ACCEPTED

Gemma Mansell, MSc, Arthritis Research UK Primary Care Centre, Primary Care Sciences,

Keele University, Keele, Staffordshire, UK

Jonathan C Hill, PhD, Arthritis Research UK Primary Care Centre, Primary Care Sciences,

Keele University, Keele, Staffordshire, UK

Steven J Kamper, PhD, Musculoskeletal Division, The George Institute for Global Health,

University of Sydney, Australia. EMGO+ Institute, VU University Medical Centre,

Amsterdam, The Netherlands

Peter Kent, PhD, Spine Centre of Southern Denmark, Hospital Lillibælt, Institute of Regional

Health Services Research, University of Southern Denmark, Denmark

Chris Main, PhD, Professor of Clinical Psychology, Arthritis Research UK Primary Care

Centre, Primary Care Sciences, Keele University, Keele, Staffordshire, UK

Danielle A van der Windt, PhD, Professor of Primary Care Epidemiology, Arthritis Research

UK Primary Care Centre, Primary Care Sciences, Keele University, Keele, Staffordshire, UK

Name and Address for Correspondence: Gemma Mansell, Arthritis Research UK Primary

Care Centre, Primary Care Sciences, Keele University, Keele, Staffordshire, UK, ST5 5BG,

Tel: 01782 734877, Fax: 01782 733911, Email: g.mansell@keele.ac.uk

ACCEPTED

The manuscript submitted does not contain information about medical device(s)/drug(s). The

National Institute for Health Research (NIHR) under its Programme Grants for Applied

Research Programme (Grant Reference Number RP-PG-0707-10131) funds were received in

support of this work. Relevant financial activities outside the submitted work: employment,

grants, royalties.

1. Introduction

In this article we consider whether adjustments to study designs, particularly of clinical trials,

might provide important information to help explain how interventions for patients with low

back pain work. Back pain clinical trials have typically shown small effects associated with

various treatments as compared to no treatment or current usual care, which could mean that

these treatments are not very effective, (1; 2)

or that the treatments are only effective for

specific clinical subgroups and interventions therefore need better targeting. (3)

However,

researchers also need to consider new methods to optimise treatment outcomes by

investigating the processes which explain how or why specific treatments work. (4)

Such

studies have been referred to as mediation analyses, (5)

and this paper seeks to highlight why

and how future studies should be designed to incorporate mediation analyses.

The primary importance of mediation analysis is that it can help us understand how to

improve treatments for patients with back pain. For example, psychological factors such as

fear-avoidance and self-efficacy are known to be predictive of patient outcomes such as

disability and return to work, but it is less clear whether targeting an intervention specifically

at reducing fear-avoidance or increasing self-efficacy is able to further improve such

outcomes. By analysing the causal pathway between treatment, potential mediators and

outcomes, we can identify the key factors to focus on during treatment and also identify

ACCEPTED

ACCEPTED

which factors do not mediate outcome, in order to further streamline and improve

intervention efficacy and efficiency.

At the XII International Low Back Pain Primary Care Research Forum in Odense, Denmark

2012, a dedicated workshop was held in which various aspects of mediation study design

were explored and methodological considerations discussed. The workshop brought together

experienced and internationally recognised low back pain researchers and clinicians to

provide an overview of the importance of mediation analyses and explore its usefulness in

testing our assumptions about how our interventions may be influencing outcomes. We

provide a summary of the workshop findings and discuss the implications for embedding

mediation analysis methodology within future interventional research. We also offer

definitions for some commonly used and misused terms, provide examples of mediation

research using different study designs and incorporate information on the methodology of

mediation analysis from the wider literature, in order to formulate recommendations for

future mediation research in the field of low back pain.

It is important to acknowledge that while the focus of this article is on mediators of specific

treatment effects, the identification of mediator variables often comes from a programme of

work which may involve different study designs. This article therefore includes a discussion

of research using observational designs to explore mediating processes, although the

limitations of evidence from such studies is noted and ultimately randomised controlled trials

(RCTs) are required to confirm which factors mediate specific treatment effects.

2. Concept Definitions

When longitudinal data from an observational study (single groups of individuals observed

over time) or an intervention study (two or more groups of individuals, exposed to different

treatment conditions) are analysed, some factors may be consistently predictive of outcome.

