What is the Best Latent Structure of Negative Symptoms in Schizophrenia? A Systematic Review.

BACKGROUND: Determining the best latent structure of negative symptoms in schizophrenia could benefit assessment tools, neurobiological research, and targeted interventions. However, no review systematically evaluated studies that assessed and validated latent models of negative symptoms.
OBJECTIVE: To identify and evaluate existing latent structure models in the literature of negative symptoms and to determine the best model.
METHOD: Systematic search of MEDLINE, EMBASE, and Scopus on July 19, 2020, for confirmatory factor analysis models of negative symptoms in patients with schizophrenia. The available evidence was assessed through 2 sets of criteria: (1) study design quality-based on negative symptoms assessment and modeling strategy and (2) psychometric quality and model fit-based on fit indices and factor definition quality.
RESULTS: In total, 22 studies (n = 17 086) from 9 countries were included. Studies differed greatly regarding symptom scales, setting, and sample size (range = 86-6889). Dimensional models included 2-6 factors (median = 4). Twelve studies evaluated competing models and adopted appropriate instruments to assess the latent structure of negative symptoms. The 5-factor and hierarchical models outperformed unitary, 2-factor, and 3-factor models on all direct comparisons, and most of the analyses derived from the Brief Negative Symptom Scale. Considering the quality criteria proposed, 5-factor and hierarchical models achieved excellent fit in just one study.
CONCLUSIONS: Our review points out that the 5-factor and hierarchical models represent the best latent structure of negative symptoms, but the immaturity of the relevant current literature may affect the robustness of this conclusion. Future studies should address current limitations regarding psychometric properties and also address biological and clinical validity to refine available models.

Introduction

Negative symptoms represent core features of schizophrenia since the first descriptions of the disorder[1,2] and refer to a diminution or absence of expected behaviors and inner experiences related to motivation, interest, or expression.[3-7] Although a major contributor to poor real-life functioning in people with schizophrenia,[3,8,9] no specific approved treatment exists for negative symptoms,[10,11] which is explained both by the insufficient knowledge about its neurobiology and by the challenging assessment of these symptoms.

To address such issues, the National Institute of Mental Health (NIMH) organized, in 2005, the NIMH-MATRICS (Measurement and Treatment Research to Improve Cognition in Schizophrenia) Consensus Statement on Negative Symptoms.[12] It acknowledged 5 constructs within the negative domain: blunted affect (decrease in the observed expression of emotion, ie, facial and vocal expression, and expressive gestures),[13,14] alogia (reduction in the quantity of speech and in its spontaneous elaboration)[4], anhedonia (diminished capacity to experience pleasant emotions)[15], asociality (reduction in social initiative due to decreased interest in forming close relationships with others)[16], and avolition (reduced initiation and persistence of goal-directed activity).[17]

Despite improvements that allowed the development of more sophisticated tools, obtaining accurate and reliable measurements of negative symptoms remains challenging due, in part, to the overlap between the symptoms that make up this dimension. A particular negative symptom may be represented by more than one item of a scale, eg, affective flattening could influence simultaneously social withdrawal and speech spontaneity, thus being accounted for multiple times. In this scenario, factor analysis methods are useful, as they reduce measurement errors by generating latent structure models that present independent unobservable variables (factors) to explain correlations between observed variables (here, the items of a scale).[18] Two possible approaches to factor analysis are exploratory factor analysis (EFA) and confirmatory factor analysis (CFA).

EFA is a data reduction technique that infers the presence of latent factors responsible for shared variance among a set of items. It can be useful to generate hypotheses about latent structures; however, EFAs do not specify an underlying structure for the observed variables but rather assume that each item could be related to each latent factor, not ensuring a true validation. EFAs using negative symptom scale items alone find support for 2 dimensions that reflect motivation and pleasure (MAP) and expression (EXP),[16,19-22] but CFA models find that 1- and 2-factor solutions usually offer a poor fit for the data.[23,24]

CFA allows testing a priori hypothesis and objective comparison with other theoretical models. This way, CFA provides more accurate conclusions about latent models, being more appropriate to evaluate the latent structure of negative symptoms. Recently, the evidence supporting the Consensus 5-factor model has been reviewed,[25] and results from CFA studies using mainly the Brief Negative Symptom Scale (BNSS) have consistently shown the robustness of the 5-factor and hierarchical models over the 2-factor or unitary models. Despite that, no study has systematically reviewed and evaluated all the literature covering CFA studies that investigated latent models for negative symptoms, even though nonsystematic reviews on the topic are available.[4,19,25-28]

Thus, we performed a systematic review of the literature searching for latent structure models of negative symptoms in schizophrenia and aimed to identify the best latent structure. To define the best model, we assessed the quality of each study, psychometric parameters, and direct comparisons between models obtained from the same sample.

