cancer incidence in the health improvement network

pharmacoepidemiology and drug safety 2009; 18: 730–736nterscience.wiley.com) DOI: 10.1002/pds.1774
Published online 28 May 2009 in Wiley InterScience (www.i
ORIGINAL REPORT

Cancer incidence in The Health Improvement Network

Kevin Haynes PharmD, MSCE 1,2,3*, Kimberly A. Forde MD 2, Rita Schinnar MPA1,2,Patricia Wong MD4, Brian L. Strom MD, MPH1,2,3 and James D. Lewis MD, MSCE1,2,3

1Department of Biostatistics and Epidemiology, University of Pennsylvania School of Medicine, Philadelphia, PA, USA2Center for Clinical Epidemiology and Biostatistics, University of Pennsylvania School of Medicine, Philadelphia, PA, USA3Center for Education and Research in Therapeutics, University of Pennsylvania School of Medicine, Philadelphia, PA, USA4Division of Gastroenterology and Hepatology, Department of Internal Medicine, Johns Hopkins Bayview Medical Center, Baltimore,MD, USA

SUMMARY

Background The utility of electronic medical record databases for clinical research relies on the validity and completeness of the recordedmedical diagnoses. This study assessed whether the recorded incidence of cancer among patients in The Health Improvement Network(THIN) database is comparable to that expected in the UK based on national cancer registry data.Methods We examined incidence rates of any cancer other than non-melanoma skin cancer and the specific cancers colorectal, lung,pancreas, and lymphoma from 1992 to 2007. Indirect standardization was used to calculate standardized incidence ratios (SIR) using age- andsex-specific rates from the UK cancer registry for England and Wales for the corresponding years.Results Recording of the incidence of all cancers combined in THIN was very close to the expected rates from 2001 to 2007, that is, SIRwithin 10% of unity. Recording of the solid cancers was less than the expected based on cancer registry data, but with SIRs> 0.80 in 2007 foreach cancer. Recording of lymphoma was close to the expected rate for the entire follow-up period. Time and experience with Vision softwareemerged as important factors in reported incidence rates for all cancers.Conclusions For all cancers combined and for lymphoma the observed rates in THIN are very close to those reported in cancer registry datafor the years 2001–2007. However, for solid cancers the observed rates in THIN are below those reported in cancer registry data. This mayreflect the use of non-specific codes to record solid cancers. Copyright # 2009 John Wiley & Sons, Ltd.

key words—cancer; incidence; epidemiology

abbreviations—AMR, acceptable mortality reporting; SIR, standardized incidence ratios; THIN, The Health Improvement Network

Received 27 March 2009; Accepted 23 April 2009

INTRODUCTION

The utility of electronic databases of primary caremedical records for pharmacoepidemiology researchrelies on the validity of the medical diagnoses and thecompleteness of recording in the medical records. Inthe United Kingdom, there are several similar data-bases collecting data from general practitioners. TheHealth Improvement Network (THIN) has made dataavailable for epidemiologic research since 2003, butcontains data from the time the practice computerized

*Correspondence to: K. Haynes, University of Pennsylvania School ofMedicine, 732 Blockley Hall, 423 Guardian Dr., Philadelphia, PA 19104-6021, USA. E-mail: [email protected]

Copyright # 2009 John Wiley & Sons, Ltd.

their medical records. THIN, like some other collec-tions of electronic medical records, is of particularinterest to pharmacoepidemiologists because of theavailable data on all prescriptions issued by the generalpractitioner. The THIN database is believed to berepresentative of the UK population, containing dataon approximately 4% of the UK population.Cancer incidence is a common question with regards

to medication safety. Most cancer registries lack dataon prior medication exposure and other confounders.Medical record databases that contain importantexposure variables such as concomitant medical condi-tions, medications, smoking, and body mass indexovercome this limitation of traditional cancer regis-tries. Additionally, large medical record databases

cancer incidence in uk primary care database 731

provide a sufficient volume of patients followed for asufficient period of time to detect rare outcomes such ascomplications of cancer therapies. However, the cancerincidence data in the medical record databases arebased on physician recording, as compared to linkageto pathology or other sources in cancer registries.Few validation studies are available regarding

