Key statistical analyses in this study include: 1) Determination of prevalence for specific musculoskeletal disorders, 2) Calculation of incidence rates for disorders, 3) Evaluation of risk factors for disorders, 4) Evaluation of interactions between various risk factors and specific disorders (e.g., carpal tunnel syndrome and lateral epicondylitis), 5) Analyses of the performance of existing ergonomic models, and 6) Building models for predicting risk(s) of DUE MSDs. Data will be analyzed in SAS 9.2 (SAS Institute, Cary, North Carolina, USA). Significant associations will be reported based two-sided statistical significance with an alpha of 0.05.
The unit of analysis in the study for primary outcomes is individuals. The final point prevalence of specific distal upper extremity musculoskeletal disorders at baseline will be calculated. Baseline prevalence of specific musculoskeletal disorders including lateral epicondylalgia, medial epicondylalgia, deQuervain’s, trigger digit and hand/wrist tendinosis will be aggregated into baseline prevalence of non-CTS DUE MSDs. Similarly, past history of specific DUE MSDs will be aggregated into a lifetime prevalence estimate of these disorders. The baseline prevalence of a specific disorder will be calculated and those worker’s hands will subsequently be excluded from incidence analyses for that specific disorder.
Some health outcomes, such as prevalence and incident cases of lateral epicondylalgia, are binary variables and will be analyzed using logistic regression models for prevalence and proportional hazards regression models for incidence. Other variables, such as impairment or severity, may be ordinal categorical or continuous (e.g., number of lost or restricted workdays) and will be analyzed using corresponding nonparametric techniques. Risk factors will be grouped according to nature (individual, psychosocial, and job physical factors) and introduced into the models. Associations between predictor variables (including existing job analysis methods) and health outcomes will initially be evaluated using univariate methods. Variables with meaningful evidence of association to the health outcomes (generally existing at p
0.20) will be considered for inclusion in multivariate models.
Incidence rates will be assessed using approaches parallel to those from the baseline analyses. Cumulative rates will be calculated using subjects remaining in the study at that time with known status. Information from subjects withdrawing from the study will be incorporated by calculating Kaplan-Meier rates of freedom from symptoms/disorder at a particular point in time using survival analysis methods.
Unadjusted univariate hazard ratios (HR) for incident cases of specific disorders and 95% confidence intervals will be determined for TLV for HAL, Strain Index score, individual ergonomic variables (e.g., force, repetition, posture) and relevant covariates using Cox proportional hazard regression with time varying covariates [58
] in SAS version 9.2 using the PHREG statements [59
Potential covariates (see Table ) will be grouped and evaluated for association with incident cases of the musculoskeletal disorder in survival analyses prior to creating multivariate models. Multivariate analyses will be performed for job physical factors (e.g., force, repetition, hand/wrist posture), as well as the TLV for HAL and Strain Index score. Covariates will be selected (from worker demographics, hobbies and physical activities outside of work, psychosocial factors, baseline prevalence of distal upper extremity musculoskeletal disorders other than the disorder being analyzed and medical history) based upon p
0.20 and biological plausibility.
Potential Covariates Considered for Multivariate Analyses of MSDs
Separate proportional hazard regression models will be fitted for each of the job physical exposure tools (SI, TLV for HAL) as well as individual job physical exposure variables such as peak force and repetition. We will then assess the hazard ratio, with 95% confidence intervals, for each factor in a proportional hazard model adjusted for confounding variables.
Changes in exposure level occurring in the course of the study will be incorporated in some of these analyses. Additional analyses to model associations with events occurring more than once in the same individual over the study period (e.g., elbow pain that recurs 6
months later) using the Andersen-Gill independent increment method [60
] and other approaches are planned. These other approaches involve fitting a basic proportional hazard model that ignores potential correlations to an appropriately define risk set, and then implementing a robust covariance estimate to adjust for correlation between events occurring in the same subject. [61
] Transformation or categorization of a predictor is an option if there are problems with model fit. Risk factors that are significant, or show a strong trend, after adjustment, will also be considered as candidates for ergonomic models incorporating the “best” independent predictors of events.
For assessment of the predictive performance of existing ergonomic models (particularly Strain Index, ACGIH TLV for HAL, and the Rapid Upper Limb Assessment), the incidence data will serve as the gold standard against which the operant characteristics such as sensitivity and specificity will be analyzed [42
For assessing the appropriateness of analysis approaches, whether dropouts are independent of study outcome will be assessed (“data missing completely at random, MCAR”, [63
]). If the missing completely at random assumption is tenable, then analyses making use of available data will generally produce valid inferences. Survival analysis methods will facilitate the use of available follow-up data in subjects who drop out of the study, under the assumption of noninformative censoring.
These analyses involve examination of several indices of exposure and several measures of two main outcomes, leading to potential for chance associations due to multiple statistical tests [64
]. This study will use a limited number of “primary analyses” that use uncorrected significance levels, given that the intended (and actual, if different) analysis plan for the study is clearly stated in reports and publications. Exploratory analyses, such as post hoc
stepwise model building, will be reported as such.
The assessment of interactions will be performed by evaluating combinations of selected risk factors. Combinations of job physical factors, particularly between pairs of force, repetition and posture will be evaluated. Individual risk factors (particularly obesity and diabetes mellitus) as well as combinations of job physical demands with individual risk factors (i.e. force and obesity, etc.) will be evaluated.