ARTICLE IN PRESS
Social Science & Medicine 65 (2007) 2357–2370 www.elsevier.com/locate/socscimed
Sequence analysis in multilevel models. A study on different sources of patient cues in medical consultations Lidia Del Piccoloa,, Maria Angela Mazzia, Graham Dunnb, Marco Sandric, Christa Zimmermanna a
Department of Medicine and Public Health, University of Verona, Italy b Medical School, University of Manchester, UK c Department of Economic Sciences, University of Verona, Italy Available online 14 September 2007
Abstract The aims of the study were to explore the importance of macro (patient, physician, consultation) and micro (doctor–patient speech sequences) variables in promoting patient cues (unsolicited new information or expressions of feelings), and to describe the methodological implications related to the study of speech sequences. Patient characteristics, a consultation index of partnership and doctor–patient speech sequences were recorded for 246 primary care consultations in six primary care surgeries in Verona, Italy. Homogeneity and stationarity conditions of speech sequences allowed the creation of a hierarchy of multilevel logit models including micro and macro level variables, with the presence/absence of cues as the dependent variable. We found that emotional distress of the patient increased cues and that cues appeared among other patient expressions and were preceded by physicians’ facilitations and handling of emotion. Partnership, in terms of open-ended inquiry, active listening skills and handling of emotion by the physician and active participation by the patient throughout the consultation, reduced cue frequency. r 2007 Elsevier Ltd. All rights reserved. Keywords: Sequence analysis; Cue; Medical consultation; Physician-patient communication; Verona Medical Interview Classification System (VR-MICS); Italy
Introduction The accuracy and completeness of patient data collected by the physician during the medical consultation, as well as a collaborative relationship with the patient are essential for the outcome of Corresponding author.
E-mail addresses:
[email protected] (L. Del Piccolo),
[email protected] (M.A. Mazzi), graham.dunn@ man.ac.uk (G. Dunn),
[email protected] (M. Sandri),
[email protected] (C. Zimmermann). 0277-9536/$ - see front matter r 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.socscimed.2007.07.010
patient care. They are increased when patients feel free to express topics of immediate importance, and when physicians pay attention to what the patient wants or needs to convey (Stewart et al., 1995). Patients may use clear expressions (‘‘I am worried about this constant pain’’) or subtle hints (‘‘y all this pain, you know’’) that demand an exploration from the doctor. These initiatives of patients have been defined as ‘‘cues’’ or ‘‘clues’’ (Butow, Brown, Cogar, Tattersall, & Dunn, 2002; Fallowfield, Jenkins, Farewell, & Solis-Trapala, 2003; Levinson, GorawaraBhat, & Lamb, 2000). The term ‘‘concern’’, instead,
ARTICLE IN PRESS 2358
L. Del Piccolo et al. / Social Science & Medicine 65 (2007) 2357–2370
tends to be reserved for more circumscribed patient expressions indicating, as in the RIAS definition (Roter, 1993), unpleasant emotions and stressful issues of current concern, regardless of whether solicited by doctor or due to patient initiative. There is a general agreement that the occurrence of cues, intended as unsolicited, spontaneous offers of new information or signals to issues that patients would want the physician to pursue further, are important elements of the medical consultation. Accordingly, research efforts have been focused on the conditions facilitating or impeding the occurrence of cues, given their implications for clinical training and practice. A number of variables have been shown to be linked to cue occurrence such as patient, physician and general consultation characteristics as well as specific patient or doctor expressions during the consultation (Zimmermann, Del Piccolo, & Finset, 2007). Patient variables, associated with an increase of cues, were identified as current emotional distress and the presence of adverse psychosocial life events and conditions (Davenport, Goldberg, & Millar, 1987; Del Piccolo, Saltini, Zimmermann, & Dunn, 2000), as well as female gender and younger age (Butow et al., 2002). Physician expressions associated with an increase were directive psychological questions, clarifications or ‘‘screening’’ questions on psychological issues (Del Piccolo et al., 2000; Goldberg, Jenkins, Millar, & Faragher, 1993), questions based on what the patient had said before, and empathic comments (Goldberg et al., 1993). In contrast, a decrease of cues was associated with an increase of doctor-centred questions or questions on physical issues (Goldberg et al., 1993), but also of patient-centred techniques of active listening such as checking, summarizing, asking for opinion or understanding (Del Piccolo et al., 2000). Physicians with a recognized ability to identify patients with emotional distress obtained more verbal and non-verbal cues per consultation than physicians known to be poor detectors of distress (Davenport et al., 1987). The General Practitioner’s ability in using communication skills appropriately emerged as another physician characteristic influencing cue emission (Goldberg et al., 1993): the same set of communication techniques increased cues, when adopted by good detectors of emotional distress, but decreased cues when applied by physicians known as poor detectors. The authors explained the finding using post hoc qualitative analyses of the consultations that demonstrated the
incoherent use of these communication techniques by the poor detectors. These last findings provide evidence of the limitations of using frequency counts of physicians’ behaviours during consultations. Wasserman and Inui (1983) described this approach to the analysis of the medical consultation as ‘‘trying to understand the game of tennis by merely counting the number of serves, slams, lobs and volleys’’, a critique which emphasizes the need to consider the sequential dynamic of doctor– patient interaction patterns. Successive developments in the analysis of sequential dyadic interactions (Bakeman & Gottman, 1997; Gottman & Roy, 1990) have allowed researchers to apply a statistical approach to doctor–patient interaction sequences such as the probabilistic dependencies between cues and preceding physician or patient talk. While the available findings summarized above suggest that different inter-dependent patient and physician variables may all contribute to the appearance of cues during the consultation, a better understanding of the weights of the contribution of different variables would enable improved targeting of research and training in doctor–patient communication. However, the introduction of sequential analysis presents challenging methodological problems (Mazzi, Del Piccolo & Zimmermann, 2003). The consideration of physician or patient behaviour immediately preceding the target behaviour ‘‘cue’’, together with other higher order variables such as patient, physician or general consultation characteristics (which in turn might contribute to modify cue offers) calls for a thoughtful use of multilevel modelling procedures. The aim of the present study is twofold: 1) to determine the relative importance of micro-variables (physician and patient speech preceding cues) and macro-variables (physician, patient and general consultation characteristics) in affecting the occurrence of cues and (2) to illustrate the methodological strategies adopted to counter problems that arise when different data sources are included in the same model. For this purpose we used an available data set of primary care consultations (Del Piccolo et al., 2000). Cue as the dependent variable of this study was defined according to the Verona Medical Interview Classification System (VR-MICS) (Del Piccolo et al., 1999, 2005) as ‘‘any expression introduced spontaneously by the patient in order to draw the physician’s attention to issues not yet discussed or
ARTICLE IN PRESS L. Del Piccolo et al. / Social Science & Medicine 65 (2007) 2357–2370
sufficiently dealt with’’. This is the case, for example, when the patient introduces topic changes, returns to a previous topic or uses metaphors or emphasizing expressions. This definition of cue excluded all patient expressions solicited by the physician responding to the asking–giving information dynamic of a medical consultation. Method Participants Patients were recruited over a two-month period in 1996 from six primary care surgeries in Verona, Italy. All consecutive patients between the ages of 16 and 74, either new or attending for a new illness episode were eligible. Participants were informed about the purpose of the study and gave written consent to be recorded. Refusal rate was 2.7%. From the original sample of 252 patients, six patients were excluded because their audiotaped consultations were incomplete, defective or referred to a third person. Patients (69% females) had a mean age of 45 years (SD 14.0) and a mean GHQ-12 score of 3.31 (SD 3.03); 46% were considered by their general practitioner (GP) to be without emotional distress, 23% to have subclinical psychological problems, 21% to have mild problems and the remaining 10% to have moderately or severe psychological distress. Twelve percent of the patients were rated by their GP as being without medical symptoms, 21% as having undefined medical symptoms, and the remaining 67% as having mild or moderately severe (46%) or severe illness (2%). Six male GPs with a mean age of 46 years (SD 6) and a mean number of years of experience in primary care of 17 years (SD 6) took part. The average number of patients on their books was 1512 (range 1300–1800). They were recruited from surgeries close to the University Hospital of Verona, and had been the first to agree to collaborate. None of them had received any training in patient-centred interviewing techniques. They did not differ in age, experience and number of patients from the population of male GPs practicing in Verona (Pini, Piccinelli, & Zimmermann-Tansella, 1995).
