Data resource profile: State Inpatient Databases

Data resource basics

Scope

The State Inpatient Databases (SIDs) comprise a collection of state-specific encounter-level administrative claims from 32/50 states in the USA. These databases are maintained by the US Agency for Healthcare Research and Quality (AHRQ) as a part of the organization’s Healthcare Cost and Utilization Project (HCUP). As part of a family of six other HCUP databases (Table 1), the SIDs represent the largest collection of longitudinal hospital care data in the USA inclusive of all insurance payers (e.g. Medicare, Medicaid, private insurance and uninsured) and all patient ages. Taken together, the SIDs capture more than 97% of all eligible hospital discharges within each state.

Table 1.

Family of Healthcare Cost and Utilization Project (HCUP) databases

Database name	Abbreviation	Content summary	Data dictionary
National US databases
National (Nationwide) Inpatient Sample	NIS	Cross-sectional subset of national inpatient stays since 1988, needs to be nationally weighted	https://www.hcup-us.ahrq.gov/db/nation/nis/nisdde.jsp
Kids’ Inpatient Database	KID	Cross-sectional subset of paediatric inpatient stays since 1997, released every 3 years, needs to be nationally weighted	https://www.hcup-us.ahrq.gov/db/nation/kid/kiddde.jsp
Nationwide Emergency Department Sample	NEDS	Cross-sectional subset of emergency department visits since 2006, needs to be nationally weighted	https://www.hcup-us.ahrq.gov/db/nation/neds/nedsdde.jsp
Nationwide Readmissions Database	NRD	Longitudinal linking within individual calendar years of inpatient claims from a subset of longitudinal SIDs since 2010, needs to be nationally weighted	https://www.hcup-us.ahrq.gov/db/nation/nrd/nrddde.jsp
State-specific US databases
State Inpatient Databases	SID	All state encounter-level inpatient claims (same hospital emergency department admissions, incoming transfers) since 1990, some states longitudinal across calendar years, AHA and SASD/SEDD external linkage possible	https://www.hcup-us.ahrq.gov/db/state/siddist/siddistvarnote2015.jsp
State Ambulatory Surgery and Services Databases	SASD	All state encounter-level ambulatory/outpatient claims at hospital-owned facilities since 1997, some states longitudinal across calendar years, AHA and SID/SEDD external linkage possible	https://hcup-us.ahrq.gov/db/state/sasddist/sasddistvarnote2015.jsp
State Emergency Department Databases	SEDD	All state encounter-level emergency department claims (non-inpatient admissions, discharged as transfers) since 1999, some states longitudinal across calendar years, AHA and SASD/SID external linkage possible	https://www.hcup-us.ahrq.gov/db/state/sedddist/sedddistvarnote2015.jsp

AHA, American Hospital Association Annual Survey Database (hospital-level information on additional institution-specific parameters).

Purpose of data collection

Originally created as the ‘Agency for Health Care Policy and Research’, AHRQ was established by the US Omnibus Budget Reconciliation Act of 1989 with the purpose of enhancing the ‘quality, appropriateness, and effectiveness of health care services […] through the establishment of a broad case of scientific research’.¹ In particular, the Act required that databases be created to support healthcare research. The HCUP-family of databases were, therefore, established as a federal-state–industry partnership sponsored by AHRQ in order to ‘enable research on a broad range of health policy issues, including cost and quality of health services, medical practice patterns, access to health care programmes, and outcomes of treatment at the national, state and local market levels’.² The HCUP SIDs maintain data expressly for research purposes dating back to 1990.

