air pollution and mortality/readmission in Medicare ADRD

01extractADRDhospital.R:

• screened 25 billing codes from 2000 to 2016
  • though FFS data only billed 10 codes before a certain year (cannot remember) and change to 25 codes afterwards so as to select ADRD people as much as possible from Medicare data
• searched with both ICD-9 and ICD-10 ADRD codes,
  • though ICD codes changed to the 10th version after 2015, in our dataset there are still some mix-ups of the 9th and 10th versions. ADRD diagnosis codes were based on Using Medicare claims in identifying Alzheimer’s disease and related dementias and Identifying Medicare beneficiaries with dementia; codes are modified from a previous one.
• export several .fst files named as ADRD'type'_'year'.fst, e.g. ADRDprimary_2000.fst
  • primary fst file contains hospitalization records with first diagnosis code (DIAG1) as ADRD, and secondary fst file contains those with any of codes (1-25) as ADRD. (secondary file contains primary);

02getEnrolledInfo.R

• qid format changed after a certain time, but there is a crosswalk file linking qids. According to Ben, QIDs in raw dataset have been converted to one single formmat
• still, some qids weren't matched. We excluded those problemetic qids based on the .csv file Ben provided.
• export EnrolledInfo.csv, only containing QID and firstADRDyr(first hospitalization admission year with ADRD), one-row-per-person

03extractDenominatorFile.R

• donominator files were prepared by research group before, should contain all Medicare enrollee's demographic info
• subset the denominator files to those with QID which appeared in EnrolledInfo.csv
• subset the person-year data to those starting from firstADRDyr (year>=firstADRDyr) based on EnrolledInfo.csv for each subject
• remove duplicates, export ADRDpeople_denom.fst

04mergeExposure.R

• PM2.5, NO2, ozone, summer ozone
• merge by calendar year and zip code
• calculate Ox
• export ADRDcohort.fst (contains all the information starting from firstADRDyr, though we don't need year==firstADRDyr rows in survival analysis itself, we do need it for correcting index event bias, should contains all the info we need in the future analysis)

05checkFollowupCompleteness.md

• follow-up data should start from the next year of first hospitalization with ADRD code, ends at the event/loss-of-follow-up/end of study period
• had considered to fill in all the missing person years, but without denominator information it will be very hard. After checking the death risk, we decided to omit those without complete follow-up
• exported omitInfo.csv with all qids that should be dropped

06cleanData.R

• when correcting the index event bias, we need the exposure info when people entering the cohort.
  • the earlist exposure info starts from 2000, then the earlist year of firstADRDyr should be 2000, and the earlist year of starting follow-up should be 2001
• clean ADRDcohort.fst: remove NAs; remove those without complete follow-up based on omitInfo.csv
• add necessary variables into the dataset: entry_age_break, race_collapsed, region, convert variables to the right class
• export ADRDcohort_clean.fst

07generateReAd.R

• ADRDcohort_clean.fst is ready to be used for time-to-death analysis
• we defined "readmission" as the second admission of ADRD people after their first hospitalization with ADRD, and the second hospitalization doesn't have to contain ADRD billing code. This is for a better capture of "how sick" ADRD people are, also once people diagnosed with ADRD, they will always have that disease, no matter it was billed or not. When people is too sick, the doctor may not bother billing ADRD instead of other fatal diseases (such as organ failure)
• instead of extracting all hospitalization records with ADRD code as in 01extractADRDhospital.R, we extracted all hospitalization records for ADRD people based on qid in EnrolledInfo.csv.
• subset the hospitalization records for each ADRD patient to the records after their first ADRD hospitalizatoin, and the minimum of admission year would be the firstRdADyr
• export ReAdissionInfo.csv, only containing QID and firstReAdyr(second hospitalization admission year), one-row-per-person
• subset ADRDcohort_clean.fst based on ReAdmissionInfo.csv for time-to-readmission analysis