Introduction

Sleep impairment is being increasingly recognised as a serious problem in modern society. Commercial activity has extended beyond traditional trading hours, shift-work has become more common, and the separation of on- and off-duty hours has increasingly blurred. Furthermore, noise and light pollution also affect the sleep quality of residents living in densely populated urban areas, where more than half of the world’s population resides. These factors contribute to sleep impairment, and studies have reported high prevalence of sleep impairment from 10.0 to 35.5 % in different communities [1–4]. In Singapore, Yeo et al. has reported that the prevalence of insomnia was 15.3 % in 1996 [5]. Sleep impairment is associated with an increased risk of hypertension, obesity, depression and cardiovascular diseases [6, 7] and affects daily activities such as driving [8] and work [9].

The Pittsburgh Sleep Quality Index (PSQI) is a 19-item self-reported questionnaire that measures the quality and patterns of sleep during the past 1 month [10]. Although the PSQI was initially developed as a clinical tool to evaluate sleep quality in psychiatric practice, it is now also widely used in general population-based epidemiologic studies. There has been much debate that the PSQI contains multidimensionality instead of being a unidimensional scale of measure. Multiple studies have shown that a scoring method based on two [11, 12] or three [13–15] factors may provide a better assessment of sleep quality. For example, Cole et al. examined the factor structure of the PSQI using a cross-validation approach and identified the presence of a three-factor structure based on a confirmatory factor analysis [14]. Another study with Nigerian students [15] also confirmed Cole’s findings.

Factor structure is influenced both by the intrinsic quality of the instrument, as well as the population on which the instrument is applied [16, 17]. Most of these factor analysis studies were conducted outside Asia namely in the USA and Europe [11–15], and their results might not apply to Asian populations such as that of Singapore. The aim of this study is to establish the factor structure of PSQI, which is administered in serial national representative cross-sectional surveys for the surveillance of the health state of the Singapore population. Two different population-based surveys were used to obtain an optimal factor model of PSQI in the Singapore population, which is a multi-ethnic Asian society representing East (Chinese), South-East (Malay) and South Asian (Indian) ethnicities.

Methods

Study design

The first study is a community-based cross-sectional study conducted in Queenstown housing estate, Singapore (Queenstown study) by medical students from National University of Singapore (NUS) as part of their training curriculum. The inclusion criterion was Singapore citizens or permanent residents aged at least 21 years who were willing to provide informed consent. A three-stage sampling method was used to randomly select the participants. In the first stage, 30.0 % of the Housing Development Board (HDB) blocks were randomly selected out of 270 HDB blocks in Queenstown estate. In the second stage, 1140 household units were randomly selected from 90 HDB blocks from the first stage. In the third stage, the interviewer randomly selected one participant from each household when there was more than one eligible participant by using the Kish grid method. Out of the 1140 participants randomly selected, 489 agreed to participate, 412 declined and 239 did not respond. The overall response rate was 42.9 %.

The second study is Singapore Health (SH) 2012, a national representative study of the Singapore adult population, consisting of a main questionnaire administered at participant’s home, a health examination and administration of sub-questionnaires (containing the PSQI) at designated health examination centres. The data from the National Database of Dwellings, maintained by the Department of Statistics, was used to draw up the sampling frame for the entire population of Singapore. Here, a two-stage randomisation was employed. In the first stage, a sampling frame of 14,200 households was randomly chosen from the database and all these households were enumerated. The enumeration response rate was 46.0 % and entailed the collection of demographic data of all the members of each household, where the inclusion criteria was Singaporeans or permanent residents aged 18 to 79 years living in these households. In the second stage, 6000 individuals were then randomly selected based on the list of individuals enumerated, with oversampling of Malays and Indians. The study overall response rate was 40.0 %, but only 1976 of them who had completed the PSQI at the health examination centres were used in the analysis.

The PSQI was administered through a standard questionnaire by trained interviewers in both studies with other information collected on demographics, medical and psychosocial conditions.

Statistical analysis

Following the scoring guidelines provided by Buysse et al. [10], the 19 items of the PSQI are grouped into seven sleep components: subjective sleep quality, sleep latency, sleep duration, habitual sleep efficiency, sleep disturbances, use of sleeping medications and daytime dysfunction. Each of these seven components are equally weighted in a 0–3 scale and aggregated into a single global PSQI score ranging from 0 to 21.