another set of 22 categories for patient speech (VRMICS/P). A description of the system is given elsewhere (Del Piccolo, Mazzi, Saltini, & Zimmermann, 2002; Del Piccolo et al., 2005). Coding is based on verbatim transcripts of the consultation, which are divided into speech units. Inter-rater reliability proved satisfactory (Saltini, Cappellari, Cellerino, Del Piccolo, & Zimmermann 1998; Del Piccolo et al., 1999). The Cohen’s kappa for the total number of physician and patient codes was 0.93 and 0.85, respectively. Construct validity (Del Piccolo, Putnam, Mazzi, & Zimmermann, 2004) was shown by means of factor analysis. For the present study, due to statistical frequency requirements, related categories were lumped into sets: six for physician and five for patient speech. The grouping was based on theoretical criteria (what is considered to be doctor and patient centred behaviour) (Epstein et al., 2005; Smith, 2002) and functional aspects of the medical consultation: Patient (Pt) speech sets
Measures The Verona Medical Interview Classification System (VR-MICS) The VR-MICS comprises 22 mutually exclusive coding categories for physician (VR-MICS/D) and
2359
Listening, agreement/disagreement (negative and positive talk, e.g. ‘‘Don’t say that’’; ‘‘Yes’’, ‘‘I see’’, ‘‘No’’). Statement: Expression solicited by physician’s question or coherently connected with the ongoing discourse topic. Statement includes here all possible content areas considered by the VRMICS (psychological, social, medical, lifestyle, personal events, impact on function and other). Expectation/opinion and question. Cue: Unsolicited new information, topic changes, emphasis, metaphors, repetitions, profanities, unusual words, comprising here, as for statements, all possible content areas. Conversation. General practitioner (GP) speech sets: Supporting, detecting and handling emotion (appraisal; reassurance, agreement, open-ended question on psychological issues). Open question and active listening (open-ended non-directive and open-ended questions on social, medical, lifestyle, personal events, impact on function and other issues, asking for understanding, asking for opinion, checking). Passive listening (facilitation, e.g. ‘‘Hmm’’, ‘‘Yeah’’ or echoing, asking to repeat). Closed question (closed questions on all content issues).
ARTICLE IN PRESS 2360
L. Del Piccolo et al. / Social Science & Medicine 65 (2007) 2357–2370
Information/instruction (transitions, giving information, giving instruction, brief answers). Conversation.
The Partnership Index (PI). The consultation may be characterized by desirable physician interventions throughout the consultation, not necessarily preceding cues, as well as other desirable patient behaviour, in addition to cues. For the physician these are open inquiry, active listening expressions, and emotion handling statements, and for the patient all expressions indicating active participation. Such expressions suggest partnership and are considered, among others, to be essential ingredients of a patientcentred consultation (Epstein et al., 2005; Zandbelt, Smets, Oort & de Haes, 2005). These expressions were summarized in an ad hoc index of Partnership (PI), expressed by the sum of the respective physician and patient behaviours (excluding cues) divided by the total sum of speech units per consultation. A similar index has been proposed by Ford and Hall (2004), while the physician skills comprised in PI parallel those considered by some of the categories of the Patient-centred Behaviour Coding Instrument (PBCI) by Zandbelt et al. (2005). PI ¼
ðsupporting=handling þ open=listeningÞ þ ðexp=opin:=quest:Þ ; total
where supporting/handling emotions corresponds to the total frequency of physician expressions of appraisal, agreement, open-ended questions on psychological issues; open questions/active listening is the sum of all open-ended directive and nondirective questions, the questions asking for patient’s understanding and opinion, all checks, summarizing expressions and demands for clarification; exp/opin/quest. refers to all patient’s questions and expressions related to personal opinions and expectations on medical and psychosocial topics. The total is obtained summing up the number of GP and patient coded specch units per consultation. PI can vary between zero (no behaviour of those included in PI was present) and one (the only verbal behaviours of the consultation were those included in the index). The closer the score is to one, the more the consultation is characterized by physician/ patient partnership behaviour.