Methods

Search Strategy

We conducted a systematic review according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyzes (PRISMA)[29,30] guidelines. Searching MEDLINE, EMBASE, and Scopus from database inception until July 19, 2020, we used the following search terms: (((((((“schizophrenia spectrum and other psychotic disorders”[MeSH Terms])) OR “schizophrenic disorders”) OR schizophrenia)) AND (((“negative symptoms”) OR (“negative symptom scale”)))) AND (((((((“confirmatory factor analysis”) OR “latent structure analysis”) OR “latent structure”) OR “latent model”) OR “validation study”)))))). We also reviewed manually reference lists of included articles and relevant reviews.

Inclusion and Exclusion Criteria

We included original, peer-reviewed articles without language restriction and excluded posters, letters, editorials, and reviews. We included samples with individuals aged ≥16 years old meeting diagnostic criteria for schizophrenia according to Diagnostic and Statistical Manual of Mental Disorders (DSM)-III,[31] DSM-IV,[32] or DSM-V,[33] although the addition of non-affective psychotic disorders was not a reason for exclusion if most of the subjects (>50%) had a schizophrenia diagnosis. We selected studies that performed CFA to evaluate the latent structure of negative symptoms. We excluded studies with EFA and principal component analysis (PCA) since such approaches do not test theoretical models of latent factors.[18]

Data Collection

Two authors (G.K. and B.H.) independently screened all retrieved records by title and abstract and assessed selected studies in full for eligibility. Any conflicts were resolved by consensus or consulting another investigator (E.D.). Additionally, the following information was extracted independently (G.K. and B.H.) from each included study: metadata, sample characteristics, intraclass correlation coefficient (ICC), negative symptom rating scales, model origin, estimation method, competing models and discriminative criteria employed, the best models, fit indices, and shortest number of items with loadings higher than 0.5 per factor.

Quality Assessment

Quality assessment was based on study design, psychometrics, and model fit.

Study Design.

The study design adequacy was evaluated according to the reliability of the assessment of negative symptoms; appropriateness of the adopted instrument to assess the latent structure of negative symptoms; and modeling strategy, which encompassed model origin, estimation method employed for the CFA, and use of competing model’s approach.

Small samples produce less stable factor analysis solutions.[34] However, there is no consensus regarding the minimum sample size for CFAs. This issue cannot be determined by rules of thumb[35,36] and would only be adequately tested if done via sensitivity power analyses for each included study,[35,36] which is far from the scope of the present study. Therefore, sample size was not included in the quality assessment.

Assessment reliability of negative symptoms was evaluated through ICC. Based on previous studies,[37-39] we considered assessments with ICC ≤ 0.5 (or with ICC not informed/performed) as “hardly reliable,” 0.5 < ICC < 0.8 as “reliable,” and ICC ≥ 0.8 as “highly reliable.”

Regarding instruments appropriateness, the Positive and Negative Syndrome Scale (PANSS) and Negative Symptom Assessment (NSA) lack the items needed to cover all constructs determined by the NIMH-MATRICS Consensus and, so, cannot properly assess the latent structure of negative symptoms.[23,40] On the other hand, the BNSS, the Clinical Assessment Interview for Negative Symptoms (CAINS), and the Scale for the Assessment of Negative Symptoms (SANS) possess the items to cover the Consensus’ domains, making them more suitable to answer our research question.[23,40]

EFA or PCA should be followed by CFA using a different sample.[41-43] Therefore, studies that performed EFA/PCA to obtain the tested model and used the same sample to perform the CFA and studies that performed CFA based on theoretical models not preceded by EFA/PCA were considered “methodologically limited.”

The selection of an estimator must be based on distributional patterns of the data and assumptions, which makes appropriate to report and justify its choice, while relying on statistical software’s default settings is not advisable.[18] Thus, studies that provided no information about the estimation method employed were also considered “methodologically limited.”

