Investigating couples’ sleep: an evaluation of actigraphic analysis techniques


Robert Meadows, Centre for Research on Ageing and Gender, Department of Sociology, School of Human Sciences, University of Surrey, Guildford, Surrey, UK. Tel.: 01483 689292; fax: 01483 689090; e-mail:


‘Blip’ analysis, fast wavelet transformations (FWT) and correlation analysis have all been used to actigraphically assess the impact one person is having on another's sleep, yet no review exists as to the differences between, and applicability of, these methods for investigating couples’ sleep. Using actigraphy data and audio sleep diaries collected from 18 couples, this paper provides such a review. This paper constructs and assesses two novel, analytical methods: Lotjonen's sleep/wake algorithm, and the partner impact on sleep wake analysis (PISWA). Both ‘blip’ analysis and correlation suggest that the strongest relationship between bed partners occurs on an epoch-to-epoch basis. However, ‘blips’ deal strictly with onset of movement and fail to incorporate strength and duration of movement. Conversely, correlation analysis incorporates some elements of strength and duration of movement but makes identification of onset problematic. FWT offer useful ‘relativistic’ pattern recognition, identifying onset, strength and duration of movement, but are difficult to quantify. Although audio diary data support the potential of Lotjonen's sleep/wake algorithm to identify sleep non-movement, sleep movement, wake non-movement (or quiet wakefulness) and wake movement, the problem remains that this method also relies on visualization. Of most promise, we argue, is the PISWA, which examines ‘impact’ of bed partners through incorporating elements of ‘blip’ analysis and the sleep/wake algorithm.


Despite the measurement of activity being used in human sleep research since 1959 (see Stanley, 2003), actigraphy has rarely been used to investigate couples’ sleep. Exceptions to this include two related studies by Pankhurst and Horne (1994) who examined actigraphically recorded concordance in movement in 46 pairs of bed partners and the differences in nocturnal movement in those with and without bed partners. The authors concluded that approximately one-third of measured movements were common to both bed partners, although couples did not seem aware of this concordance and reported sleeping better when their partner was there.

Other relevant studies include Wulff and Siegmund's (2000) correlational analysis and Wulff et al.'s (2003) fast wavelength transformations (FWT). Although more concerned with parents and children than with bed fellows, Wulff and Siegmund (2000) used cross-correlation to analyse the synchronization of 12 mother, father and child dyads. Synchronization between mother and infant was shown to be greater than between father and infant. Wulff et al. (2003) used SOPRAN, a software package designed to simultaneously analyse two or more actigraphy records, to show the similarities between the ultradian rhythms of mother and infant.

Thus, the few studies which have attempted an actigraphic investigation into bed partners or the interrelationship between parents’ and child's sleep have utilized ‘bespoke’, disparate analytical methods. This paper provides a review of their applicability. Based on actigraphy and audio sleep diary data collected from 18 couples (36 individuals), this paper compares Pankhurst and Horne's (1994)‘blip’ analysis, Wulff and Siegmund's (2000) correlational analysis and Wulff et al.'s (2003) fast wavelet transformations, examining their usefulness when investigating the impact each partner may have on the other's sleep.

The paper then constructs and evaluates two further methods. First, Lotjonen et al. (2001) developed a sleep/wake algorithm specifically applicable to the Cambridge Neurotechnology Ltd (CNT, Cambridge, UK) activity monitor, following the work of Cole et al. (1992); Sadeh et al. (1994, 1995a), Webster et al. (1982), Blood et al. (1997), Jean-Louis et al. (1996) and Stampi and Broughton (1989). Using logistic regression techniques and polysomnography as the dependent variable, Lotjonen et al. (2001) achieved an 81% agreement rate between polysomnography identified sleep/wake and the output of their actigraphy algorithm. Although Lotjonen et al. did not develop this to analyse simultaneous records, this paper assesses its potential to do so. Secondly, we developed a novel, analytical method which attempts a sophisticated form of ‘blip’ analysis – the partner impact on sleep wake analysis (PISWA). Using Lotjonen's sleep/wake algorithm to identify sleep non-movement, sleep movement, wake non-movement and wake movement, ‘impact’ analysis can be carried out in a way which identifies the state a partner is in during the epoch preceding an individual's onset of awakening.


