Responsiveness and minimal clinically important difference for pain and disability instruments in low back pain patients

This study is a secondary analysis of data from a large validation study of the Oswestry Disability Index in Danish [56, 57]. Patients from the primary sector (7 chiropractic practices) and secondary sector (an out-patient hospital back pain clinic) of the Danish health care system were recruited. In Denmark 1/3 of the LBP patients who contact a health care practitioner for treatment are seen by a chiropractor where they receive standard active and passive conservative care. These patients are comparable to patients seen by medical doctors and physiotherapists [58]. The patients seen in the out-patient hospital back pain clinic represent a broad range of chronic LBP patients with or without leg pain who have not responded to treatment in the primary sector. These patients received multidisciplinary evaluation and treatment. Inclusion criteria were: 1) age above 18, 2) presence of low back pain and/or leg pain, and 3) able to read and understand Danish. Patients were excluded if a pathological disorder of the spine was suspected (e.g., fractures, spinal infections, malignancy or inflammatory diseases). All patients received oral and written information about the project and gave their informed consent to participate in the study.

A prospective cohort study design with follow-up at one week, eight weeks and nine weeks. At baseline, one week and eight weeks follow-up, a questionnaire booklet containing sociodemographic data, medical status and outcome measures was administered to all patients. Responders at the eight weeks follow-up received a telephone interview 3–5 days after (week nine) by a specially trained professional interviewer from the Danish National Institute of Social Research. The purpose of the telephone interview was to obtain patient ratings of improvement/deterioration and the importance of such change. A detailed description of the study design can be found elsewhere [56].

The Oswestry Disability Index (ODI) version 2.1 is a self-administered questionnaire measuring "back-specific function" with reference to "today" on a 10 item scale with six response categories each. Each item scores from 0 to 5 and the score is subsequently transformed into 0–100 [2, 59, 60, 60].

The 23-item version of the Roland Morris Disability Questionnaire (RMQ) was developed specifically to target LBP patients with radicular symptoms and is a modification of the original RMQ [61]. We chose the 23-item version instead of the original 24-item version [62] for two reasons: 1) the 23-item version has been cross-culturally validated in Danish whereas the 24-item version has not, and 2) the psychometric properties of the two versions have been shown to be similar [63]. Each item is scaled as yes/no (scored as 1 and 0 points respectively) with the scale ranging from 0 (no disability) to 23 (extremely severe disability).

The Low Back Pain Rating Scale (LBPRS) was developed to measure the dimensions of pain, disability and physical impairment for patients with LBP [64]. The pain assessment index (LBPRS_pain) is measured on 0 to 10 numerical scales with 0 representing no pain and 10 representing worst possible pain. There were three 11-box numeric rating scales (pain now, worst and average pain in the last 2 weeks) for back pain and leg pain separately. Each response scale score is added giving a scale range of 0–60 points. The disability index (LBPRS_disability) comprises 15 items scaled as yes = 0 points, can be a problem = 1 points, no = 2 points, giving a total score of 0–30 points.

The SF36 is a generic 36-item questionnaire compiled from the Rand Health Insurance Long Form Health Status Scale [65]. Of the eight dimensions, we included the physical function (SF36 (pf)) and the bodily pain (SF36 (bp)) subscales. Questions are framed over a one-week period with response scales varying from dichotomous (yes/no) to six-point verbal rating scales. Each dimension is scored on a weighted 0–100 scale and an overall score is recommended [66].

Back and/or leg pain was scored on an 11-box numeric rating scale (pain now) from 0 (no pain) to 10 (worst possible pain) [67, 68].

The patients' global retrospective assessment of treatment effect (transition question) was used to assess the patients' perception of their overall change in their back condition. A 7-point Likert scale transition question (TQ) ranging from "much better" to "much worse" was used [69]. Furthermore, the importance of the change in health state experienced was measured. All patients were asked to rate the question: "How important is the change you have experienced in your back and/or leg pain since the start of the treatment?" on a 0 – 10 numeric rating scale (NRS_imp) with "very important" and "not at all important" at the extremes.