The last step of quality assessment of study design consisted of verifying whether the best model presented had been compared with other models and, if so, which were the competing models. The strongest test of a proposed model is to identify and test competing models that represent truly different, but highly plausible, hypothesized structural relationships.[36] Using the same sample to test competing theories on the latent structure of negative symptoms provides more robust evidence than testing a single isolated model. Thus, studies that compared different models obtained from the same sample were well regarded, whereas the test of a single model reduced the quality of evidence. Studies that tested models with the same number of factors obtained by different items of the same scale were penalized, as well as studies that compared a model exclusively with the null model, since it is widely accepted that negative symptoms are a separate factor in schizophrenia.[4,12,28]

Psychometrics and Model Fit.

Psychometric quality and model fit were assessed considering the following parameters: descriptive fit indices—comparative fit index (CFI), non-normed fit index (NNFI), normed fit index (NFI), root mean square error of approximation (RMSEA), weighted root mean square residual (WRMR), or goodness-of-fit index (GFI) —and factor definition quality. We also evaluated information criteria—Akaike Information Criterion (AIC) and/or Bayesian Information Criterion (BIC) —when studies compared non-nested models, with the lowest values being used to determine optimal model fit.[44] Chi-square values were not considered to evaluate model fit due to its high sensitivity to sample size as well as the ratio χ 2/df.[45,46]

The following standards for appropriate fit indices were considered:[47-49] GFI > 0.90, CFI > 0.95, NFI > 0.95, NNFI > 0.95, RMSEA < 0.06, and WRMR < 1.00. Models presenting no index with adequate fit were considered to have “poor fit”; models with at least one adequate fit index were classified as “acceptable”; and models showing all indices with adequate fit were considered “excellent.”

Factor definition quality was based on the number of items with loadings > 0.5[36] per factor. Factors with less than 3 or 4 items per factor require larger sample sizes and are more likely to provide unstable solutions.[18,50-54] On the other hand, there is support for the use of the few best indicators for the development of theoretically sophisticated models.[55] Thus, we considered factors defined by a single item as “poorly defined,” factors defined by 2 indicators as “acceptable,” and factors defined by at least 3 items as “well defined.” Models obtained through an EFA in a different sample and that used the EFA’s factor loadings to define factors, instead of CFA’s, had their quality lowered.

Additional details of quality assessment are provided in the supplementary material.

Determining the Best Latent Structure

Following the quality assessment, we selected the studies that evaluated competing models derived from the SANS, BNSS, and CAINS. After that, we compiled the models tested in a single table to verify which models performed better according to each study’s criteria. Finally, we evaluated the best models according to the predefined fit quality criteria.

Results

Our search yielded 1680 records found in the database search and 9 through other sources (figure 1). After removing duplicates, 1495 records titles and abstracts were screened, out of which 69 were selected for full-text reading, resulting in 22 included studies in our qualitative analysis.

Fig. 1.

Review flow diagram.

We extracted data from 26 latent models of negative symptoms. Table 1 summarizes the included studies, with 5 published between 1992 and 1996 and 17 between 2013 and 2020. Altogether, 17 086 individuals aged 16–79 years were included (range n = 86–6889), from 9 different countries (United States = 7, Singapore = 4, South Korea = 3, China = 2, Netherlands = 2, Germany = 1, Israel = 1, Italy = 1, and Spain = 1). Samples included chronic inpatients,[56-59] outpatients,[23,60-64] and patients in mixed settings,[24,65-74] mainly diagnosed not only with schizophrenia but also with schizoaffective disorder, schizophreniform disorder, delusional disorder, and psychosis not otherwise specified.

Table 1.