The paper compares and evaluates the above five methods for their ability to empirically measure the impact of bed partners on each other's sleep. In constructing criteria for evaluation several factors are taken into account. First, the problems associated with actigraphic methods are particularly salient when attempting to directly compare and contrast two individuals’ records. For example, relative levels of activity, rather than absolute values, are often seen as the most appropriate means of drawing inference about a person's movement, because ‘no two accelerometers of any type will give exactly the same reading for the same motion’ (CNT; The Actiwatch Activity Monitoring System: Instructions for Use and Software Manual). Debates exist as to the fundamental status actigraphy data can be afforded, especially as ‘no evidence has yet been presented that actigraphy can identify the times or durations of individual nocturnal arousals’ (Pollak et al., 2001: 965).

As a result, the success of an actigraphic analysis technique to understand the impact bed partners may have on one another is judged by whether or not it deals with absolute values or patterns, whether or not it can identify states associated with sleep and wake, and whether or not it can identify onset and duration of movement. In addition to these criteria, substantive results from each test are compared. Audio sleep diary data are also used to assess the validity of each analytical method.

Sample and procedure

Data reported here comes from a larger project investigating the ways in which couples negotiate their sleep. Twenty-two couples (male partner aged between 20 and 59 years) were interviewed together within their own home. Each partner was then asked to wear an actiwatch and complete an audio sleep diary for 1 week. All participants continued their normal sleep/wake routine throughout the week and no restrictions were placed on activities, food and drink, etc. Four couples were excluded from the present analysis for non-compliance or shift work, therefore this paper reports findings from the actigraphy and audio sleep diaries collected from 18 couples.

Activity monitoring and data management

Actiwatches (Cambridge Neurotechnology Ltd) were set to collect data at 1-min epochs. This was principally because the different methods evaluated here utilized 1-min epochs. All watches were calibrated prior to use using propriety equipment from CNT. Data management was performed on the ‘.awd’ files, and ‘off the shelf’ software was avoided as far as possible. Executable programmes, created using C++, were run on the data to check for periods of more than an hour with no movement. From this, any periods where suspicion was raised that the watch had been removed were excluded from the analysis.

Visual basic programmes were written to extract the ‘cleaned’ data from the raw files and place them into an Excel spreadsheet. The same programmes calculated the ‘blips’ and ‘sleep/wake’ dichotomies during run time. All analyses were performed using SAS version 8.01 (Cary, NC, USA), except the FWT which were performed in MATLAB (Natick, MA, USA) and the correlation coefficients were transformed and retransformed using SPSS version 10.1 (Chicago, IL, USA). The analysis window was restricted to 23:59 to 05:59 hours. This was principally because the aim was to investigate the applicability of the five methods in investigating the impact each partner has on each other's sleep and all couples were more likely to be in bed during this window. All developed programmes can be freely obtained from the first author.

Audio diaries

Audio diaries (Sony Dictaphones) were given to the subjects with the instruction that they should record, upon awakening, any information they wished about their sleep or wake during the previous night. Written instructions to this effect were provided, also asking them specifically to record bed times, any disturbances throughout the night, and number and times of awakening. Subjects were instructed that they could make multiple entries and return to the audio diary on as many occasions as they wished throughout the day. The completed audio sleep diary recordings were transcribed in full.

Lockley et al. (1999) have shown that subjective written diaries and actigraphy are poorly correlated in relation to number and duration of awakenings, and Reyner and Horne (1995) argue that there may be gender and age-related differences in the subjective assessment of sleep. Audio sleep diaries have an advantage over written diaries, as they enable subjects to offer as much information as they wish in an unstructured and user-friendly manner. Diaries within this study were further differentiated from earlier research by the fact that two people subjectively recorded the same night's sleep. Hislop and Arber (2003a,b, 2003c) successfully used audio diaries in their sociological research on mid-life and older women's sleep. The dictated diaries provided a rich source of qualitative data; often containing information which the women would not have recorded in a written diary.

