Mediation analysis of the relationship between institutional research activity and patient survival.

0.43 MB·English

by  Rochon, Justine

Rochonetal.BMCMedicalResearchMethodology2014,14:9
http://www.biomedcentral.com/1471-2288/14/9
RESEARCH ARTICLE Open Access
Mediation analysis of the relationship between
institutional research activity and patient survival
Justine Rochon1*, Andreas du Bois2 and Theis Lange3
Abstract
Background: Recentstudies have suggested that patientstreated in research-active institutions have better
outcomesthanpatientstreatedinresearch-inactiveinstitutions.However,littleattentionhasbeenpaidtoexplaining
such effects, probably because techniques for mediation analysis existing so far have not been applicable to survival
data.
Methods: We investigated the underlying mechanisms using a recently developed method for mediation analysis
of survival data. Our analysis of the effect of research activity on patient survival was based on 352 patients who
had been diagnosed with advanced ovarian cancer at 149 hospitals in 2001. All hospitals took part in a quality
assurance program of the German Cancer Society. Patient outcomes were compared between hospitals
participating in clinical trials and non-trial hospitals. Surgical outcome and chemotherapy selection were explored as
potential mediators of the effect of hospital research activity on patient survival.
Results: The 219 patients treated in hospitals participating in clinical trials had more complete surgical debulking,
were more likely to receive the recommended platinum-taxane combination, and had better survival than the 133
patients treated in non-trial hospitals. Taking into account baseline confounders, the overall adjusted hazard ratio of
death was 0.58 (95% confidence interval: 0.42 to 0.79). This effect was decomposed into a direct effect of research
activity of 0.67 and two indirect effects of 0.93 each mediated through either optimal surgery or chemotherapy.
Taken together, about 26% of the beneficial effect of research activity was mediated through the proposed
pathways.
Conclusions: Mediation analysis allows proceeding from the question “Does it work?” to the question “How does it
work?” In particular, we have shown that the research activity of a hospital contributes to superior patient survival
through better use of surgery and chemotherapy. This methodology may be applied to analyze direct and indirect
natural effects for almost any combination of variable types.
Keywords: Trial effect, Research activity, Healthcare outcomes, Mediation, Survival analysis
Background In the last years another relevant aspect of participa-
Well-conducted trials are critical to advancements in tion in clinical trials has emerged: “Do healthcare insti-
oncology. However, such trials require patients and tutions or service providers who are active in research
healthcare providers who are willing to participate in deliver better care and outcomes than those who do not
clinical research. The effect of participation in clinical participate in clinical research?” [4, p. 6]. According to
trials is controversially discussed in the literature. Re- Pater et al. [5], especially from the healthcare policy per-
views on this effect have been published in the past, but spective, this question should be the main focus of fu-
these haveusually focused oncomparisonsbetween ‘trial ture research because benefit from a research-active
patients’ and ‘non-trial patients’ and authors have come institution will affect many more patients than just the
toconflictingconclusions[1-3]. small proportion of participants in clinical trials. In
addition, this issue may also be of greater importance to
*Correspondence:[email protected] patients who have to decide where to go to receive
1InstituteofMedicalBiometryandInformatics,UniversityofHeidelberg,
healthcare and by whom. However, only one review
ImNeuenheimerFeld305,Heidelberg69120,Germany
Fulllistofauthorinformationisavailableattheendofthearticle
©2014Rochonetal.;licenseeBioMedCentralLtd.ThisisanopenaccessarticledistributedunderthetermsoftheCreative
CommonsAttributionLicense(http://creativecommons.org/licenses/by/2.0),whichpermitsunrestricteduse,distribution,and
reproductioninanymedium,providedtheoriginalworkisproperlycited.
