Inglês (pdf)
Artigo em XML
Referências do artigo
Tradução automática
Enviar este artigo por email
Citado por SciELO 
Similares em
SciELO 
Permalink
Introduction
The Patient Health Questionnaire 9 (PHQ-9) is a brief, easy-to-administer, and interprets a major depressive episode (MDE) screening instrument 1. Because of its brevity and good reliability and validity 1,2, the PHQ-9 is one of the most widely used MDE screening instruments in clinical and non-clinical settings 3,4. Two recent systematic reviews have concluded that the PHQ-9 is the most reliable and accurate tool for MDE screening, so it is enjoying increasing acceptance in the scientific community every day 5-8. Overall, the PHQ-9 has demonstrated adequate psychometric properties in both general and medical populations 1,9,10, including university populations 11-16.
The psychometric performance of the PHQ-9 has generally been evaluated using the classical test theory (CTT) framework, which focuses on the concepts of true score and measurement error 17. From the CTT perspective, the comparability of test scores has the limitation that scores are "test dependent" 18, implying that a low score may be due to a person's low ability or a difficult test 19,20. Based on an analysis of the articles included in the systematic reviews on the PHQ-9 published to date, approximately 800 articles evaluated the psychometric properties of the PHQ-9 using the CTT model 4,5,8,21-29. CTT-based techniques could provide inaccurate diagnoses and do not consider the heterogeneity in each specific item 30,31. Currently, these CCT-based methods are being complemented and, in some cases, replaced by item response theory (IRT) approaches, considered by many researchers to be the "measurement paradigm of the 21st century" 19,32-35. IRT-based models offer several advantages over CTT and are considered the most appropriate and robust methods for assessing the psychometric properties of screening scales such as the PHQ-9 36,37. The idea of IRT is that test-response behavior (i.e., solving an item or choosing a specific category) can be explained by underlying person parameters (latent ability or trait) and item parameters (difficulty) 34.
The relationship between the latent trait and probability of response to a given item can be expressed by different estimation models for both dichotomous and polytomous items 32,38. This approach estimates item and person parameters from the obtained data and is considered independent of the sample and test 39. In addition, IRT-based models provide a richer description of the performance of each item, greater detail on the accuracy of a measure, and when assumptions are met, scores are item-independent and invariant across different samples 33,34,40,41. Consequently, the use of IRT models may increase the validity and utility of depression screening when the PHQ-9 is used in culturally diverse settings 42,43. To date, only 30 published studies have used IRT models to assess the psychometric properties of the PHQ-9 and only one in college students 37,42-68.
Given the high prevalence of depression among college students 69,70 it is important to evaluate the validity and utility of the PHQ-9 in this population. The appropriate detection and management of depression in health science students could translate into significant potential benefits if the psychometric properties of screening instruments such as the PHQ-9 are optimized 71,72. Therefore, this study aimed to evaluate the psychometrics of the Colombian version of the PHQ-9 adapted for university students using IRT models and to assess possible item biases as a function of gender.
Methodology
Population and sample
This cross-sectional observational study evaluated the psychometric performance of the PHQ-9. Stratified random sampling was applied according to academic program, semester, and sex, where each stratum had several subjects proportional to its size. The participants were health science students from the University of Cartagena (Colombia). A non-probabilistic sample size of 550 participants was considered good for confirmatory factor analysis 73. Of these 550 subjects, 9 were excluded because of a significant number of failed responses.
Procedures and instruments
All the participants were informed of the research objectives and signed an informed consent form. This project was approved by the Ethics Committee of a university in Colombia. A sociodemographic characteristics instrument was applied to all participants, and all of them completed the PHQ-9 1. The PHQ-9 is a screening scale that measures the presence and severity of depressive symptoms 1 and consists of nine symptoms of the DSM-IV MDE criterion A 74. These nine items are arranged in the form of an adjectival-type scale assessing the presence of the symptom in the last two weeks ("not at all,” "several days,” "more than half of the days" and "almost every day"), which are scored from 0 to 3, reaching a score between 0 and 27 75. It can be self- or hetero-administered and is used either algorithmically to make a probable diagnosis of an MDE or as a continuous measure of scores ranging from 0 to 27 and cut-off points (CP) of 5, 10, 15, and 20, representing levels of depressive symptoms as mild, moderate, moderately severe, and severe 1. These scores can also be used dichotomously from a PC to classify subjects with or without clinically significant depressive symptoms (CSDS) 76.