Patient measures Sociodemographic data. These were age and gender. Severity of medical illness. GPs rated severity on a 5-point scale including no symptoms (0), symptoms without defined medical illness (1), mild (2), moderately severe (3) and severe illness (4). Attributed emotional distress. GPs attributed psychological disorder on a 5-point scale including no emotional distress (0), sub-clinical problems (1), mild (2), moderately severe (3) and severe (4) psychological distress. General Health Questionnaire-12 (GHQ-12). This is a self-administered screening questionnaire for the presence of emotional distress, which has been standardized for Italian primary care attenders (Piccinelli, Bisoffi, Bon, Cunico & Tansella, 1993). Procedures In the twice-weekly sessions fixed for the data collection, the GPs audiotaped the consultations and invited all patients who satisfied the selection criteria to meet the researcher next door to complete the GHQ-12. The GPs completed a form for each patient, recording the sociodemographic data and rating the severity of patient’s present medical illness and the level of emotional distress observed. The consultation transcripts were coded by the same raters of the inter-rater reliability studies (Del Piccolo et al., 1999; Saltini et al., 1998). The raters were unaware of the patients’ GHQ-12 score and GP’s attribution of emotional distress. Methodological problems The database allowed us to derive a number of macro- and micro-variables, which in isolation have emerged in the literature as important factors in affecting the manifestation of cues, but so far have not been examined simultaneously. Macro-variables contribute to cue emission at two different levels. Level 1 regards the variables specific to each consultation, here patient’s gender, age, illness severity, emotional distress, GP’s attribution of emotional distress and the PI. These level 1 variables in turn are nested within GPs, because in our data set each physician contributes with many consultations and each GP had a varying number of patients. Accordingly, GP is considered as a level 2 macro-variable. Micro-variables were defined as any patient or physician speech unit immediately preceding the patient cue. Cues can appear after a physician
ARTICLE IN PRESS L. Del Piccolo et al. / Social Science & Medicine 65 (2007) 2357–2370
expression or after a patient expression. In the latter case, the speech unit preceding the cue occurred within the same patient turn; therefore at least two lag sequences had to be considered when exploring the influence of physician speech on cue occurrence. Sample size constraints (see below) excluded the possibility of performing two-lag sequence analysis. Therefore, the data set was arranged in two different ways, condition a and condition b, which allowed us to analyse the two different cue-related sequences performing in both cases lag-1 level of analysis: Condition a considered any patient or physician speech unit (lag-1) immediately preceding the cue. Analysis of such sequences, based on speech units, would inform about the position of a cue in a communication stream (either after a patient expression within the same turn or immediately after a GP expression). Fig. 1 presents an example and shows how the micro-level variables were derived from the VR-MICS database. In condition a, the sequence was preserved as it appeared in the consultation: GP open-ended question - Pt statement - GP closed-ended question - Pt statement - GP closed-ended question- Pt positive talk - Pt expectation/opinion - Pt statement - Pt expectation/opinion - Pt cue (target present). Condition b considered any GP speech unit immediately preceding the patient turn containing the cue. This noted the identification of GPs’ behaviours that triggered cues, independently of their position in the subsequent patient turn. Patient turns
2361
were coded as cue absent or cue present and the example above became: GP open-ended question Pt cue absent - GP closed-ended question - Pt cue absent - GP closed-ended question - Pt cue present. The simultaneous inclusion of micro- and macrovariables imposed methodological constraints, related both to the structure of data and to statistical analysis assumptions: 1. Due to the research design, our data have a hierarchical structure: verbal expressions of patients and physicians are nested within consultations, which in turn are defined also by patient characteristics. Consultations are nested within physicians, who may have different communication styles. 2. Within each consultation the meaning of the talk is not given in advance, but created in cooperation throughout the conversation (Sandvik et al., 2002). Therefore there is the need to assess if the probability of a cue appearing after a particular verbal expression remains similar throughout the consultation (stationarity problem). 3. Given the different interaction dynamics of consultations, cues may appear after different patient or physician expressions. Therefore consultations can be pooled only if these appearance probabilities are similar for all consultations (homogeneity problem). 4. Cues are rare events, representing on average 10% of the total patient talk (Del Piccolo et al., 2000). This implies that there are uncertain
Consultation divided by General Practitioner (GP) and patient turns
Speech units
Code
GP: Well, how are you getting on?
1 speech unit
Open-ended non directive question
Patient: Well, I feel all right in myself.
1 speech unit
Statement
GP: You are breathing?
1speech unit
Closed-ended question
Patient: The breathing, very bad this week
1 speech unit
Statement
GP: : Is that because you’ve been tight?
1 speech unit
Closed-ended question
Patient: Might be, might be.
5 speech units
Positive talk
I don’t know what it is.
Expectation/opinion
But I feel it all across here to the stomach.
Statement
Well if has anything to do with it, I don’t know
Expectation/opinion
but about three weeks ago, I woke up in the middle of the night, I felt there
Cue
was no excuse for it, I had a cold, you know…NOSEBLEED (worried tone of voice)… And that was the first nose bleed I’ve ever had in my life!
Fig. 1. Example of sequential coding of a consultation.
ARTICLE IN PRESS 2362
L. Del Piccolo et al. / Social Science & Medicine 65 (2007) 2357–2370
estimations when fitting statistical models limiting the number of explicative variables that can be introduced into the statistical model. This calls for the need of parsimonious models. 5. A quantitative approach to communication in terms of speech sequences requires that lag sequential analysis considers the temporal proximity, or distance, between the independent variable (physician or patient speech unit) and the dependent event (cue). That is, a lag-1 effect is the effect of the unit immediately prior to the dependent event; the lag-2 effect is the effect of the unit preceding the one immediately prior to the dependent event, and so on. Given that at each lag the independent (predictor) variable may be defined by different speech categories, with increasing lags the number of predictors increases exponentially together with all possible combinations of the various categories that have to be taken into account at each lag. That is, if the predictor variable is defined by three categories (three VR-MICS speech units, a, b, c), the predicting variables will be three at lag-1 (a,b,c), nine at lag-2 (aa, ab, ac, ba, bb, bc,ca, cb, cc), 27 at lag-3 (aaa, aab,y ccb, ccc). Consequently, few incremental lags need soon a huge sample size of speech units. Data analysis methods We were concerned with modelling the relative frequency (expressed as odds) of a patient emitting a cue at any given point in a consultation, allowing for the different GPs (macro-variable level 2), the patient characteristics and the PI of the consultation (macro-variables level 1), and the speech units preceding the cues in each consultation (microvariables). Prior to fitting any models, however, uncorrected associations between preceding speech and cue emission were evaluated using transition probabilities (the classical method to study sequences), crude odds-ratio (OR) ignoring the identity of the consultation, mean OR (the mean of the crude ORs estimated for each consultation separately) and Mantel Haenszel OR estimates (allowing for consultation as a stratifying variable). This approach was used to explore how estimations change with the increasing refinement of the methods adopted to calculate ORs. In assessing the effects of the GP’s preceding speech units, the information/instruction category was adopted as reference category for the OR and the logistic
