To achieve the objectives of the study, participants responded to socio-demographic questions on their age, sex and place of residence. In addition, we administered the Spanish adaptation of the STAI (Buela-Casal, Guillén-Riquelme, & Seisdedos Cubero, 2016). This version includes 40 items (i.e. 20 for each subscale). In the Spanish adaptation, the response scale ranges from 0 to 3 points, by contrast with the original STAI, in which it ranges from 1 to 4 points (Spielberger et al., 1970). In the state anxiety subscale, item scores range between 0=nada [not at all], 1=algo [somewhat], 2=bastante [moderately] and 3=mucho [very much]. In the trait anxiety subscale, response options range between 0=casi nunca [almost never], 1=a veces [sometimes], 2=a menudo [often] and 3=casi siempre [almost always]. Although each subscale has a theoretical range from 20 to 80, the scores can be compared to those of the original scale by adding 20 to the scores obtained.

The present research was an instrumental study, as we assessed the psychometric characteristics of a psychological questionnaire (i.e. the STAI). The study was drafted following the recommendations made by Hartley (2012).

Participants were selected through non-probabilistic sampling. The STAI was administered in public places visited by a large number of people and in various organizations that cooperated in the study. Prior to administering the STAI, the experimenters verified that participants were over 18 years old and informed participants of the purpose of the study, guaranteeing its confidentiality. Next, the STAI was handed to participants and the experimenter stayed close to them to respond to any questions.

For the clinical sample, we contacted with seven psychologists. When clinicians diagnosed a patient with a general anxiety disorder, explained her or him the study and that data would be treated confidentially. If the patient accepted to participate, the clinician administered the questionnaire. We applied the following inclusion criteria: being 18 years or older and being diagnosed with generalized anxiety disorder, and have no secondary diagnosis. The research was approved by University of Granada Research Ethics Committee.

Analyses were conducted in two stages, each with a different part of the sample. First, we performed a discriminant analysis separately for the trait anxiety and state anxiety subscales using a step by step procedure using F significations. Our aim was to obtain a shorter selection of items that enabled a correct classification of subjects with and without generalized anxiety disorder. After that, we calculated the internal consistency (i.e. ordinal alpha) using a polychoric correlation matrix. This was done using R software (R Core Team, 2016).

Second, we conducted a confirmatory factor analysis to assess the fit of the short form obtained with the discriminant analysis. We also verified the fit of the short forms proposed by other authors. This allowed us to determine which of the short versions demonstrated the best fit with the present sample. To do so, we selected the robust weighted least squares means and variance adjusted (WLSMV) method, calculated by using a diagonal weight matrix. This method is appropriate for categorical measures with a sample higher than 100 (Holtmann, Koch, Lochner, & Eid, 2016). We compared the different models using the Comparative Fit Index (CFI), the Tucker-Lewis Index (TLI) and the Root Mean Square Error of Approximation (RMSEA). Although both scales of the STAI were theoretically independent, we developed path diagrams showing relations between both, since many studies have found a correlation between them. This was done by using R software (R Core Team, 2016) with the lavaan package (Rosseel, 2012) and the ltm package (Rizopoulos, 2006).

The objective of the present study was to develop a short form of both subscales of the STAI. First, we defined a four-item version for each subscale that was able to discriminate between sample of general population and a clinical group with generalized anxiety disorder. The number of items selected is similar to that of other short versions of the STAI, since most short forms include six items. A comparison of the fit of the form obtained through discriminant analysis revealed that the four item selection had higher fit indices. However, the reliability was better in the others shorts versions with more items but the test information functions were very similar between all shorts versions. Despite this, the items selected through discriminant analysis were adequate.

The short form of four items per subscale was able to correctly classify over 80% of subjects for state and 74% of subjects in trait (obtained in the discriminant analysis classification). This is an interesting finding considering that most short forms include six items. This number of items is mainly set to replicate the form developed by Marteau and Bekker (1992) or the methodology followed by them. However, these authors selected six items without explaining why they had done so. In addition, as stated by several authors, other forms with a different number of items have been found to be adequate (Fioravanti-Bastos et al., 2011; Van Knippenberg et al., 1990). For this reason, it is interesting to test scales with a different number of items and to try to set a number of items using objective and theoretical criteria. In addition, it is adequate to propose short forms of both subscales because, although state anxiety and trait anxiety are theoretically independent (Spielberger et al., 1970), many studies have shown a relationship between both factors (e.g. Guillén-Riquelme & Buela-Casal, 2011; Ortuño-Sierra, García-Velasco, Inchausti, Debbané, & Fonseca-Pedrero, 2016). This is consistent with the high correlations observed in all the models tested in the confirmatory factor analysis.

As regards the comparison between models, we observed that the version of four items per factor had the best fit. Yet, other models analyzed had very good fit indices and better reliability too. The items that discriminate the best between high and low anxiety levels have the best factor structure throughout all levels of anxiety. However, no clinical score was available in the Spanish adaptation (Buela-Casal et al., 2016). The values of the original form (Spielberger et al., 1970) were considerably higher than those of the Spanish adaptation (Guillén-Riquelme & Buela-Casal, 2011), so it would not be appropriate to use this value either. In this short version the reliability in state was adequate and adequate with some deficiencies in trait (criterion in Hernández, Ponsoda, Muñiz, Prieto, & Elosua, 2016), but we must consider that the low number of items can influence this result. In the same way, the correlations between each model total scores and total scores for complete STAI, may be influenced by number of items; even so in all models the results are upper a .85 and statistically significative.

Among the various structures analyzed, the second one that showed better fit to the sample was that of Van Knippenberg et al. (1990). This version was composed of six items per scale. By contrast, the version of 12 and 13 items proposed by Kaipper et al. (2010) showed the lowest fit indices. Thus, it is not necessary to have a large number of items to obtain a consistent structure (Marteau & Bekker, 1992; Van Knippenberg et al., 1990). In fact, it is possible to develop an adequate short form with four or five items per scale, as we did in the present study. When comparing the form developed by Van Knippenberg et al. (1990) to that proposed after the discriminant analysis, we observed very small differences of fit between both. Therefore, both versions seem to be appropriate to assess anxiety in Spanish samples. Other item combinations may also be adequate, as there seem to be several possible combinations with good fit. This explains the fact that we found a valid form of the Spanish adaptation (Perpiñá-Galvañ, Richart-Martínez, & Cabañero-Martínez, 2011; Perpiñá-Galvañ, Richart-Martínez, Cabañero-Martínez et al., 2011) based on the items of the short form proposed by Chlan et al. (2003) even though the latter was developed from form Y and the Spanish version was developed from form X. Thus, although the items did show an exact match between versions, we found a good fit based on such structure.

Despite the results obtained, this work presents several limitations. First, in the general population sample, the clinical history was not controlled to avoid including people with anxiety disorders. Furthermore, include other test or other types of anxiety measures had been useful to contribute validity evidences, and include other anxiety disorders and a large clinical sample. However, it is increasingly common to develop short forms (e.g. Kruyen, Emons, & Sijtsma, 2013) and cultural adaptations (Muñiz, Elosua, & Hambleton, 2013) of psychological assessment questionnaires with adequate reliability and validity guarantees. It is essential to determine the factor structure of personality questionnaires (e.g. Ginting, Näring, Van der Veld, Srisayektu, & Becker, 2013). In this context, proposing a short version of the STAI for a Spanish sample is an important contribution. A short version is very useful for assessments involving a large number of questionnaires, particularly in situations in which respondents cannot or must not undergo long assessments.