Cancer Medicine

Open Access
ORIGINAL RESEARCH

Identifying patterns of adaptation in breast cancer patients

with cancer-­related fatigue using response shift analyses at
subgroup level
Maxime Salmon1, Myriam Blanchin1, Christine Rotonda2,3, Francis Guillemin2,4 & Véronique Sébille1,5
1Universityof Nantes, University of Tours, INSERM, SPHERE U1246, Nantes, France
2Universityof Lorraine, University of Paris Descartes, EA 4360 APEMAC, Nancy, France
3Center Pierre Janet, University of Lorraine, EA4630 APEMAC/EPSAM, Metz, France
4INSERM CIC, 1433 Clinical eidemiology – Nancy University Hospital, Nancy, France
5Department of Biostatistics, Nantes University Hospital, Nantes, France

Keywords Abstract
Breast cancer, fatigue, patient-reported
outcome measures, response-shift, structural Fatigue is the most prevalent symptom in breast cancer. It might be perceived
equation modeling differently among patients over time as a consequence of the differing patients’
adaptation and psychological adjustment to their cancer experience which can
Correspondence be related to response shift (RS). RS analyses can provide important insights
Véronique Sébille, University of Nantes,
on patients’ adaptation to cancer but it is usually assumed that RS occurs in
University of Tours, INSERM, SPHERE U1246,
Nantes, France. Tel: +33 (0)253009120;
the same way in all individuals which is unrealistic. This study aimed to identify
E-mail: [email protected] patients’ subgroups in which different RS effects on self-­reported fatigue could
occur over time using a combination of methods for manifest and latent vari-
Funding Informaiton ables. The FATSEIN study comprised 466 breast cancer patients followed over
This work was supported by the Institut a 2-­year period. Fatigue was measured with the Multidimensional Fatigue In-
National du Cancer [grant number ventory questionnaire (MFI-­20) during 10 visits. A novel combination of Mixed
INCA_6931].
Models, Growth Mixture Modeling, and Structural Equation Modeling was used
Received: 7 June 2017; Revised: 2 September
to assess the occurrence of RS in fatigue changes to identify subgroups display-
2017; Accepted: 5 September 2017 ing different RS patterns over time. An increase in fatigue was evidenced over
the 8-­month follow-­up, followed by a decrease between the 8-­and 24-­month.
Cancer Medicine 2017; 6(11):2562–2575 Four latent classes of patients were identified. Different RS patterns were detected
in all latent classes between the inclusion and 8 months (last cycle of chemo-
doi: 10.1002/cam4.1219 therapy). No RS was evidenced between 8-­and 24-­month. Several RS effects
were evidenced in different groups of patients. Women seemed to adapt dif-
ferently to their treatment and breast cancer experience possibly indicating
differing needs for medical/psychological support.

using the MFI-­20 [4–6]. These studies showed heteroge-

Introduction
neous results with different patterns of increase, decrease,
Cancer-­related fatigue is one of the most prevalent symp- or stability in fatigue over time. Moreover, all these studies
tom [1] in breast cancer patients. Cancer-­related fatigue assumed so-­ called longitudinal measurement invariance
is defined as a complex subjective state characterized by [7] assuming that patients respond consistently on patient-­
a reduction in physical and mental abilities affecting cancer reported outcomes (PRO) and that they are directly com-
patients, from time of diagnosis and throughout treat- parable over time, which can be questioned. Indeed, in
ment, as well as survivors [2]. One of the tools for meas- breast cancer, it is likely that patients might regularly
uring cancer-­related fatigue is the Multidimensional Fatigue adapt to their illness and, as a consequence, might give
Inventory (MFI-­ 20) [3]. Several studies have explored different answers to the questionnaires over time, not
fatigue change in breast cancer patients during treatments only because their fatigue has changed, but also because

2562 © 2017 The Authors. Cancer Medicine published by John Wiley & Sons Ltd.
This is an open access article under the terms of the Creative Commons Attribution License, which permits use,
distribution and reproduction in any medium, provided the original work is properly cited.
M. Salmon et al. Reponse-­Shift In Cancer-­Related Fatigue

their perception of what fatigue means to them has changed. reprioritization) to disentangle the contributions of RS
This phenomenon is often referred to as response shift effects and latent fatigue (unobserved fatigue level of
(RS) [8] which has been hypothesized to have three dif- patients) changes to the observed change.
ferent manifestations: recalibration (change in the patient’s
internal standards of measurements), reprioritization
Method
(change in the patient’s values), and reconceptualization
(change in the patient’s definition of the measured
Data collection
concept).
In case of RS, it might be problematic or even impos- The FATSEIN study [12] included breast cancer patients
sible to distinguish, without appropriate methodology, recruited in the cancer care centers of Nancy, Dijon, and
change in fatigue from RS effects. The assessment of Strasbourg in France and followed up for 24 months after
therapeutic interventions can then lead to inappropriate surgery. Informed consent was obtained from all individu-
results, poor power to detect intervention effects and als participants included in the study. Eligibility criteria
erroneous conclusions. At the same time, one can highlight included women newly diagnosed with invasive breast
the therapeutic importance of RS itself, allowing a better cancer, undergoing breast surgery as primary treatment,
understanding of how patients adjust to their illness. age ≥18 years, without history of other cancer, no other
Indeed, RS could be one of the goals of therapy in help- major disabling medical or psychiatric conditions, no
ing patients to cope with their disease and to live with previous chemotherapy or radiotherapy, no metastases,
it. Therefore, RS can be simultaneously viewed as meas- and no inflammatory breast cancer. The study has been
urement bias as well as an indication of a possible thera- approved by the institutional review board and is regis-
peutic benefit coming from some form of psychological tered at www.clinicaltrials.gov (NTC01064427).
adaptation or adjustment. Thus, it is important to assess Patients were recruited for a 2 year-­follow-­up including
the change experienced by patients by taking into account 10 visits (Appendix 1). Socio-­ demographic and medical
RS if appropriate, but also to detect this phenomenon data were collected before surgery. Cancer-­related fatigue
and quantify it in a reliable and unbiased manner. was measured with the MFI-­ 20 [13] over the 10 visits.
Most approaches proposed for RS detection and adjust- This questionnaire explores four domains of fatigue (physi-
ment in the appraisal of change in PRO over time are cal or mental fatigue, reduction in activities or motivation)
performed at the sample-­level such as Structural Equation and provides a global fatigue score. Higher scores indicate
Modeling (SEM) [9] or Item Response Theory (IRT) [10]. more reported fatigue. Patients also completed the QLQ-­
Hence, RS is considered on the overall sample of patients C30 quality of life (QoL) questionnaire [14] and the
regardless of individual characteristics. It is thus assumed State-­
Trait Anxiety Inventory (STAI-­ State) to measure
that the majority of the sample has been affected by the transient anxiety [15] over time. The STAI-­Trait [15] and
same change in the perception of fatigue over time. the LOT (LOT) [16] were completed at inclusion to
Nevertheless, we can suspect that among a sample, only measure enduring levels of trait anxiety and optimism.
some individuals might be affected by RS and different Higher scores for the State-­Trait Anxiety Inventory and
types of RS might affect different individuals to different for the LOT indicate more reported anxiety and optimism,
extent. An alternative approach has been proposed [11] respectively.
allowing the detection of RS and its possible time of
occurrence at a subgroup level, however, it does not allow
Statistical analysis
for the identification of the form of RS that might have
occurred (recalibration, reprioritization, reconceptualiza- Several steps were performed (Fig. 1). SAS version 9.4
tion). For this purpose, this method could be comple- and Mplus version 7.2. were used. Two-­sided P values<0.05
mented with other approaches such as SEM using Oort’s were considered statistically significant.
procedure [9] for identifying the different forms of RS. Firstly, the change in global fatigue score of the MFI-­20
However, to the best of our knowledge, it has never been was assessed using a mixed model from which the residu-
done despite its potential for detailed RS identification als were retained. The final model was selected using an
at a subgroup level. ascending strategy where time effect and sociodemographic
Our objective was to identify the possible different pat- covariates were added according to Wald tests (P < 0.05).
terns of RS in a sample of breast cancer patients. A Since it was expected that chemotherapy treatment could
combination of methods was proposed to assess the occur- modify fatigue change, it was added as a fixed effect, as
rence of RS in the evolution of self-­reported fatigue over well as its interaction with time. The addition of random
time and to identify subgroups of patients regarding dif- effects on the intercept or the slope was assessed using a
ferent patterns of two RS processes (recalibration, likelihood ratio test (LRT). The studentized residuals were

