The goal of the current meta-analysis is to confirm and extend the work of two previous meta-analytic reviews (Wunderlin et al., 2021; Stanyer et al., 2022). Here, we combine studies included in these two meta-analyses (Figure 1A). Additionally, we compare the statistical measures across our study and the two prior meta-analyses and confirm a high correspondence with Wunderlin and colleagues (Figure 1B) but less so with Stanyer and colleagues (Table 1 and Figure 1C). To expand on these prior studies, we conducted a supplemental google scholar search of the last 3 years (2020–2023), using the following key-word search: (“acoustic stimulation” OR “auditory stimulation”) AND (“slow wave sleep” OR “slow oscillations”) AND (“memory” OR “consolidation”). The first 50 most relevant results are given a full-text screening for inclusion. This results in two relevant articles (Prehn-Kristensen et al., 2020; Harrington et al., 2021). Harrington et al. (2021) is likely excluded from previous meta-analyses (Wunderlin et al., 2021; Stanyer et al., 2022) as a consequence of being published after literature review. Prehn-Kristensen et al. (2020) may have been excluded as it compares subjects with and without attention deficit hyperactivity disorder (ADHD). Here, we extract the necessary effect size for the healthy control subjects for both the rewarded and non-rewarded conditions, which are included in the present meta-analysis. Additionally, we searched for studies that had cited the already-included articles and found no additional studies. In sum, we obtained k_S = 12 studies (k_ES = 14 effect sizes), and N = 206 healthy subjects of variable age (Supplementary Figure 1).

Key study design features were aligned across all studies included. First, all studies included examined the efficacy of noise-burst acoustic stimulation during slow-wave sleep on overnight memory retention in the form of a word-pair cued recall task. Two experimental sessions consisted of a SHAM condition (control, no acoustic stimulation) and a STIM condition (with acoustic stimulation) during sleep (nap or overnight). The studies all had repeated measures and cross-over designs. That is, the same subjects participated in the control (SHAM, no audio) and acoustic stimulation (STIM) conditions. Ordering of STIM and SHAM conditions was appropriately counterbalanced in all studies. Word pairs were shown to participants, and they were subsequently assessed on their immediate recall prior to the sleep session. After waking, memory was assessed on the same word pairs. Memory retention scores were the difference in recall accuracy between pre-sleep and post-sleep assessments in all studies included. Memory retention scores were then compared between STIM and SHAM conditions.

Data included in this meta-analysis was obtained from each publication or by contacting the corresponding authors. Relevant study characteristics were recorded (see Table 2) alongside effect size information. Fortunately, we were able to collect all characteristics of each moderator of each study. Most coded characteristics were directly reported in the respective manuscripts. However, slow oscillation power was not, therefore we calculated the standardized mean difference (SMD) in slow oscillation power/amplitude between STIM and SHAM. To quantify the effect of acoustic stimulation on memory retention, the standardized mean differences in word-pair retention scores between STIM and SHAM conditions were recorded. To obtain repeated measures effect sizes, correlation coefficients must be calculated between pre-sleep and post-sleep measurements. Since most studies did not report pre-sleep vs. post-sleep correlations, corresponding authors were contacted in order to acquire raw data, or data was extracted from figures containing sufficient information using WebPlotDigitizer. We were able to obtain raw data sets for seven independent samples from five studies (Ngo et al., 2013b, 2015; Weigenand et al., 2016; Henin et al., 2019; Schneider et al., 2020). Two samples were included from Ngo et al. (2015), but the second experiment had no SHAM condition therefore this sample is not used in the effect size calculations, however it is included in the reliability section of the meta-analysis. Since each study implemented a crossover design, we were able to estimate reliability coefficients for each of the seven raw data sets. Additionally, five studies reported the proportion of subjects that were aware of the stimulus at some point during the STIM condition (Ngo et al., 2013b, 2015; Weigenand et al., 2016; Prehn-Kristensen et al., 2020; Schneider et al., 2020). This proportion of subjects aware of the acoustic stimulus at any point was extracted from these five studies to examine the possibility that subjects were potentially not entirely blind to study conditions due to the no audio vs. acoustic stimulation designs employed by all (Figure 2C).

