We can start by running a regression on the unmatched data.
fit <- glm(col_anastomotic ~ icg + sex + race + inout + age + insulin + bmi +
smoke + dependent + hxcopd + ascites + hxchf + dialysis + discancr + wndinf +
steroid + wtloss + bleeddis + transfus + prsepis + pralbum + renafail + dialysis +
emergncy + optime + col_mech_bowel_prep + col_oral_antibiotic +
col_mech_bowel_prep:col_oral_antibiotic + renafail:dialysis + col_chemo + cancer + lap,
data = unmatch_df,
family = quasibinomial(link = "logit"),
maxit = 100)
exp(lmtest::coefci(fit, vcov. = sandwich::vcovHC, type = "HC1"))[2,1:2]
#> 2.5 % 97.5 %
#> 0.4485266 1.3391525The confidence interval would seem to suggest no difference. Indeed, this is confirmed by a simple chi-square test which fails to find significant difference.
table(unmatch_df$col_anastomotic, unmatch_df$icg)
#>
#> FALSE TRUE
#> FALSE 76143 597
#> TRUE 2497 13
chisq.test(unmatch_df$col_anastomotic, unmatch_df$icg)
#>
#> Pearson's Chi-squared test with Yates' continuity correction
#>
#> data: unmatch_df$col_anastomotic and unmatch_df$icg
#> X-squared = 1.8246, df = 1, p-value = 0.1768If I try to minimize this model to just significant variables:
fit <- glm(col_anastomotic ~ icg + sex + bmi +
smoke + wndinf +
steroid + wtloss + bleeddis + prsepis + pralbum + optime + col_mech_bowel_prep + col_oral_antibiotic +
col_mech_bowel_prep:col_oral_antibiotic + col_chemo + lap,
data = unmatch_df,
family = quasibinomial(link = "logit"),
maxit = 100)
exp(lmtest::coefci(fit, vcov. = sandwich::vcovHC, type = "HC1"))[2,1:2]
#> 2.5 % 97.5 %
#> 0.4465451 1.3345540It still is not much different.
And now on the matched data.
fit <- glm(col_anastomotic ~ icg + sex + race + inout + age + insulin + bmi +
smoke + dependent + hxcopd + ascites + hxchf + dialysis + discancr + wndinf +
steroid + wtloss + bleeddis + transfus + prsepis + pralbum +
emergncy + optime + col_mech_bowel_prep + col_oral_antibiotic +
col_mech_bowel_prep:col_oral_antibiotic + col_chemo + cancer + lap,
data = match_df, weights = weights,
family = quasibinomial(link = "logit"),
maxit = 100)
exp(lmtest::coefci(fit, vcov. = sandwich::vcovHC, type = "HC1"))[2,1:2]
#> 2.5 % 97.5 %
#> 0.3936455 2.1202966Both of these results would imply there is no difference in anastomotic leak rates with ICG.
table(match_df$col_anastomotic, match_df$icg)
#>
#> FALSE TRUE
#> FALSE 584 597
#> TRUE 13 13
chisq.test(match_df$col_anastomotic, match_df$icg)
#>
#> Pearson's Chi-squared test with Yates' continuity correction
#>
#> data: match_df$col_anastomotic and match_df$icg
#> X-squared = 2.7965e-28, df = 1, p-value = 1