|Home | About | Journals | Submit | Contact Us | Français|
Rapkin and Schwartz  define response shift as otherwise unexplained, discrepant change in HRQOL that is associated with change in cognitive appraisal. In this paper, we demonstrate how a Recursive Partitioning (RPART) regression tree analytic approach may be used to explore cognitive changes to gain additional insight into response shift phenomena.
Data are from the “Choices in Care Study,” , an evaluation of HIV+ Medicaid recipients’ experiences and outcomes in care (N = 394). Cognitive assessment was based on the Quality of Life Appraisal Battery. HRQOL was measured by the MOS-SF36v2 .
We used RPART to examine six-month change in MOS mental composite as a function of changes in appraisal, after controlling for patient characteristics, health changes and intervening events. RPART identified nine distinct patterns of cognitive change, including three associated with negative discrepancies, four with positive discrepancies, and two with no discrepancies.
RPART classification provides a nuanced treatment of response shift. This methodology has implications for evaluating programs, guiding decisions and targeting care.
Converging evidence shows that response shift can strongly affect how in individuals appraise their Health-Related Quality Of Life (HRQOL) [1, 4, 5, 6]. Response shift typically appears in counterintuitive findings—individuals with severe chronic illnesses reporting equal or better HRQOL scores than healthy individuals or individuals with less severe illness (e.g., [7, 8]). For example, the general public assigned a 0.39 HRQOL to dialysis whereas dialysis patients assigned their own HRQOL at 0.56 (on a 0 – 1 scale where 0 represents death and 1 represents perfect health) . This and similar paradoxical findings bring into question what HRQOL assessments are really measuring. Measurement imprecision and response bias do not fully explain the phenomena . The theory of response shift posits that it constitutes a change in the meaning of one’s self-evaluation of the QOL construct due to recalibration, repriortization, and/or reconceptualization [6, 10]. These constructs are related to work on idiographic quality of life assessment [4, 11, 12, 13]. The theoretical and measurement foundations of these constructs are well documented [5, 6, 14].
Rapkin and Schwartz  describe the assessment strategies to probe respondents on their evaluation of the meaning of QOL. They propose operationalizing response shift as change in HRQOL that cannot be explained by changes in overt health status, resources or life events, but that can be associated with change in cognitive appraisal. Their data-analytic strategy was based primarily on linear regression to estimate the extent to which residual QOL changes are associated with appraisal change . However, there is no intrinsic reason that these relationships must be linear. Rapkin and Schwartz’ notion of a final “combinatorial algorithm” that people use to summarize their experiences into HRQOL ratings explicitly posits complex interactions among constituents of appraisal. For example, an individual may report better quality of life than expected given their health status by ignoring problems, emphasizing positive experiences, selecting favorable targets for self-comparison, and/or focusing on less ambitious goals. Each of these processes may operate alone or in combination to represent distinct types of response shift.
There are obvious drawbacks to using linear regression to examine relationships involving appraisal processes that are intrinsically non-linear. Classification and Regression Trees (CART) methods are a suitable alternative to linear regression in elucidating potentially complex interactions . Using an iterative algorithm, respondents are classified into increasingly homogeneous subgroups with similar changes in cognitive appraisal profiles, allowing a more nuanced interpretation of how cognitive appraisal can influence HRQOL. The broad goal of this article is to demonstrate an empirical technique to identify prevalent patterns of cognitive changes that can account for residual variance in HRQOL change scores. Patterns of appraisal identified in this way represent different manifestations of response shift.
The study was developed by investigators at our respective institutions in conjunction with the New York State Department of Health AIDS Institute to evaluate the impact of the HIV Special Needs Plans, as part of an evaluation of patient-reported outcomes and experiences in care reported by HIV+ Medicaid recipients in New York State. Detailed data collection plans are summarized elsewhere [2, 16]. Institutional Review Boards approved the study.
Interviews were conducted in either Spanish or English, in person or by telephone, according to patient preference. The primary HRQOL assessment was the SF36.v2 , assessed at baseline (approximately six weeks post enrollment), and at 6 and 12 months post baseline. Changes in cognitive appraisal processes were assessed at these time points using the Quality of Life Appraisal Battery . The baseline interview also included measures of demographics, behavioral risks and health history.
