The Impact of DeliveryHospital on the Outcomes of Premature Infants in the Post-Surfactant Era: An Instrumental Variables Approach
Scott A. Lorch, MD, MSCE1,2,3
Michael Baiocchi4
Corinne Fager2
Dylan S. Small4
1 Department of Pediatrics, The Children’s Hospital of Philadelphia and The University of Pennsylvania School of Medicine, Philadelphia, PA
2 Center for Outcomes Research, The Children’s Hospital of Philadelphia, Philadelphia, PA
3 Senior Fellow, Leonard Davis Institute of Health Economics, University of Pennsylvania, Philadelphia, PA
4Department of Statistics, The WhartonSchool, University of Pennsylvania
Word Count: 2998
Abstract Word Count: 281
Abstract
Context: Even though prior work suggests that delivery at a high-level neonatal intensive care unit (HL NICU) is associated with improved outcomes, greater percentages of women deliver at hospitals without HL NICUs. Prior studies may provide biased assessments of deregionalization policies because many differences in casemix are not measured.
Objective: To determine the impact of delivering at HL NICUs on mortality and common complications of premature birth after controlling for unmeasured differences in casemix using an instrumental variables approach.
Design: Retrospective population-based cohort study
Setting: All hospital-based deliveries in Pennsylvania, Missouri, and Californiabetween 1993-2005.
Patients or Other Participants: Women delivering at a gestational age between 23 and 37weeks gestation (N=1,598,400)
Main Outcome Measure(s): Neonatal and fetal death, 7 complications of premature birth
Results: Women delivering at a HL NICU were more likely to have a preexisting comorbid condition or a complication of pregnancy. After controlling for unmeasured and measured factors, infants delivering at a HL NICU had 2.9 to 12.4 fewer deaths per 1000 deliveries, with similar rates of most complications studied except for lower BPD rates at Missouri HL NICUs and higher infection rates at HL NICUS in Pennsylvania and California. The association between delivery hospital and neonatal outcomes differ between the three states studied. Without accounting for ummeasured differences in casemix, HL NICUs would have significantly higher complication rates with similar neonatal mortality rates.
Conclusions: There is continued benefit to neonatal outcomes when high-risk infants are delivered at HL NICUs in the post-surfactant time period. To obtain accurate estimates of policies where patients receive care based on severity of medical condition, such as perinatal regionalization policies, either more sophisticated methodological approaches or more detailed clinical data are needed.
Introduction
Regionalization of health care may help provide high quality and cost-efficient health care by directing patients to facilities with optimal capabilities for any given type of illness or injury.1, 2 For perinatal care, a regionalized model of care was developed in the 1970s to centralize the care of the very-low-birth weight (VLBW) infant at specialized hospitals with the adequate personnel and technology. However, by the 1990s, the regional model of perinatal care began to weaken in many areas of the United States,3-6 with fewer VLBW births at regional perinatal centers.
Although several studies from the early 1990s suggest that delivery at a high volume, high technology hospital reduces neonatal mortality,7-10 this issue is worth further study. First, since the 1990s, there has been increased use of antenatal corticosteroids, routine use of postnatal surfactant and higher use of Cesarean sections, and there is little information about how this has affected the relationship between delivery hospital and neonatal outcomes. Also, no study has examined other outcomes, such as complication rates, or compared the effect of delivery hospital in different states. Additionally, these prior studies suffer from a common problem associated with observational studies: selection bias. Specialty hospitals manage sicker patients with a higher risk of poor outcomes. Typical methods cannot adjust for unmeasured or unrecorded factors in available data, such as the severity of an antenatal comorbid condition, lab results, or fetal heart tracing results. Failing to account for these unmeasured factors may result in a biased assessment of the impact of delivering at a high volume, high level NICU compared to other delivery hospitals.