Rochonetal.BMCMedicalResearchMethodology2014,14:9 Page2of8
http://www.biomedcentral.com/1471-2288/14/9
conducted by Clarke and Loudon [6] included a litera- For normally distributed mediators and outcomes and
turesearchtoidentifystudiesaddressingtherelationship in the absence of interactions or non-linear effects, nat-
between institutional research activity and patient out- ural direct and indirect effects can be estimated by a
comes. The available evidence suggests that patients stepwise approach based on standard linear regressions
benefit from being treated by practitioners or in institu- [7].Inshort,modelsarefittedfortheoutcomebothwith
tions participating in trials. The authors concluded that and without the mediator,andthe difference inthe coef-
research activity might result in better health care out- ficients for the exposure is taken as a measure of the in-
comes, although the reason for this effect remains un- direct or mediated effect. In oncology, however, the
clear. Indeed, most studies focus on answering the outcome of interest is often patient survival. Survival
pragmatic question of whether research activity has an times are known to be non-normally distributed and
effect on patient outcome. However, little attention is typically right censored. Therefore, existing techniques
being paid to the explanation of such effects (“How does formediation analysisarenot applicabletosurvivaldata,
itwork?”,i.e.theassessmentofmediation). and several authors have shown that the linear regres-
SincethepublicationoftheseminalpaperbyBaronand sion approach cannot be simply translated to logistic re-
Kenny [7], testing for mediation has been an integralpart gression orproportionalhazardmodels[10-12].
of statistical analysis in psychology. In epidemiology, both The aim of this article is to introduce a recently
practical and theoretical aspects of mediation have been developed methodology that can be used to assess medi-
considered since Robins and Greenland [8]. The import- ation on all outcome types including survival data, and
ance of mediation is well argued in Hafeman and to use this method to explain the effects of institutional
Schwartz [9] where they demanded the opening of the researchactivityonovariancancerpatient survival.
‘black box’ when investigating exposure-disease relation-
ships. The core element of mediation analysis is the esti- Methods
mation of two effects: The natural direct effect of Patients
exposure on outcome and the natural indirect effect act- Thestudyispartofanongoingqualityassuranceprogram
ingthroughaso-calledmediatorthatisassumedtoreflect (QS-OVAR) initiated in 1999 by the Arbeitsgemeinschaft
a specificcausalpathwaybetweenexposureand outcome. Gynaekologische Onkologie (AGO) Organkommission
The total effect is the aggregate of these two effects OVAR, a subcommittee of the German Cancer Society.
(Figure1A).Specifically,thenaturaldirecteffectistheef- The aim of this program is to describe the pattern and
fect one would observe if the exposure could be changed qualityofcareofpatientswithovariancancerinGermany
without inducing a change in the mediator. The natural as well as to improve their outcome. The hypothesis
indirecteffectistheeffectonewouldobserveifthemedi- is that research-active hospitals yield improved outcomes
atorcouldbechangedasitwouldwhentheexposurewas for their patients. Research activity at hospital level is de-
manipulated(withoutactuallychangingtheexposure). fined as hospital participation in prospective clinical trials
A B
Figure1Pathdiagrams.(A)PathdiagramrelatingexposureAtoamediatorMandanoutcomeYinpresenceofknownbaselineconfoundersC.
(B)Pathdiagramrelatinghospitalresearchactivitytotwomediatorsandoverallsurvivaltimeinpresenceofknownbaselineconfounders.
Rochonetal.BMCMedicalResearchMethodology2014,14:9 Page3of8
http://www.biomedcentral.com/1471-2288/14/9
conducted by one of the two German cooperative study (>500mlvs.≤500ml),comorbidity(presentvs.none),hist-
groups, the AGO Study Group Ovarian Cancer (AGO- ology(serousvs.other),andgrade(G3/4vs.G1/2).
OVAR) and the Northeastern Society ofGynecologicOn- Mediation analysis was performed using the approach
cology(NOGGO).Morespecifically,allpatientstreatedin proposed by Lange et al. [17]. This approach is based on
hospitals participating in clinical trials are compared with the counterfactual framework [18] and allows decom-
all patients treated in hospitals that did not participate in position of the total effect of a given exposure A on
anyclinicaltrialsoftheAGO-OVARandNOGGO. the outcome Y into a natural direct effect (A→Y in
The mediation analysis of the effect of research activity Figure 1A) and a natural indirect effect through a medi-
on healthcare outcomes is illustrated through QS-OVAR ator M (A→M→Y in Figure 1A). In case of a time-to-
2001. This study involved 165 hospitals and 476 patients event outcomeY,abinaryexposure A,abinarymediator
with early and advanced ovarian cancer diagnosed in the M and a number of baseline confounders C, Lange et al.