We compared the characteristics of the two study populations. For continuous data, the median and range was reported and for categorical data, proportion was used. Mann-Whitney U test (two-tailed) was performed to test the differences in medians and the chi-squared test was used to test for differences in proportion. Fisher’s exact test was used whenever the cell frequency was below 5. We assumed the two study populations did not overlap. Subsequently, an EFA and CFA were conducted separately on the two studies to evaluate an appropriate factor model for PSQI.

In EFA, two methods of estimation, principal components analysis (PCA) and maximum likelihood estimation (MLE), were used to calculate the factor loadings and to determine the correlation between each component to its corresponding factor, as well as to establish the number of factors to be retained in the model. The factor loadings were then subjected to a varimax rotation to aid in the interpretability and to evaluate against the cutoff criteria of 0.40 proposed by Steven et al. [18]. In addition, a variable factor map and individual map were plotted using the package FactoMineR in R [19] to better understand the contribution of each PSQI component to the latent factors [20]. This was followed by CFA to evaluate the fit of various models to identify the model with the best fit. We also compared our best-fit model against Cole’s three-factor model [14] and the Buysse’s single-factor model [10].

Multiple fit indices were used to evaluate the adequacy of the model fit: we considered the fit to be good if the goodness-of-fit index (GFI) and adjusted goodness-of-fit index (AGFI) were 0.90 or higher [21], comparative fit index (CFI) was 0.95 or higher [22] and the root mean square error of approximation (RMSEA) was 0.05 or lower (0.05 < RMSEA < 0.08 indicates adequate model fit) [23]. Then, the models were compared against one another based on the following: delta chi-squared (∆χ 2) statistics [24], Bayesian information criteria (BIC) [25] and consistent Akaike information criteria (CAIC) [20]. When comparing the fit indices between two models, the model with a lower χ 2 statistics, BIC and CAIC would be preferred. Furthermore, Raftery had shown in his paper that a difference of at least 10 in the BIC of two models would render the model with higher BIC to be rejected [25].

All analyses were performed using the statistical programming language R [26].

Results

Table 1 shows the characteristics of the two study populations. Of the 489 participants in the Queenstown study, 290 (59.3 %) were females and 199 (40.7 %) were males. The median age was 47 years. The majority were Chinese (73.6 %) followed by Malays (9.0 %), Indians (9.8 %) and others (7.6 %). Of the 1976 participants in SH 2012, 1021 (51.7 %) were females and 955 (48.3 %) were males. The median age was 43 years, which was slightly younger than the Queenstown study. In addition, majority of the participants were Chinese (35.1 %) and Malays (35.0 %), followed by Indians (24.4 %) and others (5.5 %). Most of the characteristics were significantly different between the two study populations except for sleeping medication use, proportion of hypertension, diabetes mellitus and cancer; this is probably due to the different study designs used in recruitment of the study populations.

Table 1 Characteristics of the study population in Queenstown study and SH 2012
Full size table

Table 2 provides the Spearman’s correlations and the descriptive statistics for the seven PSQI components in both studies. Each component score ranged from 0 to 3. Among the inter-component correlations in the Queenstown study, sleep duration and habitual sleep efficiency were found to have the highest correlation (r = 0.48) and use of sleep medications and sleep duration had the lowest correlation (r = 0.02). The mean global PSQI score was 5.04 (SD = 3.29, Cronbach’s alpha = 0.67). Using the recommended cutoff of 5 for the global PSQI score, 35.0 % of the participants have poor sleep quality.

Table 2 Pittsburgh Sleep Quality Index (PSQI) inter-component Spearman’s correlations and descriptive statistics for both studies
Full size table

Among the inter-component correlations in SH 2012, similarly sleep duration and habitual sleep efficiency had the highest correlation (r = 0.36). However, unlike in the other study, use of sleep medications and sleep efficiency had the lowest correlation (r = 0.05). The mean global PSQI score was 4.21 (SD = 2.77, Cronbach’s alpha = 0.64). Using the recommended cutoff of 5 for the global PSQI score, 24.7 % of the participants have poor sleep quality, lower than the proportion from the Queenstown study.

Exploratory factor analysis

In EFA of both datasets, both two-factor and three-factor models fitted the data well with interpretability. In the two-factor model, the first factor was labelled sleep efficiency and the second factor was labelled perceived sleep quality. Two of the components, sleep duration and sleep efficiency, were loaded highly on the first factor, while the rest of the components had good loadings on the second factor.

