Revisiting the dimensional structure of the Edinburgh Postnatal Depression Scale (EPDS): empirical evidence for a general factor

Post-partum depression (PPD) is a difficult construct to measure in practice [1]. Enabling as many health professionals as possible to make a timely first assessment of maternal mental health while leaving detailed psychiatric assessments for only those situations suggestive of PPD is an appealing approach. Similarly, applied research contexts require swift yet valid instruments. In the late 1980s Cox et al. [2] argued that a suitable instrument to evaluate depressive symptoms after childbirth was needed since available tools to assess depression in general populations put too much emphasis on somatic symptoms that could nevertheless be due to normal physiologic adaptations associated to childbearing. In an attempt to address this drawback, the authors proposed the Edinburgh Postnatal Depression Scale (EPDS), a simple and well accepted 10-item assessment tool that is easy to fill in and does not require specialized psychiatric expertise from health workers. Since its conception, the EPDS has been adapted for use in several countries [3, 4] and has become the most widely used instrument for a first approach to PPD [4, 5].

The EPDS has been extensively scrutinized and a number of studies have evaluated its psychometric properties. Several studies have focused on its dimensional structure, with at least thirteen comprising sample sizes above 150 individuals [6–18]. Although Cox et al. [2] originally proposed the EPDS as a one-dimensional measurement tool and this has been supported by a few authors [12, 13], most of the factorial analyses have shown that the EPDS would be better defined through multi-factorial structures, either by two [6, 10, 16–18] or three factors [7–9, 11, 14, 15].

Regardless of the number of uncovered factors, most studies clearly distinguished items representing 'anxiety' -- items 3 (blaming oneself unnecessarily when things [go] wrong), 4 (having been anxious or worried for no good reason) and 5 (having felt scared or panicky for no very good reason) -- from those representing low positive affect or anhedonia -- 1 ([not being] able to laugh and see the funny side of things) and 2 ([not] looking forward with enjoyment to things) -- and depression -- 9 (feeling unhappy [and] crying) and 10 ([thinking] of harming oneself). In three studies, the items on anhedonia and depression jointly loaded on a single factor forming a two-factorial structure along with an anxiety factor [16–18]. Another study showed several cross-loadings involving items 8 (feeling sad or miserable), 9 and 10 [10]. A fifth study suggesting a two-factor structure was less conclusive since a few items had been removed before the factorial analysis [19].

A distinction between the anxiety items and the others persisted in all studies showing a three-factorial solution, but with items mapping anhedonia and depression now clearly separated. In some studies, item 8 joined items 1 and 2 on anhedonia [8, 9, 11, 14, 15], whereas in others this item joined items 9 and 10 on depression [7, 13]. Items 6 (feeling that things have been getting on top of [the respondent]) and 7 (having been so unhappy that [respondent] had difficulty sleeping) showed to be the most ill-behaved irrespectively of the type of solution. They often cross loaded, in some studies joining the anhedonia/depression items and in others grouping with the anxiety items.

Despite some oddities -- which may have come about from methodological shortcomings such as inadequate sample sizes, unsuitable multivariate models (e.g., principal components analysis) and/or failing to correctly model the ordered-categorical nature of the items --, the reviewed literature shows more congruence than otherwise and, without further detailing, would be suggestive of a multi-scale measurement tool. However, considering this alleged dimensional profile in the light of the usual way the instrument has been thus far discussed methodologically and hence used in practice [3, 4], a fundamental but yet unanswered question follows. If the EPDS is really multi-dimensional, so far as to suggest an independent anxiety sub-scale [6, 18], is it appropriate to use the complete 10-item score and thereafter specify a cut-off point to define PPD as commonly done? From a psychometric stance, in principle, the answer should be 'no' since distinct although not necessarily uncorrelated dimensions of a given construct require separate empirical scales. Therefore, relying on the evidence from the literature, a call for a two- or three-tiered measurement tool would be reasonable.

A connected question concerns the use of raw scores as proxies for latent traits in applied research. Beyond assessing whether these raw scores are effectively matching the purportedly related latent factors scores, asking which factors are actually involved is also necessary. Specifically in regards to the EPDS, should raw scores relate to specific (e.g., anhedonia, anxiety and depression) factors or would a sum of all ten item scores be an adequate representation of an overall dimension covering PPD? It should not go unnoticed that separately using sums of factor-specific items may incur in missing out the full mapping [20] of an overall dimension (in tandem or instead of several specific dimensions), should it exist. Conversely, adding up item scores when items comprise a multidimensional structure would be unwarranted. Different dimensions may also have exclusive antecedents and consequents, which would not be distinguishable if all items were lumped together.

