detect effects of interest [3]. Systematic attrition in longitudinal research occurs most often in older, non-white male participants with limited education and/or multiple health problems [4]. Long duration and repeated assessments can also increase attrition due to the significant burden on participants [4]. Given the expense of longitudinal cohort studies, effective strategies that engage and retain cohort participants are critical to the integrity of research outcomes [5, 6].

In the last decade, longitudinal data collection methods and cohort retention strategies have evolved considerably. So too have participant expectations of organisations (research and otherwise) that seek information from individuals [7, 8]. Established retention strategies within longitudinal cohort studies include: cash or gift incentives, sending reminder letters to participants, re-sending surveys, and offering alternative methods of data collection (for a review, see [6]). Booker et al. [6] demonstrated that these strategies were effective in longitudinal cohort studies that used the traditional data collection methods of postal surveys, face-to-face visits (home or on-site), and telephone interviews or surveys. However, these cohort retention strategies may not be as well suited to contemporary methods of collecting longitudinal data, such as web and mobile surveys [9], wearable sensors (e.g., FitBits) [10], short message services (SMS) [11], and groupware systems (e.g., video conferencing) [12]. Novel methods of engaging participants such as web advertising [13], social media [14], and electronic reminders [15], are also now being employed in cohort studies using both traditional and modern longitudinal data collection methods.

A systematic review on the effectiveness of established and emerging cohort retention strategies in longitudinal cohort studies would provide guidance to researchers and funders on maximising cohort maintenance within these high investment programs of research. Previous reviews of retention strategies in health research include [4, 6, 16, 17]; only one of these reviews focused specifically on longitudinal cohort research designs [6]. Booker et al. [6] conducted a narrative review of retention strategies in longitudinal cohort studies, including incentives, reminders, repeat visits/questionnaires, and alternative methods of data collection, finding that incentives and reminder strategies improved cohort retention. However, this review was limited by the small number of studies identified for each retention strategy, which resulted in the identification of a restricted breadth of retention strategies and the inability to synthesise findings empirically.

Further, Booker et al. [6] did not include research completed after 2006 and thus were unable to investigate emerging cohort retention strategies. Brueton et al. [16] completed a more recent review of retention strategies that included both established and emerging digital data collection retention strategies. However, the authors specifically excluded longitudinal cohort studies and instead focused on participant retention in intervention trials. Differences between intervention and longitudinal cohort studies, such as research design factors (e.g., study duration) and the motivations of the participants in joining or withdrawing from studies, may impact the usefulness of retention strategies across both study designs [4, 6].

A review of retention strategies reported in modern longitudinal cohort studies is pertinent and timely, given the emergence of digital retention strategies alongside established retention methods. Maximising cohort retention in longitudinal research can reduce the administration costs of conducting research, improve the efficiency of research processes, and reduce outcome biases for studies by adopting an evidence-based cohort retention framework. In this review, we aimed to: (i) identify retention strategies used in recent longitudinal cohort studies; (ii) examine whether retention rate was moderated by different study or participant characteristics (i.e., number of waves, study duration, sample size, population type, gender, age, country); (iii) estimate the retention rate in studies that use specific retention strategies, and contrast this retention rate with studies that do not use specific retention strategies; (iv) examine whether retention rate is associated with the number of retention strategies used; (v) examine which retention strategies were the strongest independent predictors of retention rate; and (vi) contrast the retention rate based on whether studies utilised emerging or established strategies. Moreover, to ensure that recent innovations in retention strategies were identified, this review focused on literature published within the past 10 years.

Method

Search strategy

A systematic review was performed as per the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [18]. Two search strategies were implemented. First, the electronic databases Medline, PsycINFO, Embase, CINAHL, AMED, and the Cochrane Library, were searched in July, 2016 using search terms relevant to three themes: (i) attrition, (ii) retention, and (iii) study design (Additional file 1: Table S1). The electronic search was limited to articles published from 2006 onwards in English, to avoid duplication of literature with the previous review on this topic [6]. The search was adapted to suit each database. Second, the reference lists of all articles selected for review were manually searched.

Inclusion and exclusion criteria

The inclusion and exclusion criteria were determined prior to implementing the search strategy. Articles were included in the review if: (i) the article described a cohort study, which was defined as a representative sample of a group or population who share a common experience or condition [2], (ii) the article reported at least one wave of follow-up data collection with a participant/proxy, (iii) participant retention data were reported, and (iv) retention strategies were reported. Articles were excluded if: (i) the article was not available in English; (ii) the article was not published in a peer-reviewed publication (e.g., conference abstracts or dissertations); and, (iii) the article’s research design was cross-sectional, involved data linkage only, or the article was a clinical or non-clinical trial evaluating the effectiveness of treatment regimens or intervention/prevention programmes (for an existing review of retention in intervention studies, see (12)).