Previous meta-analyses reported different formulations of standardized mean differences (SMD) in word-pair retention scores between STIM and SHAM conditions. Wunderlin et al. (2021) used a repeated measures estimator (d_rm), whereas Stanyer et al. (2022) presumably used an average variance estimator (d_av). Our meta-analysis finds that variances in STIM and SHAM conditions violate the assumption of equal variances. This is evidenced by the average (n-weighted) variance ratio of 1.38 (Sstim2/Ssham2). This indicates that the STIM condition has a 38% higher variance than the SHAM condition. Therefore, we use the Glass' estimator (d_δ) to calculate standardized mean difference (SMD) because it does not assume equal variances. An in-depth description of various effect-size calculations can be seen in Lakens (2013). All the estimators, d_rm, d_av, and d_δ only differ in how they are standardized. Thus, they all take the form of:

Where J(n) is the small sample correction factor, J(n)=1-34n-5, and M_sham and M_stim correspond to the mean of the SHAM and STIM condition, respectively. The differences between estimators are in their calculation of the standardizer, S^*, such that:

Where S_stim and S_sham indicates the standard deviation for STIM and SHAM, respectively. The associated standard errors (se) for each SMD estimator are as follows:

Where n indicates the sample size. The point estimate and standard error of the repeated measures estimator (d_rm) requires the correlation between STIM and SHAM retention scores (r_d) which are rarely reported in repeated measure studies. Therefore, based on acquired raw data sets and plot digitizing (extracting data from figures), we calculated all available correlations. For the remaining studies where correlations were not available, the weighted mean correlation (i.e., random effects) is used to populate missing values.

To perform our new meta-analysis and confirm prior meta-analyses and assess potential sources of variation in our regression model we estimate the true effect size using a Random-effects model approach (Figure 1A). Random-effects modeling is used to estimate the mean of true effect sizes. Random-effects models allow for variation between true effect sizes (heterogeneity) across studies as should be the case when studies vary in their population, methodology, or design. We use the restricted maximum likelihood estimator to calculate between-study heterogeneity, that is, the standard deviation of true effect sizes (τ). All meta-analytic modeling is completed in R using the metafor package (Viechtbauer, 2010).

A meta-regression model was implemented to quantify the potential moderating effects of multiple study-level characteristics on the effect size (d_δ). Because our initial analysis identified a downward trend in effect size with publication date (Figure 2A), the regression model was logistically set up to examine whether independent moderators could account for this decline. Thus, we built nine different regression models to investigate whether the effect size was conditionally dependent on the study's publication date or other moderating variables (e.g., age; see Tables 2, 3). For interpretability, we also report the effects for various subgroups based off of categorical moderators (Table 4). The first model consists of a single moderator (publication date) while the other models consist of both publication date and one other moderating variable. As detailed above, the restricted maximum likelihood estimation method was used for all meta-analytic models. Additionally, we utilized the moving constant technique (Johnson and Huedo-Medina, 2011) by setting the initial moderator value to zero, this allows the intercept to be interpreted as the predicted effect size (d_δ) at the date of the original publication (Ngo et al., 2013b). To compare the nine regression models, each model is re-fit with maximum likelihood (as opposed to restricted maximum likelihood) which is necessary to conduct a likelihood ratio test between models. The likelihood ratio test assesses whether the addition of a given model parameter is warranted.

In addition, the sparsity of study-level characteristics could potentially drive our analyses to be insensitive to additive contributions of study-level characteristics to the observed heterogeneity in effect sizes. Many studies differed from Ngo et al. (2013b) methodologically, and some of those studies differed in multiple facets. To address this, we have constructed a “Methodological Deviation” factor consisting of the sum score of four other coded study design characteristics. Specifically, the sum score is taken as the sum of how study-level characteristics diverge from the seminal publication (Ngo et al., 2013b). The four characteristics used are age, semantic congruence, word count, and phase-locking. For example, if a study included subjects with similar mean age as Ngo et al. (2013b) (young adults), that component of general risk “Methodological Deviation” was coded as zero. The maximum “Methodological Deviation” is four, being relatively methodologically inconsistent with Ngo et al. (2013b), and the minimum zero being highly methodologically consistent with Ngo et al. (2013b). Subsequently, the “Methodological Deviation” factor was run with the omission of Ngo et al. (2013b). A major limitation to this index is that, since it is simply a sum score, it assumes that each methodological deviation from Ngo et al. (2013b) has identical weights.