Thus, before accepting a multi-scale usage of an instrument, an important aspect that needs scrutinizing concerns whether there is sufficient discriminatory power across the factors and whether these factors are thus appropriate descriptors of truly independent theoretical dimensions [21]. To the best of the authors' knowledge this scrutiny has never been performed in regards to the EPDS. One purpose of this study was thus to reappraise its dimensional structure with a special focus on the assessment of factor correlations and the related factor-based discriminant validity properties. This evaluation first required a reassessment of the number of constituent factors in order to examine whether the one-dimensional solution originally specified would be again refuted in the present data. Analysis would only proceed if a multi-dimensional structure were supported. Given the hypothesis that the identified factors failed to hold discriminant validity, the identification of a general factor would be further explored through a bifactor analysis [22]. An ensuing objective would then be to scrutinize the properties of raw scores as proper representations of model-based (latent) factor scores.

Mean maternal age was 25.3 years (range 13-44 years, 95% CI: 24.9; 25.8) and 22.7% (95% CI: 19.8; 25.6) of women were adolescents (< 20 years). Most of the participants had steady partners (86.6%, 95% CI: 84.2; 88.9) and up to 12 years of schooling (71.9%, 95% CI: 68.8; 75.0). Almost half (49.6%, 95% CI: 46.1; 53.0) were first time mothers and the mean age of the most recent infant was 59 days (range 3-150, 95% CI: 56.2; 61.9). The mean EPDS score was 7.8 (95% CI: 7.4; 8.2) and 24.3% (95% CI: 21.3; 27.2) women scored at or above the cut-off point of 12, which has been suggested in another study as appropriate for the Brazilian setting [23].

Four increasingly complex two- to four-factor E/CFA models were sequentially fitted. The 2-factor E/CFA model showed a RMSEA of 0.057 (90% CI: 0.045; 0.069). A formal comparison of the 3-factor E/CFA model with this simpler model yielded a χ ² value for difference testing of 49.223, which is highly significant (p ≪ 0.001 at 8 d.f.) and shows a fit improvement. Although the 4-factor E/CFA model indicated an even better fit -- RMSEA of 0.028 (90% CI: 0; 0.050) -- and tested significantly vis-à-vis the 3-factor model (χ ² = 18.876; p = 0.0086 at 7 d.f.), loadings on the forth factor were overall low (i₁ = 0.002, i₂ = 0.016, i₃ = 0.335, i₄ = 0.005, i₅ = -0.008, i₆ = 0.516, i₇ = 0.164, i₈ = 0.017, i₉ = -0.087 and i₁₀ = 0.125). Having decided for a model with three factors, we then looked into any possible residual correlations as indicated by the Modification Indices. Two correlations were suggested (i4⇔i5 = 0.425 and i3⇔i6 = 0.196), but freely estimating them in tandem showed different estimates from those projected by the MIs (i4⇔i5 = 0.193 and i3⇔i6 = 0.119). Thus, a three-factor model without residual correlations was regarded the most parsimonious. Findings are summarized in Table 1 (Model B). Factor 1 (anhedonia) encompassed items 1, 2 and 6, although the latter item showed cross-loadings on f2 and respective loadings were below the others. Factor 2 (anxiety) included items 3 to 5, and factor 3 (depression) encompassed items 7 through 10. All the three factors were highly correlated. Model fit proved to be reasonably good.

A strict CFA model was then fitted to the three-factor solution suggested in the E/CFA, with item 6 placed on factor 1 (Table 1 Model C). RMSEA (0.037) suggested a reasonable fit and loadings were fairly high. Of real interest here were the evaluations of convergent and, especially, discriminant validity. As shown in Table 2, AVE was low for f2 (0.43), indicating that convergent validity was questionable for at least one factor in the three-factor solution. Regarding factor 3, the respective square root of AVE used as benchmarks for evaluating discriminant validity was at or slightly above the related correlation, yet with quite some overlap of the confidence intervals. However, the magnitude of for factor 1 was below one of its related factor correlation (ϕ _1↔3), and for factor 2 was far below both correlations (ϕ _1↔2 and ϕ _2↔3). Overall, discriminant validity was beneath acceptable levels and results suggested that alternative options should be investigated.