third quarter of 2001. The primary outcome measure was showed that unbiased estimates for the direct and indi-
overall survival. Secondary outcomes were the receipt of rect effect are obtained from weighted Cox regression of
standard care with regard to surgery and chemotherapy. the time-to-event outcome on A, A* and C using a du-
Details on study design and results have been described plicated data set. In the first replication A* takes the ori-
elsewhere [13,14]. Here, we will focus on patients with ad- ginal value of the exposure. In the second replication A*
vanced ovarian cancer because most death events in QS- takes the opposite (‘counterfactual’) value of the expos-
OVAR2001occurredinpatientswithadvancedstagesFéd- ure. The weights are determined by Wc=P(M | A*, C) /
ération Internationale de Gynécologie et d'Obstétrique P(M | A, C), with P( · ) deriving from a logistic regres-
(FIGO)IIB–IV,whereasonlyafeweventswereobservedin sion ofthemediatorMonthe exposureandthe baseline
FIGOI–IIApatients.Furthermore,treatmentoptionsdiffer confounders [17, Appx. 4]. Assuming non-informative
markedly between early and advanced stage disease. censoring and proportional hazards, the weighted Cox
According to the German guidelines, patients with ovarian model then yields hazard ratios for A and A* that serve
cancerFIGOIIB andhighershouldreceivesurgeryinclud- as estimates for the natural direct effect and indirect ef-
ingmaximaldebulkingaswellaschemotherapywithcarbo- fect, respectively. The product of the two hazard ratios
platinandpaclitaxel[15].Wethusconsideredtwopotential yields the hazard ratio for the total effect. Standard er-
mediators. First, ‘optimal debulking’ with postoperative rors and confidence intervals can, for example, be deter-
tumor residuals up to 1 cm was regarded as standard care minedbybootstrapmethods.
withrespecttosurgery(i.e.‘optimalsurgery’).Second,each In our study, we explored two binary mediators (sur-
platinum-taxane combination was considered adherent to geryandchemotherapy)oftheeffectof hospitalresearch
treatment guidelines regarding chemotherapy (i.e. ‘optimal activity onsurvival (Figure 1B).Underthe assumption of
chemotherapy’). separate causalpathwaysthroughthetwomediators,un-
The use of data included in this study was approved biased point estimates for the natural direct effect and
by the AGO-OVAR and the AGO Ovarian Committee. the natural indirect effects related to the two mediators
All data was anonymized prior to analysis. The study were obtained by a weighted Cox regression of the out-
was approved by the Ethics Committee of the Medical come on the exposure, the baseline confounders and
Faculty of the University of Heidelberg, Germany (study two additional counterfactual variables A * and A * that
1 2
number:S-446/2013). were systematically manipulated in four replicates of the
original data [19]. Confidence intervals for mediation ef-
Statisticalanalyses fectsthataccount forclusteringofpatientswithin hospi-
All analyses were conducted in R version 3.0.2 [16]. talswereobtainedusingsimplerandomclustersampling
Continuous data were summarized with median and inter- and 10,000 bootstrap simulations. In all analyses, results
quartile range (IQR), categorical data by counts and were considered statistically significant if the 95% confi-
percentages. Survival curves were generated by the Kaplan- dence interval (CI) for the hazard ratio (HR) or odds ra-
Meier method. The Cox proportional hazards model with tio(OR)didnot include1.
robustvarianceestimatorwasusedtoassesstherelationship In the additional material (see Additional file 1), we
between hospital participation in clinical trials and overall have provided a detailed description of the analysis with
survival. Logistic regression models were fitted for ‘optimal the corresponding R code to enable future researchers
surgery’ and ‘optimal chemotherapy’ with generalized esti- to assess mediation in a survival context. There, we also
matingequationstoaccountforclusteringofpatientswithin show how we tested for interactions between exposure
hospitals. The following patient and disease characteristics and confounders, how we assessed the linearity assump-
were assumed to control for confounding: age at diagnosis tion for all variables, and how we allowed for misspecifi-
(continuous,in 5years units), Eastern CooperativeOncolo- cation in the mediators. Finally, this description includes
gy Group (ECOG) performance status (>1 vs. 0/1), ascites sensitivity analyses with only one mediator representing
Rochonetal.BMCMedicalResearchMethodology2014,14:9 Page4of8
http://www.biomedcentral.com/1471-2288/14/9
Table1Patientanddiseasecharacteristicsatdiagnosis ranged from 1 to 12, the median number was two. 219
Trialhospitals Non-trialhospitals patients (62%) were treated in 77 hospitals participating
in clinical trials, and 133 patients (38%) were treated in
No.ofpatients,N(%) 219 (62.2) 133 (37.8)
72 non-trial hospitals. However, only 59 (27%) of the
Age(years)
219 patients in trial hospitals were actually enrolled in
Median(IQR) 64 (57–73) 66 (56–74)
clinical trials which included a trial that did not show
Performancestatus,N(%) any superiority of the experimental treatment [20]. Pa-
ECOG0/1 166 (75.8) 101 (75.9) tient and disease characteristics at diagnosis for both
ECOG>1 53 (24.2) 32 (24.1) groupsarelistedinTable 1.
