Introduction

Sciatica is a term used to denote a collection of pain symptoms which may have several causes related to compression or irritation of spinal nerve roots leading into the sciatic nerve. The pain itself is typically located in the lower back, with the predominant pain radiating out into the leg, usually in only one side of the body. Pain symptoms range from mild to severe. Especially with severe pain, the disease can be debilitating, having a profound impact on physical and social functioning and lead to long periods of absence from work. The most common and well-known cause is spinal disc herniation (90% of diagnosed cases [1]). The disease is relatively common and affects 5 to 10 patients per 10000 individuals in Western countries annually [2] with disease risk increasing with known factors such as age, height, mental stress, smoking and exposure to vibration. There is no conclusive evidence of an association between risk of sciatica and gender or physical fitness. Fortunately, prognosis on acute sciatica is generally good, with 60% of patients recovering within 3 months and a further 10% within 12 weeks up to a full year [1] [2] [3]. Up to 30% of patients will however continue to experience pain for one year or longer [3] [4]. Absenteeism from work causes lumbar-spine disorders, such as sciatica, to have profound economic effects. Treatment options may be categorized into either conservative approaches (such as staying active, analgesics, non-steroidal anti-inflammatory drugs, up to multidisciplinary treatment) or invasive intervention through disk surgery. There is considerable dispute internationally about both long-term value and appropriate timing of surgical intervention, although consensus does exist that surgery should only be attempted after an initial period of conservative treatment. Few clinical trials exist to compare surgery with conservative treatment. Those that are available show no significant differences after at least 4 years follow-up [5] or even for shorter follow-up periods of two years [6] [7] [8] [9]. These trials suffer from poor compliance rates in the surgery groups, particularly for the latter trial (60% only for two years follow-up). Debate about either appropriateness or timing of surgery are further fuelled by uncertainty about disease progression and potential recovery (in the early stages) of the disease. Similar uncertainty affects our inability to screen out those patients who will eventually experience a malign and prolonged period of pain from what is otherwise a likely benign condition for most patients. Furthermore there are substantial sociocultural differences between - and even geographic variation within - countries on treatment preferences and practices. These are also part due to differences in health care provision, insurance systems and legislation related to work absenteeism. Thus for example, both the Netherlands and the United States have relatively high surgery rates for sciatica compared to other developed countries [7] [10]. The Netherlands sciatica guideline suggested surgery after a 6 weeks period of non-operative care, while other countries waited 6 months. Surgical treatment rates for lumbar discogenic syndromes for the United States are 40% higher than in any other country and more than 5 times larger than similar rates in the United Kingdom [11]. Surgery rates also seem to depend strongly on the number of neurologic and orthopaedic surgeons available per capita and preferences for surgery in other pathologies [12]. We apply both propensity score-based estimation and inverse-probability-of-treatment weighted estimation (marginal structural models [13]) to high quality data obtained in the framework of a clinical trial. The objective of these analyses is to reduce bias incurred in estimating the per-patient surgery effect. Such analysis is beyond the scope of the intention-to-treat analysis of the original randomization which could only assess the effect of treatment policy (conservative wait-and-see versus early intervention – as opposed to the individual surgery effect). The bias referred to will occur as a consequence of confounding by predictors (such as pain scores or the changes therein) which affect both the decision to take surgery as well as the outcome at end of follow-up [13] [14] [15].

Methods

Results

The wait-and-see group contains 142 patients, approximately 40% (56) of which opted into surgery during follow-up, of whom 15 within two months from randomization, 37 between two months and one year, and four patients after one year. Table 1 gives 95% confidence intervals (CI) and p-values (P) for the differences between treatment groups (Treatment) as well as association with outcome at two years (RMDQ, Likert) for gender, weight, height, age and the baseline pain scores as well the changes in these scores prior to the treatment decision. Comparisons of baseline values between treated and untreated patients at 2 years indicate significantly worse condition for RMDQ (P<0.0001), Likert score (P = 0.014) and VAS leg pain score (P = 0.011), as higher baseline scores are observed for surgically treated patients compared to non-operative and marginally significant for VAS low back pain (P = 0.12). Comparison of the change from baseline between both treatment groups shows significant differences for all pain scores (RMDQ: P<0.0001, Likert score: P<0.0001, VAS leg pain: P<0.0001, VAS low back pain: P<0.0001) with smaller pain reductions from baseline for patients who decided to take treatment. Figure 1 shows a graphical illustration of the situation for the RMDQ scores between baseline and the treatment decision. Individuals who eventually opted for treatment are clearly identified as having higher baseline scores and their profiles after baseline tend to show smaller reductions or even increases in pain score - as opposed to individuals not taking surgery. We did not find substantial evidence of the (lagged) changes in pain scores immediately preceding the decision to treat (or the final follow-up measurement if the patient was untreated) affecting the treatment decision (RMDQ: P = 0.21, Likert score: P = 0.11, VAS leg pain: P = 0.20, VAS low back pain: P = 0.027). We found evidence that the baseline pain scores predict the observed levels at outcome for the RMDQ score (RMDQ: P = 0.018, Likert score: P = 0.023, VAS leg pain: P = 0.012, VAS low back pain: P = 0.0087) but not so for the Likert score (RMDQ: P = 0.85, Likert score: P = 0.92, VAS leg pain: P = 0.75, VAS low back pain: P = 0.46). Likewise, the changes in Roland score from baseline are predictive of the observed level of the same score at end of follow-up (P = 0.0011) and similarly for the Likert score (P = 0.0091). The lagged changes in Likert score between both visits immediately preceding the point of treatment (or end-of follow up if no treatment) was found to be predictive of the observed level of the same score at end of follow-up (P = 0.0013), but not so for the RMDQ score (P = 0.14). There is no evidence of strong predictive effects of age, gender, height or weight on either the decision to treat or on final outcome (table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. The left plot in Figure 1 shows RMDQ scores for each patient separately, plotted as a connected line for each patient between the baseline measured value (“baseline”) and the point immediately prior to surgery - or end of follow-up if no treatment was taken (“treatment decision”).