Statistical analysis

As interest in activity recording has developed over the last 40 or so years the number of different devices available, such as ‘Swiss-Type’ (Pankhurst and Horne, 1994) and CNT (Wulff et al., 2001), has also increased (Gorney and Spiro, 2001). However most of the work on the identification of sleep/wake by actigraphy, and thus the analysis algorithms developed have been based around the AMA-32 (Ambulatory Monitoring Inc., Ardsley, NY, USA) (Blood et al., 1997; Sadeh et al., 1995a,b) and ‘microelectronic actometers’ (Siegmund et al., 1994). This has created three distinct activity-based methodologies: those based on time above threshold, those utilizing zero-crossing and, more recently, those relying on digital integration (Ancoli-Israel, 2000; Stanley, 1997). Within these three distinct methodologies, the signal threshold is often adapted, creating numerous sublayers of different approaches. As a result, the statistical methods utilized within these studies are not unquestionably transferable to future research or instantly comparable (Benson et al., 2004; Gorney and Spiro, 2001). The five analytical methods evaluated in this paper will be outlined.


Following Pankhurst and Horne (1994), an ‘actiblip’ analysis was carried out. This involved several components, all of which Pankhurst and Horne completed using their own software ACCORD. All epochs containing any movement following at least one epoch of no movement were tagged with an ‘actiblip’. Consecutive epochs containing movement activity were ignored with the exception of the first epoch, which was ‘actiblipped’. Concordance can then be examined between the bed partners in the following ways: Epoch (partner1) to epoch-1 (partner2); epoch (p1) to epoch (p2); epoch (p1) to epoch+1 (p2); epoch (p1) to epoch-1 (p2) + epoch (p1) to epoch (p2) + epoch (p1) to epoch+1 (p2). Statistical tests were performed using Wilcoxon paired samples test.

The present study contains two deviations from Pankhurst and Horne's (1994) original analysis. First, Pankhurst and Horne used the ‘Swiss-type’ actimeters which only react to accelerations >0.1 g and thus, they argue, are unresponsive to normal passive movements of the bed. The band pass filter with the ‘Swiss-type’ is set at 0.25–3 Hz. Within the present study the CNT actiwatches reacted to accelerations >0.05 g, with a band pass filter of 3–11 Hz. An activity count of 25 roughly equated to 1 g. Although this makes the devices in Pankhurst and Horne's study and the devices in the present study somewhat incomparable, the basic premise of identifying onset of movement and calculating concordance, can be replicated. Secondly, results here are reported as the percentage of male and female ‘hits’ which are matched. This is a slight deviation from Pankhurst and Horne, who detail proportions, using the root mean square, to eradicate any huge discrepancies between males and females.

Correlational analysis

Following Wulff et al. (2001), who also used CNT actiwatches, correlation coefficients for pairs of records were calculated for each day (within the analysis window of midnight to 6:00 hours). Due to the fact that correlation coefficients are not additive, when averaging the correlation coefficients from cross-correlation, all correlation coefficients were z-transformed. These transformed coefficients were then averaged and transformed back.

Fast wavelet transformation

Although wavelets have been used in EEG analysis (see for example Jobert et al., 1994), they have not been widely utilized in actigraphic based studies. One notable exception can be found in a novel software package developed by Wulff et al. (2002, 2003) to primarily enable the assignment of the oscillation of a rhythm to the time and day (SOPRAN).

As Wulff et al. (2003) identify, unlike the Fourier methods on which it is based, wavelet analysis can cope with time series data that contain sharp spikes and discontinuities. The concept of fixed size windows is rejected with wavelet functions (Jobert et al., 1994), which cut up time series data into its different frequency components and then analyse them with a resolution matched to its scale (see also Meyer, 1993).

Wulff et al.'s (2003) SOPRAN, utilizes FWT algorithms and the MATLAB environment and overlays transformed data, recorded from mother and infant records, enabling the time frequency distribution to be visualized. In an attempt to negate problems of artefacts (Tryon, 2004) and the fact that no two accelerometers of any type will give exactly the same reading for the same motion, the present analysis uses MATLAB and WAVELAB's De-Noising of NMR Signal algorithm (see Buckheit et al., 1995). De-Noising wavelet analysis is based on the idea of thresholding: that is, setting to zero all coefficients that are less than a particular threshold. These coefficients are used in an inverse wavelet transformation to reconstruct the data set and may enable the recovery of a ‘true signal’. The resulting graph is then visually examined for pattern matching.

