NationalDatasets
- 1. Nationally Representative and Other Large Datasets Jessica Moore Analysis and Data Management Activity (ADMA) 03/22/2012
- 2. Background • One of our goals is to estimate the burden of viral diseases among the US population by describing their association with – Hospitalizations – ED visits – Ambulatory/Outpatient visits – Deaths • Consequently, we need datasets for deriving robust, nationally representative estimates
- 3. Background Focus will be on datasets produced by two organizations within the Department of Health and Human Services • National Center for Health Statistics (NCHS) – Mission is to provide statistical information that will guide actions and policies to improve the health of the American people. • Agency for Healthcare Research and Quality (AHRQ) “ark” – Mission is to improve the quality, safety, efficiency, and effectiveness of health care for all Americans.
- 5. NCHS – Surveys National Hospital Discharge Survey (NHDS) http://www.cdc.gov/nchs/nhds/nhds_questionnaires.htm • Designed to meet the need for nationally representative information on inpatients discharged from nonfederal , short –stay hospitals in the United States – Federal, military, and VA Hospitals, as well as hospital units of institutions (such as prison hospitals), and hospitals with fewer than six beds staffed for patient use, are excluded. • Due to funding limitations, sample of hospitals reduced by half starting in 2008. • In 2009, approximately 160,000 inpatient records were obtained from 205 responding hospitals.
- 6. NCHS – Surveys National Hospital Discharge Survey (NHDS) • Variables include discharge diagnoses, procedures, length of stay, payer, outcome, gender, race and ethnicity (race missing on 19% of records) • A discharge weight (not the patient’s weight!) is provided for each record and the sum of the weights results in national estimates of the # of hospitalization associated with a particular condition • Public use data are available for download via ftp (1996-2009) – Epi share drive (cdcprojectNCIRD_DVD_EB_DATA_1nchspublic) has SAS datasets from 1996-2009 – Data collection for NHDS began in 1965 (sampling design changed in 1988). CDs available from NCHS for data 1970-1978 and 1979-2007
- 7. NCHS –Surveys National Hospital Discharge Survey (NHDS) • To conduct statistical tests, one must use the survey design variables (weight) to generate standard errors (SEs) for estimates – Calculating SEs requires SUDAAN – For access to the survey design variables, one has to prepare a project proposal for NCHS review and, w/approval, utilize the Research Data Center – Summary output only; no record-level data
- 8. NCHS – Surveys National Ambulatory Medical Care Survey (NAMCS) & National Hospital Ambulatory Medical Care Survey (NHAMCS) http://www.cdc.gov/nchs/ahcd.htm • Together – AKA “NAMCeeS/HAMCeeS” • The NAMCS is based on a sample of visits to non- federally employed office-based physicians who are primarily engaged in direct patient care. Physicians in the specialties of anesthesiology, pathology, and radiology are excluded – National estimates of doctor’s office visits • The NHAMCS is based on a national sample of visits to the emergency departments and outpatient departments of non-institutional general and short- stay hospitals. – National estimates of hospital outpatient department and ED visits
- 9. NCHS – Surveys NAMCS & NHAMCS • Demographic information, reason for visit, procedure codes, prescription information, pre-existing chronic conditions, payer, length of visit. • By using the NAMCS and NHAMCS one can generate national estimates of outpatient and ED visits. • Public use files available for download – The public use files for 1993-2009 contain sample design variables in masked form so you don’t need any special files/agreements • Still need SUDAAN for calculating SEs. • A combined SAS dataset from 1995-2009 is available in the Epi share drive for all 3 datasets cdcprojectNCIRD_DVD_EB_DATA_1nchspublic • Additional data, documentation and import code found on the NCHS site.