© 2017 The Authors. Cancer Medicine published by John Wiley & Sons Ltd. 2563
Reponse-­Shift In Cancer-­Related Fatigue M. Salmon et al.

Figure 1. Sequential steps for response shift (RS) analyses.

obtained by dividing the conditional residuals of the mixed possible clinically meaningful trajectories. When the num-
model by the estimate of their standard deviation; they ber of LC was identified, posterior probabilities of LC
were subsequently centered by removing their mean for membership were estimated for all patients. Each patient
each subject. This allowed pulling out the effect of a poor was assigned to the LC with the highest posterior prob-
fit (high or low residuals). Following Mayo et al. [11]., ability [17]. Lastly, the time points between which RS
the change in residuals quantifying the difference between could be experienced by patients were determined using
the observed global fatigue and what would have been the highest range of the residuals in each LC, and the
predicted based on all the variables significantly associated treatment history that could have triggered RS. The dis-
with fatigue change was considered as an indicator of RS. tribution of the socio-­demographic and medical variables,
It was assumed that the fluctuation in a patient’s residuals QoL, optimism and anxiety scores measured at inclusion
over time might be due to RS. Hence, we considered that and over time were compared between the LC using
a subject with centered residuals close to zero for all meas- analysis of variance (ANOVA), chi-­square tests, and mixed
urement occasions did not experience RS. models.
Secondly, different latent classes (LC) of homogeneous Thirdly, the occurrence of reprioritization, uniform/
centered residuals growth trajectories were identified as non-­ uniform recalibration RS in each of these LC and
well as the time points where RS was expected to occur its effect on observed fatigue changes measured by the
using Growth Mixture Models (GMM) [17]. A categorical scores of the four domains of the MFI-­ 20 between the
latent variable representing the different classes was previously selected time points were assessed using SEM
included in the model and LC membership was inferred [9] following a four-­step method which is briefly described.
from data. The growth parameters (intercept and slope) This procedure was performed on the overall sample and
could be different in each LC. The selection of the best in each LC allowing estimating three components of
fitting GMM and of the number of classes relied on the changes: observed changes, RS and latent change contri-
sample size–adjusted Bayesian information criteria (SABIC) butions to the observed change. Mathematical formulation
[18]. We constrained the minimum sample size to be at of the SEM model, models’ identification constraints, and
least 10% of the total sample size in all LC to enhance operationalization of RS (RS parameters) appear in
SEM convergence while keeping a maximum of different Appendix 2.

2564 © 2017 The Authors. Cancer Medicine published by John Wiley & Sons Ltd.
M. Salmon et al. Reponse-­Shift In Cancer-­Related Fatigue

Step 1 the LRT is significant, RS is assumed. In that case, the types

of RS are identified and estimated in step 3. Otherwise, overall
A “measurement model” (model 1) is first estimated
RS has not been detected. In this case, the most parsimoni-
(Fig. 2) with no across measurement constraints for RS
ous model 2 is kept, step 3 is skipped to go to step 4.
parameters: factor loadings (reprioritization RS), intercepts
(uniform recalibration RS) or residual variances (non-­
uniform recalibration RS). The common factor is measured Step 3
at two measurement occasions (Fatigue(1), Fatigue(2)); it
Starting with model 2, step-­by-­step inspection of RS param-
is assumed to explain the relationships between the observed
eters that significantly improve model fit gives model 3.
variables (domain scores at time t: physical fatigue (p.f(t)),
Across time correlations between residual factors are pos-
mental fatigue (m.f(t)), reduction in activities (r.a(t)), reduc-
sibly added followed by residual factor correlations at the
tion in motivation (r.m(t))). The model fit was considered
same measurement occasion, until satisfactory fit is reached.
acceptable if root mean square error of approximation
index (RMSEA) was <0.08, comparative fit index
(CFI) > 0.95, Tucker-­Lewis Index (TLI) > 0.95 and stand- Step 4
ardized root mean square residual index (SRMR) ≤0.08
This final step allows estimating and testing the latent
[19]; across time correlations between domain scores were
change in fatigue (model 4). The effect sizes [9] of the
possibly added to reach satisfactory fit.
latent change and of RS are computed as shown in
Appendix 2. Of note, non-­uniform recalibration is related
Step 2 to a change in residual variances that does not affect the
mean observed change; it may indicate that the change
A “no RS model” (model 2) is estimated with across meas-
in the scaling of the responses is not in the same direc-
urement constraints for all RS parameters. An overall test
tion or that it differs between patients.
for RS is performed with a LRT (model 1 vs. model 2). If

Figure 2. Measurement model in step 1. 𝜋d(t), 𝜏d(t) Δ(t)

d
: unobserved residuals variances, intercepts, factor loadings of domain d at time t, respectively.
Φ(1,2): across occasion covariance of the latent construct of fatigue.

