What sleep does for memory

You've met this claim once on the Core path, and just missed it once. How to Learn Anything said, in its last lesson, that sleep matters for learning at a smaller size than people think. Memory spent a whole course on how remembering works and didn't take up sleep at all. This lesson is where the two meet, and it is the lesson a reader arriving at a course on sleep is most likely to be waiting for.

It's built on one paper, a 2021 meta-analysis in Psychological Bulletin, and it asks two questions of it. How big is the benefit of sleeping after you learn something? And why does the answer change so much depending on how the study was run?

The second question is the interesting one. It's also a question about design, which makes it a direct descendant of lesson 1's question about instruments: what a number means depends on the machinery that produced it.

The paper, and how far this course got into it

Sabrina Berres and Edgar Erdfelder, at the University of Mannheim, pooled the literature on sleep and episodic memory.¹ Their own definition of the subject: "Episodic memory is the memory for events, experiences or information and their respective context (e.g., location, time; Tulving, 2002)." In practice, the paper says, the material has included single words, word pairs, prose passages, pictures, and combinations of them, learned in a study session and tested later.

What they pooled, in the abstract's words: "823 effect sizes from 271 independent samples that were reported in 177 articles published between 1967 and 2019."

Here's the read level, because this lesson prints a lot of numbers. This course read the authors' accepted manuscript, which is free on PsyArXiv, and read it in part: the abstract, the definitions of the study designs, the results on design and on selective reporting, the discussion of how large the effect is and what the design pattern means, and the design rows of two tables. It read the opening definition and two paragraphs of the Introduction on materials and designs, but not the long review of theories that fills most of the first half of the paper, or the other fourteen variables the authors tested. And the manuscript carries the publisher's own warning that it "is not the copy of record and may not exactly replicate the final, authoritative version of the article". Every figure below comes from that manuscript.

How big, in terms you can hold

The headline, from the abstract: "we found a moderate overall sleep benefit in episodic memory (g = 0.44)."

You met Hedges' g in Note-Taking lesson 5. It is a difference between two groups' averages, measured in standard deviations. So 0.44 says the people who slept after learning ended up, on average, a little under half a standard deviation ahead of the people who stayed awake.

That is still abstract, so here's one way to feel it. If scores in both groups are spread in the bell shape statisticians call the normal curve, a difference of 0.44 standard deviations puts the average sleeper ahead of about 67 percent of the people who stayed awake. That conversion is this course's arithmetic on the normal curve, not a sentence from the paper, and it depends on the bell-shape assumption.

So the sleep benefit sits among the effects you've already been taught to take seriously. That is the fairest single thing to say about its size. It isn't a small curiosity and it isn't a doubling.

The correction, done in the open

Lesson 3 of this course named publication bias: studies that find something get published more readily than studies that don't, so a pooled figure can run high. This paper tested for that and reported what it found, which is worth seeing because it is what a careful meta-analysis looks like.

The test didn't find clear evidence of bias. The authors adjusted for it anyway, and the estimate dropped: "the overall sleep benefit estimate was in fact only about half as large but still statistically significant in the main data set, g = 0.28".

Two things to notice about that sentence. First, 0.28 against 0.44 is a little under two thirds, not half, so the authors' own phrase, if anything, understates what survives their correction. Second, the corrected figure is uncertain in its own right, with an interval from 0.07 to 0.48, and a second check they ran put the corrected figure nearer 0.41. Their summary is balanced in both directions: "although the overall effect of sleep on episodic memory may have been overestimated in the present meta-analysis due to publication bias or other forms of selective reporting, we did not find strong evidence for selective reporting that would compromise the validity of our results."

And they add: "Importantly, a reliable sleep benefit persists even if potential selective reporting is accounted for." So this course's summary, and it's the course's, is that the benefit sits somewhere around 0.28 to 0.44, which on the normal-curve conversion puts the average sleeper ahead of about 61 to 67 percent of the wake group. Either end is a real benefit. Neither end is the thing a headline wants.

Seven ways to compare sleep with waking

You can't randomly assign people to sleep or not sleep for years. You can for a night, or an afternoon, and researchers have done it in several quite different ways. The paper sorts the literature into seven designs, and five of them matter here.

Here are the definitions, in the paper's words, because everything that follows depends on exactly what each group did.

• Natural sleep and wakefulness. "Participants in the sleep condition learn the material in the evening and are tested after a night of sleep in the morning. Conversely, participants in the wake condition learn the material in the morning and are tested after a day of wakefulness in the evening."
• Daytime nap. "Participants in the sleep condition nap during the day (i.e., before 7 p.m.), whereas participants in the wake condition stay awake throughout this time."
• Nighttime nap. The same, except that the sleep group naps "during the night (i.e., after 7 p.m.)".
• Total sleep deprivation with recovery night. "Participants in the sleep condition sleep normally, whereas participants in the wake condition stay awake in the first or second night after learning. Memory testing for sleep and wake conditions takes place after one or more recovery nights."
• Total sleep deprivation without recovery night. The same, except that "memory testing for sleep and wake conditions takes place directly after sleep deprivation".