- 10. NCHS – Surveys National Immunization Sample (NIS) http://www.cdc.gov/nchs/nis.htm • Conducted jointly by NCIRD and NCHS • List-assisted random-digit-dialing telephone survey, with mailed survey to children’s immunization providers • Target population – children 19-35 months • Rates of being up-to-date with respect to ACIP recommended number of doses • DTaP/DTP/DT, polio, MMR, Hib, HepB, varicella, pneumococcal, HepA, influenza, rotavirus • SAS datasets in EB share drive available from 2010 • 2010 dataset has records for approximately 24,000 children, with 17,000 having adequate provider data
- 11. NCHS – Surveys National Immunization Sample – Teen (NIS-Teen) • Same sampling method of NIS • Public datasets and SAS datasets in EB share drive available from 2008-2010 • Assesses vaccination coverage of teens less than 17 years old • 2010 dataset includes records for approximately 33,000 teens, with 20,000 having adequate provider data • Tables of vaccination coverage for NIS and NIS-Teen available at http://www.cdc.gov/vaccines/stats-surv/imz- coverage.htm#nisteen • Public use file does not include any provider-level weights. Providers are included in the survey through the children they vaccinate. Do not attempt provider level analysis!
- 12. NCHS- Surveys General comments from past and present experiences • NHDS is known as the ‘gold standard’ for national hospitalization estimates. Public use files are likely sufficient for some activities negating the need for design variables. – Hospitalization estimates of less common conditions can be problematic (i.e. Relative SEs>30% or < 30 unweighted records are considered unreliable. Estimates based on 30-59 records are termed ‘low reliability’). – Stringent control of design variables requires a RDC proposal • NAMCS and NHAMCS has same limitations for less common conditions – Bonus is that design variables are included in public use files • NIS/NIS-Teen – do not attempt to calculate provider-level statistics! • For all surveys, the size of the annual files are relatively small compared to other datasets – Requires less disk space and processing time
- 14. NCHS – Vital Statistics Mortality Multiple Cause of Death • Annual files captures approximately 100% of US resident deaths (i.e. it is not a sample so no need to worry about record weights). – Underlying cause of death provided – Up to 20 ICD-9 (ICD-10 beginning in 1999) recorded on death certificate – Includes demographic information for the deceased • Public use files available for download – Geographical information suppressed for counties with pops<100,00 for 1998 through 2004 – ALL geographical information suppressed beginning in 2005 • DVD has agreement with NCHS to obtain county-level geographic information.
- 15. NCHS – Vital Statistics Linked Birth and Infant Death Datasets • Information from the death certificate is linked to the information from the birth certificate for each infant under 1 year of age who dies in the United States, Puerto Rico, The Virgin Islands, and Guam • Variables include age, race, and Hispanic origin of the parents, birth weight, period of gestation, plurality, prenatal care usage, maternal education, live birth order, marital status, and maternal smoking, linked to information from the death certificate such as age at death and underlying and multiple cause of death. • Useful for identifying risk factors for infant death. – One can compare characteristics of infant deaths from one cause to infant deaths due to another cause(s) – Or compare the characteristics of infant deaths to a sample of surviving infants.
- 16. NCHS – Vital Statistics Linked Birth and Infant Death Datasets • Two designs for the linked data - period data and birth cohort data. – The numerator for the period linked file consists of all infant deaths occurring in a given data year linked to their corresponding birth certificates, whether the birth occurred in that year or the previous year. – The numerator for the birth cohort linked file consists of the deaths (current or subsequent year) linked to infants born in a given year. – In both designs, the denominator is all births occurring in the year. – Due to the different designs, the period linked files are available sooner than the cohort linked files. • Public use files available for download – Geographical information suppressed for counties with populations<250,000 for 1998 through 2004 – ALL geographical information suppressed beginning in 2005 • DVD has agreement with NCHS to obtain county-level geographical information.