© 2017 The Authors. Cancer Medicine published by John Wiley & Sons Ltd. 2565
Reponse-­Shift In Cancer-­Related Fatigue M. Salmon et al.

Results of anxiety measured by the STAI-­Trait were not significantly

different between the LC (P = 0.33) but the optimism scores
Sample characteristics of the LOT were lower in LC3 as compared to LC2 (P < 0.05).
During follow-­up, change in the observed transient anxiety
A sample of 466 women was recruited. Their mean age levels measured by the STAI-­State was not significantly dif-
was 57 years, they were usually married (75 %) with 1 ferent between the LC (P = 0.16). In contrast, the observed
or 2 children (63 %) and living with their families (80 %). changes in all the functional domains of the QoL question-
Women were mostly affected by a stage I and II cancer naire QLQ-­C30 were significantly different over time between
(84%). Most of them had a lumpectomy (78 %), a sen- the LC (Fig. 4). The observed mean scores in LC2 were
tinel node surgery (50 %), and a neo-­adjuvant treatment usually higher compared to the other LC (especially LC4)
based on radiotherapy with or without chemotherapy and and quite stable over time for the following dimensions:
hormone therapy. General Health (GH, Fig. 4A), Physical Functioning (PH,
Fig. 4B), Cognitive Functioning (CF, Fig. 4D), Social
Change in global fatigue, in residuals, and Functioning (SF, Fig. 4E), Role Functioning (RF, Fig. 4F).
identification of the latent classes Conversely, for these dimensions, in LC4, the observed mean
scores fluctuated more and usually decreased from baseline
The variables significantly associated with fatigue change to 8-­month and subsequently increased until the end of
were time, chemotherapy and marital status. Global fatigue follow-­up either reaching their initial baseline level (for the
significantly changed over time (P < 0.001). It increased GH dimension) or remaining below it (on average −8 points
from baseline to 4-­ month, remained stable between for SF, −7 for CF, −10 for PF and RF at 24-­ month as
4-­and 8-­ month and decreased during the last months compared to baseline). Moreover, the difference between the
of follow-­ up. At 4-­and 8-­ months follow-­up, married observed mean scores in LC4 and LC2 (LC4 minus LC2)
women treated with chemotherapy reported more fatigue ranged from −3 points at baseline to −22 at 8-­month, and
than married women without chemotherapy on average. was −5 points at the end of follow-­up on average. In latent
At the last follow-­up, the mean global fatigue scores were classes LC1 and LC3, the mean scores change in the GH,
close whatever the treatment group. PH, CF, SF, and RF dimensions were usually close to one
A model with four LC (LC1-­ LC4) displayed the lowest another. In contrast, for the Emotional Functioning (EF,
SABIC and satisfied the required minimum sample size in Fig. 4C) dimension, at baseline, the observed mean scores
each LC. Figure 3 shows the estimated means of the centered were lower in LC3 as compared to the other LC (−14 points,
residuals at each time of measurement for the four LC. −10, and −8 on average as compared to LC2, LC1, and
There were 154 (33%) women in LC1, 61 (13%) in LC2, LC4, respectively, P < 0.05). The mean scores subsequently
118 (25%) in LC3, and 112 (24%) in LC4. Twenty-­ one increased in LC2 and LC3 until the end of follow-­up (+16
patients could not be assigned to a LC because of missing and +20 points in LC2 and LC3 on average, respectively)
data on marital status (n = 9), treatment group (n = 9) or but remained lower in LC3 as compared to LC2 (−9 points
both (n = 3). At inclusion (Table 1), the mean ages were on average) at 24-­month.
significantly different (P < 0.05) between LC1 (58.3 years) Based on the mean trajectories of the centered residuals
and LC4 (54.4 years), and there were more patients without in the different LC (Fig. 3), and taking into account the
chemotherapy in LC1 (n = 83, 54%) as compared to other end of the chemotherapy treatment (8-­month after sur-
LC especially LC4 (n = 42, 38%). The mean enduring levels gery), baseline and 8-­month, and 8-­and 24-­month after

Figure 3. Evolution over time of the observed means of the centred residuals in each latent class.

2566 © 2017 The Authors. Cancer Medicine published by John Wiley & Sons Ltd.
M. Salmon et al. Reponse-­Shift In Cancer-­Related Fatigue

Table 1. Socio-­demographic and medical characteristics of study population at baseline in the whole sample and in each latent class.

Whole sample Number of

Characteristics N = 466 missing values LC1 (N = 154) LC2 (N = 61) LC3 (N = 118) LC4 (N = 112) P-­value
Socio-demographic
Age 57.0 (10.4) 9 58.3 (10.8) 58.3 (11.0) 56.9 (9.6) 54.4 (10.1) 0.0181
Marital status 12
Single 34 (8%) 15 (10%) 7 (12%) 5 (4%) 6 (5%)
Widowed 36 (8%) 15 (10%) 5 (8%) 9 (8%) 5 (5%) 0.414
Divorced 45 (10%) 17 (11%) 4 (7%) 13 (11%) 11 (10%)
Married 339 (75%) 107 (70%) 45 (74%) 91 (77%) 90 (80%)
Number of children 8
None 48 (11%) 14 (9%) 6 (10%) 15 (13%) 10 (9%)
1 or 2 287 (63 %) 92 (60%) 39 (65%) 74 (63%) 73 (66%) 0.813
>2 123 (27 %) 48 (31%) 15 (25%) 29 (25%) 28 (25%)
Educational level 58
End of the 213 (52 %) 72 (52%) 30 (59%) 57 (53%) 47 (48%)
compulsory school
High school degree 85 (21 %) 27 (19%) 11 (22%) 19 (18%) 27 (28%) 0.516
Higher education 110 (27 %) 40 (29%) 10 (20%) 32 (30%) 24 (25%)
Employment status 9
Employed 274 (60 %) 83 (54%) 34 (56%) 67 (57%) 80 (73%)
Unemployed 40 (9 %) 18 (12%) 4 (7%) 12 (10%) 5 (5%) 0.053
Pensioner 143 (31 %) 52 (34%) 23 (38%) 39 (33%) 25 (23%)
Medical
Type of breast surgery 3
Lumpectomy 362 (78 %) 125 (81%) 46 (75%) 89 (75%) 90 (80%) 0.596
Mastectomy 101 (22 %) 29 (19%) 15 (25%) 29 (25%) 22 (20%)
Stage of cancer 18
Stage I 239 (53 %) 87 (58%) 28 (47%) 55 (47%) 59 (54%)
Stage II 180 (40 %) 54 (36%) 29 (48%) 53 (46%) 42 (38%) 0.490
Stage III 29 (7 %) 8 (5%) 3 (5%) 8 (7%) 9 (8%)
Chemotherapy 12
No Chemotherapy 206 (45 %) 83 (54%) 29 (48%) 48 (41%) 42 (38%) 0.037
Chemotherapy 248 (55 %) 71 (46%) 32 (53%) 70 (59%) 70 (63%)
PRO measures
Trait anxiety STAI-Trait 47.9 (4.5) 40 47.7 (4.4) 47.2 (4.7) 48.3 (4.6) 48.2 (4.4) 0.333
Optimism LOT 19.8 (5.4) 77 20.4 (4.9) 21.1 (5.3) 18.7 (5.4) 19.6 (5.8) 0.0352