The psychometric characteristics of the PHQ-9, according to Kroenke et al., present a sensitivity of 88% and a specificity of 88%, an adequate internal consistency (Cronbach's α of 0.86-0.89), a test-retest score of 0.84, a concordance between the self-administered test and the one performed by the evaluator of 84%, and an area under the curve (AUC) of 0.95 1. In this study, a version adapted to this population was used 77. Preliminary analyses showed that the PHQ-9 in Colombian university students had adequate internal consistency with a Cronbach's alpha of 0.83 and McDonald's omega of 0.89 76; a two-factor model with adequate fit indicators: CFI, 0.98; NFI, 0.96; NNFI (TLI), 0.97; RMSEA, 0.045; and a metric, structural, and residual invariance by sex with a ΔCFI of 0.002, 0.001, and 0.003 (p = 0.67, 0.27, and 0.78, respectively) 11.
In Colombia, the criterion validity of the PHQ-9 in primary care has also been evaluated compared with the mini international neuropsychiatric interview (MINI), which found that for a PC of 7 or more, the PHQ-9 showed the following indicators: AUC of 0.92 (95% CI, 0.880-0.963), sensitivity of 90.38 (95% CI: 81.41-99.36); specificity of 81.68 (95% CI: 75.93-87.42) 78.
Statistical Analysis: The Rasch analysis is based on a mathematical model in which the probability of passing an item is a logistic function of the difference between the person's level of depression and the level of depression expressed by the item (item difficulty) 53,79,80. For the Rasch analysis, the following psychometric characteristics were verified.
Differential item functioning: For the analysis of differential item functioning, the following criterion was considered: the value of the difference between males and females was greater than 0.5, and the probability of the Mantel statistic (for polytomous tests) was less than 0.05 40; thus, items whose score was due to gender and not to the construct itself could be identified.
Dimensionality and local independence: Regarding the assumptions of dimensionality and local independence, the principal component analysis allowed us to establish how much of the variability in the responses was explained by the instrument 41. Here, we used the indicator of the number of eigenvalues of the non-exponential variance 41. Here, we used the indicator of the number of eigenvalues of the unexplained variance in the first contrast, which must be less than two 58,71,81. Local dependence occurs when the response to one item depends on the response to another item 67. Residual correlations above 0.3 are indicative of local dependence 82.
General fit: The general fit was studied from the joint analysis of several measures: in the first instance, the close fit (infit) and far fit (outfit). For this type of analysis, the items are expected to have a measure of fit close to one. A result between 0.5 and 1.5 indicates a measure with an appropriate balance between information and noise 83. Difficulty, measured in logits, refers to the probability that a person will respond to an item indicating a sign of depression 34. Concerning the discrimination index, each item is expected to have a positive, nonzero value of at least 0.3 and not much above one 33. Correlation refers to the association between an item and a scale. Positive correlations were expected with values of at least 0.2. Rasch’s reliability is analogous to the internal consistency indicators in the CTT. Similarly, the person separation index (PSI) was estimated, establishing acceptable values above 0.7 52, although this is relative to the sample size and the number of items. These provisions jointly show the adequacy of the data to a model that allows a consistent measurement of the construct and are the recommendations for the use of the Rasch model to establish the psychometric characteristics of instruments 81,84,85. Winsteps® version 3.80.1 was used for this analysis.
Results
Differential Item Functioning
First, a differential behavioral analysis of the items was performed. Item 2, "felt discouraged, sad, irritable or hopeless" showed a tendency toward differential behavior. This difference is close to the established limit and is because female participants tended to report experiencing the condition referred to in the item more (mean of 0.9 and for males 0.6) (Table 1).