- 17. NCHS – Vital Statistics Birth Datasets http://www.cdc.gov/nchs/births.htm • Captures all registered live births occurring in the US. Birth certificate information included (like in the linked denominator file). • Public use files available for download – Geographical information is suppressed for counties with populations<100,000 for 1998 through 2004 – ALL geographical information suppressed beginning in 2005 • DVD has agreement with NCHS to obtain county-level geographical information. • SAS datasets from 1969-2009 available in EB share drive • 2 SAS summary datasets from 1990-2009 – Summary births by year, region, state, race, ethnicity and sex – Summary births by year, region, race, ethnicity and sex
- 18. NCHS – Vital Statistics General comments • Need to closely follow the documentation as variables were added/removed/modified over time – Particularly important for the birth and linked data due to states adopting the 2003 revision of the birth certificate. • For all mortality data, need to pay attention to the conversion from ICD-9 to ICD-10 beginning in 1999. – Comparability studies found differences in how conditions were coded between the two coding versions • Birth data and linked denominator data take a bit of processing time since each calendar year has approximately 4 million live births. • Can do basic queries online using wonder.cdc.gov or by using VitalStats (http://www.cdc.gov/nchs/VitalStats.htm)
- 19. AHRQ
- 20. AHRQ Healthcare Cost and Utilization Project “H-CUP” • HCUP obtains data on inpatient and ED visits from participating states – Standardizes the format of data across states – Creates samples from the states’ 100% hospitalization and ED visit data. • Extensive documentation on HCUP and associated datasets available online (http://www.ahrq.gov/data/hcup/)
- 21. AHRQ HCUP Datasets Samples • Nationwide Inpatient Sample (NIS) • Kids’ Inpatient Database (KID) • Nationwide Emergency Department Sample (NEDS) 100% data from source • State Inpatient Databases (SID) • State Emergency Department Databases (SEDD)
- 22. AHRQ HCUP - NIS • The NIS is a very large sample of hospital discharges from HCUP participating states (44 during 2009) – Began in 1988 with 8 participating states. We have SAS datasets available from 1988 to 2009. – Database design has evolved over time but has remained relatively constant since 1998. – The NIS is approximately a 20% sample of US hospitalizations occurring in community hospitals – Similar variables to the NHDS – Provides hospital charges and cost-to-charge ratio files so one can estimate direct charges/costs – Participating states account for approximately 95% of all hospital discharges in the U.S.
- 23. AHRQ HCUP - NIS • Since the NIS is such a large sample, analyses of less common conditions can be conducted. – Also…since the NIS is such a large sample, processing time to generate subsets can be burdensome (creating subsets from 5-8 million records each year). • Like the NHDS – One must use the record weight to generate nationally representative estimates – For statistical tests, one must generate SEs for the estimates • Can use SUDAAN, SAS Survey procs, or Stata • A bonus…design variables are provided with the datasets
- 24. AHRQ HCUP - KID • The KID follows the same format as the NIS but focuses exclusively on pediatric discharges (patients aged ≤ 20 years). – Very powerful for looking at less common conditions occurring among children that lead to hospitalization. • Produced every 3 years (1997,2000, 2003,2006,2009) • Same weighting and SE calculation procedures as with the NIS • Cost-to-charge files available for 2003,2006,2009
- 25. AHRQ HCUP - NEDS • The NEDS yields national estimates of emergency department (ED) visits. • The NEDS is a brand new dataset with the 1st release being for calendar year 2006 – ~26 million and ~27 million unweighted records for 2006 & 2007, respectively. • Similar variables to the NIS and also includes ED charges • SAS datasets available in EB share drive for 2006-2009 • Race is suppressed
- 26. AHRQ HCUP – State Datasets (100% Data) • Statewide Inpatient Databases 1988-2010* (SID) – 100% hospitalization data from participating states in HCUP. Currently, we do not have direct access to these data but have current/past collaborations with HCUP investigators to conduct analyses using these data. – The gold standard for assessing rare conditions leading to hospitalization. Can calculate state specific and indirectly estimate nationally representative rates. – Increasing number of states available over time. – These are very large datasets • State Emergency Department Databases 1999-2010* (SEDD) – Currently, we do not have direct access to these data but have current/past collaborations with HCUP investigators to conduct analyses using these data. – Captures the discharge information on all emergency department visits that do not result in a hospital admission – These are very, very, very large datasets *2010 already available for some states.