Outliers can cause spurious results in small sample meta-regressions, therefore a leave-one-out cross-validation procedure was used to assess the robustness of the meta-regression model. For each iteration, one effect size is removed from the data set and the model is fit to the remaining effect sizes (Figure 2B). If the meta-regression model is to be considered robust, then each iteration of the leave-one-out-cross-validation should not greatly impact the model parameters. Specifically, each iteration of the leave-one-out-cross-validation should yield a significant slope coefficient, and no slope coefficient should significantly differ from the full model.

Reliability is a psychometric property that indexes the precision of a measurement instrument. A test-retest reliability coefficient is a method of estimating the reliability of a task by administering the same task on two separate occasions and then calculated the correlation between time points (see a diagrammatic representation in Figures 3A, B). In all the experiments within this meta-analysis, subjects underwent a cross-over design, that is, the same participants were in the control (SHAM) and treatment (STIM) conditions. Since there are multiple measurements for each participant, this allows for the opportunity to assess the test-retest reliability of the word-pair recall task used to measure memory retention. The word-pair cued recall task is characterized by three phases: (1) A learning phase where participants are presented with a list of word pairs, (2) a pre-sleep (immediate) recall phase where participants are assessed on their word-pair recall accuracy and scored on the number of word-pairs they correctly recall, (3) a post-sleep recall phase where participants are assessed on their word-pair recall accuracy and scored on the number of word-pairs they correctly recall. Memory retention is scored by the number of word pairs correctly recalled post-sleep subtracted from the number of word pairs correctly recalled pre-sleep (see Figures 3A, B):

Where the difference score, X_d is used as the final memory retention score. However, word-pair recall is an imperfect measure of memory retention, therefore we can calculate the observed difference scores of X_d in terms of true scores (T; scores indicative of actual memory retention) and error variance (E; scores indicative of random error):

In classical test theory, the reliability of a measure is a ratio of true variance to observed variance such that:

To compute the reliability of a difference score we need the reliability of pre-sleep scores (ρ_pre), reliability of post-sleep scores (ρ_post) and the correlation between pre and post-sleep scores (r_pp). The reliability of pre-sleep (ρ_pre) is the test-retest correlation between the SHAM condition's pre-sleep score and the STIM condition's pre-sleep scores. The reliability of post-sleep (ρ_post) is the test-retest correlation between the SHAM condition's post-sleep score and the STIM condition's post-sleep scores. Lastly, the pre/post correlation (r_pp) is estimated from the correlation between pre and post-sleep scores of the SHAM condition:

The estimates from each study are pooled to estimate the reliability of word-pair recall task as an indicator of memory retention (Table 5). A convenient property of Glass' estimator (d_δ) is the simplicity of the correction for task reliability. For a discussion on corrections for repeated measures effect sizes see the blog post by Pustejovsky (2023). For d_δ, the unreliability of the task introduces error variance into the total observed variance, thus inflating the observed standard deviation and subsequently reducing the observed standardized mean difference (d_{δ, obs}). This attenuation can be formally defined as:

Where d_{δ, true} is the true standardized mean difference (see Figure 3C). This attenuation is also carried over to the standard error such that:

The attenuation of the observed effect size (d_{δ, obs}) and the corresponding change in the standard error (se_{δ, obs}) altogether results in decreased statistical power with lower reliability (see Figure 3C).

Our meta-analysis examining a total of 14 studies finds a progressive annual decline in reported memory effects with acoustic stimulation during sleep. Accordingly, there is a decline in reported effects (standardized mean difference) between 2013 and 2023 (Figure 2A). Our meta-regression models examine multiple potential sources of variability contributing to this decline (Section 2) including publication date, age, gender, phase-locking, word count, and congruence of words in memory tasks, overnight sleep condition, double-blind procedures, and the physiological index of slow oscillation power (see Table 3). For effect sizes within categorical moderators, see Table 4. Using a meta-regression model (Section 2, Model 1) where publication date is the sole moderator, we find that the date of publication demonstrates a negative trend that accounts for 91.8% of the heterogeneity between studies (B = −0.166 [−0.254, −0.078], Q_m = 13.72, p = 0.0002, k_ES = 14 effect sizes, n = 206, τ_resid = 0.110, see Figure 2A and Table 3). Given the small number of effect sizes in this literature, we looked to evaluate the robustness of this trend with leave-one-out cross-validation (Section 2.7, Figure 2A) to ensure highly influential studies do not artifactually produce this trend. Our leave-one-out cross-validation (Section 2) analysis finds that the regression coefficient (beta) in each iteration remains significant and unchanged from the full model coefficient (Figure 2B). This indicates that the decline in effect size is not likely driven by outliers.

