Translation and adaptation of the German version of the Veterans Rand—36/12 Item Health Survey

The translation methodology followed a rigorous iterative forward–backward format to maintain the conceptual, functional, linguistic and cultural equivalence between the original (English) and the adapted (German) questionnaire. The translation procedure is summarized in Fig. 1. First, a German translation of the VR-36 was produced from the English original version by an experienced translator (DB). Because the VR-36 is analogous to the SF-36, the official German translation of the SF-36 items, which has already undergone rigorous translation and adaptation, served as a second translation to which we compared the forward translated VR items (German SF-36 Version 1 [8‐10] and Version 2 [43]). A reconciled German VR-36 was produced after discussion of agreements and disagreements between the forward translation, SF-36 Version 1 and SF-36 Version 2, and translated back into the source language (English) by an experienced translator who is a native speaker of English and fluent in German. The backward translation was compared to the original English VR-36. Any discrepancies between the back translation and the English VR-36 were addressed with the back translator to determine the origins of discrepancies in the first reconciled German VR-36. After this stage, a pre-final version was produced, which was tested in a cognitive debriefing process with 26 patients and finalized afterwards.

The VR-36 and the VR-12 were analyzed regarding the completeness of data on the scale-level, distributional properties, construct validity, known-groups validity, internal consistency (as one aspect of reliability), and responsiveness to change. This was done on the summary scores of the VR-36 and the VR-12 (physical component score (PCS_VR) and mental component score (MCS_VR)) as well as the eight VR-36 scales: (physical functioning (PF_VR-36), role functioning/physical (RP_VR-36), role functioning/emotional (RE_VR-36), vitality (VT_VR-36), mental health (MH_VR-36), social functioning (SF_VR-36), pain (BP_VR-36), and general health (GH_VR-36)). The VR instruments have not previously been used in German populations and normed scores have not yet been developed. Therefore, summary scores and scales were scored according to the VR-36 and VR-12 algorithms, using a t-score transformation with a mean of 50 and a standard deviation of 10 and normed to a general sample of the US population for the summary scales (PCS and MCS) [23, 60‐62]. The scoring algorithms for the VR-36 and the VR-12 impute for missing data. VR-12 extrapolates scoring based on the missing pattern; VR-36 conducts mean imputation at the subscale level if less than 50% of the subscale items is missing. In all analyses, all available data were used (available case analysis). Because the SF-36 and the SF-12 instruments are well validated across a range of populations, they were used as the comparator to the VR instruments for all analyses.

Distributional properties (such as means, standard deviations and range) for the VR instruments were analyzed on the scale and summary score levels. To compare the distributional properties of the PCS and MCS for both the VR-12 and SF-12 as well as the VR-36 and the SF-36, classical statistical indices of distribution such as mean, standard deviation, minimum, maximum, skewness (to assess and compare the type and strength of symmetry) and kurtosis (as a measure of the steepness / flatness of the frequency distribution) were assessed. Kolmogorow-Smirnov-test was used to compare the distributions of the two summary scores of the VR and the SF—i.e. PCS_VR and PCS_SF as well as MCS_VR and MCS_SF. Kernel density plots using the Epanechnikov function were used to visually examine distribution of summary scores and scales.

Construct validity refers to the degree of accuracy with which a measurement instrument captures the construct it claims to measure. To examine construct validity, Pearson correlation coefficients (r_p) between VR summary scores (PCS_VR and MCS_VR) and other self-completed health measures were assessed. We compared these to the correlations between the PCS_SF and MCS_SF with other self-completed health measures. Correlation coefficients were compared using significance tests for correlations for independent samples [63]. The correlations between PCS_VR and other self-reported physical health measures (e.g. HFAQ, CDC Physical unhealthy days, GCPS Disability) were expected to be higher (convergent validity) than with self-report measures of mental health (divergent validity). Similarly, MCS_VR is expected to be more strongly correlated with self-reported mental health measures (e.g. DASS-Anxiety, DASS-Stress, DASS-Depression, CDC Mental unhealthy days) than with physical measures. Both PCS and MCS are expected to be similarly correlated with generic self-report measures (e.g. EQ VAS, IMET) and GCPS-Pain. Correlations were interpreted as follows: r_p < 0.1 small, 0.3 ≥ r_p < 0.5 moderate, r_p ≥ 0.5 high/strong [64].

Known-groups validity is a criteria-based technique to investigate the ability of a measure to discriminate between groups known to differ in the construct of interest. For this study, known-groups were defined by clinical indication (psychosomatic, orthopedic), treatment program (“curative therapy” typically for chronically ill patients, “medical follow-up treatment” generally after joint replacement, only for orthopedic patients) age (< 45 years, 45–65 years, > 65 years), duration of rehabilitation (median), sick days in the past 12 month, self-rated health (SRH, “excellent/very good/good” vs. “fair/poor”). We examined if mean PCS_VR and mean MCS_VR scores were significantly different between those pre-defined groups using t-tests for two groups or ANOVA for more than two groups.

