The study was carried out in two phases between September 2004 and July 2005. Phase 1 involved a random-digit-dialing telephone survey to screen 19,381 adult residents (96% response) ages 18 to 59 from metropolitan, urban, and rural Georgia populations. Based on the 19,381 people from the household screening interview, 8,910 adults were randomly selected for detailed telephone interviews: 5,623 individuals completed the detailed telephone interview; 1,874 refused to participate; 141 were further confirmed to be ineligible; and 1,272 were excluded due to physical or mental inability to participate, inability to be contacted, language barriers, or because they had died. This yielded an overall response rate of 75%. Based on the detailed telephone interviews, study participants were classified into three groups:

1) CFS-like, characterized by severe fatigue lasting six months or longer that was not alleviated by rest, that caused substantial reduction in occupational, educational, social, or personal activities, and that was accompanied by at least four of the CFS case-defining symptoms.

2) Chronically unwell, having chronic (≥ six months) unwellness with or without fatigue, but not meeting the criteria for CFS.

3) Well.

In Phase 2, all 469 people with CFS-like illness were invited for clinical evaluation, and 292 (62%) participated. Of randomly selected chronically unwell participants, 286 (53%) completed the clinical evaluation. Finally, 223 individuals classified as well in the telephone interview completed clinical evaluations. They were matched to the CFS-like group based on residence (metropolitan, urban, rural), sex, race/ethnicity, and age (within three years). Overall, about 50% of invited respondents from all three groups completed the one-day clinical evaluation.

Participants completed the MFI-20 and other questionnaires during the clinical evaluation. This study involves data from 783 participants who completed the MFI-20 along with several other measures of psychosocial functioning, including the Medical Outcomes Survey Short Form-36 (SF-36), the Zung Self-Rating Depression Scale (SDS), and the Spielberger State-Trait Anxiety Inventory (STAI).

We used SAS version 9.1 (SAS Institute Inc, Cary, NC) for data analysis. Descriptive statistics (frequencies, percentages, means, standard deviations, and ranges) were generated to characterize the study sample in terms of socio-demographic parameters. We used several criteria to assess the subscale validity and reliability of the MFI-20.

Internal consistency of each of the five MFI-20 subscales was determined using three reliability tests: 1) inter-item correlation; 2) corrected-to-total (or item-total) subscale correlation; 3) and Standardized Cronbach's α coefficients (and item discrimination). The cutoff criteria for acceptance on reliability tests are as follows. First, item-total subscale correlations of not less than 0.30 and inter-item correlations of 0.30 to 0.70 were retained. Second, a fairly high reliability coefficient (Cronbach's α > 0.70) was required to assess the internal consistency reliability [30,31]. Floor/ceiling effects were considered significant if more than 15% of the subjects had either the lowest possible or highest possible score on the subscales [32]. A significant floor effect was expected in the well group.

As an indication of discriminant (known-group) validity, group differences in the five MFI-20 subscales were calculated using analyses of variance to examine the ability of the MFI-20 instrument to distinguish three groups: CFS-like, chronically unwell, and well. Using a Tukey correction, the alpha per test for each subscale was 0.01, for an overall alpha of 0.05. Two-way analyses of variance were performed to test the age and sex effects on the five MFI-20 subscales. Post-hoc analysis with Tukey p-value adjustment was performed for multiple subgroup comparisons.

To further assess construct validity of the subscales, an exploratory factor analysis was performed. A principle component analysis was used to extract factors. The obtained factors were rotated oblique using the Varimax procedure. A minimum eigenvalue of 1 was specified as the extraction criterion [33]. The desired criterion of factor loadings was set at 0.50 or above, slightly higher than the typical cutoff value of 0.40 [34].

Finally, the convergent validity of the MFI-20 was evaluated through comparisons of the MFI-20 with other instruments administered in the protocol. Pearson correlation coefficients were used to assess linear associations between the multi-item scales of SF-36, SDS, and STAI. We chose these instruments based on the association between fatigue and other measures on psychosocial functioning, such as health-related quality of life (measured by SF-36), depression (measured by SDS), and anxiety (measured by STAI) as well as the existing data from the source study.

The most valid SF-36 subscales for measuring physical health include the physical functioning, role physical, and bodily pain subscales and the physical component summary score [26]. The most valid SF-36 subscales for measuring mental health include the mental health, role emotional, and social functioning subscales and the mental component summary score [26]. For the concept of physical and mental health, we investigated correlations between MFI-20 subscales and physical and mental health as measured by the SF-36.

Table 1 summarizes subscale validity and reliability analyses for the 783 participants who completed the MFI-20 questionnaire. Of these, 37% had been classified as CFS-like based on the detailed telephone interview, 34% were chronically unwell, and 28% were considered well. The participants had a mean age of 43, were primarily female (76%), white (70%), and from rural or urban areas (83%). Nearly 95% had completed at least a high school education. Nearly 38% were unemployed, self-employed (not working for pay), retired, laid off, disabled, or students. More than 60% of participants were married or cohabitating. More than half of participants had a household income equal to or higher than the Georgia median income level of $42,679.

Only the physical fatigue subscale score differed significantly by both age (p = 0.0024) and sex (p = 0.0015). Reduced activity (p = 0.0078) and reduced motivation (p = 0.0112) scores differed significantly between age groups. General fatigue (p = 0.0003) and mental fatigue (p = 0.0272) scores were significantly worse in females than in males. The interaction between age and sex was not significant in any of the MFI-20 subscales. Although only three of the five MFI-20 subscales were significantly different by sex, descriptive statistics of all the subscales were summarized for females and males (Table S1 and Table S2, Additional file 1).