Dimensionality and Local Independence
The total variance explained by these measures was 46.8%. The residual variance in the first contrast was 1.7, which is below the undesirable level of 2.0 81. Similarly, the percentage of variance explained by the items was 22.8%, which was higher than the variance not explained by the first contrast. The residual standardized correlations had a negative sign, which is compatible with the evidence of the local independence of the items.
General Fit
All items fell within the range in which they are considered productive (between 0.5 and 1.5 in the infit - outfit measures). The correlations between item scores and the test were compatible with good test performance. There were no negative or very low correlations (Table 2).
Person separation and reliability were 1.75 and 0.75. Overall, these data indicate that the instrument allows for discrimination between the two levels. Person separation and reliability were 1.75 and 0.75. Taken together, these data indicate that the instrument allows for discrimination between the two levels. The discrimination index of the items is good, but not negative in all cases (above 0.30), although there is a tendency for most of the items to over-discriminate persons at high and low risk of depression, although in no case, the dataset indicates a possible unproductive item. The analysis of person-item fit revealed a mismatch between the degree of depression measured by the PHQ-9 in this group and the participants' responses (Figure 1).
Discussion
Evidence of one-dimensionality, local independence, adequate overall fit for the Rasch model, and moderate reliability was found for the Colombian Spanish version of the PHQ-9 among university students. These findings provide complementary information about its psychometric characteristics that reinforce the arguments in favor of its validity, but at the same time identify areas that require further research and, eventually, the development of better tests.
First, it is important to analyze the results of differential item functioning. The lack of differential item functioning (DIF) along with the Rasch model results supports the validity of an instrument across diverse cultural contexts with different characteristics when assessing depressive symptoms 40,63,68,85. DIF analysis showed that PHQ-9 responses were sex invariant in this sample of college students. These results are comparable to those reported for a sample of university students in Japan 65. Although our work item 2 of the PHQ-9 showed some tendency toward differential behavior, a previous study in Colombia from the perspective of CTT did not find invariance by gender 11. In addition, it should be considered that in this analysis, a more restrictive criterion was used (0.5) than in other studies on the PHQ-9, such as that of Jiraniramai 37 in which the decision was taken from 0.64 logists. Differential functioning findings in other research are mixed: in Thailand, in a population of healthcare workers, slight differential functioning was found for item 6 (feeling bad about oneself) according to the sex of the participants 37; in research in India with people with visual impairment 54, it was found that there was differential functioning of item 3, referring to problems with sleep when dividing the groups according to the duration of their disability; in Germany, in a sample of older adults, differential functioning of items 3 and 6 was found 53. In summary, the differential behavior attributable to the PHQ-9 items depends on the specificities of the populations studied.
When analyzing the dimensionality of the construct in our sample, it is important to consider that the variance explained by the measures was below 50% and that the residual variance was 1.7, not quite 2, but not so far from that number. This evidence could prompt further study on whether the set of items measured the same construct. This interpretation has a place given the flexibility of this type of analysis, which does not focus on one indicator but invites consideration of the data within a larger set of information. Additionally, it is necessary to consider that all psychometric analyses should be analyzed concerning contextual information: in this case, it is known that, although the test works well as a whole, there is also a somatic and a non-somatic dimension within it 11 and although one could delve from the above into the question of whether the instrument serves to assess the degree of depression or not, the reality is that the construct "depression" itself requires including behavioral, somatic and cognitive issues of a different nature, but which are expressions of that same entity 86,87. This statement is consistent with data from theory and practice, suggesting that depressive syndrome is a multidimensional construct 88.
The values of the standardized residual correlations in our study indicated that the PHQ-9 items showed local independence. This finding is compatible with those of studies on pregnant women in Peru 68 and the rural population in Brazil 45.