When adding one variable at a time using a step-wise procedure (Section 2), we find that no other study-level characteristic accounts for the annual decline in study effects (see Table 3, Models 2–9). Corrected Akaike and Bayesian Information Criterion (AICc and BIC, respectively) tests confirm that the publication date (Model 1) is the most parsimonious (according to BIC), explanatory model of the set (according to AICc). Additionally, log-likelihood test demonstrates that the addition of any alternative moderators does not improve the model fit. When we look at the predicted effect size (based on Model 1) across time, the first study (Ngo et al., 2013b) publication date has a large effect size d_δ = 0.99 (95% CI [0.49, 1.49]) that declines to a net negative effect of d_δ = −0.39 (95% CI [−0.73, −0.05]) in the year of the most recent publication (see Harrington et al., 2021, Figure 2A). This indicates that recent studies show a markedly reduced effect of acoustic stimulation on memory task performance, in spite of having similar cohort sizes. Next, we consider the possibility that there are cumulative methodological deviations accounting for the publication date effect. The relatively low number of studies deviating from Ngo et al. (2013b) in certain study characteristics (age, phase-locking, whole night) could lead to such analyses being underpowered for observing unique moderator effects. Likewise, combinations of study characteristics could drive heterogeneity in observed effect sizes. However, our “Methodological Deviation” moderator model (Section 2) indicates that cumulative methodological deviation does not account for any of the heterogeneity in effect sizes (B = −0.034 [−0.257, 0.190], p = 0.768). This indicates that a cumulative methodological deviation from Ngo et al. (2013b) does not explain the decline in effect sizes across studies or years.

In neuroscience, psychometric evaluations and measurement errors for assessing cognitive processes such as working memory and memory consolidation are rarely investigated. A potential shortcoming of standard test vs. retest memory task difference scores used in all the studies included in our meta-analysis is the loss of shared variability (Figure 3). For example, when there is no correlation between a first test (e.g., pre-sleep) vs. a second test (e.g., post-sleep) condition there is no shared variability resulting in a highly reliable difference score (Figure 3B, top row). In contrast, if subjects have similar variability across test and retest conditions there can be highly overlapped variability that is lost in the difference score measure (Figure 3B, rows 2 and 3). We estimate the reliability of the memory task difference scores as the ratio of true variance to total observed variance (Section 2, true plus error variance, Equation 10). If there is a loss in total observed variance due to overlapped variability for test vs. retest conditions this will result in low reliability for the memory performance difference score. The reliability of the word-pair association task used to measure memory retention (and consolidation) in this literature has yet to be evaluated. Thankfully, we can analyze this as investigators have provided seven raw data sets from a subset of five studies in the current meta-analysis (Section 2, Table 5). Each of these studies includes independent measures of memory task performance before and after sleep that can be used to calculate pre-sleep and post-sleep test-retest reliability (Section 2, Figure 3A). Using this model, we find that the pre-sleep memory task performance has fair reliability with a ρ_pre = 0.711 (95% CI [0.615, 0.807]). The post-sleep memory task performance shows a slightly better reliability of ρ_post = 0.835 (95% CI [0.775, 0.896]). However, the correlation between pre-sleep and post-sleep scores from the control sleep condition (SHAM acoustic stimulation) is notably strong at r_pp = 0.949 (95% CI [0.920, 0.979]). This high correlation between test-retest scores reduces the reliability of the difference score, as described above (Figure 3B). Additionally, we compute correlations for pre-sleep vs. post-sleep tests which are delayed by one week by calculating the correlation between pre-sleep scores from the STIM condition and post-sleep scores for the SHAM condition (Section 2). This delayed test-retest correlation is rpp(del) = 0.742 (95% CI [0.655, 0.830]). Based on classical test theory (Section 2, Equation 7) our estimates indicate a near zero reliability for difference scores (ρ_d = 0.006, 95% CI [−0.188, 0.199]). Even if the delayed pre-sleep vs. post-sleep correlation is used (a more liberal estimate), the reliability of the difference score is still close to zero, ρd(del) = 0.015 (95% CI [−0.178, 0.208]). This indicates that the observed changes in word-pair retention may be attributable to noise. This is surprising considering that the correlation between difference scores for STIM vs. SHAM conditions is r_d = 0.465 (95% CI [0.247, 0.682]), as the difference score correlation should be bounded by the reliability estimate. However, even if we treat this correlation as an estimate of the reliability of difference scores, the reliability would still be considered very poor. In summary, measurement error in the task used to measure memory retention is expected to bias effect sizes and decrease statistical power (Figures 3C, D). It is possible that the poor reliability of the memory task difference scores in conjunction with low sample sizes and variable study-level characteristics could drive the observed heterogeneity in effect sizes.