The goals of this study are to (1) obtain unbiased measurements of the impact on mortality of delivering at a high volume, high level NICU compared to other delivery hospitals using more recent data; (2) examine additional outcomes besides mortality; and (3) examine between states with different systems of regionalization. To control for unmeasured differences between high volume, high level NICU and other delivery hospitals, we will employ an instrumental variables study design. This study design is new to the perinatal literature, but has been used in other health policy settings.11-15
Methods
Study Design
Observational studies of perinatal regionalization need to adjust for differences in casemix between high-level NICUs and other delivery hospitals. Past studies have used regression analysis to control for those aspects of casemix that are recorded in the data set, such as the birth weight and gestational age of the infant and the co-existing medical conditions of the mother. However, the regression approach is biased if important factors are left unmeasured.
An instrumental variables approach controls for both measured and unmeasured differences in casemix between these groups of hospitals. In this study design, a variable referred to as an instrument encourages patients to deliver at a particular hospital, in essentially a randomized fashion. The instrument must have three characteristics: (i) it must be independent of unmeasured confounding variables conditional on measured confounding variables; (ii) it must influence where a patient delivers (iii) it should only influence the observed outcome of the patient through its effect on where a patient delivers. A strong and valid instrument varies where a mother delivers, while equalizing other measured and unmeasured factors.
To ensure that patients with higher and lower values of the instrument are comparable, we employ a matched pairs study. Here, we match patients on 59 measured covariates while maximizing the difference in the instrument, a design referred to as "near-far matching."16 This matching design parallels a matched pair randomized controlled trial of patients encouraged to deliver at a high level NICU vs. patients not encouraged to deliver at a high level NICU. This method reduces the influence of patients who typically deliver only at high-level NICUs in the final analysis, such as infants with severe congenital anomalies. By including both an instrumental variables approach and this matched pairs design, we improve the equality of the two study groups, which improves the accuracy of the study results (Technical Appendix 1).16
Data Population and Sources
We obtained birth certificates from all deliveries occurring in Pennsylvania and California between 1/1/1995 and 6/30/2005 and Missouri between 1/1/1995 and 12/31/2003. Each state’s department of health linked these birth certificates to death certificates using name and date of birth, and then de-identified the records. We then matched over 98% of birth certificates to maternal and newborn hospital records using prior methods.8, 17, 18 Over 80% of the unmatched birth certificate records were missing hospital, suggesting a birth at home or a birthing center. The unmatched records had similar gestational age and racial/ethnic distributions to the matched records. The Institutional Review Boards of The Children’s Hospital of Philadelphia and the departments of health in California, Missouri, and Pennsylvania study approved this study.
Infants included in this study had a gestational age between 23 and 37 weeks, and a birth weight between 400 to 8000 grams. Birth records were excluded if the birth weight was more than 5 standard deviations from the mean birth weight for the recorded gestational age in the cohort, because of the high likelihood of a recording error in one or both variables.19 Initially, 1,362,782 birth records were identified for this project; 34,650 met the exclusion criteria, leaving 1,328,132 births in the final cohort.
Definition of Study Outcomes
The primary outcome for this study was mortality. Neonatal deaths were defined as any death during the initial birth hospitalization. We examined fetal deaths because poor resuscitation around the time of delivery could convert some neonatal deaths into fetal deaths. Fetal deaths were defined in two ways. First, we included all fetal deaths in each county with either a minimum gestational age of 23 weeks or a birth weight of 400 grams. Second, we included those fetal deaths that met a prior definition of a potentially preventable fetal death by care delivered at the hospital.8 We also examined common complications of premature birth as secondary outcome measures. The complications and identifying ICD-9CM codes are listed in Table 1.
Definition of covariate variables
We included specific covariate variables in our matching algorithm based on their association with one or more study outcomes. The final models included gestational age; birth weight, grouped into 250-500 gram strata; maternal sociodemographic factors, such as race, age, education, and insurance status; maternal comorbid conditions listed in the technical appendix 2; and 49 congenital anomalies grouped by affected organ system.8
Hospital Definitions
Based on prior work,7, 8 a specialty hospital was defined as a level III or higher facility that delivered at least 50 VLBW infants, on average, per year. All levels of care were obtained from the American Academy of Pediatrics perinatal survey20 and validated using procedure codes from hospitalizations at each hospital.