In the three-factor model of the Queenstown study, the same component under each factor was identified in both the MLE and PCA methods. The factors were labelled as sleep efficiency, perceived sleep quality and daily disturbances. Components habitual sleep efficiency and sleep duration had strong loadings on the first factor sleep efficiency. Sleep latency and sleep disturbance had relatively high loading on the factor daily disturbances and the rest of the components were loaded onto the factor perceived sleep quality. A second three-factor model was also identified through the variable map: in this model, subjective sleep quality was grouped with sleep latency under the factor perceived sleep quality; sleep disturbances, sleeping medication use and daytime dysfunction under the factor daily disturbances and finally, habitual sleep efficiency and sleep duration were grouped together under the factor sleep efficiency.

In SH 2012, a similar three-factor structure was observed and each factor was labelled in the same manner as the Queenstown study. The components habitual sleep efficiency and sleep duration were loaded highly on the first factor using the PCA method, while daily disturbances, subjective sleep quality and daytime dysfunction were loaded onto the second factor; finally, the rest of the components were loaded onto the third factor. The MLE and variable mapping method also identified the same components under each of these three factors; in particular, the grouping of the components was identical to the variable mapping method as the Queenstown study.

Confirmatory factor analysis

We also compared our models with that of Cole’s three-factor model and Buysse’s single-factor model. Table 3 shows the summary of the model fits. In the Queenstown study, the three-factor model obtained from the variable map produced the best fit; hence, only this model was evaluated as the best model obtained in the EFA. For SH 2012, the three-factor model obtained from the PCA did not provide a good fit compared to the three-factor model from variable mapping and MLE; hence, only the latter was evaluated as the best model.

Table 3 Fit statistics of the models obtained from exploratory factor analysis (EFA) in study 1 and study 2
Full size table

For the Queenstown study, Buysse’s single-factor model had resulted in the poorest fit (χ 2 = 128.2, P < 0.01; GFI = 0.93; AGFI = 0.86; CFI = 0.78; RMSEA = 0.13; BIC = 214.9; CAIC = 27.5). Although applying Cole’s three-factor model did not result in the most desirable fit indices, it was more appropriate than the single-factor model (χ 2 = 47.6, P < 0.01; GFI = 0.97; AGFI = 0.93; CFI = 0.93; RMSEA = 0.08; BIC = 152.8; CAIC = 31.5). The fit of a three-factor model was clearly better than using a single-factor or a two-factor model. The best three-factor model derived from EFA had a good fit for all the indices except for RMSEA which only had an adequate fit and a significant χ 2 statistics (χ 2 = 36.1, P < 0.01; GFI = 0.98; AGFI = 0.95; CFI = 0.95; RMSEA = 0.07; BIC = 141.3; CAIC = −43.0). The difference in BIC between our best model and Cole’s model had a difference of at least 10 [25], suggesting that our three-factor model would be preferred over Cole’s model.

For SH 2012, Buysse’s model resulted in the poorest fit (χ 2 = 245.2, P < 0.01; GFI = 0.96; AGFI = 0.93; CFI = 0.85; RMSEA = 0.09; BIC = 351.4; CAIC = 124.9). Similar to the Queenstown study, Cole’s three-factor model did not have the best-fit indices, although it outperformed the Buysse’s model (χ 2 = 70.6, P < 0.01; GFI = 0.99; AGFI = 0.97; CFI = 0.96; RMSEA = 0.05; BIC = 199.6; CAIC = −23.8). The fit of a three-factor model was superior to that of a two-factor model. The fit statistics for our best three-factor model by EFA had good fit for all the indices except for a significant χ 2 statistics (χ 2 = 63.8, P < 0.01; GFI = 0.99; AGFI = 0.98; CFI = 0.97; RMSEA = 0.05; BIC = 192.8; CAIC = −30.7). However, the difference in BIC between our best model and Cole’s model had a difference of only about 7 (<10) which failed to provide adequate evidence to conclude that our best three-factor model is preferred over Cole’s model.