During the follow-up period of three years, 184 out of
Ascites,N(%)
352 patients with advanced ovarian cancer died. The
≤500ml 113 (51.6) 59 (44.4)
median overall survival time was 35 months for the 219
>500ml 106 (48.4) 74 (55.6)
patients treated in hospitals participating inclinicaltrials
Comorbidity,N(%) compared to 25 months for the 133 patients treated in
None 166 (75.8) 91 (68.4) non-trial hospitals (Figure 2). This survival benefit
Present 53 (24.2) 42 (31.6) remained stable even after adjustment for all patient and
disease characteristics listed in Table 1, as well as after
Histology,N(%)
adjustment for clustering of patients within hospitals in
Serous 167 (76.3) 96 (72.2)
a multivariable Cox model. The total effect of research
Other 52 (23.7) 37 (27.8)
activity (adjusted for known confounders) can be found
Grade,N(%) in Table 2: The adjusted HR of death was 0.58 (95% CI:
G1/2 97 (44.3) 69 (51.9) 0.42to0.79).
G3/4 122 (55.7) 64 (48.1) How much of the total beneficial effect of research ac-
tivity is mediated through optimal debulking and stan-
adherencetotreatment guidelineswith regard tosurgery dard chemotherapy? In the first step, we separately
andchemotherapyasabinary andanordinal variable. explored the relationship between hospital participation
in clinical trials and the two independent mediators. The
Results proportion of patients receiving the recommended
Altogether, two-thirds of the 476 patients diagnosed in platinum-taxane combination was 70% in research-
the third quarter of 2001 had advanced ovarian cancer. active hospitals and 59% in non-trial hospitals (adjusted
The 352 patients with FIGO IIB or higher were treated OR=1.64, 95% CI: 0.91 to 2.95). 66% of the 219 patients
in 149 hospitals. The number of patients per hospital treated in research-active hospitals had maximal tumor
100%
75%
%)
y (
bilit
a 50%
b
o
r
p
al
v
vi
ur 25%
S
Trialhospitals (N=219,Deathevents=103)
Non-trialhospitals (N=133,Deathevents=81)
0%
0 6 12 18 24 30 36
Months
Patients at risk
Trial hospitals 219 187 166 149 129 100 43
Non-trial hospitals 133 109 95 83 63 43 15
Figure2Kaplan-Meiersurvivalcurvesinadvancedovariancanceraccordingtohospitalparticipationinclinicaltrials.
Rochonetal.BMCMedicalResearchMethodology2014,14:9 Page5of8
http://www.biomedcentral.com/1471-2288/14/9
Table2MultivariableCoxregressionanalysisforoverall associated with survival. The adjusted HR of death for
survivalinpatientswithadvancedovariancancer optimally debulked patients was 0.46 (95% CI: 0.34 to
Variable N Events HR 95%CI 0.63) compared to patients with a postoperative tumor
Researchactivity residual larger than 1 cm. Similarly, we found a survival
benefitforpatientstreatedwithplatinum-taxane chemo-
Non-trialhospital 133 81 1 Reference
therapy(adjustedHR=0.42, 95%CI:0.28 to0.62).
Trialhospital 219 103 0.58 [0.42;0.79]
Finally, we used the approach of Lange et al. [17] to
Age
test if the observed positive effect of institutional re-
Continuous(5years) 352 184 1.24 [1.14;1.34] search activity onpatient survivalis at leastpartially me-
Performancestatus diated throughsurgeryandchemotherapy(Figure 1B).