Lotjonen et al.'s sleep/wake algorithms

Cole and Kripke (1989) report that in using their algorithm and Webster et al.'s (1982) methods for rescoring actigraphically identified sleep, sleep and wake can be identified. Similarly, Sadeh et al. (1994; 205) have developed a sleep/wake scoring algorithm which they believe is ‘robust and relatively unresponsive to systematic variations in input activity levels’, although they do offer the proviso that this may be different within ‘naturalistic’ studies. However, despite the promise these algorithms offer in the field of investigating couples’ sleep, they all come from actigraphs that utilize the zero crossing method of recording and, as a result, are not automatically applicable to the CNT actiwatch.

Lotjonen et al. (2001), however, have shown a 77% agreement rate between CNT actigraphy and polysomnography using Sadeh et al.'s discrimination analysis, without post-processing (Jean-Louis et al., 1996). In a later study, Lotjonen et al. (2003) specifically adapted Sadeh et al.'s (1994) algorithm to make it applicable for the CNT actiwatch. Using logistic regression techniques and polysomnographically defined sleep/wake periods as the dependent variable, the following formula gave an agreement rate between polysomnography and actigraphy of 81% (when applied to all subjects):


where s is the activity of the scored epoch; mean, mean activity in the window of seven epochs around the scored epoch; SD, standard deviation of the activity in a window of eight epochs (taken here to mean that many epochs either side of the scored epoch); nat, number of activity counts above a given threshold in the window of 11 epochs (threshold is 10 for actigraphy); log 2, is the natural logarithm of the activity in the scored epoch + 0.1. Although this cannot identify partial arousals (such as that which would be associated with an abrupt change from a deep stage of NREM sleep to a lighter stage) or microsleeps, we argue that by using this algorithm to calculate whether an epoch should be scored as sleep or wake, plots can be created of the actigraphic raw scores (colour coding them accordingly); enabling the identification of sleep movement, sleep non-movement, wake movement and wake non-movement. Further to this, the identification of these four states enables a novel form of epoch analysis to be performed.

Partner impact on sleep wake analysis

The PISWA takes Lotjonen et al's algorithm and uses it to construct a novel form of ‘blip’ analysis. PISWA identifies the onset of wake in an individual and then examines whether their partner was asleep (with movement or without) or awake in the epoch directly preceding this.



As Fig. 1a illustrates, men and women experienced comparable mean numbers of discrete movements each night [36.8 (SD = 13.5) and 33.9 (SD = 11.5)] respectively (P = 0.24).

Figure 1.

Gender and movement of bed partners.

With respect to ‘causation’Fig. 1b suggests that men are more likely to be disturbed by their bed partner than women [10% (SD = 0.08) and 8% (SD = 0.07) of movements ‘caused’ by partner respectively]. However, the mean incidence of an occurrence of a ‘blip’ for any epoch was 0.05 (that is the total number of ‘blips’ for all subjects divided by the total number of epochs examined). This figure was similar when the total number of ‘blips’ for females was divided by the total number of female epochs (0.48) and the total number of ‘blips’ for males was divided by the total number of male epochs (0.50). This suggests that approximately 50% of epoch to epoch-1 hits occurred by chance. Indeed, echoing Pankhurst and Horne's (1994) original findings, a large proportion of all of the movements common to both partners were confined to the same epoch [0.34 for females (SD = 0.137) and 0.32 for males (SD = 0.137; P = 0.39)].

Although the reliability of this method is confirmed through the findings here reflecting those of Pankhurst and Horne (1994), blip analysis fails to meet several of the criteria identified above. First, it deals with absolute values. With CNT watches activity counts greater than 0 can be caused by artefacts, such as sleeping in a waterbed (Tryon, 2004), suggesting that performing ‘blip’ analysis with CNT watches requires a threshold greater than 0. Secondly, this method of analysis fails to take into account duration and strength of movement and, with CNT watches, how these movements relate to awakenings and, actual, felt disturbances.

