The predictive value of mental health for long-term sickness absence: the Major Depression Inventory (MDI) and the Mental Health Inventory (MHI-5) compared

Data was collected as part of the DAnish National working Environment Survey (DANES) from late 2008 until early 2009. The DANES was approved by the Danish Data Protection Agency, journal number: 2008-54-0553. According to Danish law, questionnaire and register based studies do not need approval by ethical and scientific committees.

Questionnaire data was collected from a random sample of the Danish employed population, age 18–59 years. Participants could answer the questionnaire either by internet or by post. Non-responders received two reminders and were finally contacted and invited to participate in a telephone interview. The study contacted 9913 persons, 6531 responded (66%), of which 4919 were employees. We excluded participants with missing values on MHI-5 and MDI (n = 439), with missing response date (n = 150), with a sickness absence spell lasting 4 or more weeks in the preceding 3 months before or overlapping baseline (n = 177), yielding a study sample of 4153 participants. In the multivariable adjusted analyses the sample was, because of missing values to covariates, further reduced to 3713 participants. The 3713 participants are 75% of the employees that returned the questionnaire. If we assume the return rate of the employees is equal to the return rate of the full sample, i.e., 66%, the sample for the final analysis is 50% of the employed people the questionnaire originally was mailed to.

Our long-term sickness absence data came from a national register, the DREAM register [14]. The DREAM register has information on all Danish social transfer payments on a weekly basis. Employers are entitled to compensation from the municipalities if an employee has an absence spell of 22 days or longer.

The Major Depression Inventory (MDI) was developed in the late 1990s [12]. It was designed to measure depression symptoms in accordance with the symptom guidelines defined by the WHO classification for unipolar depression (ICD-10) and the American Psychiatric Association classification for major depression (DSM-IV) [12, 15]. The instrument consists of 12 questions and it includes 2 algorithms that classify participants with risk of unipolar depression (mild, moderate or severe) according to the ICD-10 definition or with risk of major depression according to the DSM-IV definition. The instruments can also be used to measure depressive symptoms scored on a scale ranging from 0 to 50, a higher score indicates a higher level of depressive symptoms. Both Bech et al. [12] and Forsell [16] have validated the MDI as a measure of depression using SCAN clinical interviews.

The questions of the MDI are shown in Table 1.

We calculated the MDI’s predictive value for long-term sickness absence when we used the MDI as: (1) a standardized scale score, (2) the scale categorized into five ordinal levels [5], (3) dichotomized into likely depression or not by the cut-off ≥20 in accordance with previous studies [5, 16], likely depression or not based on the DSM-IV algorithm, and likely depression or not based on the ICD-10 algorithm (we pooled the ICD-10 categories mild, moderate and severe) [12].

The mental health inventory (MHI-5) is a five-question subscale of the general health measure SF-36 [17]. The MHI-5 includes questions referring to both positive and negative aspects of mental health, and questions referring to both depression and anxiety [18, 19]. Berwick et al. [13], Cuijpers et al. [18] and Rumph et al. [20] have validated the MHI-5 as a measure for depression using clinical interviews as the gold standard. The MHI-5 has also been tested as a measure of anxiety, somatoform disorders, and substance use disorders, but has failed to perform equally well in these tests [18, 20]. The questions of the MHI-5 are shown in Table 1.

We scored the MHI-5 from 0 (poor mental health) to 100 (good mental health) [21]. Several cut-offs for dividing the responders into high risk and low risk group have been suggested for the MHI-5. The cut-off points are derived from ROC-curves or through another minimum misclassification criterion. The most used cut-off point is probably 52 [4, 22], other cut-off points include 54 and 74 [18] and the cut-offs 60, 68 and 76 [23]. We calculated the predictive value for long-term sickness absence using the MHI-5 as: (1) a standardized scale score; (2) categorized into 5 ordinal levels; and (3) dichotomized by four different cut-offs, i.e. the cut-offs 52, 60, 68 and 76.

Our outcome was time until onset of first episode of long-term sickness absence. In our analysis, long-term sickness absence was defined as minimum 4 weeks of consecutive sickness absence, because 4 weeks are the minimum period that is registered within the weekly based DREAM register. We followed responders in the register from 12 weeks before response date and until 60 weeks after response date. Participants, with long-term sickness absence in the 12 weeks before or overlapping the particular participant’s response date, were excluded from the study. Only sickness absence that comes after the response of the questionnaire is included in our analyses.

The following covariates were included in the analyses: Age, gender, family status, smoking, alcohol, body mass index, leisure time physical activity, social class, somatic chronic illness, self-rated health and method of data collection. With the exception of age and gender, which were retrieved from registers, the information on the covariates came from the questionnaire.