This information was collected by telephone interviews which followed a carefully planned protocol. First, all patients were told their baseline global rating of pain severity (NRS_pain) before answering the TQ to ensure optimal patient focus on the change in health rather than the present health state [70, 71]. Second, the transition question with response options was read twice. In case the patient was uncertain of which response to choose, the interviewer determined whether the patient was either better, had not changed or worse. If the interviewer decided that the patient was better the categories for being improved was read again ("much better", "better", "a little better") and similar for patients classified as worse.

All the included disability instruments have been cross-culturally adapted and validated in Danish [56, 57, 64, 72‐75]. For a complete description of the psychometric properties of each instrument we refer the reader to relevant literature reviews [2‐4, 60, 76‐78].

Patients available at the eight weeks follow-up were divided into subpopulations after either pain location or entry point in the health care system at baseline. For pain location, we looked at patients with LBP only compared to those with leg pain and/or LBP as responsiveness has been found to depend on the type of patient population studied [34, 79]. The second stratification (patient entry point) was chosen as a measure of disease severity. Back pain patients initial contact with the Danish health care system is the primary sector (general practitioners, chiropractors and physiotherapists), thus, representing mostly acute conditions (≤ 30 days of pain). Comparably, referrals to a secondary sector hospital based multidisciplinary spinal unit predominantly represents patients with more chronic conditions (> 30 days of pain) [58].

All scales were transformed to cover an interval ranging from 0 – 100 with a high score representing higher disability or pain. This makes instruments with different scoring intervals comparable despite the fact that they are not equivalent. The raw change score for each outcome measure was obtained by subtracting the eight weeks follow-up score from the baseline score. The percentage change score was calculated as follows: [Raw change score/baseline score]*100 [80].

Responsiveness was operationalised using two strategies; standardised response mean of the raw change scores (distribution-based method) and receiver operating characteristic (ROC) curves (anchor-based method). The standardised response mean of the raw change scores (SRM_raw) was restricted to patients who had changed and calculated as the ratio of the mean raw change score and the standard deviation of that raw change score [17, 81, 82]. Confidence intervals for the SRM_raw were estimated using 200,000 bootstrap samples with replacement [83]. To compare the SRM_raw of the different questionnaires within each subpopulation, we first estimated the SRM_raw using stata's regression command with group indicators and the cluster option to account for intra individual correlation between responses. The differences between SMR_raw were examined with a non-linear Wald test [84]. The same procedure was used to test the difference between "important improvement" and "no change" groups within each subpopulation.

SRM_raw was calculated for all instruments change scores according to where the patients were seen, pain location and whether the patients had experienced an "important improvement" or "no change" (see ROC analyses). The SRM_raw for the "important improvement" group addresses the sensitivity to change. On the other hand, the SRM_raw for the "no change" patients addresses the important issue of specificity to change where change without clinical relevance may occur in instrument scores.

In the second strategy we used ROC curve analyses to determine sensitivity and specificity for classifying patients as having experienced an "important improvement" or "no change" and defined "important improvement" patients from two criteria: 1) had to rate themselves as either "much better" or "better" on the TQ, and 2) had to rate the importance of the change on NRS_imp equal to or more than 7. The "no change" patients rated themselves as either "a little better", "about the same" or "a little worse" or with a rating of the importance of the change less than 7. Because of the low number of patients (n = 13) reporting deterioration, a "worse" sample was not included. The ROC curve is the sensitivity plotted against 1-specificity (false-positive rate) and shows the trade-off between the true-positive successes and the false-positive errors as each of several cut-off points in the change score is assessed [7, 85, 86]. The area under the ROC curve (ROC_auc) can be interpreted as the ability of an instrument to discriminate between "important improvement" and "no change" patients. An area of 0.5 is interpreted as no discriminatory accuracy and 1.0 as complete accuracy [87]. An omnibus statistical comparison of the area under the ROC curve within each subpopulation was carried out using a non-parametric approach as described by DeLong et al. [88].