Study selection, data extraction, and quality assessment

The search strategy resulted in 9225 articles after removing duplicates. In total, 141 articles were identified, screened, and determined to be eligible for inclusion (see Fig. 1). Data were extracted and summarised for each of the 141 articles on: (i) the research design, including baseline sample population and sample size, the number of data collection waves reported, and the duration in years between the first and last waves of data collection; (ii) the cohort demographics, including mean sample age at baseline (or age range if mean age was not reported), proportion of male participants, country of cohort participants, and whether the cohort was clinical or non-clinical; and (iii) retention data, including the retention rate between baseline and the final data collection wave reported, and the specific retention strategies. Finally, we examined the suitability of each article in addressing the current study’s research question. Articles that listed cohort attrition or retention as a research question or objective were categorised as “retention-focused”, and conversely articles that did not focus on attrition or retention were categorised as “non-retention-focused”. No articles were excluded on the basis of this quality assessment.

Fig. 1
figure 1

PRISMA procedural flow chart of the search and identification process

Full size image

Statistical method

We used meta-analysis (and meta-regression) to address the aims of the study. Meta-analyses were conducted using the Metafor package v1.9.8 [19] in R software v3.3.1 [20]. The retention rate, defined as the number of individuals who remained in the study at the last wave of data collection as a proportion of the total number of participants recruited at the baseline assessment, was the primary effect size measure of interest. All meta-analyses were conducted using inverse variance weighting, with random effects specified to account for between study heterogeneity. A binomial-normal model (with logit link) was used as the basis for analysis, which is appropriate when the effect size of interest is measured as a proportion. Where appropriate, meta-analytic effects were back-transformed to represent the median meta-analytic retention rate. We also report the I2 statistic as a measure of study heterogeneity, interpreted using the guidelines of Higgins et al. [21]. Meta-analyses were conducted when at least two independent studies contributed to the meta-analysis.

To examine the effect of gender on retention rate, we created a binary variable to denote studies as comprising a higher proportion of either male or female samples (Proportion of male participants in “male” grouping: M(SD) = 73.6%(0.20); Proportion of female participants in “female” grouping: M(SD) = 75.0%(0.21)). To examine the effect of country development level on retention rate, each study country was categorised as either high or low development level by using a mean-split of each nation’s Human Development Index – a measure of relative opportunity for longevity, education, and income, with a score range of 0 (low) to 1 (high) (Low HDI group M(SD) = 0.66(0.10); High HDI group M(SD) = 0.92(0.02)) [22]. Retention strategies were coded as either established or emerging, depending on their presence or absence in any of the earlier systematic reviews on participant retention strategies [4, 6, 16, 17]. Finally, all meta-regressions adjusted for study duration and number of waves (except when these were specifically examined as predictor variables), given these were deemed to be likely confounding variables in analyses.

Results

Cohort, participant, and article characteristics

The 141 articles identified for review described 143 cohorts (41 clinical and 102 non-clinical). Cohorts are summarised in Table 1. Overall the mean sample size reported in the first wave of each article was 3585 participants (range = 30 to 61,895). Articles reported a mean retention rate of 73.5% (SD = 20.1%), with 4.6 waves (SD = 8.0), over 4.3 years (SD = 5.0). The average baseline participant age was 30.0 years (SD = 22.0), and the average baseline proportion of male participants across samples was 40% (SD = 0.30). Studies were conducted in 28 different countries with a mean Human Development Index of 0.79 (SD = 0.15), indicating that studies were more likely to be conducted in countries with high-levels human development. Cohort attrition/retention was identified as a specific research question or objective of interest in 55 of the 141 articles, indicating that most articles were not focused on participant retention. Retention-focused articles reported significantly more retention strategies than non-retention-focused articles (non-retention-focused: M(SD) = 3.3(3.1); retention-focused: M(SD) = 11.0(7.02); t(141) = − 9.00, p < .001); however, no differences were found for the study sample size, number of waves, study duration, or retention rate. High heterogeneity was identified in all results, as expected given the diversity of research questions, methodologies, and cohorts across articles [21].