Factor 1
	ρ ve(f1)	.56	(.52; .61)
		.75	(.72; .79)
Factor 2
	ρ ve(f2)	.43	(.37; .48)
		.65	(.61; .70)
Factor 3
	ρ ve(f3)	.65	(.61; .69)
		.81	(.78; .83)
Factor correlations
	ϕ 1↔2	.74	(.67; .81)
	ϕ 1↔3	.80	(.76; .85)
	ϕ 2↔3	.81	(.74; .87)

^a Values are based on the three-factor CFA solution (Model C in Table 1) based on the evidence suggested in the 3-factor E/CFA (Model B in Table 1), namely, factor 1 comprised by items 1, 2 and 6; factor 2 by items 3-5; and factor 3 by items 7-10. In brackets: 95% confidence intervals obtained by bootstrap method (B = 1000).

In view of these results and acknowledging an adequate item-factor specification as conveyed by the interim E/CFA, a bifactor model was fitted to examine the existence of a general factor, yet still accounting for (conditional on) the three specific factors (Table 1, Model D). Model fit improved further -- RMSEA (0.026, 90% CI: 0.005; 0.041), CFI (0.997) and TFI (0.995). Formally evaluating the significance of each factor showed that f1 and f2 were highly significant; both χ ² for difference testing between the bifactor model (Model D) and the nested models in turn removing the factors under scrutiny had p ≪ 0.001. The significance of factor 3 was less marked (p = 0.042). Regardless, loadings on the specific factors were all low to moderate . In contrast, all g-factor loading were all fairly high, ranging from 0.51 to 0.81 . As expected, testing the significance of the g-factor by comparing Model D with the three-dimensional CFA (Model C) yielded p ≪ 0.001.

Figure 2 focuses primarily on the general factor as an encompassing scale mapping individuals along a continuous PPD latent dimension [20], and shows the m thresholds pertaining to the ten, four-level polytomous items. Although there are lines crossing-over item steps, there seemed to be an overall ordered gradient. The upward trends vis-à-vis θ for all items convey that women having a high level of PPD will display more symptoms and score increasingly higher item values.

The scale and items' Loevinger's H coefficients and the criteria for assessing single and double monotonicity assumptions are shown in Table 5. The numbers of absolute and relative violations per item (monotonicity and nonintersections) are also provided. All items' H _Sj coefficients are above 0.3. The scale's overall scalability coefficient is also above this cut-off point (H = 0.4208). The monotonicity assumption was not rejected since no important violation occurred and all criteria were satisfied (Table 5, columns 2 and 3). This may be visualised graphically in Figure 3. None of the item-restscore traces showed intersections. This is not the case for inter-item ISRF nonintersections assessing the assumptions of double monotonicity as judged from the number of violations and related criteria shown in columns 4 and 5 of Table 5. Although six out of 10 criteria were less than 40 (i1, i3, i5, i6, i7 and i10) and for which violations can be merely ascribed to sampling variation, three items showed criteria between 40 and 80 (i2, i8 and i9) and one item (i4) exceeded 80. As a whole, the model followed by these data is therefore only partly double monotone.

		Monotonicity		Double monotonicity
Item	H	No. of violations a	Criteria b	No. of violations a	Criteria b
i1	.4234	0		2 (.0010)	37
i2	.4086	0		3 (.0015)	46
i3	.3770	0		2 (.0010)	35
i4	.3605	1 (.0222)	13	12 (.0062)	82
i5	.4019	0		2 (.0010)	34
i6	.3880	1 (.0159)	11	4 (.0021)	38
i7	.4881	0		3 (.0015)	34
i8	.4673	0		9 (.0046)	70
i9	.4775	0		7 (.0036)	62
i10	.4201	0		0
Total (scale)	.4208	4 (.0039)	---	---	---

^a See text for explanation. In brackets: proportion of violation vis-à-vis all possible violations regarding active pairs. See Molenaar, Sijtsma & Boer [52] and Hardouin, Bonnaud-Antignac & Sebille [51] for details.

^b See Hardouin, Bonnaud-Antignac & Sebille [51], equation 5 on page 39.

As judged by the preponderance of mostly positive properties described above, the sequence of increasing intensity/severity suggested by Figures 2 and 3 could therefore be described thus: i3 (self-blaming); i4 (anxious/worried for no reason); i6 (feeling of things getting on top); i8 (sad or miserable); i5 (scared/panicky for no reason); i7 (unhappy and difficulty in sleeping); i2 (no enjoyment to things); i1 (not able to laugh and see the funny side of things); i10 (thoughts of self-harming); and i9 (unhappy and crying).

As conveyed in the introduction, while the EPDS has had a huge following over the years, there has also been quite some disagreement about its dimensionality, especially in regards to the intricacies of its factor structure. Discussions as to the number of factors and the internal distribution of items have dominated the research program. Adding knowledge to the psychometric history of the EPDS, this study sought to step forward and examine some properties that had never been previously evaluated. As a positive start, we were able to corroborate the three-dimensional structure that has been previously reported in the literature [7–9, 11, 14, 15]. More importantly, however, our study showed that, without further elaboration, this three-dimensional structure held rather poor factor-based discriminant validity. In principle, this would discourage using the EPDS as separate subscales.