We divided family status into four categories: (1) cohabitating, with children living at home, (2) cohabiting, without children living at home, (3) not cohabitating, with children living at home, (4) not cohabiting, without children living at home. Smoking was categorized into four levels; (1) never smoked, (2) ex-smoker, (3) light smoker, i.e., from occasional smoker to 14 cigarettes a day, and (4) heavy smoker ≥15 cigarettes a day. Alcohol intake was categorized into four levels following the guidelines of the Danish National Board of Health: (1) ≤ 7 units weekly, (2) 8–14 units, (3) 15–21 units, (4) > 21 units weekly. Body mass index was categorized into 4 levels according to the WHO categorization: (1) < 18.5 (underweight), (2) 18.5–25 (normal weight), (3) 25–30 (overweight), (4) > 30 (obese). Leisure time physical activity was scored on a scale from 0 to 24 points. Social class was categorized in accordance with the European Socio-economic Classification (ESeC). Somatic chronic illness was constructed as a binary variable (somatic chronic illness or not). Self-rated health was scored on a scale from 1–5 points. The self-rated health question primarily measures physical health [24, 25] and is therefore used together with somatic chronic illness to control for the responders physical health. Method of data collection was included, since responses are influenced by the data collection method [26]. Family status and social status were used as categorical variables in the analyses. All other covariates were assumed to be ordinal or linear.

We compared the MDI and MHI-5 in a head-to-head design [27]. We calculated the Pearson correlation between the MDI and the MHI-5 and we compared the distribution of MDI and the MHI-5 in two plots. We calculated the predictive value for long-term sickness absence by Cox’s proportional hazard in both a univariate regression analysis and a multivariate regression analysis. Participants were censored at the date of emigration, death, retirement, maternity leave or end of follow-up (60 weeks after baseline), whichever came first. We examined the proportional hazard assumption by visual inspection of log-log-survival plots. The predictive value of MDI and MHI-5 for long-term sickness absence was calculated with (1) measures dichotomized by different cut-offs, (2) 5-level categorical indexes, and (3) standardized scores. We standardized the two scores of the MDI and MHI-5 using mean = 0 and standard deviation = 1. The standardized MHI-5 score was reversed so high score equalled poor mental health for both scores. In order to evaluate which mental health instrument that had the best prediction of long-term sickness absence, we included both standardized scores in the same regression analysis. If one measure in such an analysis have a substantial better significance level than the other measure, it can be concluded that this measure better capture the variation in the data. To further confirm the result from this analysis, we performed a cross-validation by randomly splitting our sample into two independent samples and then carry out the analysis in each sample.

The MDI standardized score, the five-level categorized score, and the three dichotomized scores all predicted long-term sickness absence (see Table 2). The standardized score had the highest significance level. The 5-level categorization of MDI showed a steep increase in HR from the next highest to the highest level. This suggests that the best cut-off point for a dichotomization into a risk group and non-risk group is near this score, i.e., near 20 points (risk group = 7%). The risk group included 90 to 301 persons depending on which dichotomization we used, i.e., the risk group was from 2 to 7% of the employed adult population (see Table 2).

The MHI-5 standardized score, the five-level categorized score, and the four dichotomized scores all predicted long-term sickness absence (see Table 2). The standardized score had the highest significance level. The 5-level categorization of MHI-5 showed an increase in HR for each level. The increase between the next highest and the highest level of the categorization was particular steep. This suggests that the best cut-off point for a dichotomization is near 60 points (risk group = 14%). The risk group included from 304 to 1560 persons depending on which dichotomization we used, i.e., from 7 to 38% of the employed adult population.

The HRs of the standardized scores estimate the increased risk for one standard deviation increase of the standardized score, i.e., the increased risk of sickness absence for a one standard deviation of worse mental health. In the adjusted analyses, the HR of MDI was 1.27 and the HR of MHI-5 was 1.37, both were highly significant (see Table 2). With both instruments included in the same regression analysis, the MHI-5 remained significant and the MDI became non-significant (see Table 2), i.e., the MHI-5 explained a significant part of the variation in long-term sickness absence even when we controlled for the MDI. To confirm the stability of this result, we cross-validated the result by randomly splitting our sample in two independent samples and then carry out the analysis in each sample. Both results supported our initial finding (results not shown). Thus, the MHI-5 explained the variation in long-term sickness absence better than the MDI.

Figure 1 shows the distribution of MDI and MHI-5 scores within the study population. Both distributions are positively skewed, indicating that they distinguish better between degrees of reduced mental health than degrees of good mental health. The MDI score correlated with the MHI-5 score with a Pearson correlation coefficient = −0.69 and p < 0.0001. The correlation is illustrated in Figure 2. The Figure also illustrates that not all participants classified as high risk by the MDI were classified as high risk by the MHI-5 or vice versa. The MDI ‘cut-off 20’ and the MHI-5 ‘cut-off 52’ each classified 7.3% as high risk. However, only 3.8% of the ‘cases’ overlapped.