ACCEPTED

ACCEPTED

However, although such factors may show a statistical association with outcome following

treatment, the relationship can take a number of different forms: as a prognostic factor, a

treatment effect moderator or a treatment effect mediator.

2.1. Prognostic Factors

Prognostic factors are baseline characteristics that are associated with outcome regardless of

treatment, such as high baseline pain and high baseline activity limitation, that are generally

known to be associated with poor outcome in spinal pain patients across a variety of different

treatment interventions. So in an RCT, a prognostic factor might predict outcome for patients

in both the intervention and control groups. In intervention studies, prognostic factors have

also been called nonspecific predictors of treatment outcome. (5)

In statistical terms, they are

baseline variables that are associated with outcome (they have a significant main effect only)

but do not correlate or interact with treatment allocation. (5)

2.2. Treatment Effect Moderators

A treatment effect moderator is a baseline variable that identifies people who do better or

worse with a given treatment. These factors specify for whom and under what circumstances

a treatment is most effective and are therefore central to stratified care that targets specific

subgroups of people. For example, treatment A may work better for people with high fear of

movement (who) or in people seeking primary care rather than secondary care (what

circumstances) compared to treatment B. In statistical terms, treatment effect moderators

stratify or subgroup the effect of a particular treatment and are indicated by a statistical

interaction between treatment allocation (treatment A versus B), time and subgroup (e.g. high

versus low fear). (5)

In some fields of research, treatment effect moderators are also referred

ACCEPTED

ACCEPTED

to as effect modifiers or predictors of differential treatment effect, but these terms refer to the

same thing. (6)

2.3. Treatment Effect Mediators

In contrast to prognostic factors and treatment effect moderators, which are both measured

only at baseline, treatment effect mediators are factors that change in response to a specific

treatment and should therefore not only be measured at baseline but also during and after

treatment. These factors help explain the relationship between treatment and outcome, and

may therefore help identify the causal mechanisms by which treatments work. The

identification of mediators can potentially improve treatment effects by strengthening the

influence of the mediating factors involved in producing the treatment effect. Hypothetically

it might be that in order for cognitive behavioural therapy (CBT) (the treatment) to reduce

activity limitation (the outcome), a patient’s self-efficacy (the mediator) needs to be

strengthened, and that in the absence of self-efficacy improving, CBT will not be as effective

as it otherwise might be. In statistical terms, changes in mediator scores are associated with

specific treatment allocation (i.e. differ between intervention groups), and change in the

mediator has a direct association with change in treatment outcome (5)

(see Figure 1).

3. Study design in Mediation Analysis

An attractive aspect of mediation analysis is its applicability to different study designs. The

analytical process can be applied to data from cross-sectional studies, longitudinal

observational cohorts and RCTs. However, different study designs do impose different

limitations on the nature of the research question that can be answered and also the certainty

with which causal inferences can be made.

3.1. Cross-sectional designs

ACCEPTED

ACCEPTED

Mediation analysis has often been used with data from cross-sectional studies. With a

convincing biological rationale and a priori defined hypotheses, these data can provide some

evidence in support of pre-existing theory or be used to identify promising potential pathways

that can then be tested and confirmed via more robust designs such as RCTs. However,

causal inference requires a specified temporal (time-based) ordering of the investigated

variables, and in cross-sectional studies where participants are only assessed at one time

point, no temporal link can be made. Many publications that report the use of this design do

acknowledge that the use of cross-sectional data precludes the inference of causation. (7); (8); (9)

3.2. Longitudinal designs

A number of authors have stipulated that longitudinal designs are required for the assessment

of change over time. ((10); (11))

Longitudinal data enable the researcher to control for previous

levels of the mediator and outcome variables which can strengthen the claim of a potential

causal association, but do not account for the potential influence of confounding variables

that are not measured or included in the mediation analysis. Confounding is a key issue in

mediation analysis using longitudinal and cross-sectional observational designs. Confounders

are variables that are correlated with treatment exposure (i.e. show different values between

treatment groups) or the independent variable and are associated with outcome, but unlike

mediating variables are not on a causal pathway between treatment and outcome.