THIN.1,2 In anticipation of the use of THIN to conductepidemiological research, particularly studies examin-ing the association of medication exposures and cancerincidence, we designed this study to examine thequality of the data recording of cancer incidence in thisdatabase. One aspect of data quality is completeness ofrecording. Cancer diagnoses provide a unique oppor-tunity to compare national registry data with primarycare medical records. The ability to conduct incidencestudies in large electronic databases of medical recordsis predicated on accurately determining the number ofpatients with the disease and the amount of person-timecontained in the resource. The aim of this study was todetermine whether the observed and recorded incidenceof all cancers and of selected cancers (lymphoma,pancreatic, colorectal, and lung cancers) in the THINdatabase is comparable to that expected in the UKbased on national cancer registry data. These fourspecific cancers were selected to represent cancers thatwere common (lymphoma, lung, and advanced stagecolorectal), a frequent question in pharmacoepidemiol-ogy studies (lymphoma and colorectal cancer), highlylethal (lung, pancreas, advanced stage colorectal), andpotentially preventable with screening (advanced stagecolorectal). As a secondary objective, we sought todetermine whether recording practices have changedover time and whether recording rates differ amongthose THIN practices that were experienced users ofthe current software system (Vision from InPracticeSystems, introduced in 1994) or new users of thecurrent software system.

METHODS

Data source

Recruitment of general practices into the THIN data-base began in 2003. Practices that were usingelectronic medical records prior to joining THIN hadthese prior data imported into the database as well. Allparticipating practices use Vision software, and manywere using this same software for their electronicmedical record prior to joining THIN. Available datainclude patient demographics, medical diagnoses,therapy and information on preventative healthcare, tests,and immunizations on a total of more than 6 million


subjects, of which more than 2.5 million are activelyregistered with the practices and can be prospectivelyfollowed. Additionally, information on the practicesincludes the year when the practice started using Visionand the year when the practice reached an acceptablemortality reporting (AMR). The AMR, a tool devel-oped by THIN administrators, represents a qualityindicator for person-time that defines periods ofmortality reporting for each practice.3

Cohort selection criteria

All patients with an acceptable record flag wereinitially selected for the study. Cancer patients weredefined as having their first code for cancer in a givenstudy year (1992–2007). Subjects with unacceptableregistration status (e.g., not permanently registered; outof sequence year of birth or registration date; missingor invalid transfer out date; year of birth missing orinvalid; missing sex information) were excluded. Inaddition, subjects were excluded if they had a diagnosisof the specific study cancer prior to the start of follow-up.

Follow-up time

The start of follow-up for each subject was the latest ofthe following variables, all of which are included in theTHIN data: January 1st 1992, the date when thepractice implemented the computerized medicalrecord, the date when the practice started using thecurrent Vision information system if there were knownproblems with system migration from their originalsystem to Vision based on review of prescriptionrecords (as determined by the administrators of THIN),the date when the practice achieved an AMR, or thedate 6 months after registration of the subject with thepractice.4 The end of follow-up was defined as theearliest of the following: December 31st 2007, adiagnosis with the study cancer of interest, the date thatthe patient transferred out of the practice, the date ofdeath, or the last date for data collection by thepractice. If a subject had a cancer diagnosis within90 days following their recorded death date (as mightbe recorded from a hospital discharge summary), weconsidered the subject diagnosed with cancer on thedate of death.

Cancer case identification

The cancers selected for study were identified byREAD codes (available upon request). Cancer codeswere selected to be consistent with the ICD-10 codes

Pharmacoepidemiology and Drug Safety, 2009; 18: 730–736DOI: 10.1002/pds

732 k. haynes et al.

listed in the registry data.5 We elected to study allcancers as well as four specific cancers – pancreas,colorectal, lung, and lymphoma.We defined ‘‘all cancers’’ in two ways: (1) all cancer

(ICD-10: C00-C97) and (2) all cancer excluding non-melanoma skin cancer (NMSC) (ICD-10: C00-C97excluding C44). Because the cancer registry data notesthe potential for incomplete recording of NMSC in theregistry data5 we considered all cancer excluding NMSCto be the primary definition.