ECOG0/1 267 118 1 Reference The effects of hospital participation on overall survival
in terms of natural direct and indirect hazard ratios can
ECOG>1 85 66 2.02 [1.38;2.96]
be summarized as follows (Table 3): The total HR of 0.58
Ascites
was decomposed into a direct HR of research-activity of
≤500ml 172 74 1 Reference
0.67 (95% CI: 0.47 to 0.92) and an indirect HR for both
>500ml 180 110 1.77 [1.23;2.53] mediators of 0.87 (95% CI: 0.75 to 0.98). The indirect HR
Comorbidity with regard to surgery alone was 0.93, the corresponding
None 257 116 1 Reference indirect HR with respect to chemotherapy was also
0.93; thus the resulting total effect was0.67×0.93 ×
Present 95 68 1.46 [1.08;1.97]
0.93=0.58. The proportion mediated through surgery
Histology
and chemotherapy was similar and was about 13% for
Other 89 40 1 Reference
each mediator on the log HR scale. Altogether, about
Serous 263 144 1.29 [0.85;1.94] 26% of the total effect of hospital research activity on
Grade survival was mediated through the proposed pathways
G1/2 166 81 1 Reference (95% CI: 3% to 69%).
G3/4 186 103 1.10 [0.80;1.50]
Discussion
The effect of institutional research activity on patient
residuals of 1 cm in contrast to 54% of the 133 patients outcomes has not yet been investigated extensively,
treated in non-trial hospitals (adjusted OR=1.57, 95% despite its great relevance to healthcare providers,
CI: 0.92 to 2.69). Altogether, about 50% of patients in policy makers, and patients. So far, only a few studies
research-active hospitals and 38% of patients in non-trial have examined the association between patient out-
hospitals received the combination of optimal debulking comes and institutional participation in clinical trials,
with standard chemotherapy(Figure 3). as opposed to trial participation of individual patients
In the next step, we explored the effect of standard [6]. Thus, the authors of the 2011 special issue of
care on survival, again by means of a multivariable Cox Annals of Oncology entitled “Clinical Research and
model. As expected, both potential mediators were Healthcare Outcomes: A Workshop at the International
100%
Surgery–/Chemo–
80%
Surgery+/Chemo–
60%
Surgery–/Chemo+
40%
Surgery+/Chemo+
20%
0%
219 patients in trial hospitals 133 patients in non-trial hospitals
Figure3Adherencetotreatmentguidelineswithregardtosurgeryandchemotherapyinadvancedovariancanceraccordingto
hospitalparticipationinclinicaltrials.
Rochonetal.BMCMedicalResearchMethodology2014,14:9 Page6of8
http://www.biomedcentral.com/1471-2288/14/9
Table3Mediationanalysis:Total,directandindirect clinical trials was associated with improved survival. The
effectswith95%confidenceintervals effect observed in our study was large and clearly rele-
Effect(oftrialhospitalvs.non-trialhospital) HR 95%CI vant because of a relative reduction in the risk of death
Totaleffect 0.58 [0.41;0.80] by 42% in favor of research activity. Whether this effect
is in line with the literature is difficult to answer be-
Directeffect 0.67 [0.47;0.92]
cause, with respect to ovarian cancer, trial participation
Indirecteffect(throughsurgery) 0.93 [0.84;1.02]
has hardly been investigated outside Germany. Further-
Indirecteffect(throughchemotherapy) 0.93 [0.84;1.01]
more, only a few studies have examined the association
Indirecteffect(throughbothsurgeryand 0.87 [0.75;0.98] between patient outcomes and research activity at the
chemotherapy)
levelof healthcare institutions [6].
As expected, patients with ‘optimal surgery’ and ‘opti-
Agency for Research on Cancer, Lyon” agreed that mal chemotherapy’ lived longer than patients who were
further research on this topic is urgently required. If not treated according to the standard of care. In
research activity has beneficial effects on patient out- addition, patients treated in trial hospitals were more
comes, these benefits are not solely restricted to re- likely to receive optimal treatment than patients treated
search participants. in non-trial hospitals. We thus confronted the question
Investigating the relationship between research activity of how much of the effect of hospital participation in
at the level of institutions or providers and outcomes at clinical trials on patient survival was mediated through
the level of patients is not as easy as it seems at first optimal debulking and optimal chemotherapy selection.