Structure

The SIDs share a common structure, although the precise list of available data elements vary slightly by state and calendar year. Differences in key variables used to link SID data are presented in Table 2. The SIDs are organized by calendar year based on the date of hospital discharge. The data for each SID year (e.g. Florida 2014) are presented in up to five separate files: core, charges, American Hospital Association (AHA) linkage, Diagnosis Related Groups (DRGs), and disease severity. The core file contains the discharge-level data elements that make up the majority of each SID. Some core data elements are universal [e.g. AGE, DISPUNIFORM (discharge disposition), DX1-DXn (primary and secondary diagnosis codes), PR1-PRn (primary and secondary procedure codes) and LOS (length of stay)]. Others vary by state [e.g. ZIP (patient residential zip code), RACE (patient race/ethnicity), and visitLink/DaysToEvent (variables needed for longitudinal patient tracking)]. The charges file contains information needed to convert total hospital charges (core: TOTCHG) into estimated costs incurred. The AHA linkage file permits linkage to data from the AHA Annual Survey Database, which contains additional hospital-level information on institution-specific parameters such as the number of hospital beds, the number of full-time residents or cancer hospital/trauma centre designations.³ The DRGs file represents an attempt by the US Centers for Medicare and Medicaid Services to classify ICD-9-CM/ICD-10-CM diagnosis codes into meaningful diagnostic groups at the level of an inpatient stay record.⁴ For example, two ICD-9-CM codes (401.1 ‘hypertension, benign’ and 401.9 ‘hypertension, unspecified’) are combined to form a single Clinical Classifications Software category: 98 ‘essential hypertension’.⁵ Finally, the disease severity file provides data from AHRQ Co-morbidity Software using individual inpatient stays as the unit of observation to provide information on the presence or absence of pre-specified co-morbidities and overall categorical rankings of a patient’s risk of mortality and disease severity.

Table 2.

Availability of key linkage variables within the 2015 SIDs, the most complete year of data available in HCUP to date

Element	AR	AZ	CO	DC	FL	GA	HI	IA	KS	KY	MA	MD	ME	MI	MN	MS	NC	NE	NJ	NM	NV	NY	OR	RI	SC	SD	UT	VT	WA	WI	WV
AHAID	+	+	+	–	+	–	–	+	–	+	+	+	–	–	+	+	+	+	+	–	+	+	+	+	–	–	+	+	+	+	+
DaysToEvent	+	–	–	–	+	+	–	+	–	–	+	+	–	–	–	+	–	+	–	+	–	+	–	–	–	–	+	+	+	+	–
HOSPID	+	+	+	+	+	+	+	+	+	+	+	+	+	+	+	+	+	+	+	+	+	+	+	+	–	+	+	+	+	+	+
MDNUM1_R	+	+	+	+	+	+	–	+	–	–	–	+	–	+	–	+	–	–	+	+	+	+	–	+	–	+	+	–	+	–	–
visitLink	+	–	–	–	+	+	–	+	–	–	+	+	–	–	–	+	–	+	–	+	–	+	–	–	–	–	+	+	+	+	–

AHAID, permits linkage to the AHA Annual Survey Database; DaysToEvent, time to event variable using a randomly selected start date and always used in conjunction with visitLink; HOSPID, unique identifier for individual hospitals; MDNUM1_R, unique identifier for individual physicians; visitLink, unique identifier for individual patients.

Research uses

The SIDs are frequently used because of their universal coverage, ability to track patients longitudinally between hospital admissions, ability to link to AHA data, inclusion of geographic-specific information such as patient residential zip codes, and ability to identify individual US states. As the SIDs are state-specific, these data can be used to evaluate health policy interventions that take place at the state level or to compare practice patterns/outcomes that vary between states by using data from neighbouring or otherwise comparable states as regional controls. Applications in published literature have ranged from geographic mapping⁶ of access to care and variations in quality metrics to assessment of changes in practice patterns and outcomes over time⁷ to more traditional analyses of health policy utility and predictors of adverse outcomes/disparities and quality ranking.^8–10

Processing cycle and frequency of data collection

The AHRQ partners with a range of organizations in participating states, which include hospital associations and private organisations as well as the federal government, to collect data. Datasets are made available through the HCUP Central Distributor when they have been adequately processed and approved by state partner organizations. This practice introduces a delay (typically 18–24 months) before data becomes available for analysis, although such intervals vary by state.