For subscales with significant age or sex effects, we estimated partial correlations controlling for sex and age, respectively (Table 2). This had negligible effects on the correlations between the physical fatigue, reduced activity, and reduced motivation subscales.

Table 3 summarizes the results of three reliability tests for the five MFI-20 subscales. There was no item redundancy; inter-item correlations averaged 0.56 (range 0.46-0.69) for general fatigue, 0.52 (range 0.44-0.61) for physical fatigue, 0.53 (0.41-0.66) for reduced activity, 0.38 (0.17-0.56) for reduced motivation, and 0.61 (0.53-0.66) for mental fatigue. Corrected item-total correlations were higher than 0.30 for all five MFI-20 subscales. The values for standardized Cronbach's α for the five MFI-20 scales were: general fatigue: 0.83; physical fatigue: 0.81; reduced activity: 0.82; reduced motivation: 0.71; and mental fatigue: 0.86. These values were greater than the suggested criteria value of 0.70 for acceptable reliability.

Pairwise correlations between the MFI-20 subscales ranged from 0.49 to 0.74. Although the subscales are strongly related to each other, it is unclear whether an overall summary component of the MFI-20 is appropriate. Factor analysis confirmed that overall summary components accounted for 70% of the reliable variance in the five subscales. The total scale with 20 items yielded a Cronbach's α coefficient of 0.93, which is consistent with the result from the Gentile study [18].

The factor analysis solution was complex, with multiple loadings of items having factor-loading values > 0.50 across five factors (Table 4). However, the first factor, which explained 20% of the variance in the 20 items of the MFI-20, was dominated by general fatigue and physical fatigue. Six items (four physical fatigue, one general fatigue, and one reduced activity) loaded on the first factor (loadings from 0.54 to 0.83). The second factor was comprised solely of all four mental fatigue items (loadings from 0.71 to 0.81), which explained 15% of the variance in the 20 items of the MFI-20. Three of the reduced activity items fell nicely (loading from 0.52 to 0.71) on the third factor. The fourth factor was loaded by three of the general fatigue items (loading from 0.52 to 0.71) and two of the reduced motivation items (loading 0.58 and 0.64). The remaining two reduced motivation items fell nicely on the fifth factor.

The CFS-like, chronically unwell, and well groups had significantly different mean values (p < 0.0001) for all the MFI-20 subscales (Table 5). All subscales appeared to discriminate between groups, but the degree to which they discriminated varied. The CFS-like group had higher scores in all the subscales compared to the chronically unwell group (average mean difference = 2.90; range of mean difference: 2.26 points (reduced activity) - 3.54 points (general fatigue). Compared to the well group, the CFS-like group had subscale scores that were, on average, 6.01 points higher. Mean differences between these groups ranged from 4.56 points (reduced activity) to 7.96 points (general fatigue), whereas the chronically unwell group scored, on average, 3.11 points higher in the five subscales than the well group.

We observed a floor/ceiling effect in all the MFI-20 subscales in the well group, except for general fatigue, as expected. No floor/ceiling effects were detected in the CFS-like and chronically unwell groups. There were no floor/ceiling effects in the whole study sample (Table 5).

We calculated correlations between fatigue subscales and subscales measuring functional impairment (SF-36), depression (SDS), and anxiety (STAI) to evaluate convergent validity in the overall sample (Table 6). The MFI-20 subscales were substantially correlated with the eight SF-36 subscales (average: r = -0.53; range of absolute values of correlations: |r| = 0.34 - 0.83). All MFI-20 subscales were most strongly correlated [35] with the SF-36 subscales measuring vitality (average: r = -0.68; range of |r|: 0.57 - 0.83), followed by general health perception (average: r = -0.59; range of |r|: 0.48 - 0.71), and social functioning (average: r = -0.54; range of |r|: 0.50 - 0.59).

As expected, all five MFI-20 subscales were significantly correlated with depression, anxiety, and functional impairment. However, the correlations with depression and anxiety were generally lower (average: r = 0.50; range of r = 0.34-0.65) than correlations with functional impairment (the SF-36 subscales). The highest correlations were found between MFI-20 subscales and measurement of depression (SDS index) (average: r = 0.58; range of r = 0.50 - 0.65).

The study's strengths include: a rigorous study design with a large, randomly selected sample from a cross-sectional, population-based study of fatiguing illness; and the careful clinical determination of groups, selection of comparison measures, report of reading levels of the instrument, and correction of p-values for multiple testing.

This study has several limitations. Our existing data did not allow us to conduct test-retest reliability of the MFI-20. Further studies might be needed to explore test-retest reliability of MFI-20 in fatiguing illness. Another limitation is external validity/generalizability. While the study employed random sampling, the population was limited to an adult population in Georgia and could therefore differ from results that might be obtained from implementing the same study design in other regions due to the effect of regional lifestyle. Nonetheless, previous studies on MFI-20 have not identified the effect of regional lifestyle in their study populations. Our cross-sectional data precluded us from examining responsiveness (ability of the MFI-20 to detect clinically important changes over time) and obviates the possibility of eventually examining responsiveness differences due to treatments. Longitudinal studies are needed to determine minimal clinically important differences (MCIDs) of the MFI-20 subscales in fatiguing illness.

In this study, we applied a 0.01 alpha level of statistical significance to adjust for multiple testing instead of the popular standard level of 0.05. This increases our confidence in the associations that were determined to be of statistical significance but also increases the risk of failing to reject a false null hypothesis (a Type II error), and so results in less statistical power. However, the statistically significant results observed in this study are of practical significance. For example, the group mean differences in our study are greater than the generic MCID of two points in the MFI-20 subscales in Purcell's study [37]. The possibility of a Type II error should, however, be considered.