Following Rapkin and Schwartz , the Quality of Life Appraisal Battery included four components: 1) persons’ frame of reference for considering HRQOL as assessed by six probes designed to tap different motivational themes, including achievement, maintenance, prevention, problem solving, disengagement, and acceptance. For example, respondents were asked about “the main things they want to accomplish”, “problems they want to solve”, and “things they are trying learn to accept,” in order to have their best possible quality of life. Verbatim responses to these probes, or “goal statements,” were coded and analyzed to extract “goal attributes” (described below). Additionally, we assessed 2) how persons sample experiences within that frame, assessed by 13 items on, e.g., whether or not the persons evaluated HRQOL by “thinking about the worst possible moments” within that frame; 3) how persons evaluate experiences using different standards of comparison by 9 items on, e.g., whether they compared themselves with “other people living with HIV”; and 4) how persons summarize and combine evaluations to describe HRQOL by using a combinatory algorithm of 16 items on, for example, whether they were thinking about “how well you’ve been doing, how hard it has been, both or neither?”.
From the open-ended assessments of frame of reference we collected over 6,700 goal statements at baseline and 6 months (plus an additional 1,458 from our first wave of 12 months follow-up interviews, which were coded in this group, but not reported here). Content analysis of these responses was accomplished through a two-stage process which is briefly summarized below. Complete documentation of goal coding, Kappa reliability and components analysis is available from the authors upon request.
In the first step, we selected at random just over 1/3 of all responses (2,638). Each selected goal was given to two of 13 judges (students and faculty in our department), following an allocation scheme to ensure that an equal number of overlapping goals assigned to each pair of judges. Each judge independently sorted about 405 goal statements into homogeneous categories, with the sole criterion being that statements within a category must be “similar in all important ways”. Judges then recorded the “goal attributes” that they used to make distinctions among categories, including life domains, motivations, and health -relevance. After completion of independent sorting, all judges met to compare their derived dimensions. In general, there was strong agreement in the major distinctions among life domains and in prevalent fine-grain distinctions. Judges primarily differed in how specific to be in certain sub-domains (for example, to distinguish concerns about specific family members from those pertaining to the family in general). Based on this discussion, we derived a consensus set of 24 binary goal attributes. All goal statements could be characterized using combinations of these goal attributes. We calculated kappa for each of the 24 codes, to determine whether or not pairs of judges agreed on the presence or absence of each goal attribute in their initial sort of goal statements. Collapsing across dyads, we found that 11 of 24 categories exceed kappa = .70, another 4 exceed kappa =.59, 6 exceeded kappa = .35, and 3 codes (representing only 3.76% of coded statements) did not differ from chance.
Following derivation of goal attributes, the remaining 2/3 (5,148) goal statements were assigned to eleven judges. We assigned a random 20% of these goal statements (1,030) as a reliability sample, allocated evenly to all possible pairs of judges. Reliability coefficients for 13 of 24 categories exceeded kappa = .70, another 6 exceed kappa =.50, 3 exceeded an acceptable level of kappa = .39, and 2 categories (representing only 1.19% of coded statements) were not different from chance. Based on these results, final goal attributes were coded for each goal statement. Note that in final coding, we resolved disagreements among judges by assuming that differences were due to errors of omission (one judge indicated a code that the other did not).
Our next step involved combining scores across all of individual’s goal statements at baseline and separately at six-months, to characterize current priorities and concerns at each time of measurement. Our goal at this step was to achieve a parsimonious data reduction while retaining as much information as our data would permit. Our coding system yielded a binary vector describing the presence or absence of 24 different goal attributes for each goal statement. Recall that goal statements were elicited by six different motivational probes. Thus, for the 9 most prevalent codes, we calculated subtotals representing the occurrence of each goal attribute for statements elicited by each motivational theme. For example, this cross-classification allowed us to distinguish among goals about solving money problems, earning more money or learning to live more frugally. The 54 variables formed by the cells of this cross-classification of 9 major goal attribute codes by 6 motivational themes fully accounted for 74% of all responses. These represented our primary goal attributes. We reduced these 54 variables by conducting a two-stage principal components analysis, first summarizing endorsement rates of codes within each of the six motivational themes (retaining 57% to 89% of total variance in each set), and then combining the 32 first-order components from these six analyses in a single second-order principal components analysis (retaining 60% of variance among the first-order components). Second-order analysis yielded sixteen major goal attribute factors. These components are listed in Table 2.
The remaining 15 goal attribute codes were less prevalent, so we simply tallied the total number of times content codes occurred for each individual at a given time of measurement, without subtotaling by eliciting theme. Principal components analysis yielded seven relatively independent components after promax rotation, summarizing 58% of the variance among these 15 codes. These seven subsidiary goal attribute dimensions are also listed in Table 2. Substantively, we think of the primary goal content factors as capturing the individuals’ status in broadly shared areas of concern, while the subsidiary dimensions reflect more particular concerns that may nonetheless have an important influence on individuals’ appraisal of quality of life.
