Does advance contact with research participants increase response to questionnaires: an updated systematic review and meta-analysis

Background Questionnaires remain one of the most common forms of data collection in epidemiology, psychology and other human-sciences. However, results can be badly affected by non-response. One way to potentially reduce non-response is by sending potential study participants advance communication. The last systematic review to examine the effect of questionnaire pre-notification on response is 10 years old, and lacked a risk of bias assessment. Objectives Update the section of the Cochrane systematic review, Edwards et al. (2009), on pre-notification to include 1) recently published studies, 2) an assessment of risk of bias, 3) Explore if heterogeneity is reduced by: delay between pre-contact and questionnaire delivery, the method of pre-contact, if pre-contact and questionnaire delivery differ, if the pre-contact includes a foot-in-the-door manipulation, and study’s the risk of bias. Methods Inclusion criteria: population: any population, intervention: comparison of some type of pre-notification, comparison group: no pre-notification, outcome: response rates. Study design: randomised controlled trails. Exclusion criteria: NA. Data sources: Studies which cited or were included in Edwards et al. (2009); We additionally searched: CINAHL, Web of Science, PsycInfo, MEDLINE, EconLit, EMBASE, Cochrane Central, Cochrane CMR, ERIC, and Sociological Abstracts. The searches were implemented in June 2018 and May 2021. Study screening: a single reviewer screened studies, with a random 10% sample independently screened to ascertain accuracy. Data extraction: data was extracted by a single reviewer twice, with a week between each extraction. Risk of Bias: within studies bias was assessed using the Cochrane Risk of Bias tool (ROB1) by a single unblinded reviewer, across studies bias was assessed using funnel plots. Synthesis Method: study results were meta-analysed with a random effects model using the final response rate as the outcome. Evaluation of Uncertainty: Uncertainty was evaluated using the GRADE approach. Results One hundred seven trials were included with 211,802 participants. Over-all pre-notification increased response, OR = 1.33 (95% CI: 1.20–1.47). However, there was a large amount of heterogeneity (I2 = 97.1%), which was not explained by the subgroup analyses. In addition, when studies at high or unclear risk of bias were excluded the effect was to reduced OR = 1.09 (95% CI: 0.99–1.20). Because of the large amount of heterogeneity, even after restricting to low risk of bias studies, there is still moderate uncertainty in these results. Conclusions Using the GRADE evaluation, this review finds moderate evidence that pre-notification may not have an effect on response rates. Funding Economic and Social Research Council. Preregistration None. Supplementary Information The online version contains supplementary material available at 10.1186/s12874-021-01435-2.


Introduction
Questionnaires have been one of the most common methods of data collection across the social and medical sciences. For example, in epidemiology pen and paper questionnaires alone were used in 29.2% of over 2000 analytic epidemiological studies included in a review of articles published in high-impact medical journals between 2008 and 2009 [1]. Likewise, about a third of empirical research published in management and accounting journals use questionnaires, and a review of a top social psychology journal found that over 91% of empirical studies published in the second half of 2017 used some form of questionnaire [2,3].
Inherent in using questionnaires is a risk of nonresponse. Potential participants, for example, might forget to complete questionnaires, and research ethics requires a right to refuse participation. Non-response can negatively impact on studies in three major ways: Firstly, non-response can introduce selection bias [4]. Secondly, even in the absence of selection bias, because nonresponse reduces the number of participants recruited into a study, non-response increases risk of random error (i.e. reduces statistical power and precision). Finally, nonresponse increases study costs [5].
It is therefore important to minimise non-response. One potential method is for the study team to contact potential participants in advance of them receiving the questionnaire (questionnaire pre-notification). In 2009, Edwards et al. published the third update of a 2003 Cochrane systematic review of randomised control trials evaluating methods of reducing non-response in both postal and electronic questionnaires [6]. They found that pre-contact increased response when compared to no pre-contact (OR = 1.5, 95% CI 1.26-1.78, for response after first questionnaire administration, and OR = 1.45, 95% CI 1.29-1.63 for response after final questionnaire administration). However, Edwards et al. (2009) did not assess the risk of bias in or across the included studies, and is now 10 years old, so therefore does not include research published in the last decade. In addition, there was substantial heterogeneity among the study results (p < 0.000001; I 2 = 91% for the response after the first questionnaire administration, and p < 0.00001; I 2 = 89% for the response after the final questionnaire administration).
There is therefore a need for an updated review which includes recently published studies, an assessment of bias risk in and across included studies. This review will:

Update Edwards et al. (2009)'s systematic review and
meta-analysis of randomised control trials examining the effect on non-response of pre-notification relative to no pre-notification (in any population) so that it includes papers published in the last decade. 2. To carry out an assessment of the risk of bias (i) in and (ii) across included studies. 3. To examine the extent to which between study heterogeneity is explained by: (A) the delay between pre-contact and questionnaire delivery, (B) method of pre-contact, (C) if pre-contact differs from questionnaire delivery, (D) if the pre-contact includes a foot-in-the-door manipulation (required participants to do something to receive the questionnaire), and (E) differences in the risk of bias of included studies, through conducting a subgroup analysis.

Protocol and registration
The methodology of the review and analysis was approved in advance by the LSHTM epidemiology MSc course directors. A copy of this form, approved on 21/03/2018, can be found in Supplementary Table 1. However the study was not otherwise registered. This study received ethics approval from the London School of Hygiene and Tropical Medicine MSc Research Ethics Committee on 26/03/2018. This study has been written in accordance with PRISMA-2020 [7].

Eligibility criteria Inclusion criteria
Types of population: This study followed Edwards et al. (2009) in using data from "[a]ny population (e.g. patients or healthcare providers and including any participants of non-health studies). " This should maximise generalisability over different contexts. Types of interventions: interventions must include some type of questionnaire pre-contact (pre-notification, advance letter/email/text/phone call or other Page 3 of 27 Woolf and Edwards BMC Medical Research Methodology (2021) 21:265 co-referring term). No restriction is placed on the type of questionnaire pre-notification. Comparison group: Included studies need to be able to make a direct comparison of the effect of questionnaire pre-notification vs no pre-notification (i.e. include at least one arm which received identical treatment to the pre-notification arm other than not receiving the pre-notification).

Types of outcome measures:
The proportion or number of completed, or partially completed questionnaires returned after all follow-up contacts were complete. Types of study design: Any randomised control trial evaluating a method of advanced contact to increase response to questionnaires. The inclusion of only randomised control trials should on average eliminate risk of confounding biasing estimates within studies.

Exclusion criteria
There are no exclusion criteria.

Information sources
Relevant studies identified by Edwards et al. (2009 Non-English papers were translated using Google Translate.

Study selection
The eligibility assessment was conducted by one reviewer following a standardised procedure. This process was repeated on a random 10% by a second reviewer with 99.7% agreement. Citations were uploaded onto Covidence (http:// www. covid ence. org/), a website specially designed for paper screening by the Cochrane Collaboration. Covidence automatically identified duplicates of citation/abstracts, which were then manually checked for errors.
Studies were first screened based on abstracts and titles, then full text. This process was repeated for any study which was referenced by or itself cited by an included study, and on the content of any potentially relevant review identified in the search.

Data collection process
A standardised data extraction sheet (Supplementary Table 3) was developed. The sheet was pilot tested on 10 randomly chosen studies from Edwards et al. (2009). One reviewer extracted data from included studies. To minimise transcription errors, this process was duplicated by the same reviewer 1 week later. Disagreements were resolved by extracting information for a third time and using the third extraction as the definitive extraction.
To check for duplication studies which shared at least one author were compared based on similarity of study population, date, and methodology. Duplicate trials were treated as a single study in the meta-analysis.