A series of studies on people with spinal pain in Australia used longitudinal data from

observational cohorts ((12); (13))

or pooled data from an RCT (14)

to investigate the mediating

influence of psychological factors in the relationship between pain and disability. Three or

more waves of data may be useful for assessing temporality as this provides stronger

evidence in support of a causal pathway linking the variables, ((15); (16); (17))

although it is

ACCEPTED

ACCEPTED

acknowledged that this does increase the complexity of the analysis and requires more

complex modelling procedures. (17)

3.3. Experimental designs

Mediation analysis aims to identify causal associations, and experimental designs where

patients are randomly allocated to a certain intervention at a point in time provides perhaps

the strongest support for causal inference. It should be noted however that RCTs and

observational studies seek to address different types of questions in mediation research.

Studies using observational data investigate the relationships between variables and outcomes

relevant to people with a particular condition, and may support the generation of hypotheses

regarding potential mediators. Mediation studies that include allocation to one treatment or

another as the independent variable investigate the mechanism of action of a particular

treatment on outcome. It is this aspect of experimental designs which is of particular benefit

in mediation analyses as it allows the investigation of the key variables (mediators) to target

in future intervention studies. For example, studies of mediation analysis in RCTs have found

that graded exposure treatment exerted part of its treatment effect in people with chronic LBP

via reduction of catastrophising, (18)

and showed that improvement in back pain was due to an

increase in lumbar muscle endurance following an exercise program. (19)

Another advantage of experimental designs is their ability to reduce the effect of

confounding. When an RCT design is used, randomisation helps ensure that intervention

groups have similar characteristics at baseline and so any differences in outcome are less

likely to be as a result of one group differing substantially from the other to start with. ((16);

(20); (21)) This can help researchers to see whether any mediating effects would have occurred

in spite of treatment (e.g. natural resolution) or whether they are due to the treatment being

given and are therefore on a causal pathway. (21)

However, it has been recognised that the

ACCEPTED

ACCEPTED

RCT design does not necessarily account for all unmeasured variables that could impact on

outcome. (22)

This is because although participants may have been randomised to the

intervention itself, change in the mediator(s) and outcome(s) occurs after randomisation has

taken place (if the intervention is effective) (15)

meaning that there may still be factors other

than treatment that affect the relationship between the mediator and outcome variables. (23)

For example, although patients may have similar levels of fear-avoidance at baseline due to

the randomisation process, during treatment some patients may develop better therapeutic

relationships with their therapist than others which may lead to a greater change in fear-

avoidance, and therefore a greater change in outcome for these particular patients. In this case

the therapeutic relationship is a potential confounder because it may have an impact on fear-

avoidance in addition to the impact of treatment.

4. Issues around heterogeneity in primary care populations

Studies carried out in secondary or tertiary care populations, in which the condition is usually

clearly established, unambiguously identified, and of sufficient severity or impact to enable

the identification of therapeutic change, may report larger and more consistent effect sizes

than those conducted with primary care populations. (24)

A problem when investigating

primary care populations may be the greater clinical variability in symptoms and their impact

on the population under study. This could also influence mediation analyses as greater

heterogeneity in potential mediators and outcomes results in larger standard deviations in the

scores in the study population, making it harder to demonstrate a difference between

treatments. Perhaps more importantly, greater clinical variability may also imply that the

mechanism explaining treatment effect varies across patients, which can translate into

difficulty in identifying the role of mediators in explaining treatment effects. In order to

overcome this, larger sample sizes might be required so that the assessment of the effect of

ACCEPTED

ACCEPTED

mediators in different subgroups can be undertaken. A specific advantage of primary care

populations is that, from a public health perspective, the results may be more widely

generalizable.

5. Areas for Improvement to the Design of Future Mediation Studies

A number of ways to incorporate mediation analysis into future RCTs by making small

changes to their current design have been suggested. The list below outlines the main

suggestions but is by no means exhaustive; more detailed reviews of the mediation literature,

relating to both design and analysis issues, can be found elsewhere. ((11); (23); (25); (26); (27))

5.1. The importance of a theoretical foundation

Selecting mediators based on theory can help researchers choose the key factors to focus on,

(28) and will help formulate clear hypotheses regarding how a treatment might work before the

study begins, which may subsequently be supported or refuted by the results of the mediation

analysis. (29)