The other two designs deprive people of one stage of sleep rather than all of it, slow-wave sleep (the deep stage, which the paper abbreviates SWS) or REM sleep. They come back in a moment, because they're the authors' best evidence for their own reading.

That is a genuine problem, and the rest of this lesson is what the literature does with it.

What the table says

Here are the results for those five designs. Each row gives the authors' estimate for that design, adjusted for the other variables in their model, the length of the gap between learning and test among them, with its 95 percent interval and the number of independent samples behind it.

The authors' own summary of it: "With the exception of total sleep deprivation with recovery night, g = 0.14, t(11.35) = 1.14, p = .278, all sleep study designs showed a significant sleep benefit in episodic memory."¹ This analysis ran on 777 of the 823 effect sizes, from 254 of the 271 samples.

The five designs this lesson uses, drawn from the design rows of Table 2 in Berres and Erdfelder's accepted manuscript.¹ The filled circle marks the only design of these five whose two groups learn and are tested at different times of day. The bottom row, in the second colour, is the only interval that crosses zero.

The clock, and what the naps do to it

Now the worked case. Take the time-of-day problem seriously and follow it through.

If the whole benefit were the clock, then a design that keeps both groups on the same clock should find nothing. So which of these designs keep both groups on the same clock?

Read the nap definition again. The sleep group naps and the wake group "stay awake throughout this time". Both groups learn at the same hour and are tested at the same hour; the only difference is what they did in between. That reading of the definitions is this course's own, but it is a close one, and it holds for the daytime nap, the nighttime nap and both deprivation designs. Of the five here, only the natural design puts its two groups on different clocks.

Notice what just happened, because it's the most useful habit in this lesson. The confound was real and the objection was a good one. But a literature that runs several designs contains its own checks, and you didn't have to take anybody's word for the answer. You read it off the rows.

The gap that is left, and two ways to read it

So there is a gap between the night-against-day design and the daytime nap: 0.54 against 0.34. What causes it?

The authors read it through slow-wave sleep. In their discussion: "nighttime naps, natural sleep and wakefulness, and REM sleep deprivation constantly showed the largest sleep benefits compared to other sleep study designs. This pattern fits well to the assumption that SWS is important for the sleep benefit. However, it might not be SWS duration per se, but rather specific SWS characteristics such as sleep spindles that matter most".

Their best evidence is the two designs this lesson set aside. Taking away REM sleep, which leaves slow-wave sleep in place, still left a benefit of 0.43 over 10 samples. Taking away slow-wave sleep left 0.31 over 9. The one that kept slow-wave sleep did better, though the two intervals overlap a good deal (0.28 to 0.58, and 0.02 to 0.59), so this leans rather than settles. And the nighttime nap's 0.54 fits them too: it has one clock and still matches the natural design, though on ten samples with a very wide interval.

They also know the strongest objection to their own reading. Their meta-analysis found that how much slow-wave sleep people got didn't predict the size of the benefit, and they cite as "in line with" that a 2021 study by Cordi and Rasch, the one How to Learn Anything lesson 8 reported. That's why they point at features of slow-wave sleep, such as spindles, rather than its amount.

The time-of-day reading is the other one, and it's this course's. The night-against-day design is the only one of these five with two clocks and it sits at the top of the table, so some of its lead could be the clock. It's a possibility with no direct evidence behind it, because the paper didn't analyse time of day as a variable, while the authors' reading has the stage-deprivation rows. This course can't settle the gap, and if it had to lean, it would lean their way.

The design that found nothing, and what nothing means

Now the bottom two rows, which make a matched pair.

In both, one group sleeps normally and the other stays up all night after learning. The difference is when the test happens. Without a recovery night, the wake group is tested straight after the sleepless night. With one, everybody gets to sleep again before the test.

And once the recovery night is in, the benefit drops to 0.14 and stops being significant. The tempting sentence is "so an all-nighter does not cost you anything you learned". Note-Taking lesson 5 should make you suspicious of it.

Look at the interval: -0.13 to 0.42. That range contains no effect at all. It also contains 0.42, which is nearly the overall figure for the whole literature. Twenty-three samples weren't enough to tell those apart. This is a failure to find, not a finding of nothing, and it is the same distinction Note-Taking lesson 5 drew for the two meta-analyses whose intervals crossed zero.

So the fair reading of the pair is narrower than either headline. Staying up all night after learning costs you something at the next morning's test, about 0.30 on seventeen samples, and this course's reading is that some of that may be the tiredness rather than the memory; the two rows' intervals overlap too much to show it. Once you've slept again, the studies so far can't tell whether anything was lost. Neither of those is a reason to try it.

What the authors conclude, and whose conclusion it is

The paper's conclusion is confident: "In sum, although further research is needed, our meta-analysis provides convincing evidence that sleep compared to wakefulness following learning improves both storage in and retrieval from episodic memory significantly."