Table 1 Description of cohorts reported in the included articles
Full size table

Relationship between retention rate and study or participant characteristics

To examine whether retention rate was moderated by study characteristics (i.e., number of waves, study duration, sample size, study focus on retention strategies or not) or by participant characteristics (i.e., population type, gender, age, country development level), a series of meta-regressions was performed, one for each characteristic under examination. Retention rate was not moderated by: number of waves (b < 0.001; 95%CI [− 0.02 to 0.03], p = .77); study duration (b = − 0.02; 95%CI [− 0.06 to 0.02]; p = .34); sample size (b < − 0.001; 95%CI [− 0.00 to 0.00]; p = 0.48); or articles’ focus on retention strategies (b = − 0.12; 95%CI [− 0.54 to 0.30]; p = .57). Additionally, retention rate was not associated with the sample characteristics of: cohort type (clinical or non-clinical) (b = 0.04; 95%CI [− 0.42 to – 0.51]; p = .86); mean age (b = 0.02; 95%CI [− 0.01 to 0.01]; p = .74); or country development level (b = 0.11; 95%CI [− 0.46 to 0.68]; p = .71). However, gender was a significant moderator of retention rate (b = − 0.67; 95%CI [− 1.14 to − 0.20]; p < .01)). Namely, cohorts with more female participants (median retention = 81.5%, 95%CI [77.6% to 84.9%]) reported higher retention rates than articles with more male participants (median retention = 70.1%; 95%CI [60.1% to 78.5%]), after controlling for study duration and number of waves.

Relationship between retention rate and retention strategy types

A total of 95 retention strategies was identified, with an average of 6.2 strategies per article (SD = 6.2). The most common retention strategies were: cash/voucher incentives to complete a follow-up assessment (n = 59), sending a postcard or letter reminder to complete a follow-up assessment (n = 43), and offering participants alternative methods of data collection, such as completing an interview face-to-face or over the phone (n = 36).

Retention strategies were grouped into four main retention strategy domains: (i) barrier-reduction strategies, such as offering childcare services, assistance with parking and transport, and engaging a participant sub-sample to evaluate data collection methods for the next wave; (ii) community-building strategies, such as creating a recognisable study brand via logos and colour schemes, giving away study merchandise to create a sense of project community (e.g., t-shirts with study logo), and sharing study results, news and events with participants via newsletters, social media, and feedback reports; (iii) strategies to improve follow-up rates within each wave, including cash or voucher incentives for varying levels of assessment completion, and use of phone calls, SMS, house visits, mail and email reminders to participants to complete assessments; and (iv) tracing strategies, such as collecting the details of alternative contact persons for each participant at baseline, using public or non-public records to find updated contact information for participants, and collecting detailed participant contact information via a locator document (e.g. full name, address, social security number, phone numbers, email addresses, etc.). The most commonly reported category was strategies to improve follow-up rates within waves, identified 306 times within the 143 cohorts, followed by barrier-reduction strategies (adopted 268 times), community-building strategies (adopted 181 times), and tracing strategies (adopted 138 times).

Table 2 presents the retention strategies used, grouped by retention strategy domain. It compares the retention rate for those studies that did, or did not utilise a specific retention strategy type or domain. Of the 95 individual retention strategies examined, three demonstrated moderation of the retention rate. First, improved retention was associated with offering participants alternative methods of data collection (e.g., completing an interview face-to-face or over the phone) (median retention using strategy = 86.1%; median retention not using strategy = 76.3%; b = 0.24, p = .01), and having participants complete a locator document at baseline (median retention using strategy = 90.9%; median retention not using strategy = 78.1%; b = 0.49, p = 0.02). Finally, lower retention was associated with use of phone call reminders to participants to complete a follow-up wave (median retention using strategy = 72.7%; median retention not using strategy = 80.6%; b = 0.25, p = .05). There was weak evidence against the null hypothesis of no moderation effect for a further three strategies. This included having consistent research team members (median retention using strategy = 87.3%; median retention not using strategy = 78.1%; b = 0.67, p = .09); offering site and home visits for data collection (median retention using strategy = 83.9%; median retention not using strategy = 77.4%; b = 0.46, p = .07); and sending participants thank you, birthday or holiday cards (median retention using strategy = 84.9%; median retention not using strategy = 77.5%; b = 0.50, p = .07). There was no evidence to support a moderated retention rate by any other specific retention strategy type.