Correlational analysis

The correlation analysis is strongly supported by the audio diary data. As can be seen from Table 1, on average the strongest correlations between bed partners occur when a direct epoch to epoch comparison is made; with couple 12 having the highest coefficient, followed by couple 4, couple 15 and couple 6. According to the audio diaries couple 12's coefficient was high due to moving a partner because they were ‘sort of breathing onto my face’ and, on other nights, ‘having cover battles’. Couple 4 were found to be unique within the data set in that they had purposefully changed their bed times during the study week to match each other's, and couple 15 said that they would occasionally swap sides of the bed in the middle of the night. Similarly disruptive, the male partner in couple 6 said he often talked to his teenage children when they got in late at night and routinely had conversations with his wife in the middle of the night.

Table 1.  Correlation coefficients for each couple, calculated on a nightly basis, and average (Z-transformed) coefficients by couple
NightCouple 1Couple 2Couple 4Couple 6Couple 7Couple 8Couple 9Couple 11Couple 12
 Couple 14Couple 15Couple 16Couple 17Couple 18Couple 19Couple 20Couple 21Couple 22
  1. F, Female epoch correlated with preceding epoch in partner; M, male epoch correlated with preceding epoch in partner; e-e, epoch-to-epoch correlation; +, slept apart; #, one person removed watch.


As far as illustrating the impact each partner is having on the other, on 25 occasions the epoch to epoch-1 correlation give a coefficient higher than 0.3, 15 of which were the male ‘causing’ disruption and 10 of which were female caused disturbances (see Table 1). These were again supported by the audio diary entries. For the female in couple 4, night 6 was particularly restless because her partner ‘got up to go to the loo a couple of times’ and ‘had a bad cold so he was a bit restless towards morning’ (couple 4 Audio Diary Female (ADF)). In couple 11, the female suggested why she was disturbed on three nights when she proclaimed on the morning after night seven: ‘Reasonably quiet night. [Partner] didn't snore, hooray! So managed to catch up on a few extra hours I normally miss out on, because of that’ (couple 11, ADF).

As well as being supported by the subjective reports, correlation analysis meets the criteria of incorporating (if not actually identifying) strength and duration of movement. However, correlation is unable to readily offer information as to states of sleep and wake and the onset of a movement.

Fast wavelet transformation

Blip analysis and correlational analysis although concern themselves with actigraphically measured movement, the utility of wavelets is that they enable pattern recognition. Subjective reports confirmed the utility of this method. Within the audio diaries five couples gave descriptions of disturbances during the night in which both partners detailed specific times, which then were corroborated by their partners. For example, both partners within couple 8 reported an awakening on day 7 at 4:00 hours. Figure 2, illustrates how at approximately 4:00 hours, both partners experience a change in their activity pattern.

Figure 2.

Wavelet shrinkage and denoising – couple 8, night 7.

Regarding the investigation of the actual impact each partner is having on the other, SOPRAN recommends visualization. Utilizing this, Fig. 3 illustrates how overlaying the two records for a couple enables simultaneous pattern changes in the records to be observed (see for example approximately 4:50 hours) in couple 8.

Figure 3.

Wavelet shrinkage and denoising – example of visualization. *Male is green, Female is black.

As a further example, in an event corroborated by her husband, the female in couple 9 reported being restless on night 7 because of being woken by her dog at ‘about quarter to four’ and then walking the dog (couple 9, ADF). From Fig. 3b it can be seen that the female's actigraphy undergoes a change in pattern at approximately 3:15 hours. Although the male's actigraphy did not change as dramatically, there is a small increase within that period, suggesting that he was disturbed. Here strength and duration of movement are incorporated but reliance on visualization makes quantifying the impact of different sleep and wake states and generalization difficult.

Lotjonen et al.'s sleep/wake algorithm

Lotjonen et al. (2003) have specifically adapted Sadeh et al.'s (1994) algorithm for the purposes of making it applicable for the CNT actiwatch. Calculating whether or not an epoch should be scored as sleep or wake and plotting the raw scores (colour coding them accordingly) is said to offer the possibility of identifying sleep movement, sleep non-movement, wake-movement and wake non-movement (see Fig. 4: blue is sleep, red is wake).

Figure 4.

Example of the different states indentified by the sleep/wake algorithm.