Instrument
The instrument for this study is based on prior studies that suggest that women tend to deliver at hospitals near their residential zip code. For each residential zip code in this study, we calculated the difference in travel times between the nearest high-level NICU and the nearest other delivery hospital. We calculated the differential travel time between the residential zip code and the closest high level NICU vs. the residential zip code and the closest non-high level NICU delivery hospital using ArcView software from ESRI, Inc., as in our prior work.21 Women with negative differential travel times lived in residential zip codes that were closest to a high-level NICU, whereas women with positive differential travel times had to bypass a nearby hospital and travel further to deliver at a high-level NICU.
To examine validity of the instrument, we measured the distribution of measured covariates across various quartiles of the instrument and across the near and far matched pairs. Ideally, a measured covariate would have the same distribution across the quartiles of the instrument before matching, which lends credibility to the assumption that the instrument is also balancing unmeasured covariates associated with the measured covariates. However, if the measured characteristic is unbalanced across quartiles, the matched pair design will balance the characteristic. The equality of the measured covariates across quartiles before matching was assessed by calculating the standardized difference of each variable, which equals the (largest pairwise difference in means across quartiles of the instrument) ÷ (standard deviation of entire group). The equality after matching was assessed by calculating (difference in means between near and far patients) ÷ (standard deviation of entire group). A value less than 0.20 is considered adequate balance.22, 23
Data Analysis
Three analyses will be presented. First, we present a naïve analysis using unadjusted differences in each of the nine outcome measures between patients delivering a high-level NICU and other delivery hospitals. Next we performed a more sophisticated, though still inadequate, analysis where we controlled for measured differences in casemix with a matched-paired propensity score analysis. Finally, we performed the appropriate analysis which controls for measured and unmeasured differences using a matched-pairs instrumental variables analysis. Risk differences and relative risks are presented for each analysis. Confidence intervals for risk differences were calculated by standard inversion of a pivot-based test of the null, at an alpha error rate of 0.05,16 while confidence intervals for relative risks were calculated using bootstrap methods. All data are presented separately by state to allow for inter-state comparisons.
Results
Overall, women who delivered at a high level NICU were more likely to have either a preexisting comorbid condition, such as diabetes mellitus, or a complication of pregnancy, such as preterm labor or pregnancy-induced hypertension (Table 2). As a result, infants delivered at high level NICUs were smaller and had a younger gestational age.
Strength of Instrument and Equality of Population after use of Instrument
Results Appendix 1 shows the strength of the instrument. In Pennsylvania, 79.8% of the pregnancies in the first quartile delivered at a high-level NICU, compared to 23.9% in the fourth quartile. Similar strengths of the instrument were seen in California (79.6% versus 38.3% respectively) and Missouri (55.7% versus 10.1% respectively). In all three states, women in the middle two quartiles delivered at high-level NICUs at rates in the between the two extremes.
The instrument also balances measured covariates in each state (Appendix 2). Standardized differences in clinically relevant factors such as birth weight, gestational age, singleton birth, and every maternal comorbid condition and complication of pregnancy was less than 0.2 across the four quartiles of the instrument. For some socioeconomic variables, such as race and insurance status, the combination of the instrument and the matched pair analysis balanced all measured variables between those patients encouraged to deliver at a high level NICU vs. patients not encouraged to deliver at a high level NICU (Table 3).