Table 2 Median meta-analytic retention rates for each retention strategy
Full size table

To examine whether the specific strategy domains of barrier-reduction, community-building, follow-up/reminder, and tracing retention strategies were associated with retention rate, a binary variable was created for each domain that denoted whether a study did or did not utilise one or more specific strategy types within that domain. As shown in Table 2, after controlling for study duration and number of waves, studies that utilised any barrier-reduction strategy had higher retention rates than those that did not use a barrier strategy (median retention using barrier strategies = 81.1%; median retention not using barrier strategies = 70.7%; b = 0.61, p = .01). Again after controlling for the study duration and number of waves, surprisingly, articles that reported use of at least one follow-up/reminder strategy had lower retention rates when compared to studies that did not utilise any follow-up/reminder (median retention using follow-up/reminder strategies = 76.4%; median retention not using follow-up/reminder strategies = 86.1%; b = − 0.32, p < .01). No relationships were found between retention rate and the use of any community-building or tracing retention strategies.

Relationship between retention rate and number of strategies used

To examine whether the cumulative number of retention strategies was associated with retention rate, we meta-regressed retention rate on to continuous variables representing the cumulative number of strategies used across strategy domains, and then within each domain separately. Greater number of retention strategies used (across all domains) was not associated with higher retention rate (b = 0.02; 95%CI [− 0.12 to 0.05], p = .21). When examined within each domain, controlling for study duration and number of waves, we found accumulation of barrier-reduction strategies was associated with higher retention (b = 0.12; 95%CI [0.02 to 0.22]; p = .02). In separate meta-regressions, no relationships with retention were identified between number of community-building strategies (b = − 0.03; 95%CI [− 0.18 to 0.11]; p = 0.63), follow-up strategies (b = − 0.03; 95%CI [− 0.14 to 0.09]; p = 0.65), or tracing strategies (b = 0.10; 95%CI [− 0.07 to 0.28]; p = .25).

Identifying strongest independent predictors of retention rate

Three separate meta-regression models were estimated to examine strongest predictors of retention rate within strategy domains and types. Table 3-Model 1 shows that when examining retention strategy types as cumulative variables for each domain, barrier-reduction was independently associated with higher retention (b = 0.17; 95%CI [0.03 to 0.31]; p = .02) and follow-up strategies was independently associated with lower retention (b = − 0.15; 95%CI [− 0.29 to − 0.01]; p = .04) beyond the effects of other retention strategy types. By contrast, Table 3-Model 2 demonstrates that when the retention rate was regressed on to all the binary indicator variables denoting whether the study did or did not utilise at least one strategy within that domain, only the use of follow-up/reminder strategies was independently associated with reduced retention rate (b = − 0.83; 95%CI [− 1.4 to − 0.27]; p < .01).

Table 3 Meta-analytic regression results between retention strategy themes and retention rate
Full size table

Finally, we investigated whether the associations between individual strategies and retention rate remained after controlling for other effective individual strategies in a single model (see Table 3 Model 3). A meta-regression model was created by entering only individual retention strategies that were associated with a retention rate at the p < .10 level (as discussed in [23, 24]). Six individual strategies were eligible: (i) offering alternative methods of data collection; (ii) consistency in the research staff; (iii) offering site and home visits; (iv) thank you and birthday cards; (v) phone call reminders; and (vi) the use of a locator form (i.e., alternate contacts). Offering participants alternative methods of data collection was associated with improved retention, whilst the use of phone call reminders was associated with reduced retention (b = 0.59; 95%CI [0.13 to 1.05]; p = 0.01; b = − 0.72; 95%CI [− 1.18 to − 0.25]; p < .01, respectively). No associations were found between retention rates and the remaining four individual strategies.

Relationship between retention rate and emerging strategies

The final group-level analysis investigated the association between emerging retention strategies and retention rates. Within these 95 retention strategies, 44 emerging strategies were identified, including the application of social media and SMS to assist in tracing participants lost to follow-up, and the application of study websites and social media profiles for keeping participants up-to-date with the study’s news and events. Meta-regressions demonstrated that articles reporting a higher frequency of emerging retention strategies had higher retention, after controlling for study duration and number of waves (b = 0.08; 95%CI [0.01 to 0.16]; p = .03). Despite this, there was no difference in overall retention rates between those articles that did and did not report the use of emerging retention strategies (median retention using emerging strategies = 80.1%; median retention not using emerging strategies = 75.0%; b = 0.27, p = .27).

