We have employed factor analysis as a way of understanding the latent factors underlying radiographic variables in OA. We report a four-factor solution as follows: 1) IP/CMC factor (20 joints), 2) MCP factor (eight joints), 3) knee factor (four joints), and 4) spine factor including DN and OST (5 levels, 10 variables). Use of a composite (e.g., an average or a standardized score) for each of these four factors in a multiple regression model would result in representation of all 20 + 8 + 4 + 10 = 42 underlying variables (or 37 distinct joint sites) through only four explanatory variables. Taking the MCP factor as an example, the individual scores from the eight MCP joints could be added and divided by eight, resulting in a single, average value for the MCP factor rather than eight individual scores. This allows consideration of the whole body burden of rOA in a parsimonious model, resulting in reduced dimensionality and increased precision of the resultant estimates. This is particularly important when considering a systemic outcome, such as a molecular or genetic biomarker, or a performance or disability score, which is likely influenced by the overall rOA burden and not only by the specific joint of interest in a study. The results were similar by age, between men and women, between African Americans and whites, and between members of the two cohorts.
Factor analysis involves some subjective decision-making, particularly when selecting the number of factors to retain and dropping variables that do not load well. The benefits of the four-factor solution are inclusion of the LS and CMC variables, which were not part of the three-factor solution. However, the spine factor has the lowest reliability and contains one variable (osteophytes at L5/S1) with loading less than 0.4, and the reliability of the knee factor is slightly reduced in the four-factor (α = 0.85) compared with the three-factor (α = 0.87) solution. The elimination of the hip variables was robust and similar across all solutions, however, supporting the idea that the hip is not part of the generalized OA construct. Also dropping out in all solutions were the first MCP variables, for reasons that are not immediately obvious, as these joints were involved at a similar frequency to the other MCP joints.
Our results do not support a single underlying latent variable for all joint sites, and are therefore not consistent with the concept of "generalized OA." This is in agreement with a study by MacGregor, et al, who used structural equation modeling to assess relations between knee, hip, PIP, DIP, and CMC rOA (also using the KL grading system) in a population of female twins accounting for shared genetics [28
]. The authors found that, while each individual joint site demonstrated genetic influence (heritability estimates 28 to 68%), there was a moderate genetic correlation only between DIP and PIP joints, and little or no correlation between other joint sites [28
]. MacGregor et al, conclude from their study that there was "little evidence of the phenotype 'generalized OA,'" consistent with what we found using a different methodology in the present report.
The DIPs and PIPs have long been thought to group together, and bony enlargement of these sites is the basis for a clinical diagnosis of hand OA [29
]. MCP involvement is generally thought to be less common than other hand joints and to represent a secondary OA process such as calcium pyrophosphate deposition disease (CPPD) or hemochromatosis. However, as shown by several studies reporting the patterns of hand OA, MCP involvement is not infrequent in radiographic or ultrasonographic OA [6
], with prevalence estimates ranging from 8% to 36% in Caucasian populations [9
]. MCP rOA has been shown to be more common among African American compared with white women in one study [34
]. Our preliminary work on the frequency of hand rOA in the JoCo OA
showed that although African Americans were much less likely to have DIP or PIP rOA compared with whites, the frequency of MCP rOA was not different by race [35
Principal components analysis (rather than exploratory factor analysis as described in the current paper) has been used in other studies of radiographic hand OA [20
]. Marshall et al, reported on a principal components analysis of radiographic variables in the hand [20
]. They identified a four-component model where the DIPs (second to fifth digits, right and left), PIPs (second to fifth digits, right and left), and MCP (only included the second and third bilaterally) joints each grouped onto one component and the thumb joints (IP, MCP, CMC, and trapezioscaphoid) onto the fourth. In contrast to the current study, Marshall et al used binary variables in the factor analysis (rOA defined as KL ≥ 2 at each joint), which is not consistent with the assumptions of factor analysis and does not include information about KL severity. In addition, the authors of that study simplified their results to subgroup patients into finger only, thumb only, and thumb and finger together. Although such phenotyping is useful, it does not fully utilize the power of factor analysis to create meaningful composites. Hunter et al, using osteophyte and joint space narrowing scores in addition to KL scores in a principal components analysis of hand OA, chose a 10-factor solution [21
]. The DIP joints loaded on a single factor, as did the PIP joints, with other joints (CMC, individual MCPs, thumb IP) loading on separate factors, and these factors were then used in a genetic association study [21
]. We found, in common with Marshall et al and Hunter et al, that the MCPs did not load onto the same factor as the IP joints. However, the DIP and PIP joints grouped together in our study along with the CMC, and we did not identify a separate thumb factor, which may be due to our larger sample size or the above-noted differences in methodology.
Although we included data on the hips in our analysis, we found that the hip joints did not load onto any factor. There has long been controversy whether hip OA is part of the "generalized OA" construct, or a separate entity [36
]. Our attempt to develop a composite measure of multi-joint OA has shown, in agreement with other researchers [6
], that the hip is a separate entity, but it remains unclear whether this is a measurement issue or a true difference in the nature of OA at the hip compared with other sites. Arden, et al, in a study assessing different radiographic definitions of hip OA, found that composite measures such as the KL grade had superior construct and predictive validity compared with single radiographic features (such as osteophytes or joint space narrowing alone), and recommended such measures for defining incident hip rOA [42
], supporting our use of this grading system. Hip OA is less influenced by obesity in comparison with knee OA [43
], and may therefore have other unique risk factors that vary by gender and race. For example, among African Americans compared with Caucasians [46
], we have reported a higher prevalence of specific radiographic features, such as superior joint space narrowing, previously found to be predictive of progression to hip replacement [47
]. Other recently recognized aspects of hip anatomy, such as femoroacetabular impingement and hip shape as described through active shape modeling, may contribute to hip-specific OA risk [48
] while not reflecting "generalized OA."
Limitations to the current study include the lack of PFJ joint space narrowing data (although PFJ osteophytes were included due to better reproducibility and associations with knee pain [25
]), and of radiographs of other joint sites, such as the feet or cervical spine, although we did have radiographic data for many of the joint sites most commonly affected by rOA. Although we used primarily KL grades for the present analysis, use of individual radiographic features (osteophytes and joint space narrowing), could lead to different conclusions and will be the focus of future analyses. Differences in scaling can cause variables to group on a factor due to differences in scale alone, but while we had some radiographic grades that were based on a 0 to 3 scale while others were on a 0 to 4 scale, we did not observe groupings based solely on this difference in scaling. The factor structure presented here has not yet been replicated in an independent population, and should be considered specific to the JoCo OA
study until replication has been confirmed. The factors we have identified should therefore not be used in other populations without confirmation of a similar structure. We have used only radiographic data in this analysis, although other projects are underway to consider symptoms and other variables of importance to an individual's experience of OA.