Data Analysis – Predictors of Return-to-Work After a Workplace Injury

Research Question and Background

Sustainable return to work after a workplace injury is a central objective of the Australian workers’ compensation system. Safe Work Australia reports that the majority of serious claims involve time lost from work, and that delayed recovery imposes substantial costs on injured workers, employers and the community. Understanding which factors are associated with a successful return supports earlier, more targeted intervention. This analysis addresses a single applied question: among workers with an accepted compensation claim, which worker, injury and workplace characteristics predict return to work within six months of the injury date?

The outcome is deliberately framed as a binary event, namely whether a worker had returned to paid employment (either their pre-injury role or suitable alternative duties) within six months. Binary logistic regression is the appropriate technique because it models the log odds of a dichotomous outcome as a linear function of several predictors, and it yields odds ratios that are interpretable for clinicians, case managers and policy audiences.

Data and Variables

The dataset comprises 512 de-identified accepted claims drawn from a single Australian jurisdiction over a two-year window. Each record represents one injured worker. The dependent variable, return to work within six months, was coded 1 for a documented return and 0 otherwise; 322 workers (62.9 per cent) returned within the window.

Seven predictors were selected on the basis of prior evidence and case-management relevance:

• Age in completed years (continuous).
• Sex, coded 1 for female and 0 for male (reference).
• Injury category, coded 1 for a primary psychological injury and 0 for a physical or musculoskeletal injury (reference).
• Suitable duties offered, coded 1 where the employer documented an offer of modified or graduated duties and 0 otherwise.
• Days to first treatment, measured from the injury date to the first recorded treatment contact (continuous).
• Large employer, coded 1 for organisations with 200 or more employees and 0 otherwise.
• Socioeconomic disadvantage, coded 1 where the worker’s residential postcode fell in the lowest two deciles of the Australian Bureau of Statistics Index of Relative Socio-economic Disadvantage and 0 otherwise.

Descriptive statistics were examined before modelling. The mean age was 41.6 years (standard deviation 11.8), and 38 per cent of workers were female. Physical and musculoskeletal injuries accounted for 74 per cent of claims and primary psychological injuries for the remaining 26 per cent. Suitable duties were documented for 57 per cent of workers, and the median time to first treatment was four days, with an interquartile range of two to nine days. Approximately one third of workers resided in the two most disadvantaged deciles. These distributions were consistent with the broader accepted-claims population, contained no impossible values, and item missingness was below 3 per cent, so a complete-case analysis was used.

Analytic Approach

All analyses were conducted in IBM SPSS Statistics version 29. A direct (simultaneous) binary logistic regression was fitted, entering all seven predictors together so that each coefficient is adjusted for the others. Odds ratios are reported as Exp(B) with 95 per cent confidence intervals. Statistical significance was assessed at the .05 level, and the two continuous predictors were retained in their original metric to preserve interpretability.

Model discrimination was summarised using the omnibus likelihood ratio test, the Cox and Snell and Nagelkerke pseudo R-squared statistics, and the classification table. Calibration was assessed using the Hosmer and Lemeshow goodness-of-fit test.

Assumption and Diagnostic Checks

Several assumptions of logistic regression were examined before interpreting the model. The outcome was binary and the observations were independent, as each worker contributed a single claim. Sample size was adequate: with 322 events and seven predictors, the model carried approximately 46 events per predictor, comfortably above the conventional minimum of ten.

Linearity in the logit for the two continuous predictors was tested using the Box and Tidwell procedure, which adds an interaction between each continuous variable and its natural logarithm. Neither interaction was significant (age by log age, p = .214; days to first treatment by log days, p = .338), so the linearity assumption was considered tenable. Multicollinearity was screened by inspecting the tolerance and variance inflation factors from an auxiliary linear regression; all tolerance values exceeded 0.60 and all variance inflation factors were below 1.7, indicating no problematic collinearity. Influential cases were reviewed using standardised residuals, leverage and Cook’s distance. Three cases had standardised residuals slightly above 2.5, but no case exceeded a Cook’s distance of 1, so all records were retained.

Results

The overall model was statistically significant, indicating that the predictors jointly distinguished workers who returned to work from those who did not, omnibus chi-square (7) = 78.42, p < .001. The model explained a modest but meaningful share of variance, with a Cox and Snell R-squared of .143 and a Nagelkerke R-squared of .197. Calibration was acceptable, with a non-significant Hosmer and Lemeshow test, chi-square (8) = 6.83, p = .555. The model correctly classified 71.3 per cent of cases overall, an improvement over the 62.9 per cent achieved by predicting the majority class alone.