Treated individuals are shown as red profiles while untreated individuals are represented with blue profiles. To the right, three boxplots are shown for treated and untreated individuals separately and for RMDQ scores at baseline (top), RMDQ scores at “treatment decision” (middle) and for the change in RMDQ score from baseline to treatment decision point (bottom).

https://doi.org/10.1371/journal.pone.0111325.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Tabulation of predictors used in the treatment propensity model, distinguishing between baseline variables and variables subsequent to baseline.

https://doi.org/10.1371/journal.pone.0111325.t001

Stratification on propensity of surgery score

Table 2 shows stratified mean RMDQ scores at baseline and at end of follow-up for both treatment groups and across 6 strata of the propensity score, defined by cut-offs at the 15th, 25th, 50th, 75th and 85th percentile of the propensity score. As in table 1, we find that smaller propensity to have surgery is associated with lower baseline scores. RMDQ scores are slightly elevated in the delayed surgery group as compared to non-operated individuals at baseline within each stratum, indicating a small residual imbalance after stratification. While the latter effect disappears at end follow-up, an upward trend is apparent in mean RMDQ score levels at end of follow-up for untreated individuals, particularly for the last two propensity strata where patients maintain high expected mean RMDQ scores (9.78 and 9.00 for the 75–85% and 85% strata). Such trend is not observed for surgically treated individuals at end follow-up, whose mean scores are all lower than 3.3 with no clear pattern. The per-stratum treatment effect estimates (last column ‘difference’), defined as the differences between mean RMDQ scores between surgically treated and untreated individuals at end follow-up, show clear inhomogeneity of effect and a downward trend as the propensity to receive surgery increases (i.e. ever greater reductions of the RMDQ score at end follow-up). The latter effect appears predominantly due to the above described tendency for RMDQ scores to remain at high levels for untreated individuals with high treatment propensity. This trend in effect estimates is however also affected by the fact that mean RMDQ outcomes remain larger for treated individuals as compared to untreated individuals across the first 3 propensity strata. Figure 2 shows a graphical representation of the same information. The sequence of surgery effect estimates in the last column of table 2 (figure 2) has a strict monotone decreasing order, with the effect estimates actually changing sign from an increase in mean RMDQ score of 1.75 for patients at smallest propensity (due to a relatively good condition at outset of the trial) up to decreases in RMDQ scores of up to −7.93 for patients with the highest surgery propensity. The average weighted surgery effect of −1.74 which combines the stratum-specific effect estimates is in line with estimates from the primary analysis. This estimate, however, masks substantial change of the effect of surgery as the condition of the patient - and thus also the propensity to take the treatment - changes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Figure 2 shows mean RMDQ scores within propensity percentiles strata as defined in table 2, plotted versus the midpoint of each percentile stratum and for treated and untreated patients separately (left-side axis).