Linkage within the SIDs

A number of SIDs (Figure 1; Table 2) include a re-visit variable (‘visitLink’), which is a unique patient identifier that permits individuals to be tracked between discharge records from a randomly generated time origin point encoded in the variable ‘DaysToEvent’. Starting in 2003, revisit data were available for a total of 8 states. As of 2015, that number increased to 14 states, with a total of 16 states contributing re-visit data between 2009 and 2016 (Table 3). Although patients can be tracked between years within the same state, they cannot be tracked between admissions to hospitals in different states.

Figure 1.

Map of the USA showing states with State Inpatient Databases and those with 2016 re-visit variables that permit longitudinal linkage of patients between hospital admissions. The California SID has paused after the 2011 data release, although additional years may become available in future. Created with https://mapchart.net.

Table 3.

State Inpatient Databases including a re-visit variable between 2009 and 2016

State	Code	Populationa	Annual SID hospital dischargesb
Arkansas	AR	2 988 231	399 327
California	CA	39 296 476	3 933 239
Florida	FL	20 656 589	2 837 863
Georgia	GA	10 313 620	1 070 471
Iowa	IA	3 130 869	312 732
Massachusetts	MA	6 024 752	796 716
Maryland	MD	6 823 721	628 251
Mississippi	MS	2 985 415	378 494
North Carolina	NC	1 907 603	1 110 146
Nebraska	NE	2 085 432	202 658
New Mexico	NM	19 836 286	200 458
New York	NY	10 156 689	2 346 869
Utah	UT	3 044 321	296 903
Vermont	VT	623 354	53 652
Washington	WA	7 280 934	649 624
Wisconsin	WI	5 772 917	602 279

^aBased on 2016 population estimates from the US Census Bureau.⁴³

^bBased on the most recent year available for each state, which was 2016 in 12/16 (75%) SIDs³⁴.

Linkage to other datasets

Individual patients can be tracked over time between inpatient admissions (above) using SIDs in isolation. However, linkage of SIDs to State Emergency Department Databases (SEDD) and State Ambulatory Surgery and Services Databases (SASD) (Table 1) can also facilitate tracking of patients between emergency department attendances, hospital admissions and outpatient visits. Complete assessment of all emergency department visits within a state requires a combination of SID and SEDD. Many SIDs provide a variable that permits linkage to the AHA Annual Survey Database, which contains a comprehensive range of hospital characteristics that can be used to enrich individual SIDs. The AHA Annual Survey Database includes more than 1000 institution-level data elements about more than 4600 hospitals within the USA.³

Data collected

As described in the ‘Structure’ section above, the availability of data elements within a given SID depends on the state and calendar year. Some data elements (e.g. discharge destination) are available in both a ‘raw form’ as provided by individual states (e.g. ‘disp_X’) and as a second variable that has been standardized by HCUP (e.g. ‘dispuniform’). Key linkage variables are shown in Table 2. Information on other important parameters is provided below.

Demographic data

Key SID demographic data elements include information on age, sex, race, marital status, zip code (first three digits or all five depending on the SID), primary payer/insurance status (e.g. Medicare, Medicaid, private insurance or self-pay), income quartile (median household income based on residential zip code), and state of residence. Age is available in days for patients under 1 year and months for patients under 11 years. There are also data elements specific to new-borns, including birth month and year, as well as birth weight.

Admission and discharge data

Key data elements on admission and discharge data include admission time (hour, month) as well as type (e.g. emergency, urgent, elective) and whether or not the admission was during a weekend, which is defined as an admission on a Saturday or Sunday. There is also a flag to identify patients transferred in from another hospital. Discharge time, month, and quarter as well as destination are reported (e.g. transfer to a skilled nursing facility, transfer to an inpatient rehabilitation facility, died, transfer to hospice, etc.).