Previous meta-analyses (Wunderlin et al., 2021; Stanyer et al., 2022) reported small-to-moderate effects of acoustic stimulation on sleep-dependent changes in word-pair memory retention as well as a large degree of heterogeneity unexplained. We demonstrate a yearly decline in reported effects, explaining 91.8% of the heterogeneity. In light of these findings, we also presented updated analyses of moderators examined in previous meta-analyses such as age and semantic congruence in word-pair memory tasks (Table 3). As expected, the addition of all other coded study characteristics did not explain heterogeneity above solely publication date. These trends describe a yearly decline in the reported efficacy of acoustic stimulation to modulate sleep-related word-pair memory consolidation (or retention). We also report that studies consistently demonstrate significant modulation of slow oscillation power ranging from d_δ = 0.48 to d_δ = 2.91. Heterogeneity in observed effects on memory performance was not attributable to that of observed effects of acoustic stimulation on slow oscillation power.

Decline effects are a common phenomenon in psychological sciences (Schooler, 2011; Pietschnig et al., 2019; Schimmack, 2020) and other disciplines (Munafò et al., 2007). The decline effect describes the phenomena where effect sizes decrease upon subsequent replications of the initial study. The yearly decline presented here likewise explains previously reported significant findings (Wunderlin et al., 2021) for acoustic stimulation overnight memory retention. The inclusion of publication date as a covariate in Model 6 by Wunderlin et al. (2021) diminished the effect of phase-locked stimulation in non-elderly where the regression coefficient is no longer significant. A finding which, is consistent still with the addition of recently published works with phase-locked acoustic stimulation in non-elderly (children and adults; Prehn-Kristensen et al., 2020; Harrington et al., 2021). This is further demonstrated to be robust to the contribution of outliers via our leave-one-out-cross-validation analysis (Figure 2B). Highlighting the consistency of the significant effect of publication date on the heterogeneity in observed effect sizes across studies.

There are many reported mechanisms for the Decline Effect. Recent work (Pietschnig et al., 2019) has demonstrated that underpowered designs of initial studies drive observed effect declines in intelligence literature. Contrary to this, as demonstrated through our leave-one-out-cross-validation analysis of the effect of publication date on the observed effect sizes across studies, we see that no single study drives this trend. A relevant factor here is that the presently analyzed studies do not vary greatly in their sample size (n range = 11–24) and therefore statistical power. In addition, other work has argued that unpublished findings contribute to effect declines across studies (Schooler, 2011), as well as for selective reporting (Schimmack, 2020). However, while we can not confirm any previously purported mechanisms driving effect declines here, we do suspect that the poor reliability of the task being used in the present literature does decrease statistical power and bias the effect size estimates (Figures 3C, D).

The measurement properties of behavioral and cognitive tests are rarely investigated in behavioral and cognitive neuroscience. In this literature, memory retention is measured using difference scores of word-pair recall tests (post-sleep scores minus pre-sleep scores). Difference scores are sometimes necessary to isolate cognitive processes such as inhibitory control which is often measured using a color-word congruence or Stroop test where the reaction times of congruent trials are subtracted from the reaction time of the incongruent word pairs to create a difference score. However, difference scores are notoriously unreliable (Rogosa and Willett, 1983; Hedge et al., 2018), especially when the pre-test and post-test is highly correlated, which is the case in this literature. Here, we find that the word-pair task expressed as a difference score is not a reliable index of memory retention. When calculating statistical power at different “true” effect sizes, we find that low difference score reliability is associated with lower power (Figure 3B, vertical dotted line). Additionally, this low difference score reliability can result in variable and biased observed effect size estimates as illustrated for a range of simulated true effect sizes (Figure 3D). With lower reliability, effect size estimates for d_δ and d_av are consistently biased toward zero. Taken together, it is evident that the reliability of overnight changes in word-pair retention is inadequate, and drives biased effect size estimates and decreases statistical power. Given we are only able to obtain reliability estimates for a small portion of the presently analyzed study, we can not with confidence evaluate study-level differences in reliability as a potential moderator. However, low task reliability could potentially account for the large heterogeneity in effects across studies.