The red curve shows the difference in RMDQ scores between the treated and untreated patients at end follow-up across propensity percentile strata (right-side axis).

https://doi.org/10.1371/journal.pone.0111325.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Results for Roland scores stratified for percentiles of propensity scores.

https://doi.org/10.1371/journal.pone.0111325.t002

Discussion

The sciatica randomised trial is an assessment of the effect of policy (early intervention or a wait-and-see approach) at the population level. We have shown that within the wait-and-see arm of the trial, the decision to take surgery is associated with heightened pain scores at baseline and smaller reductions in those scores after baseline as compared to patients who remain free of surgery up to the end of the trial. The sciatica trial found no evidence of significant differences in RMDQ scores at either 1 or 2 years of follow-up between both trial arms. Because of confounding of the treatment decision with past pain score history within the wait-and-see arm, such results from the original trial should however not be (mis) interpreted in terms of lack of potential effect at the individual patient level. Use of inverse-probability-of-treatment weighted estimation and propensity scores in statistical analysis can reduce bias incurred in estimating the per-patient surgery effect, due to treatment predictors being confounded between pain scores at outcome and the treatment decision [14]. Re-weighting methods reduce bias by decreasing the influence of patients who are over-represented in the surgery group at end of trial, because of their heightened probability to receive surgery as a consequence of confounders, such as a relatively poorer condition at outset or relatively greater deterioration of the patient’s state during the observation period. Such variation in individual propensity is of course completely understandable and natural, but cannot be accounted for with classical analysis methods in RCT intent-to-treat analysis, when we are interested to uncover the surgery effect itself. Likewise, patients with small probabilities of receiving treatment in the group which remains free of surgery by end of trial are given higher weight, for similar reasons.

Propensity score-based approaches to treatment effect estimation solve the non-random treatment allocation based on past pain levels, by exploiting the balancing property that patients who share the same level of treatment propensity tend to have similar levels of the confounding variables.

Our application of both re-weighting and propensity-based estimation suggests that patients with markedly poor baseline condition with regard to pain and function and who do not experience sufficient improvement in condition may benefit from surgery intervention, as the decision to take treatment is strongly dependent on such pain history. Patients at the smallest levels of propensity may however be adversely affected by surgery.

These results suggest prediction methods may be devised to screen out patients most likely to benefit from surgery. Unfortunately, the propensity score cannot be used for this purpose - at baseline at least - as it depends on future observations not yet available at the baseline. Propensity scores are only retrospectively derived for the purpose of identifying individuals who are comparable with respect to potential confounders to allow for an assessment of treatment effect, after the data has been collected. Another problem with the propensity-based methodologies presented in this paper is that they are strongly dependent on all confounders between the treatment decision and the outcome pain scores to be known and available. Since our analysis is concerned with the secondary analysis of a randomised trial, this requirement may not be met. It would be helpful - even for randomized trials generally - where treatment decision may depend on prior information to record such information as accurately as possible in order to facilitate such secondary analysis as we have described in this paper, even if this is not the primary purpose of the trial. Likewise, it would be helpful if greater care could be applied in the discussion of randomised trial results to make a careful distinction between effect interpretations at either the population or patient-specific level - as and when required - as for the present trial.

The current paper does not address the issue of validation of research findings presented here. Validation could be attempted in several ways - through comparison with other similar studies - or through some form of predictive validation on new patients - both of which would however raise substantial methodological challenges – and require extensive novel data analytic work. Both of these are beyond the scope of the present paper. A recent publication [18] for example, could provide data of a similar nature as described in this paper. Application of the methodologies described in this paper to that study’s data would however involve calibrating from scratch a suitable propensity score based on that study’s data, as it is inherent in propensity score methodology that the score must be calculated separately for each study. At the time of writing, the purely methodological issues involved in causal inference comparative studies are effectively unexplored. Similarly, research is emerging on application of causal effect estimation approaches to predict future patient’s potential outcomes. Application of this in clinical research is currently however largely unexplored. Readers should beware these issues place restriction on the interpretation of findings presented here.

We should note that this RCT, as well as others, excluded patients in a fairly good state or those predominantly complaining about low back pain instead of leg pain. This results in a lack of contrast with care-as-usual and therefore one has to be careful with generalizing our results to guidelines. However, the results of this propensity-based analysis, in combination with the lack of contrast in trial circumstances as compared to realistic daily care, suggests future studies may be designed to devise prediction methods with a higher predictive value of RMDQ and VAS leg-pain scores than currently considered possible. This provides further argument to the need for careful and well thought through registration in future trials, specifically to record all indicators for surgery - at baseline - but also subsequently in the trial when patients receive surgery post baseline, as well as the duration of sciatica and time to surgery. Such studies may finally deliver medical science the screening methods capable of identifying the most acute cases who are likely to benefit from surgery, while protecting individuals whose condition would unlikely improve due to surgery from unnecessary, costly and potentially harmful intervention.

Supporting Information

Acknowledgments

We want to thank an anonymous referee for pointing out a new reference of which we were not aware at the time of writing.

Trial registry and the registration number information

Current Controlled Trials number, ISRCTN26872154. First published: Peul, W.C., van Houwelingen, H.C., van den Hout, W.B., Brand, R., Eekhof, J.A.H., Tans, J.T.J., Thomeer, R.T.W.M., Koes, B.W. (2007) Surgery versus Prolonged Conservative Treatment for Sciatica. New England Journal of Medicine, 356, 22, 2245–2256.