LC1: latent class 1, LC2: latent class 2, LC3: latent class 3, LC4: latent class 4. PRO: Patient-­Reported Outcomes; Mean (standard deviation) for con-
tinuous data, frequency (percentage) for categorical data;
1ANOVA, post hoc tests significant between LC1 and LC4;
2ANOVA, post hoc tests significant between LC2 and LC3; categorical data were compared with chi-­square tests.

surgery were considered as the two time periods where whole sample are shown as well as the effect sizes of the
RS could have occurred. contributions of RS and latent changes to the observed
changes (assuming that RS contribution + latent change
contribution ≅ observed changes). Several forms of RS
Response shift analyses were identified in each LC and in the whole sample. In
LC1, reprioritization (decrease of the factor loadings) was
Baseline to 8 months after surgery
detected. Thus, 8 months after surgery, women considered
Appendix 3 shows how the SEM fitted during the four-­ physical fatigue as less important to characterize the latent
step method proposed by Oort [9] (from model 1 to final construct of fatigue as compared to baseline. Uniform
model 4). In each LC, the fit of model 4 was satisfactory, recalibration in physical fatigue and mental fatigue was
all the fit indices were acceptable: RMSEA < 0.08, also detected (decrease in the intercepts). It indicated that
CFI > 0.95, TLI > 0.95, and SRMR ≤0.08. Table 2 shows women tended to score lower on the items of physical
the results of the RS analyses using SEM [9] in each latent or mental fatigue at 8 months compared to baseline despite
class and in the whole sample. The types of RS that have an increase in the mean level of latent fatigue (effect
been identified in the different latent classes and in the sizes = 0.48 and 0.51 for latent changes in physical and

© 2017 The Authors. Cancer Medicine published by John Wiley & Sons Ltd. 2567
Reponse-­Shift In Cancer-­Related Fatigue M. Salmon et al.

Figure 4. Change in the scores of the functional domains of the QLQ-­C30 quality of life questionnaire over the follow-­up.

mental fatigue, respectively). The effect sizes of the observed non-­uniform recalibration was evidenced with an increase
changes were thus both reduced by RS effects (effect in the residual variance of the domain “Motivation reduc-
sizes = 0.17 and 0.12 for observed changes in physical tion” at 8 months.
and mental fatigue, respectively). Finally, a decrease in the In LC3 and LC4, decreases in the factor loadings of
residual variance of the mental fatigue at 8 months was the physical fatigue and in the intercepts of the mental
evidenced (non-­uniform recalibration). fatigue were observed. Moreover, in LC4, non-­ uniform
In LC2, reprioritization was evidenced in physical fatigue recalibration occurred (increase in the residuals variances
(increase in the factor loadings); this domain was more of the domain “Motivation reduction”). In these LC, both
indicative in the construct of latent fatigue at 8 months observed mental and physical fatigue changes were reduced
of follow-­up as compared to baseline. Effect sizes of by RS effects. Thus, in all LC, observed changes in physi-
observed changes were thus higher than those of latent cal and mental fatigue underestimated changes in latent
changes. Uniform recalibration in mental fatigue was also fatigue, except for physical fatigue in LC2.
observed (decrease in the intercepts). The observed mental Furthermore, in LC1 and LC4, at 8-­month of follow-
fatigue effect size was thus reduced by RS. Lastly, ­up, reprioritization RS effects lead all factor loadings to

2568 © 2017 The Authors. Cancer Medicine published by John Wiley & Sons Ltd.
M. Salmon et al. Reponse-­Shift In Cancer-­Related Fatigue

Table 2. Response shift analyses using Structural Equation Modeling in the four latent classes and the whole sample between baseline and 8 months
after surgery.

Latent class Response shift Effect size

Latent change
R UR NUR Observed changes RS contribution contribution

LC1
Physical fatigue x x 0.17 −0.32 0.48
Mental fatigue x x 0.12 −0.38 0.51
Activity reduction 0.44 0.44
Motivation reduction 0.42 0.42
LC2
Physical fatigue x 0.72 0.11 0.61
Mental fatigue x 0.27 −0.34 0.61
Activity reduction 0.55 0.55
Motivation reduction x 0.50 0.50
LC3
Physical fatigue x 0.53 −0.11 0.64
Mental fatigue x 0.40 −0.34 0.74
Activity reduction 0.61 0.61
Motivation reduction 0.55 0.55
LC4
Physical fatigue x 0.36 −0.23 0.58
Mental fatigue x 0.38 −0.26 0.63
Activity reduction 0.54 0.54
Motivation reduction x 0.63 0.63
Whole sample
Physical fatigue x 0.41 −0.08 0.49
Mental fatigue x x 0.27 −0.25 0.52
Activity reduction x 0.59 0.14 0.45
Motivation reduction x 0.43 0.43

LC1, latent class1; LC2, latent class 2; LC3, latent class 3; LC4, latent class 4; R: reprioritization, UR: uniform recalibration, NUR, non-­uniform
recalibration, RS: response shift.