Moving beyond a basic three-factor structure, results suggested that the EPDS was capable of distinguishing a general factor representing PPD. Moreover, in tandem with the related literature on mood and anxiety disorders [44], three specific factors could also be identified, although less conspicuously. While the three factors were significant on formal testing, specific loadings were low or moderate at the most. This contrasted with the g-factor loadings, which were sensibly higher. Overall, this picture goes in hand with the relatively low contribution of three specific factors to the total and explained variance, as opposed to the clear preponderance of a g-factor in the variance partition. As a side product, the currently proposed model (D) may also have unravelled a persistent problem found since the EPDS' conception, namely, that the consistent 'split' loadings found in several other settings regarding item 6 [10, 14, 15, 53] were likely accountable to its contribution to a general factor that had not been properly specified.

It is thus laudable that the construct may be described as a single factor. The close connection between this general factor and the total raw score also suggests that the traditional practice of adding up raw component item scores to form an overall score is justifiable. To some extent, one may argue that the development process of the EPDS has 'come full circle' since the proposed 10-item score sum originally proposed by Cox et al. [2] is apparently tenable in practice as a first approach. Considering the other favourable properties that were detected (high correlation with the g-factor score, scalability and monotonicity), a total raw score may effectively be regarded as an empirical manifest of this general trait. In contrast, it is worth noting that the relatively weak correlations between specific-factor and related raw scores are indicative that separately using the EPDS sub-scales as raw scores in the postpartum period may not be so straightforward [6, 18]. This is in line with the finding of Brouwers et al. [15] who identified that both anxiety and depressive symptoms were more accurately measured when using the total 10-item EPDS score.

Thus, a general PPD dimension that is seized through indicators (also) pertaining to three specific dimensions -- anhedonia, anxiety and depression -- may apparently be supported. Theoretically, anxiety and anhedonia, represented by their respective items and factors, would be only part of the spectrum of symptoms commonly observed in the presence of depression, not least in the period following childbirth. This proposition seems to agree with previous experimental and clinical observations that within a single episode of depression it is quite common for some symptoms of anxiety precede an overt depression symptomatology [54–56]. The helplessness/hopelessness perspective proposed by Alloy et al. [54] may clarify some features of the anxiety-depression co-morbidity. Accordingly, the intertwined relationship between anxiety and depressive symptoms is explained by noting that the anticipation of helplessness that comes with anxiety is likely also to trigger certain negative-outcome expectancies, and when all efforts to exert control eventually fail, overt depression follows. Nonetheless, the understanding that depressive and anxiety manifestations are sequential, alternate or concomitant facets of a common underlying process does also bring cogent theoretical support for the present empirical evidence, which is encouraging [57–60]. The meaningfulness of the actual crescendo in intensity and severity conveyed by mingled anhedonia, anxiety and depression-like symptoms (as outlined at the end of the results section) offers further support.

As many researchers in the field of measurement development and cross-cultural adaptation, we also take a universalistic approach and strive for psychometric consistency [61]. Yet, the answer as to whether a factorial structure will be repeatable in different socio-linguistic and cultural situations requires more research. New evidences of poor factor-based discriminant validity; followed by replicating and extending the models proposed here; along with theory-based external forms of construct validation studies using the EPDS as a general factor (or perhaps its 10-item raw score equivalent) should be a constructive way forward.

Yet, albeit still pending replication, it seems reasonable to reinforce the use of the full EPDS as a 10-item score. In applied research, the raw score may be indeed used as a first-approach 'mapper' of the underlying continuum. In a clinical perspective, the currently suggested cut-off points may be applied to help identify women who would benefit from in-depth psychiatric evaluation and ensuing follow up [4]. As a reminder, though, one should not ignore that identifying heterogeneous populations bearing primarily specific mood and anxiety disorders would require more complex instruments and that a definitive diagnosis of PPD cannot be established without a specialized assessment through more accurate means.

This study attempted to take the debate on the dimensional structure of the EPDS a step forward. The present results suggest that, although a factorial analysis has again identified three factors, they lack factor-based discriminant validity and should not be used empirically as separate sub-scales. Nonetheless, albeit pending corroboration from further studies carried out in different settings, the total 10-item EPDS scale seems to be well suited as an empirical representation of a general factor representing PPD and its use in clinical practice and applied research is encouraged.