Association of delivery hospital and mortality rates
After adjusting for both measured and unmeasured casemix differences between hospital types, delivering at a high-level NICU was associated with lower mortality rates in all three states (Table 4). In Pennsylvania, there was a reduction of 7.2 neonatal deaths per 1000 deliveries (95% CI 3.7-10.7), with a relative risk (RR) of 0.27 (95% CI 0-0.59). Fetal deaths in Pennsylvania were slightly reduced at high-level NICUs. In California, neonatal deaths were only reduced by 0.7 deaths per 1000 deliveries, but fewer fetal deaths occurred at deliveries at high-level NICUs (reduction of 5.4 fetal deaths/1000 deliveries, 95% CI 3.6-6.9; reduction of 2.2 preventable fetal deaths/1000 deliveries, 95% CI 1.2-3.1). The reduction of neonatal deaths in Missouri just missed reaching statistical significance (RR 0.56, 95% CI 0.27-1.27).
Without accounting for unmeasured differences in casemix, neonatal mortality rates were statistically similar in Missouri (RR 1.01, 95% CI 0.92-1.01),Pennsylvania (RR 0.95, 95% CI 0.85-1.05), and California (RR 0.96, 95% CI 0.93-1.01). Pregnancies ending in a preventable fetal death were lower in both California and Missouri (Results Appendix 2).
Association of delivery hospital and rates of neonatal complications
In unadjusted and propensity score analyses, there were higher rates of all studied complications at high-level NICUs regardless of state (Results Appendix 2). After accounting for unmeasured casemix differences, few of these differences remained. Delivering at a high-level NICU in Missouri was associated with lower rates of BPD. Rates of other complications such as NEC and ROP were similar between the high-level NICU and other delivery hospital group. Infection rates remained significantly higher in high-level NICUs compared to other delivery hospitals, although the risk difference decreased from 5-45 extra infections/1000 deliveries to 0-14 cases/1000 deliveries (Table 4).
State Differences in Outcomes
For mortality, Pennsylvania and Missouri showed a 2-fold reduction in neonatal mortality rates with delivery at a high-level NICU, while California showed such a reduction in fetal, but not neonatal mortality rates (Table 4). The risk difference for most complications such as NEC, ROP, and ROP surgery showed some variation between states (Table 4). However, Missouri showed a large reduction for BPD rates when infants were delivered at high-level NICUs (reduction 9.5 cases/1000 deliveries), whereas Pennsylvania and California showed little change in BPD rates.
Discussion
Determining the true impact of a policy intervention such as perinatal regionalization is critical to accurately weighing the benefits and costs of the intervention. In perinatal regionalization, specialty hospitals treat sicker patients.4 After adjusting for all differences between delivery hospitals, our study suggests that there is a continued mortality benefit to delivering premature infants at high-level NICUs. This decrease was 50% larger than in previously reported studies, such as a 4-fold reduction in neonatal mortality in Pennsylvania and a 2-fold reduction in preventable fetal mortality in California. The sizes of these reductions are comparable to the overall Black-White disparity in infant mortality seen in the United States [ref]. However, if we examined this question using only factors currently measurable in most population-based datasets, such as birth weight and maternal medical conditions, we would not have found results similar to studies conducted with older data.4-9, 24-27 Increased access to clinical information, such as those data potentially available in electronic health records, would improve casemix adjustment and decrease the need for more sophisticated methods to obtain accurate estimates of the impact of health policies.
Instrumental variables approaches have been used in studies where patients with certain characteristics are more likely to receive a given treatment.11-15 In these studies, the use of the instrument gave a result that was more similar to randomized studies.12 For perinatal regionalization studies, our data found statistically significant results that were much larger than prior studies.
The strongest effect of delivery hospital was seen in the continued improvement in mortality rates. The almost 100% higher mortality rates at delivery hospitals without a high-level NICU suggest that, in the current era of widespread use of antenatal corticosteroids and postnatal surfactant, the delivery, resuscitation, and initial management of a premature infant is even more important to that infant’s survival than in prior time periods. Our work also examined the association between delivery hospital and neonatal complication rates. After adjusting for unmeasured differences in casemix, rates of important complications for long-term neurocognitive development, such as BPD, NEC, and ROP,28 were statistically similar in the two hospitals. This similarity occurred even though high-level NICUs with lower mortality rates are saving infants who would otherwise die if delivered at other hospitals.