A theoretical basis will also help with the issue of temporality described above,

as it provides a framework for showing how variables might relate to each other in a

particular order. (30)

5.2. The Design of Studies to Test Specifically for Mediation

It is suggested that where an intervention seeks to change a particular mediator or set of

mediators, measures of the proposed mediators should be included in the study. (11)

However,

investigators should limit data collection to only include mediators that can potentially be

modified by treatment and for which there is some evidence for a mediating relationship with

the outcome measure. (28)

For example, researchers might hypothesise that an exercise

ACCEPTED

ACCEPTED

programme is effective due to its action of improving aerobic endurance in patients, or

alternatively via improvements in self-efficacy. In these cases, measures of aerobic capacity

and/or self-efficacy should be incorporated into the study design to specifically test these

hypotheses.

Sample size obviously has an effect on the power of a study to detect mediating effects.

While the required sample size is dependent on the choice of statistical technique, variability

of the measures and the expected strength of the associations between the variables, a

substantial sample may be necessary.. It is recommended that sophisticated mediation

analysis methods require at least 150 (31)

to 200 (32)

participants to provide reliable results.

One review (33)

found that out of 166 mediation analysis studies published in two key journals

in their field, 113 studies (68%) included sample sizes of above 150.

5.3. Timing of Assessments

In order to assess temporality, it has been suggested that both the mediator and outcome

should be assessed frequently over the period of time that change is expected to occur. ((16);

(24)) This again shows the importance of considering treatment mediators a priori, as a clear

hypothesis regarding when a mediator is likely to change is important in planning the

optimal timing for study measurement time-points. While studies that include assessments

pre- and post-intervention may provide some useful information on what mediating factors

are related to outcome, (16)

they are not able to determine which variable (mediator or

outcome) changed first. (11)

Following the example of the exercise intervention proposed in

Section 5.2 above, while changes in aerobic capacity are likely to take several weeks to

occur, self-efficacy gains may be realised in a shorter time-frame. These considerations will

need careful thought in order to plan appropriate timing of study measurements. Also, an

intervention may work differently for different people and the effects could occur at different

ACCEPTED

ACCEPTED

times. (11)

Although attempting to untangle this issue adds considerable complexity, multiple

appropriately timed measurements are important for exploring mediators of treatment.

5.4. Inclusion of Study Measures with Appropriate Measurement Properties

Measurement error is a particularly important problem when trying to assess whether change

in a particular factor is responsible for the subsequent change in outcome, as in mediation

analysis. Measurements need to be highly reliable when measuring change (34)

to ensure

accurate estimates. Measurement error can become magnified when a raw change score

(difference between pre- and post-test) is calculated as there is likely to be error associated

with both measurements. (34)

A measure’s sensitivity to change is also important, as this

relates to whether or not a measure can detect change over time and separate this from

measurement error. (35)

The use of measures with poor reliability or responsiveness in

mediation studies may affect the user’s ability to identify effects of mediating variables.

5.5. Incorporation of Exploratory Analysis

In the workshop on mediation analysis held during the XII Forum in Odense 2012 there was

general consensus that, aside from the points mentioned above, there is a need to return to

exploratory analysis (for example mediation analyses in uncontrolled studies) and qualitative

analysis to pinpoint the concepts that are most important in investigating treatment

mechanisms. The importance of qualitative research in helping identify key factors to test as

mediators has been highlighted previously, (11)

suggesting that this can be especially helpful

in areas where no theories are available and can be used initially to explore therapeutic

processes and their impact on therapist and patient. This implies that a programme of

research will often be needed to identify key mediators, with replication of findings to

confirm the importance of those mediators being an essential part of that process. ((19); (28))

ACCEPTED

ACCEPTED

6. Summary and Conclusions

The majority of trials investigating the effectiveness of primary care interventions for back

pain have shown small or at best, moderate effects of treatment ((36); (37))

and the field is

looking for better ways to improve outcomes for patients with back pain. Mediation analysis

aims to provide better insight into the causal pathways underlying treatment effects,

explaining why treatments work or don’t work, and potentially offering new opportunities to

improve patient outcomes by optimising the content or delivery of treatment. Until recently,

mediation analysis in clinical research was often limited to a descriptive evaluation of the

processes potentially underlying the effects of treatment in trials. ((38); (39))

