Abstract
The Parental Bonding Instrument (PBI) was designed to measure parental styles as perceived by the respondents during their first 16 years [1]. The PBI is a Likert-like scale (ranging from 0 to 3) consisting of 25 items related to father and mother. These items are evaluated separately. This scale was originally devised to measure two factors: care (affection and warmth, vs coldness and rejection) and protection (control, intrusion vs encouragement of autonomy). Recently, a transcultural adaptation to the Brazilian Portuguese has been carried out [2].
The Egna Minnen Beträ ffande Uppfostran (EMBU) is another instrument that has been used to investigate parental attitude. It was developed by Perris et al. in Sweden [3]. Studies using this scale have found a four-factor structure: Emotional Warmth, Rejection, Overprotection and Favouring Subject.
The PBI is still considered to be the most consistent measure used to check parental style either in clinical or in non-clinical samples [4]. The instrument has proved to be stable over a 20 year period, and mood and life experiences have had a low impact on the stability in the perception of parental bond measured using the PBI [5]. In the last few decades the PBI has become very popular due to the fact that it is a scale of fast and easy administration, and its final score is readily calculated. It has been used in several studies that relate parental bond and psychology. Many authors, however, have found that the PBI measures three factors instead of two [6–9], as proposed by Parker et al. The second factor (protection) could be split into two other factors.
Two factors
During the PBI process of evolution Parker et al. found three factors in the structure of the instrument, different to the first versions of the PBI [1]. They decided, however, to use the two-factor structure, based on the statistical evidence that the variance calculated in the analysis of the third factor was relatively low and the items that would have a negative score on the second factor were likely to have a positive score on the third factor and vice versa, suggesting that these two factors could be developed into a single one. Another reason for the use of the two factors was that the contrast among the types of parental bond would be easily evaluated in a bidimensional model. Parker et al. concluded that the other factors had had the same source, that is to say, they had been originated from the two main variables; hence they suggested a two-factor model: care and protection.
Mackinnon et al. conducted a study on the Australian population, and they used confirmatory analysis to find the two-factor model proposed by Parker et al. [10]. Other studies have also demonstrated the two-factor structure, but those studies calculated the agreement among the coefficients for both factors in different samples [11, 12]. If a coefficient for each of the components of the factors in the model is calculated, however, the role of all factors as a whole in the model is ignored.
Three factors
In a study performed with 2147 Australian adolescents involving the PBI, Cubis et al. suggested that the three-factor model was the most adequate for their data [6]. Gomez-Beneyto et al. identified a three-factor model in the Spanish version of the PBI, which was administered to 205 women [8]. In both studies the first factor proposed by Parker et al. as care, was equally identified. As for the second factor (protection), the data were subdivided into two other factors: one was specifically related to the denial of the child's psychological autonomy; the other one was related to the discouragement of behavioral freedom. These factors were differently named in each of the studies (Table 1). In the model proposed by Gomez-Beneyto et al. three questions were excluded from the scale (items 10, 16, 23) because they did not reach the minimal score in any of the three factors proposed.
Factor models of the PBI
PBI, Parental Bonding Instrument.
(n), PBI item number.
Murphy et al. carried out a study involving three different samples: 583 US college students of psychology, 117 British high school students, and 119 British medical students [9]. The authors found better adequacy of the PBI using a three-factor model. Murphy et al. suggested that this model makes a more evident difference among the groups than a two-factor model. The data analysed using such a method would be more accurate and predictable. In the model proposed by Murphy et al., item 10 was excluded from the scale because it did not achieve the minimal score in any of the three factors (Table 1). All these three-factor models demonstrated differences regarding gender and age of the respondent, which had not occurred in the two-factor model proposed by Parker et al. [1].