The other three parameters of quality of life appraisal were analyzed by a series of principal component analyses to map the items of sample experiences to 5 factors, standards of comparison to 3 factors, and combinatory algorithms to 7 factors. Generally, principal components with eigenvalues greater than or equal to 1 were retained. Table 3, Table 4, and Table 5 summarize the total variance accounted for by the retained eigenvalues, and the rotated factor loadings of the items. Standardized factor scores were calculated and entered into the analysis. Take the combinatory algorithms scale in Table 5 as an example, respondents were prompted “When you answered today, did you think more about …” and they rated the extent to which they thought about “Things that are disappointing to you”, “How hard it has been”, “Things that make you feel worried”, and so on. These three items had high factor loadings on the first factor, which was thus labeled as “Negative Experiences, Feelings & Worries”.
Because the QOL appraisal subscales were standardized, all sub-domains were thus mapped onto a comparable scale of mean zero and unit standard deviation. A respondent with a zero “reacting to recent flare-ups” score, for example, represents an appraisal through recent disease flare-ups at the sample average. Changes in QOL appraisal were thus operationalized as changes in the standardized scores. In principle, the changes in standardized scores can be thought of as changes in effect size units [17, 18], thereby simplifying comparisons made across multiple QOL appraisal domains on arbitrary raw scales. We felt that it would facilitate interpreting the changes in appraisal by considering a set of crude by practical cut-offs. We considered a 0.75 standard deviation change a ‘moderate’ change in appraisal, a 1.0 change a ‘moderately large’ change, and a 1.5+ change a ‘large’ change. The ‘large’ change is conveniently twice as large as the ‘moderate’ change. These cut-offs are more conservative than the conventional effect size indexes  (e.g., 0.80 as a ‘large’ effect), and we believe that they help track the numerous and complex pattern of splits in the rpart analysis.
Following Rapkin and Schwartz’  formulation, in order to examine HRQOL response shift, it was first necessary to derive a discrepancy score in the HRQOL, to determine how much the observed score differed from an expected value. Response shift arises when the observed changes in HRQOL scores deviate systematically from the expected HRQOL changes due to health-related events. A simplistic example illustrates the basic conceptual premises. If a person experiences more symptoms related to HIV/AIDS, and the increased symptoms are expected to reduce HRQOL by 10% (such as the predicted HRQOL change by a statistical model derived from large-scale surveys, controlling for other validated covariates), then an observed increase of 15% suggests response shift. Thus, if change in cognitive appraisal were able to explain these systematic discrepancies, that was indicative of response shift. We decided to focus on the mental component score of the SF36v.2 rather than the physical component for this demonstration, because our prior preliminary analysis showed that it is more sensitive to response shift .
In order to provide a highly conservative test of response shift, we used an ordinary least square regression to control baseline mental composite score for a wide range of possible predictors, including demographics and personal history (e.g., history of hard drug use and involvement in the criminal justice system), baseline health status, baseline frame of reference, baseline sampling, standards, and combinatory algorithm, change in health status variables, changes in number of self-reported symptoms, and intervening events in care. Details on how these covariates are assessed can be found in [2, 16]. Standardized residual scores controlling these predictors were computed and entered into the rpart analysis [20, 21, 22] to determine whether and how changes in cognitive appraisal could be used to explain these discrepancies.
The rpart [20, 21, 22] model fitted the standardized residual MCS scores in SF36v2  with 38 predictors representing the changes in appraisal variables between baseline and 6-months assessment: changes in 16 primary and 7 subsidiary goal content dimensions, 5 predictors on the sampling of experiences, 3 predictors on standards of comparisons, and 7 predictors on the combinatory algorithms. For ease of interpreting the magnitude of response shift, we divided our sample by the MCS discrepancy scores to three categories: 40% with the largest positive residuals (deemed “Positive” response shift), 40% with the largest negative residuals (“Negative” response shift), and 20% with residuals close to zero (“No Change”).
We followed the general approach in rpart analysis: first grow a complex tree and then prune the tree back by cross-validation [20, 21, 23, 24, 25, 26, 27, 28, 29]. Feldesman  is a highly-accessible tutorial on the different statistical computations for continuous and categorical outcome variables; it also outlines a few default model specifications in the complex tree: 1) stopping rule for a terminal node (< 20 observations); 2) criterion for tree pruning (“cost-complexity parameter”, CP=0.01); 3) validation by 10-fold cross-validation (1-SE rule for pruning by CP); 4) specification of priors (proportional to data counts) and 5) missing data are handled by surrogate splits. Details can be found in textbooks and are omitted here [15, 28, 30, 31].