Internal consistency (IC) is a measure of reliability. A scale is considered reliable if its items are homogeneous—i.e., highly correlated because they measure the same underlying construct [65]. In this study, Cronbach's alpha was used as a measure of IC with α ≥ 0.7 interpreted as acceptable, α ≥ 0.8 as good, and α ≥ 0.9 as excellent.

Responsiveness refers to a self-assessed health instrument’s ability to capture changes in health over time [66]. The raw difference of SF and VR summary scores from t1 to t2 were divided by the pooled standard deviation of change to produce standardized response means (SRM), or divided by baseline standard deviation to produce standardized effect size (SES). As we assess patients before and after an intensive treatment, analysis were restricted to respondents who reported stable (t1 = t2) or improved (t1 < t2) health on a single SRH item (n = 133) to assess responsiveness to health improvements. We further checked improvement (from t1 to t2) for all PCS- and MCS-scores of all four instruments using paired t-tests. The magnitude of changes in scores (expressed as SRM and SES) was interpreted as following: values of < 0.3 were considered as small, values between 0.3 and 0.59 were considered as medium, and values ≥ 0.6 were considered as large [67]. Since there are different methods to estimate the magnitude of change within groups, and consensus is lacking on their interpretation [68], we are calculating both SES and SRM for comparison purposes. Due to the repeated measurement design the measurements are correlated, which was shown to affect the magnitude of SRM [69]; to account for this, we additionally correlated both measurements (Pearson correlation coefficient, r_t1/t2).

Data were analyzed using IBM SPSS Statistics 24 and STATA SE 13. Wherever applicable, analyses were stratified by clinical indication (orthopedic or psychosomatic rehabilitation).

There were no major problems found in the forward–backward-translations. Reconciliation of the items did not lead to problems. The field test yielded that most of the questions (except for RE and RP instructions, response scales and questions) of the VR-36 are clear and simple to both rehabilitation patients (n = 15, 4 male, 11 female, 30 to 80 years (mean 55.3 years)) and patients from general practice (n = 11, 25 to 77 years (mean 57.4 years)) of all ages. Additional file 1 shows the key issues that were discussed during the translation process (forward–backward translation, reconciliation and cognitive debriefing) and how the items were reconciled. Besides the already described adaptation needs identified during the cognitive debriefings, adaptations to the cultural context were needed. The German SF-36 was used as a guide in these decisions. For example, playing golf (used as example in one item) is a less popular activity in Germany than for the USA. In the considerations for a culturally appropriate counterpart, hiking and walking were found to be appropriate but not practicable. We therefore removed the example as was also done for the German SF-36. In two items (BP2, SF1), for purposes of international equivalency, the right-most response category “extremely” was translated into German as “sehr” (English: “very much”), which is also used by the German version of the SF-36.

During the translation process, some double negatives were introduced as a result of combining the questions with their response choices (e.g. “[…] nicht so lange […]” (part of the question) “nein, nie” (response option)). As these double negatives also exist in the English version of the instruments, they were left in the German translation. However, nearly every third field-test participant had problems with the double negatives. Therefore, “yes” and “no” were omitted for these response categories to clarify the language. From a linguistic point of view, these revised response categories resemble the English SF-36 Version 2 and the German SF-36 (versions 1 and 2).

The final German VR-36 is conceptually identical to the English original.

At t1, data are available from n_t1 = 399 orthopedic (response: 99.8%) and n_t1 = 292 psychosomatic (73%) rehabilitation patients. From n_t2 = 378 of the 399 orthopedic (94.7%) and n_t2 = 248 of the 292 psychosomatic (84.9%) patients data are also available for follow-up. Due to block-randomization, number and sample characteristics of participants were balanced across all four groups (n_VR-36 = 169, n_SF-36 = 174, n_VR-12 = 177, n_SF-12 = 171, Table 1). Study participants were on average 53 ± 10.6 (20–89) years old; 67.7% were women and 48.3% were fully employed. About every fourth participant (26.8%) completed high school. Average duration of inpatient rehabilitation (for their primary diagnosis) was 22 days for orthopedic and 35 days for psychosomatic patients (overall mean = 27.5 days). There were no systematic differences in the self-reported health status at baseline between the four study arms (CDC general health status p(χ²) > 0.05). Socio-demographic and clinical characteristics were comparable across the four arms of the study, which allowed for indirect comparisons (Table 1). The most common primary diagnosis were diseases of the musculoskeletal system and connective tissue (ICD-10: M00-M99: 48.9%), affective disorders (ICD-10: F30.0-F39-0: 19.8%) and neurotic, stress and somatoform disorders (ICD-10: F40.0-F49.0: 13.6%).