Over the past few

years, interest in the use of more sophisticated approaches to mediation analysis has

increased, often guided by methods described in the psychological literature. In this article,

we have summarized the concepts and different designs used in mediation analysis and

explained the importance of experimental designs when investigating mediators of treatment

effect. We have also emphasized the importance of other considerations such as defining and

understanding constructs, selecting study measures with appropriate measurement properties,

and ensuring study measurement time-points are appropriately selected, in order to

investigate the longitudinal associations between mediating and outcome variables. We have

also outlined the relevance of observational and qualitative research in identifying potential

mediating factors. Based on the discussions during the 2012 workshop and supported by the

literature we have proposed a set of recommendations to support and improve the design of

mediation analysis in back pain research (Box 1), with the ultimate aim to improve the design

and delivery of intervention studies and optimize outcomes for patients with back pain.

ACCEPTED

ACCEPTED

Acknowledgements: Chris Main would like to thank Dr Kelvin Jordan for his statistical

advice.

References

(1) Van der Windt D, Hay E, Jellema P, et al. Psychosocial interventions for low back pain in

primary care: Lessons learnt from recent trials. Spine 2008;33: 81-9.

(2) Guzman J, Esmail R, Karjalainen K, et al. Multidisciplinary rehabilitation for chronic

low back pain: Systematic review. BMJ 2001;322: 1511-6.

(3) Haldorsen EMH, Grasdal AL, Skouen JS, et al. Is there a right treatment for a particular

patient group? Comparison of ordinary treatment, light multidisciplinary treatment, and

extensive multidisciplinary treatment for long-term sick-listed employees with

musculoskeletal pain. Pain 2002;95: 49-63.

(4) Shaw WS, Linton SJ, Pransky G. Reducing sickness absence from work due to low back

pain: How well do intervention strategies match modifiable risk factors? J Occup Rehabil

2006;16: 591-605.

(5) Kraemer HC, Wilson GT, Fairburn CG, et al. Mediators and moderators of treatment

effects in randomized clinical trials. Arch Gen Psychiatry 2002;59: 877-83.

(6) Kraemer HC, Stice E, Kazdin A, et al. How do risk factors work together? Mediators,

moderators, and independent, overlapping, and proxy risk factors. Am J Psychiatry 2001;158:

848-56.

(7) Woby SR, Roach NK, Urmston M, et al. The relation between cognitive factors and

levels of pain and disability in chronic low back pain patients presenting for physiotherapy.

Eur J Pain 2007;11: 869-877.

(8) Arnstein P, Caudill M, Mandle CL, et al. Self efficacy as a mediator of the relationship

between pain intensity, disability and depression in chronic pain patients. Pain 1999;80: 483–

91.

ACCEPTED

ACCEPTED

(9) Gheldof ELM, Vinck J, Van den Bussche E, et al. Pain and pain-related fear are

associated with functional and social disability in an occupational setting: Evidence of

mediation by pain-related fear. Eur J Pain 2006;10: 513–25.

(10) Liu LC, Flay BR, Aban Aya Investigators. Evaluating mediation in longitudinal

multivariate data: Mediation effects for the Aban Aya Youth Project drug prevention

program. Prev Sci 2009;10: 197-207.

(11) Kazdin AE. Mediators and mechanisms of change in psychotherapy research. Annu Rev

Clin Psychol 2007;3: 1-27.

(12) Kamper SJ, Maher CG, Costa L da CM, et al. Does fear of movement mediate the

relationship between pain intensity and disability in patients following whiplash injury? A

prospective longitudinal study. Pain 2012;153: 113-9.

(13) Costa L da CM, Maher CG, McAuley JH, et al. Self-efficacy is more important than

fear of movement in mediating the relationship between pain and disability in chronic low

back pain. Eur J Pain 2011;15: 213–9.

(14) Hall AM, Kamper SJ, Maher CG, et al. Symptoms of depression and stress mediate the

effect of pain on disability. Pain 2011;152: 1044–51.

(15) Collins LM, Graham JW, Flaherty BP. An alternative framework for defining

mediation. Multivar Behav Res 1998;33: 295-312.

(16) Lockhart G, MacKinnon DP, Ohlrich V. Mediation analysis in psychosomatic medicine

research. Psychosom Med 2011;73: 29-43.