become closer to each other meaning that women gave in LC1 as compared to the other LC for “physical
the same importance to all domains of fatigue at follow- fatigue” (−0.32 in LC1 and ranging from −0.11 to +0.11
up. The parameters estimates of model 4 for LC1 are
­ in the other LC) and “mental fatigue” (−0.38 in LC1
shown in Figure 5. Parameters separated by a slash rep- and ranging from −0.34 to −0.26 in the other LC).
resent first and second measurement occasion estimates; The only positive effect size of RS contribution was
all other parameters were equal across measurement occa- observed in LC2 for “physical fatigue” which became
sions. Three types of RS were detected indicated in bold: more indicative of the fatigue construct (reprioritization)
reprioritization in physical fatigue (decrease in factor but it was counterbalanced by uniform recalibration in
loadings), uniform recalibration in mental and physical “mental fatigue” which reduced the observed change in
fatigue (decrease in the intercepts) and non-­ uniform this dimension (0.27) which was lower than the observed
recalibration in mental fatigue (decrease in the residual changes in LC3 (0.40) and LC4 (0.38). Finally, uniform
variances). recalibration in “mental fatigue” was the lowest in LC4
In summary, the changes in the mean level of latent in absolute value (RS contribution = −0.26) as compared
fatigue in all dimensions increased in all LC (all the to the other LC (RS contribution ranging from −0.34
effect sizes were positive). However, it seems that LC1 to −0.38).
corresponded to the patients with the lowest increase Lastly, in the whole sample, reprioritization was detected
in fatigue (effect sizes ranging from 0.42 for “reduction for physical fatigue (decrease in the factor loadings), uni-
in motivation” to 0.51 for “mental fatigue”) as compared form recalibration for mental fatigue (decrease in the inter-
to the other LC (effect sizes ranging from 0.50 for cepts) and reduction in activities (increase in the intercepts).
“reduction in motivation” in LC2 to 0.74 for “mental Non-­ uniform recalibration was also detected for mental
fatigue” in LC3). Moreover, it also seems that the con- fatigue and reduction in motivation with a decrease and
tribution of RS was more important (in absolute value) an increase in the residual variances, respectively.

© 2017 The Authors. Cancer Medicine published by John Wiley & Sons Ltd. 2569
Reponse-­Shift In Cancer-­Related Fatigue M. Salmon et al.

Figure 5. Parameters estimates of model 4 in latent class 1 between baseline and 8 months after surgery.

might receive less chemotherapy and that they could be

Eight months after surgery to the
more prone to RS [20, 21]. However, the reverse has
end of the follow-­up
also been reported regarding the association between RS
The test of overall RS was not significant for all LC and and younger age [22] and more research is needed on
the whole sample; hence RS was not evidenced between this topic. LC2 was the smallest latent class and comprised
these two time points. The latent and observed changes patients with the same age mean as those in LC1, the
were therefore the same. A decrease in fatigue was observed highest mean optimism score (P < 0.05 as compared to
with the following effect sizes: −0.19, −0.03, −0.41, −0.58, LC3), usually the highest levels of QoL over the follow-
and −0.29 in LC1, LC2, LC3, LC4, and in the whole ­up and for whom reprioritization RS went in an opposite
sample, respectively. direction as compared to the other LC indicating that
“physical fatigue” became more indicative of the fatigue
construct at 8-­ month compared to baseline but with
Discussion
recalibration RS in “mental fatigue” also being evidenced
Different patterns of RS effects on self-­reported cancer-­ as in the other LC. LC3 comprised patients with the
related fatigue were identified in subgroups of breast cancer highest increase in latent fatigue (“physical,” “mental
patients. Four LC were identified where different forms fatigue,” and “activity reduction”), the same recalibration
of RS and effect sizes were evidenced between baseline RS effects in “mental” fatigue” as compared to LC2 but
and 8 months after surgery (1 month after the last cycle with a lower mean optimism score and a lower QoL at
of chemotherapy). In contrast, no RS was evidenced baseline regarding emotional functioning as compared to
between 8 and 24 months after surgery. In all LC and the other LC. During the second time period (from 8 to
in the whole sample an increase in the latent fatigue was 24 months), a decrease in fatigue was observed in all LC
evidenced over the 8-­ month follow-­ up, followed by a but it was more marked in LC4 (effect size = −0.58) as
decrease between 8-­and 24-­month. During the first time compared to the other LC where the effects sizes ranged
period (from baseline to 8 months), the LC can be ten- from −0.03 in LC2 to −0.41 in LC3. Hence, during this
tatively characterized as follows. LC1 corresponds to the time period after the end of treatments, patients in LC4
oldest patients receiving less chemotherapy and showing who were younger seem to have a better ability to recover
the lowest increase in latent fatigue and the highest RS from fatigue.
effects. In contrast, LC4 corresponds to the youngest During the first time period, in all LC and the whole
patients receiving more chemotherapy, showing the lowest sample, uniform recalibration RS was evidenced in mental
recalibration RS effects, and usually the poorest QoL fatigue. This led women to score lower 8 months after
between 4 and 8 months. Other studies using latent class surgery as compared to baseline on this domain, despite
analysis to identify subgroups of cancer patients with dif- a mean increase in latent fatigue. Uniform recalibration
fering symptom experiences have also reported that, as was also detected for physical fatigue in the same direc-
compared to older patients, younger patients often display tion but only in LC1. Such uniform recalibration RS could
the highest levels of symptoms and the lowest QoL levels indicate some form of psychological adjustment such as
which could be explained by the fact that older patients adaptation [23] to the mental and physical fatigue