Focusing on two case studies it can be seen that, in examining the accuracy of the sleep/wake identification, the female in couple 9 reported awakening at about 3:45 hours, and the subsequent walking of the dog (see Fig. 3b above) do show as wake movements (see Fig. 4a). However, she offers no information as to the other suggested wake periods within her record, other than her reported feelings of restlessness throughout the night (couple 9, ADF). Similarly, the female in couple 8 reports frequently ‘waking in the night’ and her reported bed time of ‘half one’ is roughly corroborated by the algorithm (couple 8, ADF).

For couple 6, night 2 seemed unusually active for both partners. The male recorded in his diary that he was disturbed at 2:30 hours for about 5 min and that it then took him about an hour to go back to sleep. He woke again at about 5:30 hours and then dozed until about 6:00 hours (couple 6, ADF). The resulting sleep/wake graph, shows matches for awakenings between 12:00 and 1:00 hours, and again at 2:30 hours. Taking about an hour to go back to sleep, and awakening at 5:30 hours are also matched (see Fig. 5a). Similarly, the female partner's audio diary recorded awakenings at 2:30 hours, going to the toilet at about 3:15 hours and getting up for a drink at about quarter to six. Figure 5b illustrates that these periods of wakefulness can also be seen in the sleep/wake algorithm. According to the algorithm analysis, she also had an ‘unmentioned’ awakening at about 1:30 hours, which was the same time her male partner awoke. These sleep/wake algorithm results also suggest that between the 2:30 hours awakening and going to the toilet at about 3:15 hours, the female was lying in bed but awake.

Figure 5.

Visual comparison of partner records using the sleep/wake algorithm – couple 6, night 2.

The algorithm's potential to capture wake non-movement can be seen in Fig. 6. The female in couple 21 reported on night 1 that she ‘went to bed about 20 past 11, little girl was up at 4 o'clock in the morning and … it did take me a little while to get back to sleep actually, but I was just lying there and felt quite rested’ (couple 21, ADF: see Fig. 6). It should be noted however, that this wake period does include some scores above zero.

Figure 6.

Sleep/wake algorithm – example of wake non-movement.

Partner impact on sleep/wake analysis

We demonstrate that sleep/wake algorithms allow for quantitative analysis of the actual impact which each partner is having on the other (for example through simple rolling correlations between wake/sleep in both partners). Table 2 gives the results from our novel form of ‘impact’ analysis, developed by the authors specifically for the actigraphic investigation of couples’ sleep. This has the potential to evaluate the impact bed partners are having on each other, whilst taking into account the ‘felt impact’ of the disturbance and the type and duration of causative movement. Following the application of Lotjonen et al's sleep/wake algorithm, all ‘wake epochs’ following at least one epoch of ‘non-wake’ were tagged with an ‘actiblip’. For each person, epoch-1 comparisons were carried out in order to identify what ‘state’ the partner was in preceding this epoch.

Table 2.  Partner impact on sleep wake analysis (PISWA). Proportion of onset of wake epochs preceded by partner wake, sleep movement and sleep non-movement
WakeSleep movementSleep non-movementNumber of awakenings during the nightWakeSleep movementSleep non-movementNumber of awakenings during the night
70.500.000.50120.730.000.2715 8 70.330.000.6612
30.500.000.50 8
40.000.330.66 60.500.000.50 4 80.500.000.50 2
15.10.330.000.66 60.730.000.2711 70.500.120.3818
50.630.000.37 80.710.000.2914 90.670.000.3312 90.670.000.3312
16.10.720.000.28180.560.000.44 9
40.940.000.06170.290.140.57 7
60.400.000.60 50.500.000.50 8 80.380.130.50 8
18.10.330.000.67 5 90.400.000.6010
30.500.000.50 90.390.000.6118
50.420.000.58120.620.000.3813 70.460.000.5413
70.550.000.45110.220.000.78 9 7
30.500.000.50 20.500.000.50 2
60.500.000.50 20.300.000.7010
70.000.750.25 8
20.10.500.000.50 4
20.400.200.40 50.400.000.60 5
30.330.000.67 30.330.000.67 3
40.500.000.50 60.300.000.7010
60.670.000.33 30.830.000.17 6
60.400.000.60 50.420.170.4212
70.690.000.31130.450.090.4511 3 8 9
40.210.720.07140.200.000.80 5 6
70.770.000.23 90.500.000.5010
Average (variance)0.36 (0.86)0.05 (0.01)0.58 (0.08) 9 (22.5)0.30 (0.63)0.05 (0.01)0.64 (0.07) 9 (23.5)

As shown in Table 2, for the female, on average 36% of their discrete wake epochs followed an epoch of wake in their partner. On average 30% of men's discrete wake movements followed an epoch of wake in their partner. Although not significant (P = 0.12), this difference between males and females is somewhat inconsistent with the results from the earlier ‘blip’ analysis based on Pankhurst and Horne's (1994) methodology which reported that a higher proportion of men's movements were ‘caused’ by their partner.