glance.Researchersinterestedinthisrelationshiparecon- To answer this question, we used a recently developed
fronted with several practical challenges, such as getting methodology for assessing mediation in the context of a
data from less research-active institutions, as well as with survival analysis [17]. Taking into account several known
severalmethodologicalchallenges,forinstance,thechoice baseline confounders, the overall hazard ratio (total ef-
oftheappropriatestudydesign[5].Inaddition,solefocus fect) of 0.58 was decomposed into a direct effect of re-
on establishing effects and measuring the potential bene- search activity of 0.67 and two indirect effects of 0.93
fits of research activity on healthcare outcomes is often each mediated through surgery and chemotherapy. The
considered insufficient [6]. The assessment of variables aggregate indirect effect through both mediators was
mediating the effects may be helpful in explaining the ef- 0.87, that is, about 26% of the beneficial effect of re-
fects of exposure or in investigating the reasons why an search activity was mediated through both surgery and
exposure failed to yield an expected outcome. More im- chemotherapy.Inconclusion,trialparticipationofahos-
portantly,knowledgeofsuchmechanismsenablesspecific pital contributed at least partially to a superior outcome
measurestoimprovehealthcareattheindividualleveland throughthebetterqualityoftreatment provided.
at the institutional level—and such measures may be im- The probability of surviving ovarian cancer depends
plementedevenintheabsenceofexposure.Krzyzanowska on (1) patient characteristics (e.g.age), (2) tumor biology
et al. [21] described a conceptual framework for under- (e.g. stage), and (3) the quality of treatment (e.g. surgical
standing how institutional research activity might lead to outcome, selection of chemotherapy regimen). The first
better outcomes, even for patients who are not partici- two factors are hard to change but the quality of treat-
pants in a research project. The authors pointed out that ment is susceptible to direct influenceand thus seems to
the processes of care received by patients may have a be of utmost relevance when considering efforts to im-
strong impact on outcomes and that such processes may prove the outcome of this disease. The implementation
systematicallydifferbetweenresearch-activeandresearch- of standards into clinical routine in our study was not
inactive settings. For example, institutions that actively satisfactory and still needs improvement; however, taken
participate in research may be more likely to follow clin- both mediators together, treatment standards were more
icalguidelinesforcancertreatment.Inaddition,participa- strictly implemented in trial hospitals than in non-trial
tion in research may facilitate early access to new hospitals. The good news is that clinicians can influence
treatment approaches, allowing the faster implementation the implementation of standards, regardless of whether
ofnewevidenceintopractice. they areemployed inresearch-activehospitalsornot.
The main purpose of the present work was to explore The main limitation of our study is the lack of ran-
the mechanisms underlying the association between in- domization of patients into research-active and research-
stitutional research activity and patient outcomes. We inactive hospitals. Such a randomized controlled trial
first re-examined the association between institutional would facilitate a clear causal interpretation but would
research activity and survival described by du Bois et al. also be hard to implement. When randomization is not
[13]. Using the data of 352 patients with advanced ovar- possible,well-conductedobservationalstudiescanprovide
ian cancer, we showed that hospital participation in the necessary data to guide the future development of
Rochonetal.BMCMedicalResearchMethodology2014,14:9 Page7of8
http://www.biomedcentral.com/1471-2288/14/9
clinical research and healthcare. In particular, mediation mechanisms. This tool allows analysis to go beyond the
analysis can help explore the mechanisms by which re- question “Does it work?” In particular, mediation ana-
search activity leads to the outcome of interest. The key lysis has shown that the research activity of a hospital
assumptionformediationanalysisisthatallrelevantcon- contributes to superior patient survival through better
founders are included into the analysis. To be more pre- adherence to treatment guidelines with regard to surgi-
cise, the approach proposed by Lange et al. [17] requires cal outcome and selection of chemotherapy. Yet, the
that the considered variables are sufficient for controlling analysis of the ovarian cancer study presented in this
the confounding of 1) the exposure-outcome relation, 2) article demonstrates only one way of applying this
the exposure-mediator relation, and 3) the mediator- methodology, but the potential for this technique is
outcome relation. We addressed this issue by incorpora- much wider. Thismethodology, which can be conducted
ting all established prognostic factors into the models for with standard software, may be applied to analyze direct
the mediator and the outcome. In particular, age at diag- andindirecteffectsforalmostanycombinationofvariable
nosis, FIGO stage, ECOG performance status, volume of types(inparticular,themediatorandoutcomecanbebin-
ascites, histology, grade, and comorbidity are well-known ary, ordinal, categorical, or continuous, and the outcome
prognostic factors for survival in ovarian cancer [22-24]. can also be survival time). Researchers interested in the
Thesefactorsalsoinfluencetreatmentrecommendations. question “How does it work?” are thus strongly encou-
There must be no other variables (measured or un- ragedtousemediationanalysisintheirinvestigations.