Summary of the Main Findings

Note: AIC, Akaike information criterion; aBIC, adjusted Bayesian information criterion; BIC, Bayesian information criterion; aBIC: BNSS, Brief Negative Symptom Scale; CAINS, Clinical Assessment Interview for Negative Symptoms; CAINS-item 1, Motivation for Close Family/Spouse/Partner Relationships; CAINS-item 2, Motivation for Close Friendships/Romantic Relationships; CAINS-item 5, Motivation for Work and School Activities; CAINS-item 6, Expected Pleasurable Work and School Activities—Next Week; CFA, confirmatory factor analysis; ECVI, expected cross-validation index; CFI, comparative fit index; EXP, Expression; G5, Mannerisms and Posturing; G7, Motor Retardation; G13, Disturbance of Volition; G16, Active Social Avoidance; GFI, goodness-of-fit index; ICC, intraclass correlation coefficient; IFI, incremental fit index; K-CAINS, Korean version of the Clinical Assessment Interview for Negative Symptoms; N1, Blunted Affect; N2, Emotional Withdrawal; N3, Poor Rapport; N6, Lack of Spontaneity and Flow of Conversation; MAP, Motivation-pleasure; ML, maximum likelihood; MLR, robust maximum likelihood; NFI, normed fit index; NIMH, National Institute of Mental Health; NNFI, non-normed fit index; NOS, non otherwise specified; NSA, Negative Symptom Assessment; NSA-item 6, Affect: reduced modulation of intensity; NSA-item 16, Slowed movements; PANSS, Positive and Negative Syndrome Scale; PCA, principal component analysis; RMR, root mean square residual; RMSEA, root mean square error; SANS, Scale for the Assessment of Negative Symptoms; SANS-item 4, Poor eye contact; SANS-item 6, Inappropriate affect; SANS-item 9, Poverty of speech; SANS-item 10, Poverty of Content of Speech; SANS-item 11, Blocking; SANS-item 14, Grooming and Hygiene; SANS-item 19, Sexual activity; SANS-item 23, Social Inattentiveness; SRMR, standardized root mean square residual; WLSMV, weighted least square mean and variance; WRMR, weighted root mean square.

**Models in bold outperformed other models.

Study Design

Negative symptom assessment was hardly reliable in 7 studies,[56,57,59,61,67,70,73] reliable in 2 studies,[68,71] and highly reliable in 13 studies.[23,24,58,60,62-66,69,72,74,75]

Negative symptom scales varied greatly. Four models were based on the NSA,[56,57,63,66] 7 on the BNSS,[23,24,62,64,73] and the others were equally divided between the SANS,[23,58,59,65,67] the PANSS,[60,61,68,70,71] and the CAINS.[23,69,72,74,75]

Regarding models origins, most of the studies based the CFA on a PCA, EFA, and/or CFA from previous studies; [56,61,62,64,65,68-70,73-75] 5 studies obtained the CFA model from a PCA or EFA performed in a different sample,[24,60,63,66,72] while 3 studies used the same sample of the CFA; [59,67,71] and 3 studies performed the CFA based on a theoretical model not preceded by a PCA or EFA.[23,57,58]

The most employed estimation method was the maximum likelihood,[56-58,60,68,73,74] followed by a combination of robust maximum likelihood and robust weighted least squares,[23,24] and the robust weighted least squares alone.[61] Twelve studies did not provide information about the estimator.[59,62-67,69-72,75]

The competing model’s approach was adopted by 18 studies. The most used discriminative criteria were based on the analysis of descriptive fit indices and information criteria,[23,24,64,67,73,75] followed by the chi-square difference test combined with the analysis of descriptive fit indices,[56-58,65,71] the analysis only of descriptive fit indices,[62,63,66,70,72] and factor definition quality.[59]

Psychometrics and Model Fit

Dimensional models ranged from 2[60,61,68-71,74,75] to 6[56] dimensions (median = 4). Two and 5-factor[23,24,57,58,62,73] models were the most depicted among studies, followed by hierarchical,[62,64,73] 3-factor,[59,65-67] and 4-factor[63,72] structures. The hierarchical model consists of 2 second-order factors reflecting EXP and MAP and 5 first-order factors reflecting the domains of the NIMH consensus development conference.[12]

We observed high heterogeneity in terms of models and factors composition. For example, none of the studies that evaluated the 3-factor model using the SANS presented a common factor composed of the same indicators. In addition, none of the models derived from the NSA presented the same number of factors. On the other hand, models obtained from the PANSS, BNSS, and CAINS showed greater uniformity.

Fit indices used and their values varied. To assess model fit, older studies used the chi-square test and few descriptive fit indices, whereas more recent articles used modern fit measures and information criteria. The summary of descriptive fit indices is presented in table 2. Regarding factor definition quality, 9 models were considered well-defined,[56,60,61,68-71,74,75] 14 were adequate,[23,24,58,59,62-67,73] and 3 were poorly defined.[23,57,72]

Table 2.

