# Surgical Complications

Project Overview

This project involved an overseas client in a private hospital. Stonewall evaluated the viability of surgical interventions in 74 elderly patients given certain variables associated with each patient. The variables associated with the patient include age, gender, length of stay (adjustor), number of comorbidities (adjustor), and surgical complications (indicator). The outcome variable used in the model included surgical viability (i.e., whether the surgery improved the patient's symptoms).

Analysis

A binomial logit model was used to answer the main research question of evaluating variables associated with the patient in determining surgical viability. The outcome variable (dependent variable) was improvement in symptoms. Originally, four values were presented with this scale (0 = missing, 1 = no change, 2 = improved, 3 = worsened). As missing values are excluded from the model, it was decided the best course of action was to reduce the remaining values for this variable and turn them into a binomial value (i.e., did symptoms improve or not). This outcome variable was then evaluated against the other variables associated with the patient (gender, age, length of stay, comorbidities, and surgical complications). Prior to any statistical model, we always produce figures for the variables under study and perform basic descriptive statistics. The following boxplots demonstrate some of the key variables examined in this analysis.

A binomial logit model presents values in terms of log-odds. As interpreting a coefficient based in terms of log-odds is challenging, model coefficients can be exponentiated so coefficients are presented in log-odds. Odds have close relation to probability, and presenting model coefficients in terms of odds is quite common in medical research. All coefficients were compared against an alpha of 0.05. The logit model is as follows (note all coefficients are presented in log-odds). The model is first evaluated in log-odds to determine whether each coefficient impacts the outcome variable. In this case, no variable significantly impacts the outcome variable (as a result the estimates are not exponentiated).

When evaluating model coefficients, no coefficients have a probability value less than 0.05. Comorbidities and length of stay have marginal significance (i.e., they have significance close to 0.10 or less), but it is highly likely that length of stay is correlated with surgical complications, which could impact overall interpretation.

Conclusion

While this model did not reveal any significant differences in terms of variables associated with patients and surgical viability, there are a few shortcomings that could be corrected with further research, which may demonstrate a significant difference does in fact exist. First, there were a number of missing variables associated with viability and other variables. In most frequency-based statistics, when variables are missing they are excluded from the analysis (termed list-wise deletion). In this specific case, the main logit model excluded some observations. If the sample size were increased with fewer missing cases, then it is possible significant differences may exist. Second, some of the variables used as adjustors (e.g, length of stay, surgical complications) are potentially highly correlated with one another. When performing frequency-based statistics with highly correlated variables and fewer observations (30 or less), coefficients could behave in a manner that prevent meaningful inference. One way to overcome correlated variables in this analysis would be to increase the sample size.

Return to Our Work.