models for two reasons: (1) because it is the most frequent of all physician categories, as shown in Table 1, it eliminates the probability of many structural zeros and (2) because it has about the same probability of appearing in any part of the consultation, as previously demonstrated (Goss, Del Piccolo, Rimondini, Mazzi, & Zimmermann, 2005), it avoids the use of a distorting reference category for the assessment of stationarity. Multilevel modelling allows for the testing and estimation of covariate (fixed) effects (i.e. patient sex, age, GHQ-12 score, GP attribution of emotional distress and illness severity, the PI and the sequence of speech units that precede a cue) and, simultaneously, for unexplained random variation at the level of the consultation and the GP. Multilevel logit models, using the presence/ absence of the cue as dependent variable, were adopted to take into account the nested structure of the data. Variables were assigned to each level depending on where the variation of their values occurred. The micro-level corresponded to the stream of speech sequences. The macro-level 1 accounted for the individual characteristics of the patient (gender, age, GHQ-12 score, attributed emotional distress and illness severity) and the PI, which varied between, but were constant within consultations. The macro-level 2, defined by GP membership, varied between groups of consultations (each GP contributing with a variable number of consultations) and was accounted for as random factor. To understand the contribution of the different levels of variables that play a role in patient’s cue emission in conditions a and b, the likelihood ratio chi-squares, calculated on the basis of differences in goodness of fit, as measured by the scaled deviance (a generalization of the residual sum of squares), were compared in a hierarchy of multilevel models: 1. A basic model allowed for random variation at two levels: between consultations and between GPs. The model contained also the fixed effect of the GHQ-12 score (covariate) on patient’s cue emission. The GHQ-12 score was included from the beginning, being known for its association with cue emission (Del Piccolo et al., 2000). 2. An intermediate model added patient and consultation characteristics (patient gender and age, GP attribution of emotional distress and illness severity, PI) to the ‘‘basic model’’ as fixed
ARTICLE IN PRESS L. Del Piccolo et al. / Social Science & Medicine 65 (2007) 2357–2370
2363
Table 1 Transition probability and crude odds ratio speech sets preceding a cue in condition a (GP and patient antecedents) and condition b (GP antecedents only). Coded speech unit
Condition a
Condition b a
GP categories Supporting, detecting, handling emotions Open question and active listening Passive listening Closed question Information/instruction Conversation Patient categories Listening, agreement/disagreement Statement Expectation/opinion and Question Cue Conversation Total
Transition probability
Crude odds ratio
12.5 (33/264) 3.8 (67/1766) 13.8 (191/1387) 2.8 (101/3596) 8.0 (388/4857) 6.4 (48/755)
1.66 0.46 1.84 0.33 Reference 0.79
26.8 34.1 33.3 30.0 24.0 10.7
4.23 6.00 5.77 4.94 3.69
(88/328) (478/1402) (98/294) (104/347) (25/104) (1621/15100)
b
Transition probabilitya
Crude odds ratiob
19.5 (51/262) 9.8 (172/1764) 17.5 (242/1384) 10.9 (392/3596) 11.7 (568/4839) 9.6 (72/748)
1.82 0.81 1.59 0.92 Reference 0.80
a Transition probability in each cell represents the frequency of chains that begin with the event indicated as ‘‘coded speech unit’’ and end with a cue (the ratio of the cue and the coded speech unit in brackets). b The crude odds ratio is the ratio between the coded speech unit and the reference category, calculated on the total sample of speech units without stratifying by interview.
factors. In order to create a parsimonious model, these explanatory variables were added and removed from the model. The magnitude of their impact on cue emission was reflected in the estimated coefficients, and their significance was indicated by the likelihood ratio chi-squares. 3. A final model added sequence information as a covariate to the significant macro-variables found in the intermediate model.
The intra-class correlation (ICC) was calculated for the two random variables in each of the three multilevel models, according to Rabe-Hesketh and Skrondal (2005). The statistical procedure was similar to that using regular logistic models. However, prior to the main analyses, we verified the conditions of stationarity and homogeneity as suggested by Gottman and Roy (1990). ‘‘Stationarity is a concern with the stability of our parameters of sequential connection over time’’ (Gottman & Roy, 1990, p. 60), and responds here to the question of whether cues have the same probability of appearing after a coded speech set, independently of the consultation phase. This means that, after dividing the consultation into halves, the distribution of the odds of the speech set
to precede a cue is the same when the first half is compared to the second one (Sign rank test). ‘‘Homogeneity is a concern with the stability of our parameters of sequential connection across subjects’’ (Gottman & Roy, 1990, p. 60). This means that all 246 consultations are comparable, when considering the odds for a cue. To estimate how variable the ORs are across all consultations, between consultation heterogeneity has been analysed applying meta-analysis methods (Sutton, Abrams, Jones, Sheldon, & Song, 2000). The odds of the speech sets in each consultation were considered as effect size: consultations were therefore considered as different studies and the metaanalysis method was repeatedly applied for each of the 10 sets of doctor and patient speech. The Heterogeneity test evaluated whether the differences between the odds were due to random variation and not attributable to systematic differences between consultations. Being a weighted estimator, the heterogeneity index has been weighted by the length of the consultation (expressed by the total number of coded speech units). Statistical analyses were performed with STATA 8.2 (StataCorp, 2005). The STATA GLLAMM (generalized linear latent and mixed models) command (Rabe-Hesketh & Skrondal, 2005) has been applied for multilevel logit models.
ARTICLE IN PRESS 2364
L. Del Piccolo et al. / Social Science & Medicine 65 (2007) 2357–2370
Results Consultation Characteristics GPs contributed a varying number of consultations (range 15–60) and showed a high variability in the frequency of specific verbal behaviours, as already shown in a previous analysis (Del Piccolo et al., 2002). Eleven consultations were without cues, but were included in the data set, being long enough to represent the general sequence of a consultation. The 246 medical consultations generated 29,851 speech units (mean per consultation 121.3; SD 54.4, range 23–331), 15,100 of these were patient generated. There was a mean number of 6.6 cues per consultation (range 0–30). The first time a cue appeared was on average at the 23rd speech unit (range 1–136). The PI varied between 0.12 and 0.56, with a mean value of 0.35 (SD 0.08). The index showed a normal distribution and discriminated among GPs (ANOVA F(5, 240) ¼ 5.57, po0.0001). Which speech units precede a cue Table 1 shows the probability of different sets of speech units preceding a cue and reports their transition probabilities and crude ORs for conditions a and b. Transition probabilities give an idea of the magnitude of the phenomenon we are observing; for instance, in condition a, the cue was subsequent to support, detecting, handling emotions or passive listening in about 13% of the sequences. For the reference category information/instruction this percentage was 8%. Translated in terms of crude OR, support, detecting, handling emotions and passive listening expressions of GPs showed a positive effect (OR ¼ 1.66 and 1.84, respectively) on cue emission, whereas speech categories with a lower occurrence probability than the reference category hindered cues (e.g. open questions and active listening, OR ¼ 0.46). Table 1 shows that in condition a, most cues are preceded by another patient expression, particularly statements and expectation/opinions and questions, but also by other cues. The comparable transition probabilities and crude ORs obtained for all patient speech sets means that all have similar predictive power on cue emission. The cue thus may be considered as a spontaneous expression that emerges after any type of patient expression. Conversely, GP speech sets differed in their