Model performance was evaluated by a 3-class classification performance metric based on overall error rate in a confusion matrix, and also by pairwise area under the ROC curve analysis (AUC) [32, 33]. Our three-class classification was separated into six binary comparisons after the rpart classifier has been carried out. We calculated the AUC on “Positive” vs. “No Change” response shift, “Positive” vs. “Negative” response shift, and so on for all 6 pairwise ROCs. A single average AUC index was calculated, called the M function , to represent the overall model performance. We also entered the same 38 predictors in a multinomial logit model for comparison.
At this time of analysis, 619 individuals were recruited to this study, of which 443 were due for the 6 months assessments and 394 completed them (89%, follow-up data collection ongoing). Table 1 summarizes participant characteristics. Men and women were approximately evenly distributed, with diverse race and ethnicity backgrounds, low socioeconomic status, and an average age of 47.1 years and 11.6 years since the identification of HIV.
Figure 1 (a) shows the fullest rpart dendrogram, derived by accepting the default settings. The 10-fold cross-validation suggested pruning the tree back to only 9 terminal nodes (Figure 1 (b)). This was based on the 1 SE rule [20, 21], plotted in Figure 2, to find the least complex tree within 1 standard deviation of the minimal cross validation error. The pruned tree showed the lowest cross-validation error, beyond which tree complexity entailed no additional improvement.
Table 3 shows the confusion matrix of the 9-node tree and the model performance AUC measures of three alternative models. The 9-node tree made 243 correct classifications (62% accuracy, 95% CI = 39% – 80% by bootstrapping), which was superior to the 36% chance accuracy by the marginal 40-20-40 split. The pairwise AUC indexes show comparable performance between the 9-node rpart tree and the multinomial logit, with an overall AUC of 0.72. The 24-node tree consistently outperforms the pruned tree as well as the multinomial logit model. However, the cross-validation argued against it because of the low generalizability.
We now discuss the nine terminal nodes in Figure 1 (b) from left to right by interpreting the distinctions among groups which emerged in this analysis. The first group of 78 individuals in node 1 stood out because they reduced the salience of negative experiences by a moderately large amount. This group tended to have a high prevalence of positive discrepancies (47 out of 78). For the remaining 316 individuals in nodes 2 through 9, the salience of negative experiences in evaluating HRQOL was either maintained or increased (>= −1.04 sd). For succinctness we interpret non-large reduction in splits as roughly maintenance or possible increase. Reduction in salience of negative experiences alone was not sufficient to affect discrepancies in HRQOL. A combination of other cognitive variables comes into play. The second, third and fourth nodes were distinguished from the rest of the sample based on a moderate reduction in the extent to which they compared themselves to others. Group 4 represented a small subgroup that differed from groups 2 and 3 by a moderately large increase in goals related to solving problems associated with living situations and work. The majority of individuals in this small group demonstrated little or no discrepancy from expected change in psychological well-being, although three displayed negative discrepancies. For individuals in nodes 2 and 3, discrepancy was associated with moderate changes in goals related to independent functioning. Group 2 maintained or increased goals related to independence, which was associated with predominately positive response shift. Conversely, group 3 markedly reduced goals associated with independence, contributing to more negative psychological well-being than expected. It is noteworthy that groups 2, 3, and 4 were all affected by changes in specific goals related to problems with work or with maintaining independence. Such changes in frame of reference interact with changes in standards of comparison to produce a range of response shifts.
Nodes 5 – 9 either maintained or increased their tendency to compare themselves with others, as well as the salience of negative feelings and experiences. On the far right node 9, the largest group of 145 individuals identified in this analysis sample, stood out from the others because of maintained or increased concerns about monetary obligations and other external demands. This combination, greater salience of negative feelings and comparison of self to others along with increased demands, was clearly associated with marked negative discrepancies in reported quality of life. However, for individuals without concerns about reducing demands and obligations, other factors came into play. Node 8 represented a group that had moderately high reduction in goals on maintaining relationships by accepting others and improving their own outlook. This group tended to report changes in psychological well-being close to values predicted by baseline factors, health changes and events in care.
Individuals in the remaining groups 5, 6, and 7 all tended to increase or sustain their concerns about maintaining relationships and achieving a positive outlook. Again moving in from the right, node 7 contained a preponderance of individuals with negative response shift. Interestingly, this group reported a marked decrease in goals related to preventing or avoiding interpersonal and monetary concerns. Conceivably, these individuals wanted to stave off certain problems at baseline but later realized that this was untenable by 6 months, thus contributing to a negative response shift.
