The aim of our study was to formally assess the factor structure of the Turkish IPQ-R and to rigorously test the measurement model proposed by Moss-Morris et al. (2002) [8
] in Turkish patients with diabetes and cardiovascular disease residing in Turkey. Aside from this study, only three other studies have examined the dimensional structure of non-English language versions of the IPQ-R by means of a hypothesis-testing framework. The original structure of the second part of the IPQ-R could not be confirmed for the population in our study. While the emotional representations and illness coherence factor showed satisfactory construct validity, several modifications had to be applied to the other factors to achieve a good model fit. A poor fit of the original measurement model was also observed in other CFA studies. This comprises an evaluation of the Chinese IPQ-R applied to hypertensive patients in Taiwan [17
], an evaluation of the Swedish IPQ-R applied to patients recovering from myocardial infarction [18
] as well as several CFA studies on the English version of the instrument applied to patients with different chronic diseases [11
], including a recent study on populations of African origin with type 2 diabetes [14
]. In these studies, several areas of ill-fit were identified and substantial changes to the measurement model such as the deletion of items and the respecification of indicators had to be applied to achieve acceptable model fit.
In our study, items purported to load on the personal control, treatment control and timeline cyclical factor had to be deleted from further analysis due to low factor loadings. Two of these four items (17 and 19) are negatively worded and, for the same reason, were also dropped from the analysis in the studies by Chen et al. [17
] on the Chinese and Cabassa et al. [19
] on a shortened 27-item Spanish version of the IPQ-R. Though retained in the model, item 17 also showed a low factor loading (standardized λ=0.25) in the Swedish validation study of the IPQ-R [18
]. Item 19 was also deleted by Abubakari et al. [14
] in their study of the IPQ-R in type 2 diabetes patients of African origin because of a low loading on its respective factor. The problems encountered with these two items may be due to a method bias. It is known from scale development research that negatively worded items may introduce method effects because they can be misunderstood by respondents [26
Similar to our study, item 20 was also considered problematic by Abubakari et al. [14
]. Based on consultations with subjects from their study population, the authors argue that this may be due to a mismatch between own beliefs of the patients (that are also shaped by supernatural causal attributions) and information they receive from others (e.g. doctors). This may be also true for Turkish patients who have been reported to have strong beliefs in fate and divine influence [27
The four items (17, 19, 20, 31) deleted in our study should be thoroughly scrutinized in further applications of the Turkish IPQ-R and their omission should be considered. Researchers should also thoroughly evaluate the reliability of the timeline cyclical and treatment control factor. Though still above the threshold that indicates acceptable reliability, the composite reliability of both factors was lower than for the other five factors of the model.
Modifications indices strongly suggested to add two error covariances between items 7 and 8 and between items 33 and 34, indicating that these item pairs shared measurement error. Similar to the reversed coding of items 17 and 19, the error covariance between items 7 and 8 may result from the negative wording used in item 8. The error covariance between items 33 and 34 most likely results from a very similar phrasing (see Table
) which in the Turkish version only differs by a weakly discriminating adjective (“Hastalığımı düşündüğüm zaman çökkün
oluyorum” vs. “Hastalığımı düşündüğüm zaman üzgün
oluyorum”). In addition, unlike üzgün
the word çökkün
is rather rarely used in the Turkish language which bears potential for additional bias.
We recommend an alternative wording for item 33 such as “Hastalığımı düşündüğüm zaman bunalıyorum” or “Hastalığımı düşündüğüm zaman karamsar oluyorum” (likewise literally meaning “When I think about my illness I get depressed“) that may be better able to discriminate and achieve equivalence with the original questionnaire.
In our study, item 6—originally purported to load on the consequence factor—was related to the timeline acute/chronic factor. Presumably this is because of the item’s wording used in the Turkish version (“Ciddi bir hastalığım var”). Although the adjective ciddi means “serious”, it is often associated with chronicity in the context of illness. The adjective ağır (severe, serious) may be better able to attain semantic equivalence with the original version and should be tested in future evaluations of the Turkish version.
The intercorrelations observed between the factors of the revised final model support the discriminant validity of the latent dimensions and are mostly in accordance with those reported in other studies on the IPQ-R [8
]. The largest correlations indicated that patients who perceive their treatment as effective also hold stronger beliefs about personal abilities to control their illness. Similarly, individuals viewing their illness as a condition with serious consequences also show a more intense emotional response to their illness. Both findings are in line with assumptions of the self-regulatory model [3
The present study has limitations. First, our results cannot be generalized to patients with diseases other than diabetes and cardiovascular disease. Second, the post hoc modifications of the model that we made guided by modification indices and theoretical considerations have to be considered an exploratory type of analysis [10
]. Thus, the changes we suggested are a starting point for further analysis that should aim to cross-validate and confirm the revised measurement model for other populations. Third, we are aware that a sample size of 300 is smaller than suggested by classical rules of thumb that are usually guided by the ratio of subjects to model parameters [25
]. However, these recommendations have to be reconsidered in light of newer simulation studies [21
]. They show that characteristics such as the number of indicators in the model, the number of indicators per latent variable and the average size of factor loadings are of greater importance to attain precise parameter estimates. Given the high factor loadings in the present study (on average factor loadings in the measurement model were ≥0.70) and a ratio of indicators per latent variable of at least 3:1, the estimation in the present study can be considered robust. In order to further support the precision of our results we conducted an additional Monte Carlo simulation using the model estimates as population values following suggestions by Brown [10
]. The results showed that with the given sample size and missing values proportion, the percentage of parameter and standard error bias, the degree of coverage and average values for goodness-of-fit indices were within acceptable ranges [10