predictive power: Supporting, detecting, handling emotions and passive listening showed the higher probability to induce cue occurrence. Similar findings emerged for condition b. Crude ORs confirmed the positive effect of these two speech sets and the hindering effect of both closed questions and open questions and active listening, although the last effect was less pronounced than in condition a. Cues represented 10.7% of all patient speech sets. Stationarity and homogeneity of consultations Stationarity: When applying condition a, stationarity was different for two of the 10 speech sets considered: passive listening and patient conversation. In both cases, cues subsequent to these speech sets prevailed in the first part of the consultation (z ¼ 2.09, p ¼ 0.04 and z ¼ 2.44, p ¼ 0.02, respectively). In condition b the odds of a cue after any of the six sets of GP speech units were the same in the first and the second part of all 246 consultations. This means that one chunk of the consultation is like another in terms of odds when considering GP’s verbal behaviours preceding a patient turn in which a cue was present/absent (condition b), whereas some differences emerged between the first and the second half of the consultation when considering the sequence of both patient and GP speech units (condition a). Homogeneity: Table 2 shows the mean OR and the Mantel Haenszel estimation for a cue of each of the speech sets considered in conditions a and b, and the Heterogeneity index for all combinations of subsamples of the 246 consultations. None of the consultations with valid OR was an outlier. The 246 consultations were homogeneous, that is, the null hypothesis that the odds were the same among all consultations and the assumption that variability between consultations was random were confirmed. Mean ORs and Mantel Haenszel estimations slightly differed from crude ORs (Table 2). Other than crude ORs and transition probabilities, they stratify by consultation and, in addition, Mantel Haenszel OR estimations take into account consultation length. In condition a the effect size of each of the 10 sets of speech units was significant. Cues were confirmed to follow any other expression of the patient, and the same pattern reported in Table 1 was found for GP speech sets, with the exception of GP conversation that passed from 0.79 (Crude OR) to a significant positive predictive value of 1.50 (Mantel Haenszel OR). Condition b findings
ARTICLE IN PRESS L. Del Piccolo et al. / Social Science & Medicine 65 (2007) 2357–2370
2365
Table 2 Mean odds ratio, Mantel Haenszel odds ratio estimation and heterogeneity index in condition a (GP and patient antecedents) and in Condition b (GP antecedents only) Coded speech unit
Condition a
Condition b
Mean odds ratio OR Mantel (range) Haenszela GP categories Supporting, detecting, handling emotions Open questions and active listening Passive listening Closed question Information/instruction Conversation Patient categories Listening, agreement/ disagreement Statement Expectation/opinion and question Cue Conversation
Heterogeneity w2 (df)b
Mean odds ratio OR Mantel (range) Haenszela
Heterogeneity w2 (df)
1.34 (0.0–32.0)
2.11
53.70 (100)
1.24 (0.0–20.0)
2.08
49.96 (110)
0.45 (0.0–6.8)
0.74
73.51 (168)
1.03 (0.0–20.0)
0.97
130.18 (205)
1.81 (0.0–34.0) 0.49 (0.0–17.3) Reference 0.73 (0.0–10.6)
1.95 0.46 Reference 1.50
126.71 (177) 97.57 (166) Reference 76.01 (146)
1.94 (0.0–32.0) 1.04 (0.0–17.3) Reference 0.93 (0.0–19.0)
1.70 0.90 Reference 1.31
147.87 (195) 167.76 (213) Reference 110.14 (168)
3.88 (0.0–45.0)
4.76
79.13 (132)
7.95 (0.0–120.0) 4.85 (0.0–84.0)
5.42 5.82
172.63 (206) 76.55 (121)
3.22 (0.0–28.0) 1.53 (0.0–20.7)
3.77 4.11
66.52 (114) 28.13 (49)
a
hypothesis under test is that OR ¼ 1. df ¼ degrees of freedom, which correspond to the number of consultations with valid odds ratio (not null or incalculable) minus one. pp0.05. pp0.01. pp0.001 b
showed the same wide range of mean ORs observed for condition a, and confirmed the predictive power of giving support, detecting, handling emotions, passive listening and GP conversation on cue emission. In conclusion, stationarity and homogeneity proved that the 246 consultations were stable over time in terms of GP expressions (condition b stationarity) and comparable in terms of odds (both, conditions a and b homogeneity). This last finding indicated that it was appropriate to analyse the pooled sample of all 246 consultations. What contributes to cue emission? We used multilevel modelling techniques to explore sources of variation that affect cue emission: patient’s emotional state, gender, age and severity of illness, the PI, the attribution of emotional distress by GP and the expressions that precede a cue (conditions a and b). Table 3 shows the results of fitting the same three models for each of the two conditions. The basic model allowed for random variation between consultations (level 1) and between GPs
(level 2) and contained the fixed effect of the GHQ12 score (ranging from 0 to 12). The estimated variance components for consultation and GP showed no significant change from one condition to the other (0.29 and 0.02 in condition a and 0.30 and 0.03 in condition b, respectively). Most of the variation occurred at the level of the consultation (ICC ¼ 9%), compared with that at the level of GP (ICC ¼ 1%). The GHQ-12 score had a significant positive effect (pp0.001) on cue emission: an increase of one score corresponded to an estimated OR of 1.08, an increase of five to 1.5 and an increase of nine doubled the probability of cue emission. In probabilistic terms, with a GHQ-12 score of 0, cue frequency was 7%, whereas with a GHQ-12 score of 12, the frequency became 16%. The intermediate model contained the same basic model effects but included, in addition, the PI as covariate. Different alternative models, with the remaining patient variables (gender, age, illness severity, attribution of emotional distress) as covariates, were tested, but none of these variables reached significance. Table 3 illustrates the significant effect of the PI in both models: the higher the
ARTICLE IN PRESS 2366
L. Del Piccolo et al. / Social Science & Medicine 65 (2007) 2357–2370
Table 3 Intermediate and final multilevel logistic models compared with the basic model in conditions a (GP and Pt antecedents) and b (GP antecedents only) Condition a models
Random effect Consultation General practitioner
Condition b models
Basic
Intermediate
Final
Basic
Intermediate
Final
Variance (SE) 0.29 (0.05) 0.02 (0.02)
Variance (SE) 0.27 (0.05) 0.03 (0.03)
Variance (SE) 0.26 (0.05) 0.05 (0.04)
Variance (SE) 0.30 (0.05) 0.03 (0.02)
Variance (SE) 0.27(0.05) 0.03(0.03)
Variance (SE) 0.29 (0.05) 0.05 (0.04)
Fixed effect Odds ratio (SE) GHQ 1.08 (0.02) Partnership Index Speech units GP: Support, detect, handle emotions GP: Open question–active listening GP: Passive listening GP: Closed question GP: Information/instruction GP: Conversation Pt: Listening, agreement/ disagreement Pt: Statement Pt: Expectation/opinion-Question Pt: Cue Pt: Conversation Log-likelihood 5027.44 Likelihood ratio test
Odds ratio (SE) Odds ratio (SE) Odds ratio (SE) Odds ratio (SE) Odds ratio (SE) 1.07 (0.02) 1.08 (0.02) 1.08 (0.02) 1.08 (0.02) 1.08 (0.02) 0.18 (0.11) 0.36 (0.24) 0.19 (0.12) 0.22 (0.15) 1.68 (0.29) 0.78 (0.08) 1.60 (0.14) 0.86 (0.06) Reference 0.87 (0.12)
1.44 (0.29) 0.42 (0.06) 1.78 (0.17) 0.30 (0.04) Reference 0.85 (0.14) 4.13 (0.58)
5023.68 7.52 (1)
6.11 (0.50) 6.17 (0.88) 3.93 (0.53) 4.01 (0.99) 4362.03 4483.01 1323.30 (10)
4479.77 6.48 (1)
4434.14 91.26 (5)
pp0.05. pp0.01. pp0.001.