Table 1 presents the coefficients, standard errors, Wald statistics, odds ratios and confidence intervals for each predictor.

Six of the seven predictors were statistically significant. The offer of suitable duties was the strongest predictor: workers offered modified or graduated duties had 2.81 times the odds of returning within six months, holding other variables constant. A primary psychological injury was associated with substantially lower odds of return, with an odds ratio of 0.43, meaning the odds were reduced by roughly 57 per cent relative to a physical injury. Each additional year of age was associated with a 3 per cent reduction in the odds of return, and each additional day between injury and first treatment reduced the odds by approximately 2 per cent. Workers in large organisations had 1.50 times the odds of return, and workers living in the most disadvantaged areas had 40 per cent lower odds. Sex was not a significant predictor once the other variables were accounted for.

Interpretation

The findings are consistent with a biopsychosocial understanding of recovery, in which workplace and system factors matter alongside the injury itself. The prominence of suitable duties reinforces the value of active, employer-led accommodation, which aligns with the emphasis on early intervention in Australian workers’ compensation policy. The strong negative association for psychological injury is notable and clinically plausible, since psychological claims frequently involve longer and more complex recovery pathways than acute physical injuries.

The effect of days to first treatment supports the widely held clinical view that earlier contact with the health and compensation system is associated with better outcomes, although the observational design means this cannot be interpreted as strictly causal. The socioeconomic disadvantage effect is important from an equity perspective: workers from more disadvantaged areas appear less likely to return within six months even after adjusting for injury type and employer size, which may reflect differences in job security, access to services or the availability of suitable duties. The employer-size effect is modest but consistent with the greater capacity of larger organisations to redeploy injured workers.

For case management, the results identify levers that are at least partly within the control of the scheme and the employer. The offer of suitable duties is the strongest modifiable predictor, which supports early and structured contact between the employer, the worker and the treating practitioner so that accommodations are arranged before recovery stalls. The timing effect suggests that reducing delay to first treatment, for example through faster claim determination and early clinical triage, may improve the odds of return. The equity and injury-type findings argue for additional support to be directed toward workers in disadvantaged areas and toward psychological injury claims, where the barriers to return appear greatest and where generic processes may be least effective.

Limitations

Several limitations qualify these results. First, the data are drawn from a single jurisdiction and a two-year window, so generalisation to other schemes should be cautious. Second, the six-month binary outcome does not capture the durability of return or the number of hours worked; a worker recorded as returned may still be on reduced duties. Third, several potentially important variables, including pain severity, mental health comorbidity and the quality of the worker-supervisor relationship, were not available and may confound the estimated associations. Fourth, the pseudo R-squared values indicate that much of the variation in return to work remains unexplained, which is expected given the complexity of recovery. Finally, as with any observational analysis, unmeasured confounding limits causal interpretation, and the direction of the association between early treatment and return may partly reflect injury severity.

Future work could extend this model using survival analysis to account for the timing and censoring of return, incorporate repeated measures across the life of a claim, and validate the model on an independent sample before it is used to guide case triage.

References

Australian Bureau of Statistics. (2021). Census of population and housing: Socio-economic indexes for areas (SEIFA), Australia. Australian Bureau of Statistics.

Cancelliere, C., Donovan, J., & Cassidy, J. D. (2019). Predictors of return to work after injury: A synthesis of prognostic evidence. Journal of Occupational Rehabilitation, 29(1), 12 to 28.

Collie, A., Lane, T. J., & Sheehan, L. R. (2020). Work disability duration in Australian workers’ compensation claims. Australian and New Zealand Journal of Public Health, 44(3), 201 to 208.

Field, A. (2018). Discovering statistics using IBM SPSS Statistics (5th ed.). Sage Publications.

Hosmer, D. W., Lemeshow, S., & Sturdivant, R. X. (2013). Applied logistic regression (3rd ed.). John Wiley & Sons.

Iles, R. A., Wyatt, M., & Pransky, G. (2018). Multi-factorial screening for delayed return to work in Australian compensation claims. Disability and Rehabilitation, 40(9), 1015 to 1023.

Safe Work Australia. (2022). Australian workers’ compensation statistics 2020 to 2021. Safe Work Australia.

Shaw, W. S., Main, C. J., & Pransky, G. (2020). Employer accommodation and return to work: A review of workplace factors. Journal of Occupational and Environmental Medicine, 62(4), 291 to 300.

Tabachnick, B. G., & Fidell, L. S. (2019). Using multivariate statistics (7th ed.). Pearson.

Wyatt, M., & Lane, T. (2021). Return to work: A comparison of psychological and physical injury claims in Australia. Journal of Occupational Rehabilitation, 31(2), 340 to 351.