*2.6. Variables*

In this study, the following variables were collected: hematocrit level, age, sex, occupation, history of allergy, injury mechanism, fracture classification, hypertension, diabetes, coronary heart disease, arrhythmia, hemorrhagic stroke, ischemic stroke, cancer, associated injuries, dementia, chronic obstructive pulmonary disease (COPD), hepatitis, gastritis, age−adjusted Charlson comorbidity index (aCCI), and time from injury to admission.

The dependent variable was preoperative DVT, and the independent variable was the level of the hematocrit. Other variables were confounding factors.

## *2.7. Statistics Analysis*

Descriptive statistical analyzes were performed using standard reporting methods. Continuous variables are reported as mean ± standard deviation (normally distributed data) or median (interquartile range) (nonnormally distributed data). Categorical variables were reported as percentages. Chi−square (categorical variables), one−way ANOVA (normal distribution), or Kruskal–Wallis H tests (skewed distribution) were used to detect differences among different levels of the hematocrit at admission.

We analyzed the association between Hct level and preoperative DVT. Univariate and multivariate binary logistic regression models were used to test the association between Hct levels and preoperative DVT using three distinct models. Model 1: No covariates are adjusted. Model 2 was a minimally adjusted model, adjusted only for sociodemographic covariates. Model 3 was fully adjusted for all covariates. We performed a sensitivity analysis to verify the robustness of the results. We converted admission hematocrit into a categorical variable according to the quintiles, calculated *p* for the trend to verify the results of admission hematocrit as a continuous variable, and examined the possibility of nonlinearity (Q1–Q5 groups).

To account for the nonlinear relationship between hematocrit and preoperative DVT, we also used a generalized additive model and smooth curve fitting (penalized spline method) to address nonlinearity. If nonlinearity was detected, we first calculated the inflection point using a recursive algorithm and then constructed a two−piece logistic proportional hazard regression model for each side of the inflection point.

All analyzes were performed using the statistical software packages R (http://www.Rproject.org, R Foundation, Vienna, Austria) (accessed on 25 September 2022) and Empower-Stats (http://www.empowerstats.com, X&Y Solutions Inc., Boston, MA, USA) (accessed on 25 September 2022). Statistical significance was established by a two−sided *p*−value, where *p* < 0.05 (two-sided) was considered to be statistically significant.