PI, the lower was the number of cues. The GHQ-12 contribution remained the same in the presence of the added covariate (1.07 in condition a and 1.08 in condition b). The introduction of the PI influenced the estimates of the two random effect variables, but not dramatically so (for consultation level ICC ¼ 8%, for GPs level ICC ¼ 1%). In condition a, when GHQ-12 ¼ 0 and PI ¼ 0.1, the frequency of cues was 7%, and became less than 4% when PIX0.5. Applying the same consultation conditions to a distressed patient (GHQ-12X5) the frequency of cues ranged between 9% and 15% (GHQ12 ¼ 5–12) when PI ¼ 0.1 and between 4% and 11% (GHQ-12 ¼ 5 to 12) when PIX0.5. Predictions were very similar when the same conditions were applied to condition b model. The final models for conditions a and b contained the same intermediate model effects and, in addition, included the speech units that immediately preceded a cue (Table 3). The models eventually chosen contained the two random effects, the two significant covariates and, in addition, the effects of
the odds of the speech units that preceded a cue, using information/instruction as reference category. In condition a the introduction of the sequence of speech units influenced the estimate of the PI, which became non-significant. The contribution of the GHQ-12 score remained about the same, corresponding to 1.08, as did the estimates of the two random effect variables. All patient speech sets solicited cue emission, with the higher contribution of statements and expectations/opinions and questions (OR ¼ 6.11 and 6.17, respectively, that is, compared to the reference category, there was a sixfold increase in cue emission). Cues were also directly facilitated by GPs’ use of passive listening (OR ¼ 1.78) whereas closed questions and open questions and active listening contributed to reduce immediate cue emission when compared to information/instruction (OR ¼ 0.30 and 0.42, respectively). According to this model the patients without emotional distress (GHQ-12 ¼ 0) had a cue frequency of 2% and 12% after a closed question and a passive listening expression by GP and of 32% after
ARTICLE IN PRESS L. Del Piccolo et al. / Social Science & Medicine 65 (2007) 2357–2370
their own statement. These percentages reached the maximum of 6%, 26% and 55% when GHQ12 ¼ 12. Thus, in the presence of emotional distress about one-fifth of facilitating expressions by GP and half of all patient statements were followed by a cue. In condition b the model maintained the significant contribution of GHQ-12 score (OR ¼ 1.08, pp0.001) and PI (OR ¼ 0.22, pp0.05). As in condition a cues were hindered by closed questions and open questions and active listening (OR ¼ 0.86 and 0.78, respectively) and facilitated by GPs’ expressions of passive listening (OR ¼ 1.60), and, in addition, by support, detecting and handling emotions (OR ¼ 1.68) According to this second model the patient without emotional distress (GHQ-12 ¼ 0) showed an estimated frequency of 17–18% to express a cue after passive listening or after support, detecting and handling emotions, and 9–10% after closed question and open questions and active listening, when PI ¼ 0.1. The same percentages became less than 11% and 6%, respectively, when PIX0.5. This last condition almost halved the likelihood of cue appearance. The significant likelihood ratio test (Table 3) showed that when comparing the intermediate with the basic model and the final with the intermediate model, the variables added in each of the subsequent models contributed to improve their goodness of fit. That is, with increasing complexity the models increase their power in explaining the variability in cue emission. Discussion Our study aimed to get a better understanding about what makes patients take the initiative in the medical consultation to give new, unsolicited information, here defined as cue. To pursue this aim we had to pass several methodological hurdles inherent to the study design, to the related statistical assumptions and to the data set characteristics. The methodological approach adopted and described here in detail is an important part of this study. We consider this approach as essential for studies that want to grasp the complex interplay of micro-level (critical speech sequences) and macro-level variables (the partnership, the consultation, the physician or the patient makeup) that operate in the consultation context. One main result was strictly connected to the definition of cue that we adopted. As new information spontaneously introduced by the patient, cues
2367
appeared very often within a patient speech turn, preceded by other expressions of the patient. Therefore, cues emerge when patients have space to talk. Another finding was related to what GPs actually said: closed-ended inquiry and active listening expressions compared to information/instruction hindered the appearance of cues, whereas expressions of support, detecting and handling emotions and passive listening induced cues. Thirdly, apparently in contrast with this last finding, cue emission decreased the more the consultation was characterized by partnership in terms of open-ended inquiry, active listening skills and emotion handling of the physician and active involvement of the patient. Limitations and strengths of the study To our knowledge, this is the first study that applies lag sequential analysis adopting a multilevel approach with the goal of taking into account different contextual variables that might contribute to modify a very specific but relevant patient behaviour as is cue emission. Previous studies based on lag sequential method directly inspected the contribution of specific physician expressions to cue (Zimmermann, Del Piccolo & Mazzi, 2003) or concern emission (Eide, Quera, & Finset, 2003; Eide, Quera, Graugaard, & Finset, 2004; Langewitz, Nu¨bling, & Weber, 2003) without including other contextual, confounding, variables in the analysis. These studies omitted testing the basic assumptions for stationarity and homogeneity as recommended by Gottman and Roy (1990). In the present study these assumptions were tested before analysing the data on the pooled sample of consultations, and the consultations were found almost stable over time and comparable in terms of odds. Multilevel modelling obliged us to group VRMICS codes into a manageable number of speech category sets in search of a balance between interpretability and complexity of statistical models. The creation of sets precluded the evaluation of the effect of single speech categories on the occurrence of cues, but their inclusion would have reduced the possibility of introducing different confounding variables into the multilevel model. The two ways (conditions a and b) adopted to study the contribution of patient and physician speech to cue emission permitted us to identify different processes: condition a informed us as to where the cue was prevalently placed, namely within a patient turn, preceded by any other patient expressions, condition
ARTICLE IN PRESS 2368
L. Del Piccolo et al. / Social Science & Medicine 65 (2007) 2357–2370