The MCID was determined by an optimal cut-point analysis using both the raw (MCID) and percentage (MCID_%) change scores. The optimal cut-off change score was identified as the cut-point with equally balanced sensitivity and specificity [89] and this was considered an expression of the MCID. First, we calculated overall MCIDs, quarter-specific MCIDs by dividing the scale range into four equally sized score groups [17, 32], and MCIDs specific for pain location and patient entry point. Categories with less than 10 patients were excluded from the analysis.

Second, ODI and RMQ quarter-specific MCID_% were graphed for each subpopulation. Third, we adjusted the dependence of the MCID on the baseline score by a weighted linear regression. As the number of patients in each baseline score strata was different the regression was weighted by the number of persons used to detect each cut-point (all patients only).

All statistical calculations were analysed using the statistical package STATA^® v. 9.2 SE (StataCorp) and statistical significance was accepted at the P < 0.05 level.

The mean raw change scores and SRM_raw for the two study samples stratified according to pain location are shown in Table 2. As expected the raw change scores for the SeS sample (chronic patients) were lower in comparison to PrS sample (acute patients). To convert the transformed raw change scores to original scale scores please refer to Table 3.

The RMQ proved to be the most responsive disability measure for patients with LBP only (both PrS and SeS samples) and this was statistically significantly different from the other disability measures in the PrS patients (P < 0.001). For patients with leg pain the ODI and RMQ was equally responsive in PrS patients (P = 0.2) as was the case between the disability measures in the SeS patients. Of the 3 pain measures, the SF36 (bp) had the highest responsiveness in all subpopulations. This was statistically significant in the LBP only subgroup for both PrS and SeS patients (P < 0.002).

Lower change scores and SRMs were found for SeS patients. This is because the SRM is dependent on both the effectiveness of the treatment and the patient population characteristics and therefore expected to vary in a study using two distinctly different patient populations [17, 21, 90].

The ODI and RMQ have been compared in several studies, and reported SRMs for the ODI range between 0.2 and 1.9 [26, 36, 37, 46, 61, 91‐93] with a similar range for the RMQ (0.5–2.0) [31, 41, 46, 61, 92, 93]. We found that the RMQ was most sensitive to change in patients with LBP only (significantly different in PrS patients, Table 2) whereas the ODI was slightly more responsive in leg pain patients when considering both SRM and ROC_auc. Several authors have argued that the RMQ is more sensitive to change at lower levels of disability compared to the ODI which is sensitive to change at higher disability levels [60, 91, 94]. Indeed, our data showed a statistically lower mean initial disability scores in patients with LBP only compared to leg pain patients supporting this finding. Furthermore, we found the RMQ to have significantly larger differences in SRM_raw between "important improvement" and "no change" patients in all subpopulations (Table 4).

The LBPRS has not been psychometrically tested for responsiveness until now, and it has been unknown how the responsiveness of this instrument compares to other functional status questionnaires [4, 64]. For the disability subscale we found lower responsiveness using both SRM and ROC_auc in comparison to the other instruments. Second, the responsiveness was conflicting depending on which strategy was used. The smaller SRMs resulted from five outliers in our dataset who showed an improvement in disability and pain on all other instruments, however, rated themselves as getting worse on the LBPRS_disability scale. We suspect these patients have misunderstood the answer categories of the scale thus reversing a positive change score to a negative. A reanalysis omitting the outliers produced SRMs of more comparable magnitude to the rest of the disability measures. Due to the discrepancies in responsiveness according to index used and the effect of the outliers we conclude that the responsiveness of the the LBPRS_disability is inconclusive.