Summary of Fit Indices

Fit Index	Range	Median	Reference	Used (nº of Studies)	Adequate (nº of Studies)	Adequacy Rate (%)
CFI	0.870-1.0	0.975	>0.95	20	14[23,24,59–64,68,70,72–75]	70.0
NNFI	0.806-0.999	0.961	>0.95	19	11[23,24,59,61–64,68,70,73,75]	61.1
RMSEA	0.0082-0.140	0.074	<0.06	17	6[60–62,67,71,75]	35.3
WRMR	0.430–1.0	0.827	<1.0	5	5[23,61,62,70,72]	100.0
GFI	0.828-1.0	0.920	>0.9	5	3[61,67,71]	60.0
NFI	0.810-0.910	0.880	>0.95	4	0	0

Note: CFI, confirmatory factor analysis; GFI, goodness-of-fit index; NFI, normed fit index; NNFI, non-normed fit index; RMSEA, root mean square error of approximation; WRMR, weighted root mean square.

Six models presented poor fit,[56-58,65,66,69] 14 presented acceptable fit,[23,24,63,64,67,68,70-74] and 6 models met requirements for excellent fit.[59-62,75]

The Best Latent Structure of Negative Symptoms

Table 3 summarizes the studies that used the SANS, BNSS, or CAINS to assess competing models of negative symptoms. The 5-factor and hierarchical models based on the NIMH-MATRICS Consensus outperformed other models in all 5 comparisons using the BNSS and in a single comparison both with the CAINS and the SANS.

Table 3.

Competing models of the SANS, CAINS, and BNSS

Among these studies that directly compared the Consensus 5-factor and hierarchical models with other models, only Ang et al[62] obtained models with excellent fit according to the fit criteria proposed, whereas the others presented models with acceptable fit. Ang et al[62] used the BNSS and, based on models generated by Strauss et al[23] and Ahmed et al[24], concluded that the hierarchical model and the 5-factor model outperformed the 2-factor model, with the hierarchical model having an advantage over the 5-factor model. Regarding the studies with acceptable fit, according to the predefined criteria, the superiority of the hierarchical model in relation to the 5-factor model was not unanimous.

After contrasting the best 2 models, Strauss et al[23] concluded that the 5-factor model obtained a better fit both with the BNSS and the CAINS, based on the information criteria, but found that the hierarchical model performed slightly better with the SANS. Ahmed et al[24] considered both models adequate, based on descriptive fit indices and information criteria, but extended their analyzes of cultural invariance of the BNSS exclusively to the 5-factor model. Mucci et al[73] and Jeakal et al[64] also used the BNSS and considered both models equally adequate, but Jeakal et al[64] only provided the data about the hierarchical model.

Regarding the studies that did not directly compared the Consensus 5-factor and hierarchical models with other models, only Sayers et al[59] and Xie et al[75] presented models with excellent fit. Sayers et al[59] obtained excellent fit both with a 3-factor and a 5-factor SANS model but considered the latter less adequate because it had a factor composed of 2 items. Xie et al[75] used the CAINS and concluded that the 2-factor model by Kring et al[16] outperformed the unitary model and the 2-factor model by Chan et al[76].

Discussion

We analyzed 22 publications addressing the dimensionality of negative schizophrenia symptoms, 12 of which used the SANS, the BNSS, or the CAINS to assess competing models of negative symptoms. The Consensus 5-factor and hierarchical models outperformed unitary, 2-factor, and 3-factor models on direct comparisons, and our results suggest that they currently represent the best latent structure of negative symptoms. However, the reduced number of available studies may affect the robustness of this conclusion.

Twelve studies directly compared models using appropriate instruments to assess the latent structure of negative symptoms, of which only 5 compared the Consensus 5-factor and hierarchical structure to other models, resulting in 5 comparative analysis derived from the BNSS, 1 from the CAINS, and 1 from the SANS. Of note, Strauss et al[23] assessed competing models from the BNSS, CAINS, and SANS using a different sample for each instrument. Peralta and Cuesta[58] and Sayers et al[59] also tested SANS 5-factor models, but both studies presented latent structures with an inattention factor (no longer included in the negative dimension) and grouped anhedonia and asociality into a single factor. Moreover, among the 5-factor and hierarchical Consensus-based models, only those obtained by Ang et al[62] showed excellent fit according to the proposed criteria. In sum, direct comparisons favored the Consensus 5-factor and hierarchical model, but psychometric properties are limited.