2570 © 2017 The Authors. Cancer Medicine published by John Wiley & Sons Ltd.
M. Salmon et al. Reponse-­Shift In Cancer-­Related Fatigue

experienced by women over time leading them to score patients in LC4 who were younger and received more
lower on these domains after the last cycle of chemo- chemotherapy displayed the lowest recalibration RS and
therapy despite an increasing level of fatigue. Recalibration poor QoL. Some limitations and paths for future research
RS in fatigue was similarly observed in other studies in can be outlined. The change in residuals was considered
breast cancer patients where the then-­test approach [22, as an indicator of RS [11], but residuals could change
24] was mostly used. This approach uses a retrospective over time for other reasons such as poor model fit.
(then-­test) self-­
assessment of the pretest level initially Nevertheless, the residuals of the mixed model were close
measured at baseline along with a posttest assessment to zero at each time so the model had a good fit
(e.g., 8 months after surgery). In this approach, the dif- (Appendix 4). The choice of the number of LC relied
ferences between pretest and then-­ test assessments and on the SABIC that has been shown to be more efficient
between posttest and then-­ test assessments are assumed than several other criteria such as AIC and BIC to detect
to represent recalibration RS and latent change, respec- the true number of LC in simulation studies [18]. Yet,
tively. Due to its simplicity, this approach is still used women were assigned to a LC according to their highest
but has some pitfalls such as requiring an additional posterior probability. The uncertainty of classification was
questionnaire, detecting only recalibration RS, being prone therefore not taken into account [31]. There have been
to recall bias [25, 26]. Thus, SEM [9] or other approaches some recommendations in the mixture modeling literature
such as IRT [10] are sometimes now more often to deal with this issue [32, 33] but their feasibility and
preferred. performance are unknown to date in SEM for RS analyses.
In the 1st, 3rd and 4th LC, reprioritization RS was All estimation methods (ML and REML) assumed ignor-
detected in the physical fatigue domain which became able missing data [34]. In case of non-­ignorable missing
less indicative of the latent construct of fatigue at month data, the probability of missingness depends on unobserved
8 compared to baseline whereas it went in the opposite data (patients might be too tired to fill in the fatigue
direction in LC2. Reprioritization RS was also observed questionnaire) which can lead to biased estimates of RS
in other studies focusing on QoL in prostate and breast and change in PRO [35]. Although the underlying missing
cancer patients [27, 28] where change in priorities and data mechanism is uncertain, in our sample, the com-
importance given to social or physical domains were evi- parison between patients who had missing data or not
denced over time as being more or less indicative of QoL did not reveal any significant differences on covariates
or subjective well-­being. distributions.
In all these studies RS was investigated at a sample The results of our study showed that RS was experienced
level, assuming that this phenomenon is experienced in in all LC and in the whole sample but with different
the same way for the majority of the sample which might forms and extent. Moreover, women also experienced strong
not be very realistic. One can indeed expect that there tiredness and probably had to adapt to this situation which
is significant variability between patients due to different might be related to the uniform recalibration RS effects
experiences and personality traits and that RS, often related that were evidenced especially in older patients receiving
to patients’ adaptation to illness might not have similar less chemotherapy. Furthermore, in all LC, if RS had not
manifestations in all individuals. To date, quite a few been taken into account, observed changes in physical and
studies examined RS at more individual level aside from mental fatigue would have underestimated changes in latent
Mayo et al. [11]. who used mixed models and GMM in fatigue, except for physical fatigue in LC2. In conclusion,
a sample of stroke patients but without distinguishing this study confirmed that RS can occur in different ways
the different forms of RS and Blanchin et al. [29]. who within a sample. The proposed approach allows taking
used IRT and Guttman errors [30] to detect discrepancies into account several aspects of RS by distinguishing several
in respondent’s answers compared to some expected groups of women who might have adapted differently to
response pattern (e.g. no RS) in a sample of hospitalized their treatment and illness possibly indicating differing
chronically ill patients. In the latter, RS was investigated needs for medical/psychological support. Special attention
at item-­level within a single dimension of the SF-­36 ques- might be given to younger patients with more chemo-
tionnaire (General Health). therapy treatment who might have a poorer fatigue experi-
In our study, different forms of RS could be assessed ence and QoL during therapy and could benefit from
at a sub-­ sample level on several dimensions of fatigue. psychological support helping them to cope and to better
Patients in LC1 seemed to be more prone to recalibration adjust with their symptom experience.
RS in physical and mental fatigue which could be associ-
ated with a better adaptation to illness and to its symp-
Conflict of Interest
toms. This might related to the fact that these patients
were older and received less chemotherapy. In contrast No conflict of interest disclosures from any authors.

© 2017 The Authors. Cancer Medicine published by John Wiley & Sons Ltd. 2571
Reponse-­Shift In Cancer-­Related Fatigue M. Salmon et al.