Discussion and Conclusions

Using the novel pairing of actigraphy data and audio sleep diaries, this paper has reviewed the applicability of ‘blips’, correlational analysis and FWT for the actigraphic investigation of couples’ sleep. In addition, two novel analytical methods were constructed and assessed. First, Lotjonen et al's sleep/wake algorithm was used to identify sleep non-movement, sleep movement, wake non-movement and wake movement. Secondly, PISWA was then performed to identify what state a partner was in during the epoch preceding an individual's onset of a wake period.

Results illustrate how ‘blip’ analysis focuses solely on onset of movement. Correlation analysis incorporates elements of strength and duration but makes the identification of onset problematic. Both also rely on absolute values. This is somewhat problematic when it is acknowledged that no two accelerometers can be guaranteed to give the same value. Overcoming this last point, pattern recognition via FWT offers a ‘relativistic’ way of examining actigraphic records; illustrating strength, duration and onset of movement. Although the method's utility for identifying these patterns is supported by the audio sleep diaries, FWT is difficult to quantify and generalize from. The suggested FWT analysis method of overlaying two actiwatch records and using visualization does not overcome this and statistical methods, such as correlating the wavelet coefficients, fall foul of some of the arguments discussed above. Wavelet-based correlational analysis would, like the correlational analysis discussed above, incorporate elements of strength and duration but cannot identify the onset of movement or the ‘felt’ impact of bed partners.

In relation to the two methods for investigating bed fellows which are constructed and assessed here, Lotjonen's sleep/wake algorithm is based on methods which Sadeh et al. (1994) claim are unresponsive to systematic variations in input activity levels. Data from the audio sleep diaries support the accuracy of Lotjonen et al.'s (2003) algorithm and illustrate how the four states of sleep non-movement, sleep movement, wake non-movement and wake movement can be identified. As Sadeh and Acebo (2002) suggest, wake non-movement or ‘quiet wakefulness’ is often miscoded as sleep; decreasing the validity of actigraphy as a measurement tool. Practice points in overcoming this and other difficulties are said to include only using algorithms which are appropriate for the specific task and documentation of sleep/wake and artefact-related behaviour (Sadeh and Acebo, 2002). Audio diary reports of sleep/wake behaviours suggests Lotjonen et al.'s (2003) sleep/wake algorithm has the potential to identify quiet wakefulness.

This algorithm also provides the building blocks for an ‘impact’ analysis of sleep/wake which can quantitatively investigate the ‘effect’ within a couple that each partner may be having on the other's sleep. The PISWA approach incorporates onset duration and strength of movement in a way that can be quantified. Support for this measure also comes from the correlation results. Further studies will be required to test the full potential of the PISWA analytical method for investigating the impact of bed fellows on each other, itself an underinvestigated field (Strawbridge et al., 2004).

The underinvestigated nature of this field means that situating present methodological findings in relation to earlier ones is problematic. What the present study does suggest, however, is that a method can be devised which can fulfil the requirements of identifying onset, strength and duration of movement, identifying sleep and wake and the actual ‘felt impact’ of disturbances and which relies on relative pattern recognition and not actual figures. This needs to be taken into account when investigating the impact bed partners may have on each other (Pankhurst and Horne, 1994) or the synchronization of mother, father and child dyads (Wulff and Siegmund, 2000; Wulff et al., 2003).


The authors would like to thank Jean Dawson for downloading the watches and offering her vast experience of actigraphy and Dr Sigurd Johnsen for numerous discussions and references on statistical analysis techniques. This research is part of a larger study ‘Negotiating Sleep: Gender, Age and Social Relationships amongst Couples’ (Economic and Social Research Council, grant number RES-000-23-0268).