(17) Maric M, Wiers RW, Prins PJM. Ten ways to improve the use of statistical mediation

analysis in the practice of child and adolescent treatment research. Clin Child Fam Psychol

Rev 2012;15: 177-91.

ACCEPTED

ACCEPTED

(18) Leeuw M, Goossens MEJB, van Breukelen GJP, et al. Exposure in vivo versus operant

graded activity in chronic low back pain patients: Results of a randomized controlled trial.

Pain 2008;138: 192–207.

(19) Mulroy SJ, Winstein CJ, Kulig K, et al. Secondary mediation and regression analyses of

the PTClinResNet database: Determining causal relationships among the international

classification of functioning, disability and health levels for four physical therapy

intervention trials. Phys Ther 2011;91: 1766-79.

(20) Bullock JG, Green DP, Ha SE. Yes, but what’s the mechanism? (Don’t expect an easy

answer). J Pers Soc Psychol 2010;98: 550-58.

(21) Murphy R, Cooper Z, Hollon SD, et al. How do psychological treatments work?

Investigating mediators of change. Behav Res Ther 2009;47: 1-5.

(22) Judd CM, Kenny DA. Process analysis: Estimating mediation in treatment evaluations.

Evaluation Rev 1981;5: 602-19.

(23) Emsley R, Dunn G, White IR. Mediation and moderation of treatment effects in

randomised controlled trials of complex interventions. Stat Methods Med Res 2010;19: 237-

70.

(24) Grissom RJ, Kim JJ. Effect sizes for research: Univariate and multivariate applications.

(2nd

Ed.) New York: Taylor & Francis (Routledge); 2012.

(25) Hayes AF. Beyond Baron and Kenny: Statistical mediation analysis in the new

millennium. Commun Monogr 2009;76: 408-20.

(26) MacKinnon DP, Fairchild AJ, Fritz MS. Mediation analysis. Ann Rev Psychol 2007;58:

593-614.

(27) MacKinnon DP, Fairchild AJ. Current directions in mediation analysis. Curr Dir

Psychol Sci 2009;18: 16-20.

ACCEPTED

ACCEPTED

(28) MacKinnon DP. Analysis of mediating variables in prevention and intervention

research. National Institute on Drug Abuse Monograph 1994;139: 127-53.

(29) MacKinnon DP. Integrating mediators and moderators in research design. Res Social

Work Pract 2011;21: 675-81.

(30) Mathieu JE, Taylor SR. Clarifying conditions and decision points for mediational type

inferences in organizational behaviour. J Organiz Behav 2006;27: 1031-56.

(31) Holbert RL, Stephenson MT. Structural equation modelling in the communication

sciences, 1995-2000. J Hum Commun Res 2008;28: 531-51.

(32) Tomarken AJ, Waller NG. Potential problems with “well fitting” models. J Abnormal

Psychol 2003;112: 578-98.

(33) Fritz MS, MacKinnon DP. Required sample size to detect the mediated effect. Psychol

Sci 2007;18: 233-9.

(34) Streiner DL, Norman GR. Health measurement scales: A practical guide to their

development and use.( 4th

Ed). Oxford: Oxford University Press; 2008.

(35) Terwee CB, Bot SDM, de Boer MR, et al. Quality criteria were proposed for

measurement properties of health status questionnaires. J Clin Epidemiol 2007;60: 34-42.

(36) Keller A, Hayden J, Bombardier C, et al. Effect sizes of non-surgical treatments of non-

specific low-back pain. Eur Spine J 2007;16: 1776-88.

(37) Machedo LA, Kamper SJ, Herbert RD, et al. Analgesic effects of treatments for non-

specific low back pain: A meta-analysis of placebo-controlled randomized trials.

Rheumatology (Oxford) 2009;48: 520-7.

(38) Jellema P, van der Windt DAWM, van der Horst HE, et al. Why is a treatment aimed at

psychosocial factors not effective in patients with (sub)acute low back pain? Pain 2005;118:

350-9.

ACCEPTED

ACCEPTED

(39). Thorn BE, Day MA, Burns J, et al. Randomized trial of group cognitive behavioural

therapy compared with a pain education control for low-literacy rural people with chronic

pain. Pain 2011; 152: 2710-20.

ACCEPTED