The physical function subscale of the SF36 has been investigated in chronic LBP patients and reported SRMs range from 0.2 – 0.6 [26, 37, 46] and from 0.7 – 0.8 in improved patients [37, 92]. It has also been suggested that the SF36 (pf) is less responsive compared to back-specific questionnaires [37, 46, 92]. Our results suggest that the SF36 (pf) has poorer responsiveness in patients with leg pain compared to the ODI and RMQ when considering both responsiveness indices. However, in LBP only patients the physical function scale showed lower responsiveness in SeS patients while this, remarkably, was approximately equivalent in PrS when compared to the back-specific questionnaires. Thus, we conclude that responsiveness of the SF36 (pf) is dependent on the subpopulation it is applied to.

Overall, the RMQ showed superior responsiveness and discriminative abilities in patients with LBP only which represent the more acute conditions (58% had pain ≤ 30 days) and this was irrespective of where in the health care system they were seen. However, the ODI seemed marginally superior to the RMQ in patients with leg pain +/- LBP corresponding to the more chronic conditions (66% had pain > 30 days) in both PrS and SeS patients. The LBPRS_disability generally demonstrated lower responsiveness in comparison to the other disability measures; however, the responsiveness was conflicting according to which strategy was used.

For the pain measures, we found comparably higher SRMs for the SF36 (bp) in all subpopulations (range: 0.6 – 1.4) which is somewhat higher than previously published values (0.7–1.0) [26, 37, 92]. This finding questions the finding that the NRS_pain is the most responsive pain scale [24, 67]. However, the relatively large SRMs seen in the "no change" group signifies that some patients who indicated "no change" by the external criterion in fact changed modestly on the SF36 (bp) subscale. Reanalysing our data with a less stringent external criteria (including the "a little better" patients in the important improvement group) only altered the mean change score of the "no change" patients slightly and the SRMs remained the same (data not shown). Thus, one may question whether the specificity of the SF36 (bp) subscale is adequate when using a combined external criteria as a golden standard.

The LBPRS_pain showed differing sensitivity to change according to which responsiveness index was used. Using SRMs the LBPRS_pain was equally responsive to the NRS_pain; however, using ROC_auc it was the most responsive pain instrument in LBP only patients. Thus, we conclude that the LBPRS_pain scale is responsive and probably preferable to the NRS_pain as it provides more information about the pain dimension.

In summary, we found that all pain measures demonstrate similar responsiveness and this was in turn comparable to the disability measures. We recommend using the LBPRS_pain as it is easy to use and provides more information about the patients' pain.

The optimal design and analytic strategies for a responsiveness study are topics of much debate with little or no consensus [16, 95‐98]. However, a recent article suggests that analytic strategies in studies of responsiveness should be based on the chosen study design and their corresponding sample change characteristics [99]. In our design we included both PrS and SeS patients to allow for subpopulation analysis and the patient composition can therefore be viewed as heterogeneous with identifiable subgroups of patients who change by different amounts. Stratford et al. argues convincingly that the proper analysis for this design would be either the area under ROC curve or Norman's S_repeat, and our inclusion of SRMs may therefore seem obsolete. We have chosen to include both analyses as most researchers and clinicians are familiar with interpretation and application of effect sizes in comparison to ROC curves. Furthermore, the overall conclusions about responsiveness would not change (with the exception that the LBPRS_disability subscale would have comparable responsiveness) using the ROC curve analysis alone.

The concept of the MCID defines the smallest meaningful change score for outcome measures. An assumption behind this concept is that the instruments can indeed detect this change. Ultimately, one may question the ability of well established outcome measures to determine the smallest meaningful change as the "true" MCID is unknown. Further, the variability of the MCID is large as it is context-specific and not a fixed attribute [96].

Published MCID values for the included instruments range from: 4 – 16 points (ODI) [27, 37, 41, 52, 92, 100], 3 – 5 points (RMQ) [31, 32, 41, 60, 61, 63], 7 – 16 points (SF36 (pf)) [61, 101], and 2–3 points for the NRS_pain [80, 102, 103]. MCIDs specific to LBP patients for the SF36 (bp) and LBPRS could not be located in the literature. Our overall MCID estimates fall within reported ranges for all the instruments apart from a slightly higher MCID for the SF36 (pf). The MCIDs were generally lower for all subpopulations compared to the overall MCID, however, only minor differences were found between stratification layers (except for the subscales of the SF36). We were surprised to find similar MCIDs in the PrS and SeS samples since the perception of disease (and thus the need for improvement) has been shown to differ [58].