Missing values were acceptable (< 7%) for the VR-36 and comparable to missing data patterns of the SF-36 (Table 2). The scale GH had the lowest percentage of completion for both the SF-36 (93.1%) and VR-36 (93.5%). As expected, there is a tendency of missing values to increase with increasing age and lower education.

Table 3 gives the distributional properties of the PCS and MCS for the VR and SF short and long form versions. Means, standard deviations and ranges for PCS were very similar between SF-36 and VR-36, SF-12 and VR-12. For MCS, mean differences (e.g. mean VR-36: 36.2, mean SF-36: 39.0), skewness, kurtosis, minimum and maximum of the distribution were larger between the SF-36 and VR-36 than between the SF-12 and VR-12.

For the long and the short form versions of the VR and the SF, the PCS has normal distributions (p = 0.057 to 0.097) while the MCS does not (p < 0.05, Table 3). The findings do not substantively change when stratified by study arm and clinical indication (results not shown).

The VR-36 scales distribute toward slightly lower scores than the SF-36 on the MCS, but not for the PCS. Kernel density plots show that the four instruments were more similar in PCS for orthopedic and MCS for psychosomatic patients. The distributions were more similar between the SF-12 and the SF-36 than between the SF and the VR instruments in PCS for psychosomatic and MCS for orthopedic patients (Fig. 3a). Differences were observed after stratifying by clinical indication. For the scales of the instruments, kernel plots of the VR-36 and the SF-36 are comparable for PF and BP, RP and RE, while kernel plots of SF_VR-36, VT_VR-36 and MH_VR-36 are slightly more left-skewed compared to the SF-36 (Fig. 3b).

Table 4 presents the correlations between VR and SF component scores and other self-reported measures. Moderate to strong correlations were observed between convergent measures with similar correlations observed across the VR-12 and SF-12, and VR-36 and SF-36. Differences (∆) of correlations between corresponding measures (PCS: HFAQ, CDC healthy days physical unhealthy days; MCS: DASS, IRES-VT, CDC mental unhealthy days) were below r_p = 0.090 and with one exception (IRES-VT vs. MCS for the short versions) statistically not significant (p > 0.5).

The PCS_VR had moderate to strong correlations (r_p = 0.33 to r_p = 0.62) with generic measures and strong correlations (r_p = -0.55 to r_p = 0.76) with physical health measures. The MCS_VR had moderate correlation with generic health measures (r_p = 0.32 to r_p = 0.49) and strong correlations with mental health measures (r_p = -0.60 to r_p = 0.78).

At r_p =  -0.5 (PCS_SF-12) and r_p = -0.6 (PCS_VR-12), the correlation between the short versions of the PCS and the GCPS Pain was greater than for the long versions (both PCS_SF-36 and PCS_VR-36 r_p = -0.4). The MCS of all versions was almost uncorrelated with the GCPS Pain (r_p = -0.172 to r_p = 0.013).

Table 5 illustrates the PCS_VR-36 and MCS_VR-36 scores in sub-samples of known groups. Lower mean PCS_VR-36 was found for orthopedic patients while lower mean MCS_VR-36 was found for psychosomatic patients. In line with our hypothesis, higher mean PCS_VR-36 and MCS_VR-36 scores were found in younger patients with fewer sick days in the last year and a shorter duration of rehabilitation. As expected, at baseline, orthopedic patients reported better mental health compared to psychosomatic patients and the other way around for mental health, which is reflected by higher mean MCS_VR-36 scores in orthopedic and higher mean PCS_VR-36 scores in psychosomatic patients. Results were similar for VR-12 and VR-36 suggesting that both instruments perform similarly with respect to known-groups validity (Table 6): all MCS and PCS scales differentiated groups based on clinical indication, duration of rehabilitation and self-rated health, PCS_VR-12 additionally for sick days. As this is only applicable for orthopedic patients, both PCS scales additionally differentiated for type of therapy.

Except for GH (acceptable), IC was good to excellent for both VR and SF scales and with one exception (MH) always higher for the VR scales (Table 7).

Responsiveness to change analysis included the n = 50 to n = 88 cases with no deterioration in SRH from t1 to t2, stratified as necessary by study arm (Table 8). For PCS, SES varied from 0.102 (VR-36 psychosomatic) to 0.398 (SF-12 orthopedic) and SRM varied from 0.127 (VR-36 psychosomatic) to 0.695 (VR-12 orthopedic) with better responsiveness across all instruments for orthopedic patients. Effect sizes of the short versions (VR-12, SF-12) were larger than those of the long versions (VR-36, SF-36). In psychosomatic patients, responsiveness to change of MCS was at least twice as large as responsiveness of PCS, while in orthopedic patients there were less obvious differences in responsiveness to change between PCS and MCS. Responsiveness of the PCS_VR-36 for psychosomatic patients was smaller than the other instruments. Score improvements for all four instruments were statistically significant at p < 0.001 (paired t-tests).