The reliability and separation indices were low in this study, although the reliability of individuals was within an acceptable limit above 0.7 for this type of analysis 41,89. According to Linacre 83, this indicates that the test can discriminate between two levels of people, which, in principle, would be sufficient. However, this indicates that the test could be improved to better discriminate between different levels of a major depressive episode. In this sense, the analysis of the location of the persons and the items along the measure (logist) shows that the test measures the depressive episode at a high level of occurrence for a non-clinical population, with item 9 "Have you thought you would be better off dead or have you thought about hurting yourself or hurting yourself in some way" corresponding to the highest level of the construct and the items referring to the feeling of tiredness and loss of appetite, those that measure a lower level. In this section, several readings are possible. On the one hand, the test does not appropriately measure the construct in people with low levels of depression. This is important because a mild form of depression precedes a moderate form and can develop into a severe form when an individual's coping strategies fail. Screening tests are useful to the extent that they detect a mild form of depression, but it is believed that the PHQ-9 did not capture this well enough in this population, as has been observed in other non-clinical populations 37.
However, the fact that the PHQ-9 is better able to identify more severe levels of the construct probably does not affect whether individuals with significant levels of depression are adequately detected and subsequently diagnosed by the professional in charge. Findings related to items measuring feelings of tiredness and changes in appetite could be explained by the nosological relationship between anxiety symptoms and depressive symptoms 90,91; for example, some somatic symptoms of depressive disorder, such as easy tiredness and fatigue, are part of the diagnostic criteria of other disorders, such as generalized anxiety disorder 92. A gap between death ideation and movement impairment, as measured by logist distance, was also evident. This observation raises the possibility that some intermediate psychopathological manifestations are missing between death ideation and movement disturbance, which could be interpreted as a failure of the test to measure some aspects of a construct that we accept exists and that is accepted by the scientific community.
This raises questions about further refinement of the PHQ-9: is there an aspect of the depressive syndrome on this continuum that perhaps the test is not measuring, and can the PHQ-9 uniformly measure all persons who would meet the criteria for a depressive episode? In terms of reliability, inter-person separation was not excellent, although it was acceptable (0.74), while Cronbach's alpha was 0.83. This difference could be explained by the fact that Pearson's separation index is lower than Cronbach's alpha because the reliability in this model is based on a linear scale at the interval level when a good fit between the model and the data is observed, whereas alpha is based solely on the assumption of linear measures 93.
To our knowledge, this is the first study to report the validity of the PHQ-9 using Rasch analysis on a sample of Spanish-speaking university students. The large sample size results in high statistical power, which could technically increase the probability of model misfit or DIF detection. The results of this study should be analyzed considering several limitations. First, we did not analyze the concurrent validity of the PHQ-9 with other screening tests for depression or using a reference criterion. Second, we could not effectively control for response bias. Because the respondents were health science students, it is possible that some underestimated their actual symptoms for fear of stigmatization. Third, we did not assess the test-retest reliability, which can be useful when repeated measures of the instrument are required. Finally, it is worth noting that the Rasch model is a one-parameter IRT model that assumes that discrimination is the same for all items 36,57,94. Under this assumption, easy-to-endorse items (e.g., Item 4, low energy) discriminate as well as difficult-to-endorse items (e.g., Item 9, suicidal ideation) for subjects with a lower level of depressive symptoms 95, which is not always true.
Based on the limitations of this study, future research should verify the construct validity of the PHQ-9 by contrasting the results of the application of the test through a diagnosis of depressive disorder using structured psychiatric interviews. Further research is needed to verify whether invariance is maintained in the successive measurements of the instrument. This is important because there is work proposing the use of the PHQ-9 to assess the severity of a depressive episode and to monitor response to treatment 96. To further advance construct validity, future studies using two- or 3-parameter IRT models are warranted to validate the PHQ-9 in university populations.
Considering the above results, it can be concluded that the Rasch model analysis of instruments such as the PHQ-9 offers a detailed perspective on psychometric behavior, with evidence of its strengths and weaknesses. A way for the improvement of the instrument could be an analysis of the construct and the construction of items that represent a more continuous measure of depression in its different levels of severity.