measured) that confound the mediator-outcome relation
which are themselves affected by the exposure; this last Additional file
assumption essentially ensures that each of the consi-
dered pathways between exposure and outcome does Additionalfile1:Adetaileddescriptionofthemediationanalysis
indeed represent a unique causal mechanism. In obser- withthecorrespondingRcode.
vational studies, these assumptions are inherently un-
testable and must instead be justified by means of Competinginterests
knowledge about the biological processes under conside- Theauthorsdeclarethattheyhavenocompetinginterests.
ration. In our study, for example, we did not collect in-
Authors’contributions
formation on socioeconomic factors (e.g. income or
JRandAdBjointlydesignedthestudy.JRandTLcarriedouttheanalysis.JR
insurance status). Patients with a higher socioeconomic draftedthemanuscript.Allauthorscontributedtorevisionsofthe
status (SES) may find it easier and hence choose to manuscript,andreadandapprovedthefinalversion.
travel to specialized or better rated hospitals or institu-
Funding
tions that are research-active and participate in clinical
TheAGOqualityassuranceprogramQS-OVAR2001wassupportedwitha
trials. In contrast,cancer patientswith alowSESmay be grantbyBristol–Myers–Squibb,Germany.
less likely to choose research-active hospitals because of
Authordetails
the lack of corresponding health care information. The
1InstituteofMedicalBiometryandInformatics,UniversityofHeidelberg,
total effect of research activity would then at least be ImNeuenheimerFeld305,Heidelberg69120,Germany.2Departmentof
partially due to better general survival rates of patients GynecologyandGynecologicOncology,KlinikenEssen-Mitte,Essen,
Germany.3DepartmentofBiostatistics,UniversityofCopenhagen,
with a high SES. Lower socioeconomic status has occa-
Copenhagen,Denmark.
sionally been associated with lower likelihood of recei-
ving surgery and chemotherapy but does not seem to be Received:1August2013Accepted:14January2014
Published:22January2014
an independent prognostic factor for survival in ovarian
cancer [25,26].
References
Finally, the mediation analysis is only valid if the logistic 1. BraunholtzDA,EdwardsSJ,LilfordRJ:Arerandomizedclinicaltrialsgood
regressions are adequate descriptions of the mediators. forus(intheshortterm)?Evidencefora‘trialeffect’.JClinEpidemiol
2001,54:217–224.
Sensitivityanalysesshowedthatmisclassificationoftheme-
2. PeppercornJM,WeeksJC,CookEF,JoffeS:Comparisonofoutcomesin
diator will bias the estimates of the indirect effect towards cancerpatientstreatedwithinandoutsideclinicaltrials:conceptual
one and the direct effect away from one. In our study, frameworkandstructuredreview.Lancet2004,363:263–270.
3. VistGE,BryantD,SomervilleL,etal:Outcomesofpatientswhoparticipate
measurementerrorwasminimizedbyobjectivelyassessing
inrandomizedcontrolledtrialscomparedtosimilarpatientsreceiving
thequalityofchemotherapyandbyevaluatingsurgicalout- similarinterventionswhodonotparticipate.CochraneDatabaseSystRev
come(i.e.tumorresidual)inastandardizedmanner. 2008,3,MR000009.
4. SelbyP,AutierP:Theimpactoftheprocessofclinicalresearchonhealth
serviceoutcomes.AnnOncol2011,22(Suppl7):vii5–vii9.
Conclusions 5. PaterJ,RochonJ,ParmarM,SelbyP:Futureresearchandmethodological
Mediation analysis, which is regularly and successfully approaches.AnnOncol2011,22(Suppl7):vii57–vii61.
6. ClarkeM,LoudonK:Effectsonpatientsoftheirhealthcarepractitioner’s
applied elsewhere in health care, has a great potential to orinstitution’sparticipationinclinicaltrials:asystematicreview.
be used as an instrument for investigating underlying Trials2011,12:16.
Rochonetal.BMCMedicalResearchMethodology2014,14:9 Page8of8
http://www.biomedcentral.com/1471-2288/14/9
7. BaronRM,KennyDA:Themoderator-mediatorvariabledistinctionin
socialpsychologicalresearch:conceptual,strategic,andstatistical
considerations.JPersSocPsychol1986,51:1173–1182.