While the above analyses substantially explain a large degree of heterogeneity found in this body of literature, there remain other factors that could be of interest. There is great evidence suggesting that the rhythmic coordination of oscillations during slow-wave sleep is an essential component of memory processes during slow-wave sleep (Isomura et al., 2006; Staresina et al., 2015; Rothschild et al., 2017; Navarrete et al., 2020). It is our view that this is the theoretical and physiological basis for all the studies analyzed in the present study: Ngo et al. (2013b), Ngo et al. (2015), Ong et al. (2016), Leminen et al. (2017), Henin et al. (2019), Choi et al. (2019), Prehn-Kristensen et al. (2020), Schneider et al. (2020), Diep et al. (2020), and Harrington et al. (2021). Accordingly, the degree to which acoustic stimulation enhances slow-oscillation power and its temporal coupling to spindles could determine the corresponding memory consolidation. The frequency distribution of spindle oscillations and coupling of slow waves and spindle oscillations change during development through adulthood (Purcell et al., 2017; Hahn et al., 2020; Kurz et al., 2020). Additionally, slow wave activity declines with aging in adults (Mander et al., 2013, 2017). In theory, the efficacy of acoustic and other forms of stimulation to enhance slow waves, spindles, and memory consolidation could vary across development and aging. Hence, there is a strong rationale to constrain age to strengthen initial future studies validating an approach. Studies such as that of Schneider and colleagues find older adults can have reduced pre-stimulus slow wave oscillations compared with young adults and from a physiological standpoint this may impact the ability to enhance the slow waves (Schneider et al., 2020). However, the vast majority of studies fail to report the phase-locking estimates and their standard deviations for the arrival of acoustic stimuli relative to slow-oscillations, with the exception of three studies (Ong et al., 2016; Leminen et al., 2017; Papalambros et al., 2017). Similarly, sleep spindles are known to be coupled with slow oscillations and spindle density correlates with memory retention and general cognitive ability (Reynolds et al., 2018; Ujma, 2018) and spindles are considered biomarkers of the memory reconsolidation process (Hennies et al., 2016; Denis et al., 2021). Additionally, studies indicate slow and fast spindles play distinct roles in memory consolidation (Barakat et al., 2011; Mölle et al., 2011; Ayoub et al., 2013; Cox et al., 2017). In line with this, several studies included in this meta-analysis find distinct time scales of effects on slow vs. fast spindles (Ngo et al., 2013b; Ong et al., 2016; Weigenand et al., 2016; Henin et al., 2019) while other studies simply reported effects on fast spindles (Ngo et al., 2015; Leminen et al., 2017; Schneider et al., 2020; Harrington et al., 2021). However, several studies also reported reductions in sleep spindle power during periods of slow-wave sleep without acoustic stimulation (Weigenand et al., 2016; Diep et al., 2020; Schneider et al., 2020). Taken together, studies varied greatly in the extent to which they broadly characterized changes in these features during sleep. Inconsistent reporting of characteristic changes in slow wave and spindle oscillations and their relationships to one another and behavioral measures makes meta-analytically addressing these problems infeasible with such small sample sizes. In line with this, the sparse distribution of study-level characteristics in combination with these small sample sizes could lead to a lack of clear variability between study subgroups, and moderator analyses being insensitive to contributions of study-level characteristics to the heterogeneity. Thus, better tracking and reporting of these metrics in future studies with larger sample sizes would greatly facilitate future studies and meta-analyses.

In conclusion, our results demonstrate a yearly decline in the reported efficacy of acoustic stimulation to modulate overnight word-pair retention. The presence of such a trend in small bodies of studies investigating novel interventions is unsurprising (Schimmack, 2020). For the studies presented here, this finding is accompanied by another highly relevant characteristic across studies: the memory tasks used are largely unfit for the present study designs. While several study characteristics were unable to be accounted for in this meta-analysis, and the present literature analyzed is small, we find the results presented here suggestive of the following five recommendations:

The application of these recommendations to future studies would greatly reduce the risk of small-sample bias, low statistical power, and confounding psychological effects. They would facilitate future research-synthesis efforts by providing the greater statistical power necessary for analyses of behavioral and physiological effects of acoustic stimulation of slow oscillations.