Data items
Information extracted for each included trial comprised 5 domains: 1) Information on the inclusion criteria: The study design, nature of the control arm, information on the intervention arm(s), information about the outcome measurement (the number of responses, and/or the response rate, in each arm). 2) Information on risk of bias: how the allocation sequence was generated, information of allocation concealment, blinding of participants and personnel, blinding of outcome assessors, any incomplete outcome data, information on other possible sources of bias (e.g. source of funding). 3) Information on the participants: the total number of participants, numbers in each arm, setting, country. 4) Information on the outcome: number of items returned, or response rate, in each arm. 5) Other information: the time from the sending of prenotification to questionnaire, if it includes a foot-inthe-door manipulation, the type of questionnaire administration, the type of pre-contact.

Risk of bias in individual studies
Assessment of risk of bias within each study was conducted by one unblinded reviewer. Information on risk of bias was extracted twice with a one-week gap between each extraction, and conflicts were handled by using the results of a third extraction. Authors included in the 2018 search were contacted for extra information about study bias risk, and still existent copies of communication from Edwards et al. (2009) were examined. Bias was evaluated using the Cochrane Risk of Bias tool [8]. The tool involves rating the risk of bias across 7 domains (random sequence generation, allocation concealment, blinding of participant and personnel, blinding of outcome assessment, incomplete outcome data, selective reporting, and other biases) at the outcome level. Within each domain, the studies were ranked as either high or low risk of bias, depending on the description of the study provided. If insufficient information was provided to form a decision, studies were designated as 'unclear' risk of bias. Studies were classified as at a low risk of bias if they had a low risk in all domains, at a high risk of bias if at a high risk in one domain, and were otherwise classified as having an unclear risk of bias. A full description of the tool can be found in chapter 8 of the Cochrane Handbook [8].
Results are stratified based on Risk of Bias score.

Summary measures, and planned methods of results synthesis
The primary summary measure of association estimated was the ratio of the odds (OR) of response in the treatment groups compared with the odds of response in the control group.
In line with Edwards et al. (2009), the meta-analyses were performed by comparing the ORs using a randomeffects model. The analysis was performed on an intention-to-treat basis. Outcomes were only included if they occurred within the period of follow up.
The results were synthesised in a meta-analysis conducted using STATA 15, using the 'metan' command [9]. To be consistent with Edwards et al. (2009), a random effects meta-analysis was used. Heterogeneity was assessed using the Cochran-Q Chi [2] statistical test for heterogeneity, and the I 2 statistic [10]. Results were presented using a forest plot.
To test the hypothesis that heterogeneity is explained by 1) the length of time between pre-contact and questionnaire, 2) method of pre-contact, 3) if pre-contact and questionnaire delivery differ, 4) if the pre-contact includes a foot-in-the-door manipulation, four planned subgroup analyses were conducted by separately stratifying the meta-analysis on these factors. Studies in which participants were not all assigned to the same type of pre-notification were excluded.

Risk of bias across studies
Risk of bias across studies was assessed with funnel plots. Asymmetry was investigated informally, by visually assessing how symmetrical the plots are around the effect estimate, and formally, using Harbord's test. Funnel plots were created using the 'metafunnel' command in STATA. Because ORs are naturally correlated with their standard error, response rates were used instead of ORs [9].

Assessment of certainty in the body of evidence
Outcome level limitations were evaluated using the GRADE approach [11] for both the overall estimate, and the estimate for studies at low risk of bias.

Study section
A total of 103 papers, reporting a total of 107 trials, were identified for inclusion in the review. The search resulted in a total of 35,931 citations, including 14,207 duplications. Eight reviews (Supplementary Table 4) were included in the search and checked for citations. The reasons for exclusions are stated in Fig. 1 and Supplementary Table 5. The numbers identified and excluded at each stage are described Fig. 1. After re-reading the reports, and contacting study authors, five studies (Temple-Smith 1998 [12]; Waisanen 1954 [13]; Wright 1995 [14]; Wynn 1985 [15]) which were included in Edwards et al. (2009) were excluded for not having randomised participants to receive or not receive a pre-notification. No duplicates were identified during data extraction. Overall, the updated review now includes 60 more studies than Edwards et al. 2009; increasing the number of participants from 79,651 to 364,527.