Kendler proposed a three-factor model of the PBI, which consisted of 16 items for epidemiological research (Table 1) [7]. The author excluded some questions from the instrument for two main reasons: first, because he did not consider some of them appropriate for the sample under study; second, some questions (9, 20, 22, 24, and 25) did not present the same classification proposed by the three-factor model [5]. Sato et al. administered the PBI to 418 Japanese workers, and compared their data to the five models proposed (Parker et al., Cubis et al., Gomez-Beneyto et al., Murphy et al. and Kendler), obtaining a better performance of the confirmatory analysis in the Murphy and Kendler models [13]. The Kendler model, however, was better because it presented relative invariance in the subgroups regarding gender and age. Recently, Cox et al. conducted a research study for the National Comorbidity Survey in a psychiatric sample, and they also found that, among the five models proposed, the one that best adapted to the data analysis was the model proposed by Kendler [14]. This model seems to be relatively invariant regarding gender and age. Taking into consideration the importance of this questionnaire, which has been widely used in many countries and has also been recently administered in Brazil after being transculturally adapted, the aim of the present study was to perform likelihood confirmatory factor analysis to compare the five different factor structures of the PBI in terms of model fit in a female Brazilian population.
Method
Sample
The sample consisted of 257 women (average age = 32.47 years old, SD = 15.64). One hundred and thirty-nine of them were outpatients referred for psychiatric treatment at the university hospital located in Porto Alegre (Hospital de Clínicas de Porto Alegre; HCPA), Rio Grande do Sul, Brazil, and 118 were controls selected from the HCPA employees and the medical or psychology students at Universidade Federal do Rio Grande do Sul.
Diagnoses were made according to the DSM-IV-TR criteria [15] by a psychiatrist who was under the supervision of an experienced psychiatrist. Regarding the controls, the absence of a diagnosis was investigated using the Brazilian version of the Mini International Neuropsychiatric Interview [16]. The subjects completed the Brazilian Version of the PBI [2]. From among 257 subjects, 250 completed the PBI for mother and 241 completed the PBI for father. All patients gave informed written consent so that the information they provided could be used for research purposes. The present study was approved by the local Ethics Committee of the HCPA, according to the International and National Guidelines and Norms, specifically Resolution 196/96 and the complementary resolutions of the National Health Council.
Analysis
The whole confirmatory analysis was conducted using Amos 5.0.1 (Amos, Chicago, IL, USA). Data on paternal and maternal PBI were analysed separately throughout the present study. The five different factor models (Table 1) were compared to each other in terms of model fit. For each model tested, the following criteria were used for evaluation: χ2 goodness-of-fit index (GFI) [17], adjusted goodness-of-fit index (AGFI) [17], and Tucker–Lewis index (TLI) [18, 19]. The following criteria were used to indicate the goodness of fit for a particular model [19–21]: GFI >0.85, AGFI >0.80, TLI >0.90. Non-significance of the χ2 statistical test indicates a good fit. This statistical test, however, is sensitive to the sample size; it should, therefore, be used with caution, because good-fitting models can fail to reach non-significant levels, especially if the sample size is large (n > 200) [21]. The TLI is believed to be relatively independent of the sample size [21]. The invariance of the five competing factor models was investigated across younger and older subgroups. In the analysis, the hypothesis that only a similar factor structure is invariant across the subgroups tested was raised. This procedure was used in the Sato et al. study of Japanese non-clinical adults [13]. Additionally, the invariance between the four three-factor competing models was investigated across control and cases subgroups.
Kendler eliminated items that did not fit his model well. An additional analysis of all models using the Kendler item subset was therefore performed.
Results
The breakdown of primary diagnosis was mood disorders (22.5%), of which 17.1% was major depressive disorders and 5.4% was bipolar disorder, followed by eating disorders (8.9%), anxiety disorders (6.6%), and major depressive disorders comorbid with anxiety (5.8%). Personality disorder comprised 10.1% (all borderline cases), and 45.91% of the sample had no axis I and II diagnoses (controls).
Table 2 presents the results of the confirmatory factor analysis for each of the five models. The Parker et al. two-factor model produced the smallest GFI and AGFI regarding both paternal and maternal PBI, compared with the other four three-factor models.
Fits provided by the competing factor models
AGFI, adjusted goodness-of-fit index; GFI, goodness-of-fit index; PBI, Parental Bonding Instrument; TLI, Tucker–Lewis index.; ∗χ2 statistics for model A were significantly larger than those for the other four models.; χ2 statistics for model E were significantly smaller than those for the other four models (p < 0.001).