b informed us as to which of GP’s expressions specifically prompted cues, that is passive listening, emotion detecting and handling and conversation. The participating GPs were males. Male physicians are known to be more doctor-centred than female physicians and to engage in less partnership behaviour (Roter, Hall & Aoki, 2002). This might apply also to the Italian male GPs of our study, and could have resulted in a frequency of cue expressions not representative for patients of female GPs. Expressions of concern were unrelated to physician gender (Hall & Roter, 2002), but whether this also holds for cues has yet to be examined. Variables influencing patient’s cue emission There were about seven cues per consultation, representing a 10th of all patient speech units. The first appearance of a cue was highly variable as shown by the wide range of the speech unit position in the consultation. A cue is therefore a rare but possible event (there is at least one cue per consultation) in the great majority of the consultations. As expected on the basis of previous studies (Davenport et al., 1987; Del Piccolo et al., 2000) there was a positive association between cue emission and patients’ self-rated emotional distress (GHQ-12), while GP ratings of the patient’s emotional distress had no effect on the cue frequency of their patients. Patient gender, age and illness severity also were unrelated to cue occurrence. To our knowledge illness severity in relation to cue offers has not been studied, whereas the few studies that examined age and gender in relation to cues (Butow et al., 2002) or concerns (Ishikawa et al., 2002; Timmermans et al., 2005) report contradictory findings. The two final models showed that cue emission was influenced by most of the micro- and macrolevel variables considered, although differently in conditions a and b. In condition a, where the exact sequence of physician and patient codes of speech units was preserved, GPs’ passive listening showed the higher probability of inducing cues in the succeeding patient turn, while closed inquiry and open question and active listening reduced their appearance. Within patient turns, cues appeared as spontaneous information anywhere in the consultation, after many different expressions, and with increased frequency after reports of symptoms or problems (statements) or personal ideas and doubts
(expectation/opinion and question), but also after other cues or after patient conversation. The same GP speech sets that had shown a significant effect on cues in condition a, were operant in condition b, but in addition, support, detecting and handling emotions emerged as a cue facilitating and the PI as cue impeding variable. This apparently conflicting finding suggests that consultations where physicians display open inquiry skills and ability to handle emotions and patients are actively involved, have fewer cues, but if a cue appears, this is after a passive listening or an emotion-focused expression by the physician. These latter results parallel findings reported in literature: empathic comments (Goldberg et al., 1993; Maguire, Faulkner, Booth, Elliott, & Hillier, 1996) and facilitation were shown to induce cues (Zimmermann et al., 2003) and concerns (Eide et al., 2004; Street, 1992); doctor-led questions or questions on physical issues (Goldberg et al., 1993), particularly in distressed patients (Del Piccolo et al., 2000), and the use of active listening techniques (Del Piccolo et al., 2000) hampered cues. The change in significance of the PI in predicting cues when patient speech sets are present (condition a) or absent (condition b) needs to be commented on. The patient speech set ‘‘expectation/opinions and questions’’ is part of the PI, but is also one of the cue preceding sets considered in the sequence analysis of condition a and accordingly has absorbed part of the variance explained by the PI. This was not the case in condition b where only GP speech sets were considered as cue predictors, enabling the PI to become significant. Conclusions The quality of a medical consultation increases with the amount of relevant information provided by the patient. Patients may offer information spontaneously, either explicitly or as hints, and we defined this way to inform as cue. We studied the phenomenon in terms of sequence of speech, applying the methodological procedures required for these data. Moreover, we also had to take into account the nested structure of the different data sources. Accordingly, the statistical analyses became a lengthy and painstaking process, but crucial for the soundness of our findings. Communication is a complex phenomenon to study, where many different factors interact simultaneously. It is therefore difficult to describe the
ARTICLE IN PRESS L. Del Piccolo et al. / Social Science & Medicine 65 (2007) 2357–2370
consultation process realistically or entirely accurately using any statistical model. These models are never more than approximations to reality, although our findings showed that with increasing complexity of our multilevel models we can expect the model to better describe and, potentially, explain the sequence of utterances or behaviours preceding a cue. Ideally, qualitative and/or narrative analyses should complete the picture by adding important information on the effects of meaning and intentionality during communication. Cues, if not ignored, contribute to complete the physician’s data collection and often permit the physician to understand patients’ underlying emotional difficulties as evidenced by the positive correlation between number of cues and GHQ-12 score. Cues were interspersed throughout a patient’s talk within a turn and were less likely as a first response to GP speech. However, by passive listening and emotion detecting and handling skills physicians succeeded in increasing the likelihood of a cue in the subsequent patient turn, regardless of whether it appeared as immediate and first response or whether it occurred in midst of other patient expressions. We found that an overall increase of partnership behaviour throughout the consultation decreased cue emission as did closed inquiry. In light of these results, the number of cues cannot be considered as proxy measure of a patient-centred consultation approach, as suggested in earlier studies (Davenport et al., 1987, Goldberg et al., 1993). From our data we conclude that passive listening, together with supporting and emotioncentred expressions, activate cue emission by encouraging the patient to add new information or to direct the physician’s attention to issues of importance, whereas physician closed questions tend to suppress cue expressions. On the other hand, to solicit a patient’s expression of personal needs by open inquiry and active listening, to acknowledge and to sensitively handle their cues, as physician components of partnership, will satisfy these very needs and lower cue offers. Cue inducing skills and partnership documenting exchange between physician and patient throughout the consultation, should be an important target for communication skills training. Neither of these two aspects should be preferred to the other on the assumption that a general approach is more important than specific skills or that the mere use of cue inducing skills is sufficient to explain the quality of a consultation.