In the three-factor model for Queenstown study, each PSQI component score and its corresponding factor had relatively large standardised path coefficients that ranged from 0.37 (sleep disturbance to daily disturbances factor) to 0.72 (subjective sleep quality to perceived sleep quality factor, and habitual sleep efficiency to sleep efficiency factor), while the correlations between the three factors ranged from 0.27 to 0.89. In the three-factor model for SH 2012, each PSQI component score and its corresponding factor had relatively large standardised path coefficients that ranged from 0.26 (sleep medication use to daily disturbances factor) to 0.67 (habitual sleep efficiency to sleep efficiency factor), while the correlations between the three factors ranged from 0.39 to 0.90 (Fig. 1).

Fig. 1
figure 1

Best three-factor model obtained in EFA with standardised path coefficients between factor solution and the PSQI component in the a Queenstown study and b SH 2012

Full size image

Discussion

We have shown that there exists a three-factor structure of the PSQI underlying an individual’s sleep experience in the Singapore population. This same conclusion is reached from the analyses of two separate studies conducted in Singapore with two distinct study populations recruited using different study designs. For the Queenstown study, the study population was mostly Chinese, while for SH2012, the three ethnicities were equally sampled. Despite differences in the distribution of ethnicities, we obtained the same best three-factor structure in both studies, indicating that there is unlikely to be any bias in the evaluation of sleep quality among the three ethnicities. This also supports the theory of PSQI containing multidimensional factor structures instead of a unidimensional one.

The Cronbach’s alpha [27] which is a measure of the internal consistency, usually taken to be at least 0.80 [28], was relatively low in both of our studies (0.67 in the Queenstown study, 0.64 in the SH 2012). Our low Cronbach’s alpha was similar to Mariman et al. [29] of 0.64, and this is not surprising given our findings of a three-factor structure is in sharp contrast with the Cronbach’s alpha (0.83) reported by the team that developed the PSQI [10].

Our finding is consistent with most other studies in the USA and Europe which show that contrary to the intentions of the developers of the instrument, the PSQI is not unidimensional [13–15]. Our three-factor model derived from EFA appears to perform better than Cole’s model. The main difference was that the component sleeping medication use was loaded onto the factor perceived sleep quality in our model instead of daily disturbances in Cole’s model.

A large proportion of participants in both our studies reported that they did not use any sleeping medication during the past month (94.9 % in the Queenstown study, 96.7 % in the SH 2012 study). Insufficient numbers of participants with sleep medication use may render the analyses of this component unreliable. The standardised path coefficients for sleeping medication use to its factor is one of the lowest in both of our studies (Fig. 1), suggesting a weak dependency between the component and its factor. This difference could be due to a lack of power from insufficient sample sizes. Alternatively, it may also be due to underlying inherent differences in the study populations, in particular their perception of the effect and consequences of the use of sleep medications, the diverse nature of sleep problems, and the varying sleeping behaviours in different populations. Possible future research work may involve validating the three-factor model against a gold standard and to confirm if the factor structure differs between Western and Asian populations.

Strengths and limitations

The strengths of this study are the following: we evaluated the factor structure of PSQI in two different population-based studies with different study designs. The statistical analysis performed separately on the two studies had arrived at the same finding of a distinct three-factor structure, suggesting that a three-factor model of the PSQI in a multi-ethnic Asian population is robust. Our study is also the largest that we know of, using data from a total of 2465 participants. We also used data from healthy community-dwelling individuals rather than from disease cohorts. Disease and its sequelae such as pain may affect the experience of sleep, in that results from these cohorts may not apply to the general populations.

We highlight limitations that may affect the interpretation of our results. First, the overall response rates for the Queenstown study and SH 2012 were 42.9 and 40.0 %, respectively. While these were comparable to other population-based studies in Singapore [30], the relatively low response rate may affect the generalisability of our results. Second, our three-factor model showed only an adequate fit criterion with RMSEA, and the χ 2 statistics were statistically significant in the Queenstown study. However, other fit indices such as GFI, AGFI, CFI, BIC and CAIC suggested that our model had a good fit with the data.

Conclusion

In summary, this study evaluates the structure of PSQI in the multi-ethnic Asian city-state of Singapore. We show that PSQI has a distinct three-factor structure (sleep efficiency, perceived sleep quality and daily disturbances) in the Singapore population, using data from two distinct population-based surveys. In population-based studies that use the PSQI to evaluate sleep quality, validating cutoffs and scoring the PSQI along these three factors may provide a more accurate understanding of the relationship between sleep impairment and health conditions rather than using a single global score.