Discussion

This study aimed to identify retention strategies employed in longitudinal cohort studies during the past decade, and to examine their effectiveness. We identified 143 longitudinal cohort studies that described retention strategies and outcomes, resulting in 95 different retention strategies. We then investigated whether study or participant characteristics moderated retention, the relationship between retention rate and retention strategy type, and whether new cohort retention strategies have emerged since previous reviews. In so doing, this study is the first meta-analysis of retention strategies conducted in longitudinal cohort studies. This research particularly complements the previous narrative review that investigated cohort retention strategies in longitudinal research [6], and the wider literature investigating participant retention strategies across health research designs (e.g., 4,16,17). Such research has important implications for maximising cohort retention and reducing research administration costs, which will subsequently improve the efficacy and quality of health research.

We first investigated how study or participant characteristics may influence cohort retention. Study characteristics included sample size, study duration, number of waves, and country development level - none of which were associated with retention rate. Participant characteristics included mean age at baseline, cohort type (clinical or non-clinical), and gender. We found that cohort studies with a higher proportion of male participants had lower retention rates than studies with a higher proportion of female participants; no associations were found for participants’ age or cohort type. While difficulties in retaining male participants are well-documented in previous research (e.g., 4,20,21), our study noted that cohorts with a higher proportion of male participants were also more likely to be clinical samples than cohorts with a higher proportion of female participants. In addition, cohorts with a higher proportion of male participants were also disproportionately focused on high-risk groups, such as substance use and men who have sex with men (e.g., the Bangkok Men who have Sex with Men Cohort Study (BMCS) [25] and the International Multicenter ADHD Genetics (IMAGE) study [26]). Thus, the difficulties in retention reported in this study and the wider literature could potentially be attributed to the differential impact of these clinical issues that affect men more than women. Researchers working with hard-to-retain populations, such as men in particular clinical groupings, may benefit from investigating what retention strategies work within their specific populations and settings beyond the core retention strategies identified in this review.

Second, we investigated the relationship between retention rate and retention strategies. We identified 95 different retention strategies, grouped thematically into four classes: barrier-reduction, community-building, follow-up, and tracing. Specific strategies associated with improved retention rates included the barrier-reduction strategy of offering alternative methods of data collection to participants (e.g., completing an interview over the phone or in person); and the tracing strategy of collecting detailed contact information from participants at baseline via a locator document. Further, weak evidence was found for one community-building and two further barrier-reduction strategies: (i) sending participants thank you, birthday or holiday cards; (ii) having consistent research team members, and; (iii) offering site and home visits for data collection.

Overall, barrier-reduction strategies emerged as the strongest predictor of improved retention. Barrier-reduction strategies may be particularly useful in longitudinal research given participants are likely to experience significant changes in their capacity to remain involved across the study’s duration (typically years). Follow-up/reminder strategies, such as incentives and reminders, were associated with significantly poorer retention. This result was surprising, given that the previous review investigating retention strategies in longitudinal cohort studies found the opposite, that use of these follow-up/reminder strategies resulted in improved retention rates [6]. The lack of support for follow-up/reminder strategies found in the current review could be due to a number of extraneous variables including: (i) timing: studies may have implemented this strategy after other retention efforts proved ineffective; (ii) participant burden: the studies using follow-up/reminder strategies may have involved a high data collection burden (e.g., long surveys); (iii) sampling: studies using follow-up/reminder strategies may be over-represented in studies of difficult-to-retain populations, such as men. However, these explanations are unlikely, given that follow-up/reminder strategies were identified in most of the cohorts included in this review (111 out of the 143 cohorts), and the cohorts employing follow-up/reminder strategies did not differ by research design (sample size: t(141) = .67, p = .50; no. waves: t(141) = −.43, p = .67) or participant characteristics (age: t(141) = −.11, p = .91; gender: χ2(2, n = 143) = .37, p = .85; HDI: χ2(2, n = 143) = .01, p = .97). Differences were observed only for study duration (any M(SD) = 3.9(4.4); none (SD) = 5.8(6.4); t(141) = 2.00, p = .05). Alternatively, participants may perhaps view follow-up/reminder strategies as the research team “badgering” them to complete assessments, thereby damaging rapport. This negative perspective of follow-up/reminder strategies may be further exacerbated if the research team has not implemented sufficient barrier-reduction strategies to help make it easier for participants to remain involved in the study. Future research could consider investigating participants’ perspectives of retention strategies in longitudinal cohort studies, ensuring that both active and inactive participants are included, to better understand the costs and benefits of different approaches.