Diagnoses and procedures

Diagnoses are recorded using ICD-9-CM diagnostic and procedure codes from the beginning of data collection until 30 September 2015 when the USA switched to ICD-10. Since 1 October 2015, the SIDs have used ICD-10-CM diagnostic and ICD-10-PCS procedure codes.¹¹ For each discharge record, up to 15 diagnostic and 15 procedure codes are reported as are up to 4 E-codes for designated trauma patients. By convention, the primary diagnosis and procedure codes occupy the first-named variable, i.e. ‘DX1’ and ‘PR1’, respectively. A ‘present on admission’ variable can be used to distinguish between co-morbidities present at the time of admission and those that developed in hospital. However, the present on admission variables are not always consistently reported. Researchers have often relied more heavily on clinical experience as to which diagnoses are likely to constitute a pre-existing medical condition when incorporating information from secondary diagnosis codes. For example, chronic diseases (e.g. diabetes) might be assumed to represent a pre-existing condition whereas this is less likely to be true for acute diagnoses such as acute respiratory distress syndrome. Researchers have previously used user-generated software such as CHARLSON¹² and ICDPIC¹³ that can translate ICD-9-CM codes into standard disease/injury scores such as the Charlson Comorbidity Score, Elixhauser Co-morbidity Index, and Injury Severity Score.^8,10,14,15 The SIDs also include diagnosis related groups (DRGs) for each disease diagnosis that are defined using the Medicare DRG Grouper¹⁶ algorithm during HCUP processing.

The date of each coded procedure is available with reference to the day of admission. For example, a record with ‘3512’ under variable ‘PR1’ and ‘3’ under ‘PRDAY1’ would indicate that the primary procedure was a mitral valve replacement (ICD-9-CM 35.12 ‘Open heart valvuloplasty of mitral valve without replacement’), which took place on the third day after admission.

Possible outcome measures

Variables that have previously been used as outcomes include documented occurrence of specific complications, in-hospital mortality, receipt of specific interventions, scores on pre-defined quality metrics, discharge destination, length of stay, total hospital charges in US dollars for the corresponding calendar year (convertible into total hospital costs), and routine vs non-routine discharges. Longitudinal linkage variables enable assessment of readmissions, prior hospitalizations, repeat operations and/or repeat need for hospitalizations under similar or changing diagnoses, and corresponding times to all of these events. Although SID data have been used to identify readmissions, there is no single data element that can be used to determine whether a readmission was ‘unplanned’.¹⁷ The data also lack information on complications or fatalities that occur outside of a hospital setting. Some complications and other outcomes are particularly susceptible to surveillance bias, i.e. ‘the more you look, the more you find’.¹⁸ Although this problem can be addressed in some datasets by controlling for diagnostic procedures, this is difficult for researchers using the SIDs as some low-cost procedures are not routinely captured in the billing-driven HCUP dataset suite.¹⁹

Data resource use

There has not been a definitive bibliography of research papers published using the SIDs. However, a search of PubMed using “state inpatient database*”[tw] OR “state inpatient databases” identified n = 536 publications as of December 2018 (Figure 2). Figure 2 shows that SID use has increased every year since the first publication appeared in 2000.

Figure 2.

PubMed-indexed peer-reviewed research publications using data from HCUP State Inpatient Databases from January 2000–December 2018.

The ability to track patients between admissions over time in SIDs has been utilized by researchers interested in hospital readmissions.¹⁵ SIDs containing identifiers for individual physicians and hospitals have been used to determine annual case volumes and search for associations between volume and patient outcomes.^10,14 SIDs that can be linked to the AHA Annual Survey Database have been analysed for reported associations between hospital characteristics (e.g. bed size, teaching status, rural/urban location) and outcomes.^20,21 SIDs have also been used to describe the burden of disease,²² evaluate disparities,^8,23 challenge measures of hospital quality,^24,25 describe inter-hospital variation,²⁶ and evaluate health policy initiatives,²⁷ including differences in implementation at the state level.²⁸