Indeed, few studies fulfilled all the standard fit requirements proposed, and the RMSEA was the fit index with the lowest adequacy. Since it assesses how well a model fits in the population,[18] the RMSEA determines how a proposed model is apart from a perfect model.[77] It represents one of the most informative fit indices[78] for being sensitive to the number of estimated parameters in the model and relatively little influenced by sample size.[18] RMSEA seems to improve—indicating better fit—as more variables are added to the model, which means that the lower the number of items per factor, the worse the RMSEA.[79] As most of the instruments have few items to assess each factor, the psychometric criteria did not fully favor the 5-factor and hierarchical model, which should not be considered a criterion for rejecting these models but rather an example of the contrast between requirements for good fit and available instruments. In fact, the only fit index that separated the models obtained by Ahmed et al,[24] Strauss et al,[23] Mucci et al,[73] and Jeakal et al[64] from an excellent fit, according to the proposed quality criteria, was the RMSEA, whose cutoff value of 0.08 is also accepted by some authors.[18,80]

Despite the superiority of the 5-factor and hierarchical models over the 2-factor model in studies that performed comparative analyzes, the neurobiological and clinical evidence for specific constructs of the 5-factor model is still limited, although promising. Previous investigations regarding the underlying neurobiology of negative symptoms may have failed to find neural correlates for the 5 domains due to assessments based on scores of constructs with fewer factors.[25] Similarly, past unsuccessful attempts to promote targeted treatments for negative symptoms may, in fact, have been effective for 1 or more of the 5 domains, with positive results being masked by assessments not specific enough for each domain.[25] Further studies could fill these gaps assessing negative symptoms with greater granularity, which could, in turn, provide domain-specific therapeutic and neurobiological advances.

Although the present study strengthens the thesis that negative symptoms should be evaluated with greater granularity matching clinical experience, conclusions derived from only 5 studies. Therefore, the field may benefit from additional studies investigating unanswered questions: the 5-factor/hierarchical models are invariant across different disease stages or over time? Other 5-factor models could fit similarly or best? Can a multilevel structure affect the results?

Psychometrics, clinical experience, and neurobiology have convergences and tensions. No single source of evidence can be sufficient. The balance between seeking psychometric and clinical relevance has been especially controversial. We acknowledge that the rejection of models solely based on one single fit index can be excessive. However, merely disregarding fit indices as inadequate can overlook opportunities for new research questions. In the end, we believe that data and clinical utility provide the best guides to the current discussion. In this sense, our results identify some caveats when framing available evidence on the dimensional nature of negative symptoms.

Future investigations of latent structures for negative symptoms should adopt strict procedures regarding the study design, such as calculating the minimum sample size, using modern instruments (BNSS and CAINS), ensuring the highest possible reliability for the assessment of negative symptoms by performing and reporting the ICC, not using the CFA sample in a previous EFA or PCA, reporting the estimation method and justifying its use, and adopting the competing models’ approach. Additionally, we encourage future studies to provide comparative analysis between the CAINS 2-factor, 5-factor, and hierarchical models and, ideally, also between the BNSS and the CAINS using samples jointly assessed by these instruments.

The limitations of this review include the high heterogeneity among studies’ samples and designs (eg, assessment tools and modeling strategies), the small number of studies that analyzed competing latent models of negative symptoms using appropriate instruments, and the lack of a validated method to measure the quality of factor analysis studies. Nonetheless, we operationalized a unified quality assessment according to criteria previously recommended by methodological papers, which enabled a systematic comparison among the included studies. We adopted standard and widely used cutoff points but that may be considered overly rigorous and not completely unanimous in literature. Psychometrics represents an important perspective to validate a construct but does not provide the final answer alone. Thus, we added additional perspectives, ie, the study design and instruments used, to bring a comprehensive view of the field.

Overall, we conclude that the 5-factor and hierarchical models are currently the best conceptualization of negative symptoms. We believe that these results may guide future psychometric studies and facilitate the search for biological and clinical validity of the negative dimension in schizophrenia.

Click here for additional data file.