2369
References Bakeman, R., & Gottman, J. M. (1997). Observing interaction. An introduction to sequential analysis 2nd ed. Cambridge, UK: Cambridge University Press. Butow, P. N., Brown, R. F., Cogar, S., Tattersall, M. H., & Dunn, S. M. (2002). Oncologists’ reactions to cancer patients’ verbal cues. Psycho-Oncology, 11, 47–58. Davenport, S., Goldberg, D., & Millar, T. (1987). How psychiatric disorders are missed during medical consultations. Lancet, 2, 439–441. Del Piccolo, L., Benpensanti, M. G., Bonini, P., Cellerino, P., Saltini, A., & Zimmermann, C. (1999). The Verona Medical Consultation Classification System/Patient (VR-MICS/P): the tool and its reliability. Epidemiologia e Psichiatria Sociale, 8, 56–67. Del Piccolo, L., Mazzi, M., Saltini, A., & Zimmermann, C. (2002). Inter and intra individual variations in physicians’ verbal behaviour during primary care consultations. Social Science & Medicine, 55, 1871–1885. Del Piccolo, L., Mead, N., Gask, L., Mazzi, M. A., Goss, C., Rimondini, M., et al. (2005). The English version of the Verona Medical Consultation Classification System (VRMICS). An assessment of its reliability and a comparative cross-cultural test of its validity. Patient Education and Counseling,, 58, 252–264. Del Piccolo, L., Putnam, S. M., Mazzi, M. A., & Zimmermann, C. (2004). The biopsychosocial domains and the functions of the medical consultation in primary care: Construct validity of the Verona Medical Consultation Classification System. Patient Education and Counseling, 53, 47–56. Del Piccolo, L., Saltini, A., Zimmermann, C., & Dunn, G. (2000). Differences in verbal behaviours of patients with and without emotional distress during primary care consultations. Psychological Medicine, 30, 629–643. Eide, H., Quera, V., & Finset, A. (2003). Exploring rare patient behaviour with sequential analysis: An illustration. Epidemiologia e Psichiatria Sociale, 12, 109–114. Eide, H., Quera, V., Graugaard, P., & Finset, A. (2004). Physician–patient dialogue surrounding patients’ expression of concern: Applying sequence analysis to RIAS. Social Science Medicine, 59, 145–155. Epstein, R. M., Franks, P., Fiscella, K., Shields, C. G., Meldrum, S. C., Kravitz, R. L., et al. (2005). Measuring patient-centered communication in patient-physician consultations: Theoretical and practical issues. Social Science & Medicine, 61, 1516–1528. Fallowfield, L., Jenkins, V., Farewell, V., & Solis-Trapala, I. (2003). Enduring impact of communication skills training: Results of a 12-month follow-up. British Journal of Cancer, 89, 1445–1449. Ford, S., & Hall, A. (2004). Communication behaviours of skilled and less skilled oncologists: A validation study of the Medical Interaction Process System (MIPS). Patient Education and Counseling, 54, 275–282. Goldberg, D. P., Jenkins, L., Millar, T., & Faragher, E. B. (1993). The ability of trainee general practitioners to identify psychological distress among their patients. Psychological Medicine, 23, 185–193. Goss, C., Del Piccolo, L., Rimondini, M., Mazzi, M. A., & Zimmermann, C. (2005). Information-giving sequences in
ARTICLE IN PRESS 2370
L. Del Piccolo et al. / Social Science & Medicine 65 (2007) 2357–2370
general practice consultations. Journal of Evaluation in Clinical Practice, 11, 339–349. Gottman, J. M., & Roy, A. K. (1990). Sequential analysis. A guide for behavorial researchers. Cambridge, UK: Cambridge University Press. Hall, J. A., & Roter, D. L. (2002). Do patients talk differently to male and female physicians? A meta analytic review. Patient Education and Counseling, 48, 217–224. Ishikawa, H., Takayama, T., Yamazaki, Y., Seki, Y., Katsumata, N., & Aoki, Y. (2002). The interaction between physician and patient communication behaviors in Japanese cancer consultations and the influence of personal and consultation characteristics. Patient Education and Counseling, 46, 277–285. Langewitz, W., Nu¨bling, M., & Weber, H. (2003). A theory-based approach to analysing conversation sequences. Epidemiologia e Psichiatria Sociale, 12, 103–108. Levinson, W., Gorawara-Bhat, R., & Lamb, J. (2000). A study of patient clues and physician responses in primary care and surgical settings. Journal of the American Medical Association, 8, 1021–1027. Maguire, P., Faulkner, A., Booth, K., Elliott, C., & Hillier, V. (1996). Helping cancer patients disclose their concerns. European Journal of Cancer, 32, 78–81. Mazzi, M. A., Del Piccolo, L., & Zimmermann, C. (2003). Eventbased categorical sequential analyses of the medical interview: A review. Epidemiologia e Psichiatria Sociale, 12, 81–85. Piccinelli, M., Bisoffi, G., Bon, M. G., Cunico, L., & Tansella, M. (1993). Validity and test–retest reliability of the Italian version of the 12-item General Health Questionnaire in general practice: A comparison between three scoring methods. Comprehensive Psychiatry, 34, 198–205. Pini, S., Piccinelli, M., & Zimmermann-Tansella, C. (1995). Social problems as factors affecting medical consultation: A comparison between general practice attenders and community probands with emotional distress. Psychological Medicine, 25, 33–41. Rabe-Hesketh, S., & Skrondal, A. (2005). Multilevel and longitudinal modeling using Stata. Stata Press Publication. College Station, TS: StataCorp LP. Roter, D. L. (1993). The Roter method of interaction process analysis. The Johns Hopkins University: Baltimore, MD.
Roter, D.L., Hall, J.A., & Aoki, Y. (2002). Physician gender effects in medical communication: A meta-analytic review. J. Am. Med. Assoc. 756–64. Saltini, A., Cappellari, D., Cellerino, P., Del Piccolo, L., & Zimmermann, C. (1998). An instrument for evaluating the medical consultation in general practice: VR-MICS/D. Epidemiologia e Psichiatria Sociale, 7, 210–223. Sandvik, M., Eide, H., Lind, M., Graugaard, P. K., Torper, J., & Finset, A. (2002). Analyzing medical dialogues: strength and weakness of Roter’s interaction analysis system (RIAS). Patient Education and Counseling, 46, 235–241. Smith, R. C. (2002). Patient-centred consultation: An evidenceBased method. Philadelphia: Lippincott Williams and Wilkins. StataCorp. (2005). Stata Statistical Software: Release 8.2. College Station, TX: Stata Corporation. Stewart, M., Brown, J. B., Weston, W. W., McWhinney, J. R., McWilliam, C., & Freeman, T. R. (1995). Patient-centered medicine: Transforming the clinical method. Thousand Oaks, CA: Sage Publications, Inc. Street, R. L. (1992). Communicative styles and adaptations in physician–parent consultations. Social Science & Medicine, 34, 1155–1163. Sutton A. J., Abrams K. R., Jones D. R., Sheldon T. A., & Song, F. (2000). Methods for meta-analysis in medical research. Wiley: Baffins Lane, Chichester, England. Timmermans, L. M., van der Maazen, R. W., Verhaak, C. M., van Roosmalen, M. S., van Daal, W. A., & Kraaimaat, F. W. (2005). Patient participation in discussing palliative radiotherapy. Patient Education and Counseling, 57, 53–61. Wasserman, R. C., & Inui, T. S. (1983). Systematic analysis of clinician–patient interactions: a critique of recent approaches with suggestions for future research. Medical Care, 21, 279–293. Zandbelt, L. C., Smets, E. M., Oort, F. J., & de Haes, H. C. (2005). Coding patient-centred behaviour in the medical encounter. Social Science & Medicine, 61, 661–671. Zimmermann, C., Del Piccolo, L., & Finset, A. (2007). Cues and concerns by patients in medical consultations. A literature review. Psychological Bulletin, 133, 438–463. Zimmermann, C., Del Piccolo, L., & Mazzi, M. A. (2003). Patient cues and medical consultation in general practice: Examples of the application of sequential analysis. Epidemiologia e Psichiatria Sociale, 12, 115–123.