Strengths and weaknesses

Very few of the papers identified in the search (above) sought to validate or critically evaluate SIDs, although this is likely to change as the databases are increasingly used by health services researchers. The AHRQ has undertaken a number of methodological exercises aimed at improving the quality of HCUP data.²⁹ Recognized limitations of SIDs include the potential for absent or incorrect coding of relevant information and a lack of nuanced clinical detail. Databases are not identical across states, do not track information across state lines, and do not capture information that occurs outside of a hospital setting—most notably including a lack of links to death registry data. The use of research checklists, particularly those designed to improve analysis of HCUP datasets, may help mitigate these limitations.^30–33

Coverage

A major strength of the SIDs is the comprehensive coverage, with approximately 97% case capture.^14,29 The broad coverage is particularly important in the USA, where health datasets are often based on payer status and so may not be representative of the broader population. However, as there is substantial variation among individual states, the SIDs are less well-suited to generate national estimates. The large size of some SIDs (e.g. almost 4 million discharges per year in California³⁴) can help to study rare diseases or outcomes, particularly when multiple SIDs are combined.

Longitudinal and horizontal linkage

The ability to track patients over time is a valuable feature of the SIDs, particularly as linkage to other HCUP datasets can permit follow-up through future emergency department attendances and day-case surgical procedures, as well as hospital admissions. This linkage permits researchers to capture a range of outcomes such as future operations and healthcare utilization.

Out of hospital deaths

Although longitudinal and horizontal linkage through healthcare episodes are particular strengths of SIDs, an important weakness is the inability to capture deaths outside of the hospital setting. The re-visit identifiers are encrypted by HCUP and data recipients must agree not to link visits to other datasets. As SID records cannot be linked to death registration data, researchers are limited to using in-hospital death, which is known to be affected by early discharge.³⁵ For example, in-hospital hip fracture mortality in the California SID (2007–2011) is 1.8%, with a corresponding reported median length of stay of 5 days with a US-care-specific high rate of discharge to post-operative rehabilitation.¹⁴ In the UK, where index hospitalization and rehabilitation are usually combined, in-hospital mortality is much higher at 8.1%, with a reported median length of stay of 15 days.³⁶ Lacking a constant follow-up point (e.g. 30-days post-injury) such comparisons can be fraught with difficulty.^17,35 There is evidence to suggest that in-hospital mortality can be used as a surrogate for 30-day mortality,³⁷ but this assumption needs to be explicitly considered by researchers undertaking comparative analyses. The SIDs can, however, account for patients who survive the initial hospitalization but who die during a subsequent readmission within a set number of days or who are discharged but then return to hospital as ‘dead on arrival.’

Patients re-admitted within other states

Although a number of states permit longitudinal tracking of patients between hospital admissions (Figure 1; Table 3), the unique identifiers do not permit tracking between SIDs. It is therefore possible that patients living near state borders might be re-admitted to a hospital in a different state and not be captured within the SID where they initially presented for care. At one extreme, the National Association of Health Data Organizations (NAHDO) reported that 4.8% of patients in the Maryland SID were residents of neighbouring states and that 1.4% of Maryland residents discharged from Maryland hospitals were initially treated in a different state.³⁸ It should be noted that Maryland is the ninth smallest state by total area in the USA and that it shares close geographic boarders with Virginia and Washington DC. As such, it is not surprising to see a relatively high level of cross-state exchange within the Maryland SID. By contrast, the NAHDO found that only 0.4% of Florida residents were discharged from a hospital in Georgia, which is one of the southern state’s only two immediate neighbours.