Stratford and Riddle have shown a large increase in MCID with increasing raw baseline score for the RMQ [32, 33]. We found this pattern to be true for the overall MCID for all outcome measures and for PrS and LBP only patients (acute patients). However, using the percentage change scores of the ODI and RMQ, the MCID_% was more or less independent of the baseline entry score for all subpopulations (figure 2). This suggests that patients relate to a percentage change in their condition rather than to an absolute change when quantifying an important improvement. Interestingly, the percentage change signifying an important improvement was dependent on the severity of the condition. PrS and LBP only patients (less severely affected) had to change significantly more (65%–81%) compared to the more severely affected SeS and leg pain +/- LBP patients (28%–36%). Maybe the more disabled leg pain patients have learned not to have too high expectations to the outcome of treatment?

Since the meaning of change varies according to baseline entry score, it seems reasonable to assume that other baseline characteristics may affect the MCID [96]. The present study examined the effect of patient entry point into the health care system (primary or secondary sector) and pain location (LBP only or leg pain) on the MCID, and found these factors to be of minor importance for most of the included disability and pain measures. An exception was the physical function and bodily pain subscales of the SF36 which showed large variations in MCID according to patient entry point and pain location.

In conclusion, we found that the overall MCID varied only slightly when stratifying patients according to point of entry into the health care system (i.e. acute vs. chronic patients) and pain location (LBP vs. leg pain +/- LBP) with the two subscales of the SF36 as an exception. Furthermore, increasing baseline entry scores resulted in greatly increased MCIDs in PrS patients and patients with LBP only. However, the dependence on baseline entry score was not monotonous for all measures in the SeS and for patients with leg pain.

The results of this study should be interpreted in light of several potential limitations. The classification of the ODI and RMQ as purely disability instruments may be misleading as virtually all items in each questionnaire inquire about functional activities in relation to pain [2, 3, 60]. Comparing these instruments to the SF36 (pf) which only measures function of daily living and to the LBPRS_disability which partly measures pain related function (33% of the items) and function of daily living may be problematic. Second, we reported overall responsiveness and MCIDs for a broad spectrum of care-seeking LBP patients' receiving treatments ranging from simple advice to intensive multidisciplinary rehabilitation. Consequently, responsiveness and MCIDs for specific subgroups of LBP patients are likely to vary depending on such factors as entry point into the health care system, pain location, treatment received and possibly psychosocial factors, as indicated by our subgroup analyses. Statistical power issues prevented us from further sub-dividing the sample, and estimates presented are to be regarded as an overall guideline. Therefore, we recommend that researchers calculate MCIDs relevant for their individual study populations and use this when reporting the proportion of improved patients and numbers needed to treat in a clinical trial [104]. Third, the validity of using a global retrospective appraisal of change has been challenged especially with respect to recall bias [22, 105]; however, this may be a minor problem [36, 106]. The validity of combining two different dimensions (improvement and importance) may also be a problem since little is known about its psychometric properties. The combination was used because both improvement and importance is central to the concept of the MCID. Further, the cut-point used to describe who has improved or stayed the same was arbitrarily set for both dimensions. However, our results showed correlation coefficients greater than 0.63 (recommended threshold of 0.5 [105]) between the change scores and the transition question for 5 out of the 7 instruments and an expectedly lower correlation between the change scores and the rating of importance (data not shown) and between the transition question and the rating of importance (0.43). Fourth, the decision of having at least 10 patients in each baseline entry score category was arbitrarily set before the analysis was carried out. Most categories had more than 20 patients making the analysis more reliable. Lastly, some of the MCID cut-points resulted in poor sensitivity or specificity reducing the discriminative ability and validity of the cut-point. However, this occurred in only 10% of the calculations and we consider this acceptable.