Study characteristics
Of the included studies, 32 (31.1%) were factorial designs. 60 (58.3%) were conducted in North America, 33 (32.0%) in Europe. Two (1.9%) were conducted in East Asia (Hong Kong and Thailand), 7 (6.7%) in Australia, one study did not state where it was conducted, and none were conducted in South America or Africa. One study was not written in English. 85 (79.4%) of the pre-notifications were posted. 19 (17.8%) of the others were telephone, with a few delivered by email (n = 7, 6.5%) or text message (n = 7, 6.5%). Only 17 (15.9%) trials reported a pre-notification which included a foot-in-the-door manipulation. 28 (26.2%) trails had a delay of less than 1 week, 33 (30.8%) had a delay of 1 week, 11 (10.3%) of 2 weeks. One (0.9%) for delays of 3 weeks, 5 weeks and 6 weeks. 70 (65.4%) trails administered the questionnaire by mail, 24 (22.4%) over the phone, 12 (11.2%) by email or online, and one used interviews. The characteristics of the included studies are described in detail in Table 1.

Risk of bias within studies
Judgments formed for each domain of the Cochrane Risk of Bias tool in each study are represented graphically in Fig. 2. The supporting evidence can be found in Supplementary Table 6. Overall, 8 studies were at high risk, 21 at low risk and 78 were at unclear risk. The proportions of studies at each level of risk is presented in Fig. 3.

Sequence generation
Thirty-three studies described the process used to generate the random sequence, or confirmed the use of randomisation in correspondence. Seventy-four studies have an uncertain risk of bias.

Allocation concealment
Thirty studies described concealment, or confirmed it in communication. Five confirmed that they had not used allocation concealment in communication. The remaining 72 studies provided insufficient information to reach a judgment, and so are of unclear bias.

Participant and personnel blinding
Participant and personnel blinding was not reported most trials. However, the design of many trials ensured that a degree of blinding did occur. A common design was to randomise participants to receive or not to receive a pre-notification without prior consent. The pre-notification itself would also often not explain that the participant had been allocated to receive it randomly. Thus any effect of treatment could not be due to the effect of knowing that they had been specially selected for an intervention which others had not got. Although the participant still knew they had received the pre-notification, this knowledge is part of the effect of a pre-notification -and therefore does not introduce any risk of material bias. Similarly, although most did not describe any blinding procedure for personnel, its absence was often unlikely to lead to bias in estimates. In studies using a pre-written pre-contact (e.g. e-mail, letters, SMS) unblinded study personnel do not have the ability to influence the experience or perceptions of potential participants, as their only means of communication with each other is through a pre-written pro-forma message. This, however, is not true for studies which used a telephone pre-notification, in which the personnel and potential participants can have a genuine interaction. No study with telephone prenotification reported no blinding of personnel.
Overall 92 studies were regarded as being at low risk of bias, and 15 at unclear risk.

Blinding of outcome assessment
Outcome assessment blinding was reported in 8 studies. However, the outcome (whether the questionnaire had been returned) is objective, and unlikely to be influenced by whether the outcome assessor knows the group assignment. Because the analyses are a comparison of two proportions, data analysers were unlikely to have enough researcher degrees of freedom for bias to be introduced in the analyses. All studies were therefore judged as being at low risk of bias for this domain.

Incomplete outcome data
One hundred three provided enough information to ascertain the total number of participants randomised in each arm and the total number of questionnaires returned in each arm. However, 4 are at unclear risk because they did not report sufficient detail to estimate per protocol rates, or state if the rates were intention to treat or per protocol, and one study at high risk.

Selective reporting
There was little evidence of selective reporting. All studies reported information on the relevant outcomes of interest. However, study protocols were not examined.