Standardized coding

The pre-processing of state datasets ensures consistency of variables across the SIDs. All SIDs used ICD-9-CM diagnostic and procedure codes from their inception until 30 September 2015. When the USA transitioned to ICD-10-CM diagnostic and ICD-10-PCS procedure codes on 1 October 2015, all subsequent SID datasets used ICD-10. The 2015 SID files are split into two separate coding profiles representing quarters 1–3 (January–September) as ICD-9 and quarter 4 (October–December) as ICD-10.¹¹

Sensitivity to admission thresholds and other changes over time

As only admitted patients are included, the SIDs are likely to be sensitive to changes in admission thresholds.³⁹ For example, a trend towards treating haemodynamically stable patients with pulmonary embolism (PE) on an ambulatory pathway⁴⁰ would manifest as fewer PE cases in the SID over time. Concurrent use of SEDD is one means of quantifying such changes. Other disorders (such as hip fracture) are always likely to result in hospitalization and so should be unaffected by admission thresholds. Changing practice patterns have also been known to yield similar effects, as has been reported for perforated appendicitis admissions.²⁸ Similarly, changes in surgical care have moved a number of common operations (e.g. appendicectomy, cholecystectomy) from the inpatient to the outpatient setting.⁴¹

Missing data

The levels of missing data in the SIDs are usually much lower than for other large administrative databases; although some variables are poorly coded. One report using the Michigan SID noted that missingness was <1% for most data elements. However, a small number of variables had high levels of missingness (e.g. total charges 19.8% missing, race 17.2%, and physician ID 4.5%).⁴² The HCUP website and instructional literature for each SID often indicate why values are missing in these cases (e.g. some hospitals intentionally do not report race) and can help to determine how missing data should be handled in analyses.

Data resource access

Data releases are available on application to the HCUP Central Distributor (https://www.distributor.hcup-us.ahrq.gov). Applicants are required to complete online ‘data use agreement training’, which takes approximately 15 minutes. Applicants must then complete a number of agreements online and provide a ‘Statement of Intended Use’. Once the application has been approved and payment made, the data files are made available for download using a separately emailed protected password. Historically, some SIDs had to be sent by post. The cost of data files varies by state and year, with specific information available online from the HCUP Central Distributor. Prices typically vary depending on whether or not the applicant is a student, has an educational or not-for-profit affiliation, or represents a commercial interest. For example, the Arizona SID costs $35–50 per year for students, whereas the Utah SID can cost upward of $3200 per year for private organizations. The process from application to download of data can be completed within a few days but could require slightly longer if individual states require additional proposal review. There are no geographical limitations on the data and so it is not necessary for applicants to be working in a specific state for which they request data or even to be working in the USA.

Profile in a nutshell

• The SIDs contain data on almost all inpatient discharges from hospitals in participating states. They were established from a range of data sources for the specific purpose of undertaking health-related research.

• The first SIDs are available starting in 1990.

• SID data elements vary by year and state, but typically include demographic characteristics, admission and discharge details, diagnoses and procedures, charge/cost and payment details, and outcomes such as in-hospital death. They are linkable to information contained in the AHA Annual Survey Database.

• Some SIDs permit longitudinal linkage so that individual patients can be tracked between hospital admissions. Patients can often also be tracked through emergency department visits through linkage to other state-level HCUP databases.

• Access to data is swift and at reasonable cost, particularly for non-profit and student researchers.

Conflict of interest: E.R.H. reports support by contracts CE-12–11-4489, DI-1603–34596, and PCS-1511–32745 from the Patient-Centered Outcomes Research Institute (PCORI), by grant 1R01HS024547 from the Agency for Healthcare Research and Quality (AHRQ) and grant R21HL129028 from the National Heart, Lung and Blood Institute of the National Institutes of Health (NIH). E.R.H. reports receiving support by contract W81XWH-16-R-BAA1 from the Department of Defense. E.R.H. receives book royalties from Lippincott, Williams & Wilkins for Avoiding Common ICU Errors. E.R.H. received compensation for authoring a paper commissioned by the National Academies of Medicine titled ‘Military Trauma Care’s Learning Health System: The Importance of Data Driven Decision Making’. The corresponding author confirms on behalf of all other co-authors that there have been no involvements that might raise the question of bias in the work reported or in the conclusions, implications, or opinions stated.