The AGFI was 0.692 with respect to the paternal PBI and 0.729 regarding the maternal PBI, indicating insufficient fit. The χ2 values for the Parker et al. model were significantly larger (p < 0.001) than those for the four three-factor models. This indicated that the Parker et al. original solution provided a significantly poorer fit to the data than the three-factor solutions. Among the three-factor models, the Kendler model produced the best fit to the data. The χ2 values for this model were significantly smaller (p < 0.001) than those for the three-factor models. The GFI for the Kendler model was higher than 0.85 concerning both paternal and maternal PBI, indicating a good fit to the model. The second-best fit was provided by the Murphy et al. model, although the GFI was lower than 0.85 (GFI father = 0.803; GFI mother = 0.805). All the other GFIs for the two three-factor models were also lower than 0.85.
When subdivided into younger and older subgroups, 130 respondents were younger than 26 years old, and the remaining 127 were ≥26 years old. Table 3 shows to what degree the competing factor models were invariant across younger and older subgroups. The fits in Table 3 were generally poorer than those in Table 2, indicating that the factor structure provided by the five models was not strongly common across the subgroups. Similarly to the results in Table 2, the poorest fit was obtained by the Parker et al. model, and the best fit was obtained by the Kendler model, which produced an AGFI >0.8 and a GFI >0.85 in regard to both paternal and maternal PBI and while testing the invariance across age subgroups. None of the other factor models produced AGFIs >0.8, or GFI >0.85, and an invariance hypothesis of factor structure was supported only by the Kendler model. The Murphy et al. model, however, also supported an invariance hypothesis of factor structure (Table 4) across the case and control subgroups.
Invariance of the competing models across younger and older subgroups
AGFI, adjusted goodness-of-fit index; GFI, goodness-of-fit index; PBI, Parental Bonding Instrument; TLI, Tucker–Lewis index.
∗χ2statistics for model A were significantly larger than those for the other four models (p < 0.001).
Invariance of the competing models across case-control subgroups
AGFI, adjusted goodness-of-fit index; GFI, goodness-of-fit index; PBI, Parental Bonding Instrument; TLI, Tucker–Lewis index.
†model did not converge for the calculation of values.
Using the Kendler item subset, all models performed better than the original subset (Table 5).
Fit provided by 16-item models
AGFI, adjusted goodness-of-fit index; GFI, goodness-of-fit index; PBI, Parental Bonding Instrument; TLI, Tucker–Lewis index.
Discussion
The present study was carried out using a female sample, but the results support other studies that have emphasized that the original protection factor should be divided into two factors [6–9, 14, 22]. The present results indicate that the three-factor models generally fit the data better than the two-factor model. All three-factor models proposed, however, do not seem to provide similar good fit. Among the three-factor models, only the Kendler model (1996) fitted the data. In the Sato et al. study either the Kendler model or the Murphy et al. model fitted the data [13]. In the present study, however, when considering invariance of the factor structure across gender subgroups, the Kendler model was the only one acceptable. Sato et al. suggested that the Kendler model may be the only acceptable factor structure when considering invariance across subgroups [13]. Murphy et al. implied that an appropriate factor structure of the PBI may differ between subgroups [9]. They suggested that some modifications to the original PBI might be needed to obtain an invariant factor structure across subgroups. Therefore, the present results, which were obtained from a female sample, are in agreement with the results found by Sato et al. [13] and Murphy et al. [9].
One important question raised by Sato et al. [13] was whether the best fit of the Kendler model had to some extent a connection with the smaller item set. When comparing the different competing models using only the Kendler item subset, all models performed better than the original subset. This suggests the possibility of excluding those items proposed by Kendler, to enhance the performance of other models. This should be addressed in further studies.
In addition, the validity of a factor structure should not be discussed only in terms of statistical consistency; it also needs to be underpinned by the fact that the three-factor model structure of the PBI has a strong ability to predict outcome variables such as the occurrence of mental disorders and the prognosis of mental disorders [13].
Further research is needed to obtain confirmatory factor analysis of the PBI in both female and male Brazilian samples. Despite the gender limitation, the present study contributes to a better understanding and use of the PBI in Brazilian samples.