Other biases
Three of the factorial studies had significant interaction effects.

Results of individual studies
The results from individual studies are presented in a forest plot, Fig. 4. Fifty-nine studies had 95% confidence intervals which were incompatible with the null hypothesis, of which 55 implied that pre-notification increased response rates. There were a number of studies which appeared to have extreme results (Stafford 1966 [51]; Kulka 1981 [80]; Gillpatick 1994 [40]; Rodgers 2018 [112]; Sakshaug 2019 [113]; Taylor 1998 [53]). The extreme result of Rodgers appears to be due to the unusually high overall rate of response (97.1%). The other apparent outliers all were at high or unclear risk of bias.

Synthesis of results
Information on response was available in all trials, thus data from all trials was used. These randomised a total of 338,429 participants, and had 174,323 returned questionnaires. The pooled estimate shows an increase in response for the final follow-up after questionnaire prenotification (OR = 1.33, 95% CI: 1.20-1.47, p < 0.001), compared to an increase of 1.45 (95% CI 1.29 to 1.63) for Edwards 2009 (Supplementary Table 7). There was  Table 8). However, studies with low risks of bias and which send the pre-notification online had 95% confidence intervals which were compatible with the null hypothesis and appears to have reduced I 2 (67.4 and 65.1% respectively).

Risk of bias across studies
To explore the possibility of small study bias, funnel plots were created for the outcome, Fig. 5. Visual assessment implies that there is no major asymmetry. However, more studies than expected fell outside the 95% confidence limits. In addition, a formal assessment of asymmetry, using Harbord's test, did not find evidence to reject the null hypothesis of no asymmetry (p = 0.749).

Effect of risk of bias within studies on the pooled results
Seventy-eight studies were at unclear risk, 21 at low risk, and 8 at high. When stratified by risk of bias, there was no longer evidence against the assumption of a pooled association across studies which were of low bias (OR = 1.09, 95% CI: 0.99-1.20, Fig. 6).

Risk of Bias
Across domains, high risk of bias was uncommon. However, few studies provided sufficient information to be assigned low risk of bias. The interpretation of the overall results is therefore downgraded.

Imprecision
Due to the large number of participants in each arm, even after stratification by bias risk, confidence intervals were relatively narrow. GRADE suggests additionally assessing he 'optimum information size' (i.e. have the number of participants a randomised trial needs to have sufficient power to answer the question) [118,119]. Because larger sample sizes are required to detect smaller estimates, we calculated the optimum information size using information from the meta-analysis of studies at a low risk of bias (see Supplementary Table 7). Around 2500 participants would be required for each arm, for a 90% power and 5% alpha, which was obtained for both estimates.

Indirectness
There was generally little indirectness in the review. All studies were randomised control trials examining the effect of pre-notification on questionnaire response, so directly answered the review's question.

Publication bias
Visual inspection of the funnel plots and formal testing with Harbord's test both imply that small study bias was unlikely. As high questionnaire response is important to non-academics, e.g. polling companies, an unassessed grey literature will probably exist.

Heterogeneity
There was substantive heterogeneity within the review, and in all stratified analyses. We therefore downgraded the evidence due to the unexplained heterogeneity. Future studies should consider further explanations.

Overall GRADE evaluation
After two downgrades, there is low certainty in the overall estimate, but, with only one downgrade, moderate certainty in the estimate for studies at low risk of bias.

Summary and interpretation of evidence
This meta-analysis and systematic review of randomised control trials examined the effect of pre-notification compared to no pre-notification on questionnaire response rates. Pre-notification led to 1.33 (95% CI: 1.20-1.47) times greater odds for response. However, this was greatly reduced after restricting to studies of low risk of bias, OR = 1.09 (95% CI: 0. 99-1.20). This low OR implies that researchers should be cautious when using pre-notification as they may not lead to improvements in participant response rates. Specifically, in instances where pre-notification would be an expensive addition to a study, we believe that there is too much uncertainty to recommend the use of a pre-notification. One potential implication of the remaining unexplained heterogeneity is that there are unmeasured effect modifiers which cause pre-notification to work in some circumstances but not other. Therefore, if pre-notification would have a negligible impact on the cost of recruiting participants, nesting a high-quality randomised control trail could help reduce the uncertainty around the potential benefits of pre-notification in a specific setting.