References
European Organization for Research and Treatment of
1. Berger, A. M., L. H. Gerber, and D. K. Mayer. 2012. Cancer QLQ-­C30: a quality-­of-­life instrument for use in
Cancer-­related fatigue: implications for breast cancer international clinical trials in oncology. J. Natl Cancer
survivors. Cancer 118:2261–2269. Inst. 85:365–376.
2. Stone, P. C., and O. Minton. 2008. Cancer-­related 15. Spielberger, C. D. 1989. State-trait anxiety inventory: a
fatigue. Eur. J. Cancer 44:1097–1104. comprehensive bibliography. Consulting Psychologists
3. Smets, E. M., B. Garssen, B. Bonke, and J. C. De Haes. Press, Mountain View, CA.
1995. The multidimensional Fatigue Inventory (MFI) 16. Scheier, M. F., and C. S. Carver. 1985. Optimism, coping,
psychometric qualities of an instrument to assess and health: assessment and implications of generalized
fatigue. J. Pshychosom. Res. 39:315–325. outcome expectancies. Health Psychol. 4:219–247.
4. Liu, L., M. Rissling, L. Natarajan, L. Fiorentino, P. J. 17. Muthen, B., and L. K. Muthen. 2000. Integrating
Mills, J. E. Dimsdale, et al. 2012. The longitudinal person-­centered and variable-­centered analyses: growth
relationship between fatigue and sleep in breast cancer mixture modeling with latent class analysis. Alcohol.
patients undergoing chemotherapy. Sleep 35:237–245. Clin. Exp. Res. 24:882–891.
5. Rissanen, R., C. Arving, J. Ahlgren, and K. Nordin. 18. Tofighi, D., and C. K. Enders. 2008. Identifying the
2014. Group versus individual stress management correct number of classes in growth mixture models.
intervention in breast cancer patients for fatigue and Pp. 317–341 in G. R. Hancock, K. M. Samuelsen,
emotional reactivity: a randomised intervention study. (eds.). Advances in latent variable mixture models.
Acta Oncol. 53:1221–1229. Information Age, Greenwich, CT.
6. Molassiotis, A., J. Bardy, J. Finnegan-John, P. 19. Schermelleh-Engel, K., H. Moosbrugger, and H. Müller.
Mackereth, D. W. Ryder, J. Filshie, et al. 2012. 2003. Evaluating the fit of structural equation models:
Acupuncture for cancer-­related fatigue in patients with tests of significance and descriptive goodness of fit
breast cancer: a pragmatic randomized controlled trial. measures. Methods Psychol. Res. 8:23–74.
J. Clin. Oncol. 30:4470–4476. 20. Astrup, G. L., K. Hofsø, K. Bjordal, M. G. Guren, I.
7. Widaman, K. F., E. Ferrer, and R. D. Conger. 2010. Vistad, B. Cooper, et al. 2017. Patient factors and
Factorial invariance within longitudinal structural quality of life outcomes differ among four subgroups of
equation models: measuring the same construct across oncology patients based on symptom occurrence. Acta
time. Child Dev Perspect. 4:10–18. Oncol. 56:462–470.
8. Sprangers, M. A., and C. Schwartz. 1999. Integrating 21. Kober, K. M., B. A. Cooper, S. M. Paul, L. B. Dunn, J.
response shift into health-­related quality of life research: D. Levine, F. Wright, et al. 2016. Subgroups of
a theoretical model. Soc. Sci. Med. 48:1507–1515. chemotherapy patients with distinct morning and evening
9. Oort, F. J. 2005. Using structural equation modeling to fatigue trajectories. Support. Care Cancer 24:1473–1485.
detect response shifts and true change. Qual. Life Res. 22. Andrykowski, M. A., K. A. Donovan, and P. B. Jacobsen.
14:587–598. 2009. Magnitude and correlates of response shift in
10. Guilleux, A., M. Blanchin, A. Vanier, F. Guillemin, B. fatigue ratings in women undergoing adjuvant therapy for
Falissard, C. E. Schwartz, et al. 2015. RespOnse shift breast cancer. J. Pain Symptom Manage. 37:341–351.
algorithm in item response theory (ROSALI) for response 23. von Blanckenburg, P., U. Seifart, N. Conrad, C. Exner,
shift detection with missing data in longitudinal patient-­ W. Rief, and Y. Nestoriuc. 2014. Quality of life in
reported outcome studies. Qual. Life Res. 24:553–564. cancer rehabilitation: the role of life goal adjustment.
11. Mayo, N. E., S. C. Scott, N. Dendukuri, S. Ahmed, and S. Psychooncology 23:1149–1156.
Wood-Dauphinee. 2008. Identifying response shift 24. Dabakuyo, T. S., F. Guillemin, T. Conroy, M. Velten,
statistically at the individual level. Qual. Life Res. D. Jolly, M. Mercier, et al. 2013. Response shift effects
17:627–639. on measuring post-­operative quality of life among breast
12. Rotonda, C., F. Guillemin, F. Bonnetain, and T. Conroy. cancer patients: a multicenter cohort study. Qual. Life
2011. Factors correlated with fatigue in breast cancer Res. 22:1–11.
patients before, during and after adjuvant chemotherapy: 25. Schwartz, C. E., and M. A. Sprangers. 2010. Guidelines
the FATSEIN study. Contemp. Clin. Trials. 32:244–249. for improving the stringency of response shift research
13. Gentile, S., J. C. Delarozière, F. Favre, R. Sambuc, and using the thentest. Qual. Life Res. 19:455–464.
J. L. San Marco. 2003. Validation of the French 26. Blome, C., and M. Augustin. 2015. Measuring change in
‘multidimensional fatigue inventory’ (MFI 20). Eur. J. quality of life: bias in prospective and retrospective
Cancer Care (Engl).12:58–64. evaluation. Value Health 18:110–115.
14. Aaronson, N. K., S. Ahmedzai, B. Bergman, M. 27. Gerlich, C., M. Schuler, M. Jelitte, S. Neuderth, M.
Bullinger, A. Cull, N. J. Duez, et al. 1993. The Flentje, M. Graefen, et al. 2016. Prostate cancer

2572 © 2017 The Authors. Cancer Medicine published by John Wiley & Sons Ltd.
M. Salmon et al. Reponse-­Shift In Cancer-­Related Fatigue

patients’ quality of life assessments across the primary on van Leeuwen et al.. Qual. Life Res.
treatment trajectory: ‘True’ change or response shift? 2013:2625–2629.
Acta Oncol. 55:814–820. 32. Bolck, A., M. Croon, and J. Hagenaars. 2004.
28. Tessier, P., M. Blanchin, and V. Sébille. 2017. Does the Estimating latent structure models with categorical
relationship between health-­related quality of life and variables: one-­step versus three-­step estimators. Polit.
subjective well-­being change over time? An exploratory Anal. 12:3–27.
study among breast cancer patients. Soc. Sci. Med. 33. Vermunt, J. K. 2010. Latent class modeling with
174:96–103. covariates: two improved three-­step approaches. Polit.
29. Blanchin, M., V. Sébille, A. Guilleux, and J. B. Hardouin. 2016. Anal. 18:450–469.
The Guttman errors as a tool for response shift detection at 34. Little, R. J. A., Rubin, D. B.2002. D. B. Statistical
subgroup and item levels. Qual. Life Res. 25:1385–1393. analysis with missing data. 2nd ed. John Wiley & Sons,
30. Sijtsma, K., and I. W. Molenaar. 2002. Introduction to Inc., Hoboken, New Jersey.
nonparametric item response theory(1st ed., Vol. 5). 35. Swartz, R. J., C. Schwartz, E. Basch, et al. 2011. The
Sage: Thousand Oaks. king’s foot of patient-­reported outcomes: current
31. McIntosh, C. N. 2012. Pitfalls in subgroup analysis practices and new developments for the measurement of
based on growth mixture models: a commentary change. Qual. Life Res. 20:1159–1167.

APPENDIX 1

Visit 0 Visit 1
Visit
Visit 2 Visit 3 Visit 4 Visit 5 Visit 6 Visit 7 Visit 8 Visit 9
3
End of
RT RT

CT CT4 CT6
CT1 CT2
group

Inclusion 7 8 12 18 24
months months months months months
2
1 month
months
4 6
No CT months months
group End of
RT RT

Flowchart of the treatments timeline. CT: chemotherapy; CT #: evaluation at the first day of the cycle of chemotherapy
n°#; RT: evaluation just before the beginning of radiotherapy; End of RT: evaluation during the last week of
radiotherapy.

© 2017 The Authors. Cancer Medicine published by John Wiley & Sons Ltd. 2573
Reponse-­Shift In Cancer-­Related Fatigue M. Salmon et al.