Interestingly, the current study found that simply adding more cohort retention strategies did not result in higher retention rates. These results contradict the findings of Robinson et al. [17] and Davis et al. [4], who both found that the use of more retention strategies across multiple classes was associated with improved retention rates. However, neither study specifically examined participant retention in longitudinal cohort studies, and both synthesised their retention results using a narrative rather than meta-analytic approach. Given that the implementation of retention strategies can be costly in terms of both time and money, the overall number of strategies employed is important to evaluate. The interaction of quantity of retention strategies used and provision of flexibility needs to be better understood, given research protocols that accommodate the changing lives of participants should remain a key focus of retention efforts.

Finally, we examined whether studies utilising new or emerging retention strategies had improved retention compared with studies using established strategies. Of the 95 retention strategies described in the included articles, 44 were identified as an emerging retention strategy that had not yet been described in extant systematic reviews examining participant retention [4, 6, 16, 17]. Emerging strategies included using social media and SMS to assist in tracing participants lost to follow-up, and the use of study websites and social media profiles for keeping participants up-to-date with study news and events. Emerging retention strategies were endorsed by only a handful of studies, and the use of a single emerging strategy was not significantly associated with retention rate. However, we found that studies that employed more emerging retention strategies were associated with improved retention rates. Importantly, emerging strategies were identified across all four retention strategy domains (barrier-reduction, community-building, follow-up/reminder, and tracing), demonstrating that the association between emerging strategies and improvements in retention are due to the use of modern technology to help achieve core cohort engagement goals. Thus, we recommend that researchers continue to innovate their retention efforts, particularly where such strategies may reduce participant burden.

The current study has a number of limitations. First, the number of articles that focused on reporting retention strategies in detail was proportionally low compared to the number of articles that did not focus on reporting retention strategies. Although retention strategies were identified within 143 longitudinal cohorts, only 55 included cohort retention as a key focus area. Very few articles (n = 12) were identified that reported strategy-specific retention rates within the longitudinal cohort studies. The number of retention strategies reported by articles ranged from one to 32, with 35 of the 141 articles describing only one retention strategy. Longitudinal cohort studies should aim to publish protocol papers that delineate their cohort retention strategies, and ensure that the protocol is updated as retention efforts evolve.

Second, net retention rates were calculated by the difference between the first and last wave of data collection reported in the article. Where specified, ineligible participants (e.g., participants recruited after the first wave, or deceased participants) were excluded from the retention rate calculation. However, some articles did not provide detailed information on the eligibility of the sample at the final wave, and thus it is possible that the retention rates calculated for some studies may be slightly inaccurate. This limitation could be addressed by researchers providing details on the eligibility of their samples at each wave.

Third, high levels of heterogeneity were reported for most analyses in this study. This may best be explained by two factors. First, we expected to identify high heterogeneity given the diversity of research questions, methodologies, and cohorts reported across articles. Second, only a small number of studies were eligible for most meta-regressions in this paper, which reduces the precision of heterogeneity estimates [27]. This limitation could be addressed in future work, which could aim to investigate the effectiveness of different retention strategies within different subgroups.

Finally, by nature of synthesising retention results across different samples and settings, the current study is unable to disaggregate nuanced effects of various retention strategies across specific contexts and populations, given results are pooled across multiple studies. The current study did address this broadly by investigating the effects of study and sample characteristics on retention.

A final point to note is that available to researchers are a range of statistical or methodological approaches that can minimise potential biases introduced with attrition. Whilst beyond the scope of this paper, these approaches include formal statistical methods for addressing missingness due to attrition such as multiple imputation or full information maximum likelihood methods [28, 29]. Moreover, researchers may address attrition methodologically by using replacement sampling approaches that recruit new participants into a study to replace those who have dropped out, based on shared characteristics measured in the original sampling frame [30, 31]. All these methods provide useful avenues to address attrition once any employed retention strategies have been used to retain the largest proportion of the original sample as possible.

Conclusions

Overall, this study has important implications for the retention efforts of longitudinal cohort studies. Combined, these results suggest that researchers need to be strategic in choosing how to invest their resources to better target participant retention, rather than simply increasing the number of strategies applied. Projects should invest both time and funding into matching retention strategies to the sample prior to implementation, including careful consideration of unintended burden for participants. Finally, given the high number of emerging retention strategies identified, longitudinal research methods clearly continue to evolve. Longitudinal cohort studies may benefit from open and regular protocol revision to incorporate new strategies, particularly where these strategies may offer greater flexibility to participants.