Missing data imputation

Similar to before, we don’t have much missing data, with only a few patients missing gestational age and birth weight

We’ll impute those now

Defining categories

First, we’ll collapse Race into larger categories, since we really have 3 common races recorded (including a large “other” group). Therefore, any race that shows up fewer than 10 times will be lumped together with “Other”. We’ll also recode BCx to where CONS is considered “other” like the other contaminants

Next we make a new wt_for_GA category that compares the patients gestational age & weight to the national Z-scores. Data comes from Talge et al., 2014, and uses less than the 10th percentile for SGA and greater than the 90th percentile for LGA. Because so few patients are LGA, this category is combined with AGA. We’ll also make categories for LBW (less than 2500 gram, regardless of GA) and prematurity (GA < 37).

Length of Stay

Bivariate analysis: t-test

DOT

Bivariate analysis: Mann-Whitney U Test (nonparametric d/t non-normality)

LOS model

LOS is simple, just use a Poisson GLM

The table above is transformed to show incidence rate ratio (IRR), which can be thought of as the relative risk². Thus the coefficient can be thought of as follows:

The LOS was 1.16 times longer in the Pre-PCR era, compared to the Post-PCR era, when adjusting for the above variables

This can be demonstrated by looking at the conditional means in the two eras

The results of the regression graphically:

And conditional mean LOS by each covariate:

DOT model

For DOT we have a more complicated picture, since we have lots of zeros³ in the response variable. We’ll use a zero-inflated model, which is kind of a 2-in-1 solution that addresses two questions:

1. Who got any antibiotics? Called the zero-inflated model, which is a logistic regression
2. Among those who did get antibiotics, how much did they get? Called the conditional model, which works just like to LOS model but is only applied to patients with non-zero DOT’s

Here’s our model:

##  Family: genpois  ( log )
## Formula:          DOT ~ Era + wt_for_GA + Simple_Dx + BCx + Sex
## Zero inflation:       ~Era + wt_for_GA + Simple_Dx + Age
## Data: subj_rounded
## 
##      AIC      BIC   logLik deviance df.resid 
##    954.9    988.2   -465.5    930.9      106 
## 
## 
## Overdispersion parameter for genpois family (): 19.4 
## 
## Conditional model:
##                Estimate Std. Error z value Pr(>|z|)    
## (Intercept)     3.63194    0.14455  25.125  < 2e-16 ***
## EraPre_PCR      0.66856    0.11776   5.677 1.37e-08 ***
## wt_for_GASGA   -0.06666    0.15094  -0.442  0.65875    
## Simple_DxViral -0.25768    0.12196  -2.113  0.03461 *  
## BCxContaminant  0.44643    0.16345   2.731  0.00631 ** 
## SexMale         0.13253    0.11807   1.122  0.26165    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Zero-inflation model:
##                Estimate Std. Error z value Pr(>|z|)   
## (Intercept)    -1.06751    0.44348  -2.407  0.01608 * 
## EraPre_PCR     -0.49930    0.44432  -1.124  0.26112   
## wt_for_GASGA   -0.54999    0.63718  -0.863  0.38804   
## Simple_DxViral -0.83465    0.46436  -1.797  0.07227 . 
## Age             0.02879    0.01012   2.846  0.00443 **
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

The same results, shown as a figure:

Looking at the results of the model above⁴, we see that Age is significant for the logistic regression. The odds ratio of not receiving any antibiotics was 0.43 for those who were ultimately diagnosed with a viral infection, but this wasn’t significant. Of note, the OR for those in the Pre-PCR era (OR = 0.61) is not significant.

Moving up to the conditional regression, we can see that the DOT was 1.95 times longer in the Pre-PCR era, compared to the Post-PCR era. We also can see that patients with contaminants on their blood culture had a longer duration of antibiotics, which makes sense.

The conditional means for each era are presented below:

Methods

We performed a retrospective chart review of infants younger than 90 days old who were admitted to the pediatric ward for concerns of fever or sepsis from 1/1/2015 to 9/30/2019. Subjects with documented bacterial infections were excluded. Our primary outcomes were differences in DOT and LOS between patients admitted before and after PCRs became available (Pre-PCR era and PCR era).

Results

118 subjects were included in our analysis, with 52 (44.1%) coming from the Pre-PCR group. Descriptive statistics, stratified by era, are available in Table 1. The unadjusted mean LOS and DOT decreased by 8.4 (P = .003⁵) and 32.8 (P < .001⁶) hours, respectively.

Multivariate regression demonstrated that LOS was 1.16 times longer in the Pre-PCR era (95% CI: 1.1 - 1.22; P < .001). Similarly, patients in the Pre-PCR era had DOT that was 1.95 times longer (95% CI: 1.55 - 2.46; P < .001) when adjusting for covariates.

Tables / Figures

Fig 1:

Table 1

Table 2

Fig 2