- 27. AHRQ HCUP – General Comments • Extensive online documentation available for the datasets. http://hcup-us.ahrq.gov/ • Access to sample data requires a brief online course and a signed DUA. – It’s easy. • Processing time can be a bit daunting. – Workarounds include remote submitting to branch server, CSP (pending), or using Citgo SAS (recommended) • Online tool (HCUPnet) is a good start for assessing the numbers that might be available for a study of interest. – http://hcupnet.ahrq.gov/ • Specifically for the NIS, changes in dataset format over time can make assessing trends challenging. HCUP has released trend files to facilitate these types of analyses. • AHRQ datasets include 15 possible diagnosis codes per record
- 28. MEDSTAT • MarketScan data warehouse • Contains individual-level healthcare claims, health risk assessments, absence, short-term disability, workers’ compensation, and hospital discharge info from large employers, managed care organizations, hospitals, Medicare and Medicaid programs • We have access to data from 1993 through 2010. Pre-2004 is in Citgo. Some of post-2004 data is located in CDC network land and some is located in CSP (consolidated statistical platform). • Very large datasets with extensive admission and discharge data, including cost and charge information, i.e. deductibles, coinsurance, other payments, etc. • Once you pull data from various sources and locations it’s a powerful dataset for estimating burden of various disease.
- 29. What about rates? Numerators are nice but what about denominators? • Bridged-race census data available for 1990-2009 – cdcprojectNCIRD_DVD_EB_DATA_1census – Goal of bridged race data is to make population estimates by race/ethnicity comparable for 1990-1999 to >2000. • Longer-term project is to also create intercensal population estimates datasets. • Remember that for the sample datasets, you have to factor in the SEs (i.e. compute relative standard errors) for your rates to do comparisons or to calculate 95% confidence intervals
- 30. Data Processing • Many of the national datasets are very large, i.e. NEDS approx. 27 million unweighted records per year! • While these are powerful datasets for estimating burden of disease, especially for rare conditions, the size of the datasets make them cumbersome to analyze. • We suggest, if working with these data in SAS, to pull the data from the share drive and run on SAS in Citgo. We tested various methods and this had the shortest run time, as well as least cpu usage. • Since the national datasets we use are complex survey designs, you need to analyze them in SUDAAN to calculate weighted estimates. SAS-callable SUDAAN is available in Citgo. • Other suggestions for larger datasets: – SGIO=yes – Compress=binary
- 31. Limitations • Hospitalizations, ED, outpatient visits – Have to rely on ICD-9-CM codes to identify cause – Primary versus secondary diagnosis versus any-listed • Was the condition of interest the ‘cause’ of the healthcare visit • Mortality – Rely on ICD-9/ICD-10 codes • Conversion to ICD-10 does result in changes in national estimates for a lot of conditions • Can be difficult/impossible to ascertain comorbid conditions • Survey datasets are encounter based rather than case based – Multiple visits by same individual can not be determined
- 32. Limitations • Public-use data good for general picture – Suppressed information (especially since 2005) is problematic. • Access to suppressed elements or design information requires annual agreements with NCHS. – Requires RDC proposal for access to design information. • Working with the data can be challenging, especially the birth, linked, and HCUP sample datasets • Caveats on producing SEs for the sample datasets – Requires knowledge of SUDAAN, SAS Survey procs and how does one know they got it right? • Check published reports for NHDS, or HCUPnet for the HCUP samples.
- 33. Conclusions • Between the NCHS and AHRQ, one can obtain a very complete picture of the healthcare burden associated with viral diseases • There is extensive online documentation and queries to assist one’s initial efforts in assessing the viability of a new project • All datasets (except Marketscan) are located in the Epidemiology Branch Data1 share drive: cdcprojectNCIRD_DVD_EB_DATA_1 • The ADMA team can help pull subsets of data as needed for various projects. – If interested in the data it is important to do the initial legwork first to determine if the project is feasible. Once you determine feasibility, the ADMA team is available to answer questions as needed.
- 34. Questions?
Editor's Notes
- #7: 6 diagnosis codes included Want estimate of length of stay? Multiply each record weight by the number of days admitted and then add the sum of products
- #12: -Statistical analyses require only data from children with adequate provider data (PDAT=1) along with their final provider sampling weights (PROVWT/PROVWTVI)
- #26: 15 diagnosis codes available
- #31: SGIO = scatter-read/gather-write input/output