8. RobinsJM,GreenlandS:Identifiabilityandexchangeabilityfordirectand
indirecteffects.Epidemiology1992,3:143–155.
9. HafemanDM,SchwartzS:OpeningtheBlackBox:amotivationforthe
assessmentofmediation.IntJEpidemiol2009,38:838–845.
10. ColeSR,HernánMA:Fallibilityinestimatingdirecteffects.IntJEpidemiol
2002,31:163–165.
11. KaufmanJS,MaclehoseRF,KaufmanS:Afurthercritiqueoftheanalytic
strategyofadjustingforcovariatestoidentifybiologicmediation.
EpidemiolPerspectInnov2004,1:4.
12. VanderWeeleTJ:Causalmediationanalysiswithsurvivaldata.
Epidemiology2011,22:582–585.
13. duBoisA,RochonJ,LamparterC,PfistererJ:Patternofcareandimpact
ofparticipationinclinicalstudiesontheoutcomeinovariancancer.
IntJGynecolCancer2005,15:183–191.
14. RochonJ,duBoisA:Clinicalresearchinepithelialovariancancerand
patients’outcome.AnnOncol2011,22(Suppl7):vii16–vii19.
15. BauknechtT,BreitbachGP,duBoisA,etal:MaligneOvarialtumoren.In
DiagnoseundTherapiemalignerErkrankungen.Kurzgefassteinterdisziplinaere
Leitlinien.EditedbyHermanekP.Muenchen:ZuckschwerdtVerlag;
2000:301–318.
16. RCoreTeam:R:Alanguageandenvironmentforstatisticalcomputing.
Vienna,Austria:RFoundationforStatisticalComputing;2013.
17. LangeT,VansteelandtS,BekaertM:Asimpleunifiedapproachfor
estimatingnaturaldirectandindirecteffects.AmJEpidemiol2012,
176:190–195.
18. PearlJ:Causality:Models,Reasoning,andInference.NewYork,NY:Cambridge
UniversityPress;2009.
19. LangeT,RasmussenM,ThygesenLC:Assessingnaturaldirectandindirect
effectsthroughmultiplepathways.AmJEpidemiol2014,179:513–518.
20. PfistererJ,WeberB,ReussA,etal:RandomizedphaseIIItrialoftopotecan
followingcarboplatinandpaclitaxelinfirst-linetreatmentofadvanced
ovariancancer:AgynecologiccancerintergrouptrialoftheAGO-OVAR
andGINECO.JNatlCancerInst2006,98:1036–1045.
21. KrzyzanowskaMK,KaplanR,SullivanR:Howmayclinicalresearchimprove
healthcareoutcomes?AnnOncol2011,22(Suppl7):vii10–vii15.
22. ThigpenT,BradyMF,OmuraGA,etal:Ageasaprognosticfactorin
ovariancarcinoma.TheGynecologicOncologyGroupexperience.
Cancer1993,71(2Suppl):606–614.
23. ClarkTG,StewartME,AltmanDG,etal:Aprognosticmodelforovarian
cancer.BrJCancer2001,85:944–952.
24. SperlingC,NoerMC,ChristensenIJ,etal:Comorbidityisanindependent
prognosticfactorforthesurvivalofovariancancer:aDanish
register-basedcohortstudyfromaclinicaldatabase.GynecolOncol2013,
129:97–102.
25. SugiyamaVE,ShinJY,KappDS,etal:Theeffectofsocioeconomicstatus
onthesurvivalofovariancancerpatients.JClinOncol2008,
26(Suppl):5557.
26. BristowRE,PowellMA,Al-HammadiN,etal:Disparitiesinovariancancer
carequalityandsurvivalaccordingtoraceandsocioeconomicstatus.
JNatlCancInst2013,105:823–832.
doi:10.1186/1471-2288-14-9
Citethisarticleas:Rochonetal.:Mediationanalysisoftherelationship
betweeninstitutionalresearchactivityandpatientsurvival.BMCMedical
ResearchMethodology2014,14:9.
Submit your next manuscript to BioMed Central
and take full advantage of:
• Convenient online submission
• Thorough peer review
• No space constraints or color figure charges
• Immediate publication on acceptance
• Inclusion in PubMed, CAS, Scopus and Google Scholar
• Research which is freely available for redistribution
Submit your manuscript at
www.biomedcentral.com/submit