Nodes 5 and 6 share many features, including the salience of negative experiences, a tendency to compare oneself to others, goals to maintain relationships, improve one’s outlook, and avoid interpersonal and monetary problems, but not to reduce obligations or demands. The majority of these individuals, in node 5, tended to report more positive psychological well-being than expected as a result of efforts to manage communications during the interview. This factor, from the sampling experience domain reflects individuals’ efforts to edit their responses by not complaining too much, by giving their first reaction, and by trying to convey the seriousness of their situation. Increasing or sustaining this response set was associated with positive discrepancies in well-being. Alternatively, individuals in node 6 had reduced or abandoned efforts to manage communication during the interview. The majority of individuals in node 6 demonstrated negative response shift.
The rpart-derived model was useful in identifying aspects of response shift that was hard to detect through linear analysis. Rpart performed equally well as a multinomial logistic regression of the same predictors. Rpart provided a straightforward method that yielded more clinically interpretable results for identifying subgroups of response shift, which appeared to be mostly influenced by changes in emotion (e.g., ‘negative experiences and feelings’) and subjective norms (‘comparing to others’). Invoking comparisons with others played an important role in explaining discrepancies in SF36v2-MOS change which only became apparent when examined in conjunction with the salience of negative events. Thereafter, frames of reference and individual concerns came into play in response shift, including 5 of the 16 primary goal factors. Our findings also shed light on the potential for individuals to sample and report experiences that affect their HRQOL selectively, to manage communication during the interview. Permitting the interplay of cognitive change variables provides additional, complementary information about QOL response shift.
The present findings have bearings on application of QOL measures and on understanding and meeting patients’ needs. Individuals continually encounter new challenges and new opportunities, and factor these into how they self-evaluate their well-being. Individuals living with a chronic, life-threatening illness encounter many such challenges, and the stakes are high: An interpersonal conflict may interfere with an important source of social support; monetary or housing problems may strain an individual’s limited resources; challenges to independence may invoke personal fears of premature debilitation and mortality. These challenges may be even greater in the Medicaid HIV/AIDS population. Additionally, we show that some individuals adapt to challenges by altering their cognitions, by relinquishing goals or modifying expectations for what they are seeking (e.g., node 9 by ‘reducing practical & monetary obligations/demands’ and node 7 by ‘avoiding interpersonal & monetary concerns’). The distinction between nodes 7 and 9 is subtle, with node 9 emphasizing more on practical concerns such as reducing the financial obligations of paying bills. Individuals may selectively disengage from situations that they can no longer manage. These processes necessarily play out over time. It is not surprising that individuals’ cognitive criteria for the appraisal of psychological well-being and distress is quite fluid.
Cognitive assessment provides a way to take these wide variations in quality of life appraisal into account . We can use these methods to control response shift effects in evaluations of programs or treatments. For example, we might observe improvement in an individual’s emotional well-being if we take into account that they are presently engaged in solving housing problems and in boosting independence (e.g., note 2). Similarly, apparent reduction in emotional well-being might be reinterpreted in light of an individual’s selective emphasis on reducing practical and monetary demands (e.g., node 9). More fundamentally, it may be important to interpret the impact of disease and treatment on measures of cognitive appraisal. As our analysis demonstrates, there is a complex interplay among measures of appraisal and quality of life. It is important to understand when and how increased contact with the health system is associated with a sense of greater dependence, and when it is associated with increasing expectations and standards for self-evaluation. Change in the content and process of cognitive appraisal is a worthy patient-reported outcome domain in its own right.
Our results support the Rapkin and Schwartz model , in that cognitive variables helped to account for substantial HRQOL response shift in ways that were interpretable and consistent. However, several methodologic challenges remain. The QOL appraisal battery generates considerable, detailed descriptive data about the appraisal process. It is quite challenging to operationalize the process of describing the intermediaries of response shift, as we have attempted. There are inherent problems in CART methods , which originate from the fact that one predictor may win a particular split by only a small margin. This makes such splits somewhat arbitrary; errors cascade into subsequent splits, highlighting the importance of tree pruning by cross-validation . New methodological developments are available [35, 36, 37, 38], and may be helpful in future research on response shift. We hope that our study will prompt further theoretical and empirical work to improve on the description of the response shift phenomena.
Funding acknowledgment: the authors would like to acknowledge funding from the New York State Department of Health AIDS Institute (US Health Resources and Services Administration grant: 2X07 HA 0025-17) (BR) and the Weill Cornell Medical College Clinical and Translational Science Award (NIH UL1-RR024996) (YL).