Table 4.
Logistic regression analysis to assess the association of the tumour CD274 expression score (predictor) with histological lymphocytic reaction (outcome)
| Univariable OR (95% CI) | Multivariable OR (95% CI)† | ||
|---|---|---|---|
| Model for Crohn’s-like lymphoid reaction (n=681, as a binary outcome variable*) | |||
| Tumour CD274 expression score | 0 | 1 (reference) | 1 (reference) |
| 1 | 0.68 (0.37–1.25) | 0.79 (0.40–1.57) | |
| 2 | 0.86 (0.47–1.56) | 1.06 (0.54–2.09) | |
| 3 | 0.53 (0.28–0.98) | 0.65 (0.32–1.31) | |
| 4 | 0.59 (0.22–1.56) | 0.68 (0.23–1.98) | |
| Ptrend‡ | 0.09 | 0.30 | |
| Model for peritumoural lymphocytic reaction (n=808, as a binary outcome variable*) | |||
| Tumour CD274 expression score | 0 | 1 (reference) | 1 (reference) |
| 1 | 0.45 (0.19–1.04) | 0.43 (0.18–1.03) | |
| 2 | 0.33 (0.14–0.77) | 0.35 (0.15–0.82) | |
| 3 | 0.65 (0.27–1.54) | 0.58 (0.24–1.42) | |
| 4 | 0.50 (0.16–1.61) | 0.43 (0.13–1.40) | |
| Ptrend‡ | 0.87 | 0.59 | |
| Model for intratumoural periglandular reaction (n=812, as a binary outcome variable*) | |||
| Tumour CD274 expression score | 0 | 1 (reference) | 1 (reference) |
| 1 | 0.53 (0.23–1.25) | 0.54 (0.22–1.28) | |
| 2 | 0.40 (0.17–0.92) | 0.42 (0.18–0.99) | |
| 3 | 0.93 (0.38–2.27) | 0.88 (0.35–2.19) | |
| 4 | 0.79 (0.22–2.86) | 0.71 (0.19–2.62) | |
| Ptrend‡ | 0.51 | 0.69 | |
| Model for tumour-infiltrating lymphocytes (n=811, as a binary outcome variable*) | |||
| Tumour CD274 expression score | 0 | 1 (reference) | 1 (reference) |
| 1 | 0.68 (0.40–1.13) | 0.78 (0.42–1.43) | |
| 2 | 0.59 (0.35–0.99) | 0.66 (0.36–1.24) | |
| 3 | 0.42 (0.25–0.72) | 0.54 (0.29–1.01) | |
| 4 | 0.59 (0.26–1.34) | 0.64 (0.25–1.65) | |
| Ptrend‡ | 0.004 | 0.049 | |
Abbreviations: CI, confidence interval; OR, odds ratio.
Since the proportional odds assumption was not satisfied in the ordinal logistic regression model, we used the binary logistic regression model to assess the independent association of the tumour CD274 expression score with each histological lymphocytic reaction pattern.
The multivariable logistic regression analysis model initially included age, sex, year of diagnosis, family history of colorectal carcinoma in any parent or sibling, tumour location, microsatellite instability, CpG island methylator phenotype, KRAS, BRAF, and PIK3CA mutations, and LINE-1 methylation level. A backward elimination with a threshold of P=0.05 was used to select variables in the final models.
Ptrend value was calculated by the linear trend across the ordinal categories of the tumour CD274 expression score (0 to 4, as an ordinal predictor variable) in the binary logistic regression model for each histological lymphocytic reaction pattern (a binary outcome variable [absent vs. low/high]). Because we assessed eight primary outcome variables, we adjusted two-sided α level to 0.006 (=0.05/8) by simple Bonferroni correction.