Acute low back pain in primary care: can emotional distress explain high healthcare use?

13th May 2015 

Adrian C Traeger, Markus Huebscher, Hopin Lee, Christopher M Williams, Chris G Maher, Nicholas Henschke, G Lorimer Moseley, James H McAuley

Our aim is to investigate the relationship between emotional distress and healthcare overutilization in patients with acute low back pain. We hypothesise that after the index visit:

  • Emotional distress has an independent effect on healthcare use in both the short- and long-term;
  • Emotional distress moderates the effect of pain and disability on healthcare use after the index visit.


Data sources

This study used data from 2 previous studies on a total of 2891 patients consulting primary care for their acute low back pain in Sydney, Australia. The aim of the randomised trial (Study 1) was to test the effect of paracetamol on acute low back pain. Healthcare use was recorded in the 3-months following the index visit. The cohort study (Study 2) aimed to describe the course of acute low back pain. Healthcare use in the observational study was recorded 12-months after the index visit.

Primary outcome

The primary outcome will be the number of primary healthcare visits measured in the 3-months and 12-months following the index primary care visit.

Data analysis

PART I – The total effect of distress on healthcare use in the short- and long-term.

To test our first hypothesis, we will model the association between emotional distress and healthcare use using a regression analysis. The dependent variable will be healthcare use measured as the number of primary care visits at 3- and 12-months. Because Study 1 measured short-term healthcare use and Study 2 measured long-term healthcare use, we will build a separate model for each data set. To determine the appropriate statistical model to use, we will examine the distribution of the healthcare use data. If the data approximates a Poisson distribution, we will perform a Poisson regression analysis. If data appears normally distributed, we will perform a multiple linear regression analysis. If the healthcare data is neither normal nor Poisson distributed, a negative binomial model will be used. Although it is common to dichotomise healthcare use data into high- and low-usage, we will retain the visits as count data. Retaining the count data avoids the substantial loss of information and statistical power that results from dichotomising variables. The independent variable, emotional distress, will be measured using two single item questions. These two items, “How tense or anxious have you felt in the past week” and “How bothered by feelings of depression have you been in the past week” are measured on a 10-point scale, and are taken from a valid and reliable acute low back pain screening questionnaire.

In order to identify the most important covariates to include in our model, we have constructed a directed acyclic graph according to the method of Shrier. DAGs use a graphical approach to identify, a priori, important sources of confounding. This approach is suggested to be superior to traditional methods used to identify potential confounders, eg automatic stepwise methods and ‘the relative change in the adjusted estimate’ method. These traditional statistical approaches have been criticized for their atheoretical foundation and for their undesirable ability to introduce, rather than reduce, bias. Consequently, epidemiological researchers have recommended that for aetiological research questions, covariates should be chosen on the basis of an underlying causal structure rather than on statistical associations.

According to our DAG, adjusting for important predisposing factors (age, gender, socioeconomic status, past history, cultural background) and illness factors (pain, disability, co-morbid illness), is likely to reduce bias in our estimate. One potentially important confounder, marital status, was not measured in either data set and will have potential to introduce bias in the final model.

We will test for multicollinearity in the distress variables using collinearity diagnostics (Tolerance and Variation Inflation Factor) in SPSS. If we detect multicollinearity between different types of emotional distress (anxiety and depression), a composite score will be produced by factor analysis. If there is no multicollinearity detected, we will force both factors into the first block of the model.

Co-variates to adjust for when modeling the effect of emotional distress on healthcare use:

  1. Pain intensity (0-10)
  2. Leg pain (y/n)
  3. Disability (Roland-Morris Score or 0-10 scale)
  4. Co-morbid illness (Self-rated general health 1-5 likert scale)
  5. Age
  6. Gender
  7. Socioeconomic status (post code)
  8. Past history of low back pain (y/n)
  9. Compensation status (y/n)
  10. Cultural background (Born in Australia y/n)*

* only measured in Study 2.

PART II – The moderating effect of distress on the relationship between symptoms (pain/disability) and healthcare use

To test our second hypothesis, we will model the association between symptoms and healthcare use in both the short- (Study 1) and long-term (Study 2). To select important confounders of the relationship between pain/disability and healthcare use we will use the same approach and causal diagram from PART I. If no multicollinearity is detected in measures of distress, we will investigate potential moderating effects of anxiety and depression using the two scales separately. We will include the following interaction terms in a regression model: Anxiety*Disability, Anxiety*Pain, Depression*Disability, Depression*Pain. Moderation analysis will be performed using the PROCESS macro in SPSS, written by Andrew F. Hayes. The PROCESS software helps with moderation analysis by centering continuous variables and producing interaction plots to interpret interaction effects.

Covariates to adjust for in the second model:

Predisposing factors

  1. Age
  2. Gender
  3. Socioeconomic status (post code)
  4. Past history of low back pain (y/n)
  5. Cultural background (Born in Australia y/n)*

* Only measured in study 2.