What they mean by storage and retrieval, from the same paragraph: "sleep benefits storage in memory due to memory consolidation and, in addition, retrieval from memory by reducing interference." In other words, sleep may both strengthen what was learned and protect it from being crowded out by whatever you'd otherwise have done while awake. That isn't the storage-and-retrieval-strength distinction How to Learn Anything lesson 1 taught, which is about how well something is learned against how easy it is to recall right now. And about this part the same paragraph is more careful: it calls it "encouraging initial support".

Two parts of that sentence carry different weight. That sleep after learning helps later recall is what the 271 samples show, across designs, at a moderate size, and surviving the authors' own bias correction. That it does so by improving both storage and retrieval is the authors' theoretical account, argued in the half of the paper this course didn't read. Report it as theirs.

And the sentence they give the design result is one worth keeping whole: "To conclude, our findings urge caution in generalizing results across different sleep study designs."

What this adds to the two earlier courses

Memory taught you that remembering is reconstruction, that retrieving one thing can suppress its neighbours, and that a finding depends on its sample. This lesson does not re-teach any of it. What it adds is one variable those lessons left alone: what happens between learning and testing, and specifically whether you slept.

How to Learn Anything taught retrieval practice and spacing, and in its last lesson told you that sleep matters at a smaller size than popular accounts claim. This lesson puts a pooled figure and a design breakdown under that sentence. The figure is moderate, it is about the same size as the testing effect, and it is not a reason to change your study method. It is a reason not to trade the night before an exam for another pass through your notes, which is the one practical sentence this lesson thinks the evidence supports, and it is a judgement rather than a finding: no design in the paper compared a night's sleep with a night's extra study.

Everything here is about laboratory memory for lists, pairs, pictures and passages, tested in most studies about half a day later and never more than a month later. None of it is about exam results, skills, or anything measured over months.

Four things people get wrong

"Sleep consolidates memory, and that is settled." The benefit is well supported across nearly a century of studies. How it works is argued. The authors propose storage and retrieval both, and features of slow-wave sleep as important, while reporting that the amount of it didn't predict the benefit in their own data. Settled is the wrong word for the mechanism.

"An all-nighter wipes out what you learned." Tested straight after, the sleepless group does worse, and on this course's reading some of that may be tiredness. Tested after recovery sleep, the design found no significant difference, on too few samples to say there is none.

"A nap is as good as a night." The nighttime nap matched the night-against-day design, on ten samples with a very wide interval. The daytime nap came in lower, at 0.34. Neither comparison is a nap against a full night directly.

"A g of 0.44 is too small to matter." It's about the size of the testing effect, one of the two techniques How to Learn Anything rated highest. Small and unimportant aren't the same word.

Practice

Connections

Back. Lesson 1 said every number carries the limits of its instrument. This lesson says the same about designs, and the time-of-day problem is lesson 4's clock question turning up inside a memory study. Lesson 3's publication bias is corrected for here in the open. Note-Taking lesson 5 taught Hedges' g and interval charts; Memory lesson 3 printed the same retrieval-induced forgetting figure this paper uses as a yardstick.

Forward. Lesson 8 asks you to read a claim about your own sleep. The design question from this lesson goes with it: before you ask how big an effect is, ask what was compared with what.

Go deeper

• The sleep benefit in episodic memory, Berres and Erdfelder's accepted manuscript, free on PsyArXiv. This course read it in part: the abstract, the design definitions and results, the selective-reporting results, and the discussion of size and design. The design results are one paragraph of the Results and one of the Discussion. The version of record is in Psychological Bulletin at doi:10.1037/bul0000350.
• How to Learn Anything lesson 3 (the testing effect) and Memory lesson 3 (retrieval-induced forgetting) for the two yardsticks, and How to Learn Anything lesson 8 for the slow-wave study.

Sources

1. Sabrina Berres and Edgar Erdfelder, "The sleep benefit in episodic memory: An integrative review and a meta-analysis", Psychological Bulletin 147(12), 2021, pp. 1309 to 1353, doi:10.1037/bul0000350. Read in part, from the authors' accepted manuscript on PsyArXiv (osf.io/r2an6), which carries the publisher's notice that it may not exactly match the version of record. Read: the abstract, the opening definition and the Introduction's paragraphs on materials and designs, the design definitions, the results on design and selective reporting, the discussion of size, of design and of sleep parameters, the conclusion, and the design and retention-interval rows of Tables 1 and 2. Supports: every figure and quotation in this lesson. The testing-effect and retrieval-induced forgetting figures are the authors' citations of Rowland (2014) and Murayama and colleagues (2014); this course read neither for this lesson, though Memory lesson 3 read the second.
2. The course's own constructions, each labelled where it appears in the body. The conversion of g into "ahead of about 67 percent" is this course's arithmetic on the normal curve. The reading of the nap and deprivation definitions as keeping both groups on the same clock is this course's. The reading of the recovery night as separating the memory from the tiredness of the person retrieving it is this course's. The observation that 0.28 is nearer two thirds of 0.44 than half is this course's arithmetic, and "somewhere around 0.28 to 0.44" is this course's summary. The plain-language gloss of storage and retrieval is this course's. The time-of-day reading of the gap between designs, and the judgement that it has less behind it than the authors' reading, are this course's. The practical sentence about the night before an exam is a judgement, and the body calls it one.