Limitations of the evidence included in the review
Level of certainty in the evidence The level of certainty in both the overall and low risk of bias estimates were downgraded because of high unexplained heterogeneity. Exploring other factors could be a topic of other reviews. The large number of high and unclear risk of bias studies lead to the overall estimate being downgraded an additional time.
The number of studies with an unclear risk of bias could have potentially been reduced if studies in the 2021 search were contacted for further information. However, the age of many of the remaining studies made communication difficult, e.g. due to address change, and information not being available for studies where contact could be made. In addition, between the beginning of the project and its end Cochrane released an updated version of the Risk of Bias tool. The new tool changed the structure  of the evaluation and by allows reviewers to come to a qualitative decision about the probability of bias risk in each domain. Most studies with an unclear risk of bias have it because they did not describe randomisation and/or allocation concealment in sufficient detail. It is likely that many of these studies could have been either upgraded or downgraded when evaluated using ROB2 based of covariate balance. We would therefore expect fewer studies to have an unclear risk of bias if we had used ROB2.

Generalisability
There are very few studies from lowor middle-income countries. The review's results may not generalise to any population, especially given the heterogeneous effect.

Limitations of the review process
Search strategy Cochrane recommends that the literature searching be done by two independent reviewers, while this review only used one [120]. In addition, the search lacked specificity, and some extra publications might have been found by contacting authors to see if they had published other studies on the question. However, citation searching is not always common in systematic reviews, although it proved an effective way of detecting new studies.

Data extraction and risk of Bias assessment
Cochrane recommends that data extraction should be done by two independent reviewers [121]. Although this review only used one reviewer to extract data and conducted the risk of bias assessment, both were done twice by this reviewer, which should also reduce transcription errors. There is still, however, some risk of bias due to the reviewer being unblinded.

Strengths and weaknesses in relation to other studies
The updated review more than doubled the number of included studies, even with four old studies were excluded for poor methodology (Supplementary Table 5). The overall results of the two studies are relatively similar, with overlapping confidence intervals overlap the results of the two studies might be consistent. However, restricting to low risk of bias studies implies that this estimate may be due to study bias. Therefore, while Edwards et al.
(2009) concluded that pre-notification does improve response rates, this review would conclude that there is moderate evidence that pre-notification may not improve response rates to questionnaires.
Both Edwards et al., and this study, might be criticised for their choice of outcomes. Response rate does not entail response quality [5]. For example, a questionnaire might not have been fully completed, or completed inaccurately. In addition, to be a useful intervention for researchers pre-notification needs to be cost effective. However, neither of these outcomes are examined in the reviews.
The conclusion was also different from two other systematic reviews which explored a similar question. Both Lacy et al., and van Gelder et al., concluded that pre-notification did improve response rates (with OR = 1.45, 95CI 1.01 to 2.10, and OR = 1.12, 95%CI 1.12 to 1.22 respectively) . However, the 95% CI of both of these studies is compatible with the results of this study, and neither of these studies stratified their metanalyses by risk of bias.

Conclusions and implications for further studies and practice
This systematic review and meta-analyses of randomised control trials examining the effect of pre-notification on questionnaire response found evidence which supports the use of pre-notification. However, after excluding studies at high or unclear risk of bias the effect of the intervention was greatly reduced, and is probably no longer of relevance. The quality of evidence among low risk of bias studies was downgraded due to substantial unexplained heterogeneity. Future reviews could consider exploring other explanations. In addition, studies originated from a limited set of settings, such as generally high-income countries. Future studies could explore if the results generalise to new settings.