APPENDIX 2
( )
Ω(1,1) Ω(2,1)
Ω=
Ω(1,2) Ω(2,2)
where Ω(11) and Ω(22) are the residual factor variance-­
Mathematical formulation of the SEM covariance matrices at two measurement occasions, and
model, models’ identification in the Ω(12) and Ω(21) are the residual factor variance-­covariance
different steps, RS parameters, and matrices across occasions. Usually, no correlation is
effect sizes assumed between residuals at a given time but they can
be added to improve model fit.
Underlying model
This model is based on SEM. For the ith patient, we can Identification of the models
write:
In order to identify model 1 in step 1 the means of the
Wi = 𝜏 + Δ𝜉i + 𝜋i common factor are set equal to zero, and the variances
Where Wi is the vector of observed domain scores (physi- of the common factor to 1, at all measurement occasions.
cal and mental fatigue, reductions of activities and moti- All other parameters are freely estimated across time. To
vation), τ the matrix of intercepts, Δ the matrix of loading identify model 2 in step 2, the mean and variance of the
factors, ξi the vector of unobserved common factor (the common factor are constrained to be equal to 0 and 1
latent construct fatigue), πi the vector of unobserved at the first measurement occasion, respectively.
residuals factors.
In SEM, models are usually written as mean and covari- RS parameters
ance patterns as follows:
RS is operationalized in the following way where 𝜋d(t) , 𝜏d(t) , Δ(t)
d
Mean(W) = 𝜇 = 𝜏 + Δa are the unobserved residuals variances, intercepts, factor
where a is the vector of common factor means. loadings of domain d at time t, respectively. Φ(1,2) is the
The covariance of the fatigue domains can be written across occasion covariance of the latent construct of fatigue.
as Reprioritization corresponds to Δ(1) d
≠ Δ(2)
d
for domain d,
∑ because a change in factor loadings indicates that a domain
Cov(W (1) ,W (2) ) = = Δ(1) ΦΔ(2) + Ω of fatigue has become less or more indicative of the latent
fatigue construct during follow-­up. Respondents could have
where Φ is the covariance ( matrix ) of the unobserved com- also changed their interpretation of the response scale op-
mon factor (Φ = Cov 𝜉 (1) , 𝜉 (2) ), and Ω the covariances
tions during follow-­up showing recalibration. If this change
matrix ( of the unobserved residual factors
) occurs in the same direction and to the same extent for
(Ω = Cov 𝜋 (1) , 𝜋 (2) at the two measurement occasions (1)
all items of a domain, a uniform recalibration is assumed
and (2).
and corresponds to 𝜏d(1) ≠ 𝜏d(2) If the change occurred in
The Δ and τ matrices can be written as:
different directions or extents, non-­uniform recalibration
( (1)
Δ 0
) ( (1) )
𝜏 is assumed and corresponds to 𝜋d(1) ≠ 𝜋d(2)
Δ= and 𝜏 =
0 Δ(2) 𝜏 (2)
Effect sizes
where τ(1) and τ(2) are the vectors of the intercepts and
Δ(1) and Δ(2) the two vectors of the factor loadings at The estimated changes in observed fatigue (𝜇̂ (2) − 𝜇̂ (1)) is
the two measurement occasions. divided into three components.
The mean and variance structures of the common 𝜇̂ (2) − 𝜇̂ (1) = (𝜏̂ (2) − 𝜏̂ (1) ) + (Δ ̂ (1) )̂a(2) + Δ
̂ (2) − Δ ̂ (2) × â (2)
factors are:
where 𝜏̂ (2) − 𝜏̂ (1) represents the contribution of the uniform
( (1) )
a
( (1,1)
Φ Φ(2,1)
) recalibration to the observed changes in fatigue, (Δ ̂ (1) )̂a(2)
̂ (2) − Δ
a= and Φ = the contribution of the reprioritization, Δ ̂ × â the con-
a(2) Φ(1,2) Φ(2,2)
(2) (2)

tribution of the latent changes with a(2) representing the

where a(1) and a(2) are the means of the two common change in the construct of fatigue across time (since a(1)=0).
factors (the means of the fatigue), Φ(11) and Φ(22) the The effects size of each component is computed by dividing
two variances of the common factors construct at each it by the estimated standard deviation of the observed changes
measurement occasions, and Φ(12) and Φ(21) the covari-
√
𝜎̂ (1,1) + 𝜎̂ (2,2) − 2 × 𝜎̂ (1,2) where 𝜎̂ (1,1) and 𝜎̂ (2,2) are the estimated
ances of the common factors (Φ(12) = Φ(21)). variances, and 𝜎̂ (1,2) an estimate of across time covariances
The matrix of residuals factor variances can be written as of the observed change in model 4.

2574 © 2017 The Authors. Cancer Medicine published by John Wiley & Sons Ltd.
 M. Salmon et al. Reponse-­Shift In Cancer-­Related Fatigue

APPENDIX 3

Fit of the Structural Equation Models in the different latent classes and in the whole sample

Fit index

RMSEA CFI TLI SRMR

LC1 Model 1 0.04 0.99 0.99 0.04

Model 2 0.12 0.91 0.91 0.20
Model 3 0.04 0.99 0.99 0.06
Model 4 = Model 3 0.04 0.99 0.99 0.06
LC2 Model 1 0.10 0.97 0.95 0.06
Model 2 0.16 0.87 0.86 0.04
Model 3 0.08 0.97 0.97 0.08
Model 4 0.07 0.98 0.97 0.08
LC3 Model 1 0.09 0.98 0.96 0.05
Model 2 0.11 0.94 0.94 0.15
Model 3 0.12 0.91 0.91 0.20
Model 4 0.07 0.98 0.97 0.05
LC4 Model 1 0.04 0.99 0.99 0.04
Model 2 0.07 0.97 0.96 0.20
Model 3 0.04 0.99 0.99 0.06
Model 4 = Model 3 0.04 0.99 0.99 0.06
Whole sample Model 1 0.08 0.97 0.96 0.04
Model 2 0.12 0.91 0.91 0.19
Model 3 0.07 0.98 0.97 0.04
Model 4 = Model 3 0.07 0.98 0.97 0.04

LC, latent class; RMSEA, root mean square error of approximation; CFI, comparative fit index, TLI, Tucker-­Lewis index; SRMR, standardized root
mean square residuals.

APPENDIX 4
1.5

1
Centred residuals

0.5

–0.5

–1

–1.5
0 5 10 15 20 25
Time after surgery (months)

Figure: Mean of the centered residuals of the mixed model

© 2017 The Authors. Cancer Medicine published by John Wiley & Sons Ltd. 2575