Statistical analysis

Cancer incidence rates were calculated for age- andsex-specific strata, categorizing age into 5-year agegroups through age 80, and then 85 years and older.Using the age- and sex-specific rates from UK cancerregistry data for each study year,5 we calculated thestandardized incidence ratios (SIRs) in THIN, alongwith Poisson 95% confidence intervals, using indirectstandardization methods. We extended the analysis toinclude 2007 using registry data from 2006.We conducted a sensitivity analysis adding in codes

consistent with history of cancer or treatment forcancer (radiologic or chemotherapy administration).An additional sensitivity analysis added in codes forneoplasm of uncertain significance. To address con-cerns over practice list inflation with patients no longerseeking care at the practice,6 a sensitivity analysiscensored follow-up 2 years after the patient’s lastrecorded entry in the therapy or medical file. Becausespecific cancers could be listed with a non-site-specificmorphology code (e.g., adenocarcinoma), for pancrea-tic, colorectal, and lung cancer we also searchedpreviously depersonalized free text comments forevidence of a specific cancer site for non-specificcancer diagnosis codes. Of importance, more than 80%

0.5

11.

5

SIR

1990 1995 2000 2005 2010Year

(a) (b

Figure 1. (a) Standardized incidence ratios (SIRs) for any cancer other than NM1992–2007


of the free text comments coded with non-specificcancer diagnostic codes have not been depersonalized,and as such were not available.We then used direct standardization, using the European

population standards, to calculate age-standardizedincidence rates for all cancers combined excludingNMSC using our primary definition for each practice ineach year and modeled these rates with calendar yearand the following practice specific variables: Visionsoftware utilization, practice size, and constituentcountry (England, Wales, Scotland, Northern Ireland).Given that each practice contributed an adjusted ratefor each year from 1992 to 2007 and to explore factorssuch as Vision experience on reporting; a mixed effectsmodel was utilized to analyze this form of longitudinaldata.7,8 Becausewewere interested in a limited numberof variables, all variables were included in the finalmixed effects model.All statistical analyses were performed using STATA

10.0. The study protocol was approved by the Universityof Pennsylvania’s Institutional Review Board.

RESULTS

A total of 3 714 021 subjects met the inclusion criteriaat some point during the follow-up period for theanalysis of all cancer excluding NMSC. The averagefollow-up time was 7.8 years. There were 147 851subjects with a cancer diagnosis, of which 127 701subjects had a cancer diagnosis excludingNMSCbetweenJanuary 1, 1992 and December 31, 2007 that met inclu-sion criteria. Over the study period there were 3048,12 790, 13 986, and 5669 subjects with pancreatic,colorectal, lung, and lymphoma cancer, respectively.The SIRs for the all cancer analysis were close to

unity for all study years with some minor fluctuationsin the late 1990s (Figure 1a). Inclusion of NMSC

0.5

11.

5

SIR

1990 1995 2000 2005 2010Year

)

SC in THIN 1992–2007. (b) SIRs for any cancer including NMSC in THIN



resulted in slightly higher SIRs (Figure 1b). For both ofthese definitions, SIRs for the all cancer analyses werewithin 10% of unity for the years 2001–2007, with theexception of 2004 for the analysis including NMSC(SIR¼ 1.12).In a sensitivity analysis, adding in subjects with

codes for neoplastic conditions of uncertain signifi-cance and adding in subjects with codes for neoplasticconditions of uncertain significance and codes indi-cating cancer treatment administration the SIRs weregreater than unity for most years and were higher thanin the primary analysis (Table 1). When patients werecensored 2 years after the final medical encounter orprescription record the results were not appreciablyaltered (Table 1).In the three solid organ cancers studied, the observed

recording rates in THIN were lower than expectedbased on the cancer registry data, although the SIRsapproached unity in later years, particularly after 2004(Figure 2). In contrast, the SIRs for lymphoma includedunity for all years studied (Figure 2d). Analysis withand without the available depersonalized free text didnot result in differences in the SIRs for colorectal,pancreas, or lung cancer (data not shown).In mixed effect modeling, year, Vision software

experience, and constituent country emerged as signi-ficantly associated with the reported adjusted incidencerates for all cancers. Practice size did not significantlyimpact reported rates. In univariable analysis, each oneunit increase in year since 1992, resulted in a 0.00007increase in the adjusted incidence rate (p< 0.001).When examining year and Vision software experience

Table 1. Sensitivity analyses

Year All cancerexcluding NMSC

All cancerexcluding NMSC

and codes for neoplasmof uncertain significance

1992 0.91 (0.88, 0.94) 0.98 (0.95, 1.01)1993 0.94 (0.91, 0.96) 0.99 (0.96, 1.02)1994 0.87 (0.84, 0.90) 0.93 (0.90, 0.95)1995 0.88 (0.86, 0.91) 0.94 (0.91, 0.96)1996 0.85 (0.83, 0.88) 0.91 (0.88, 0.93)1997 0.83 (0.80, 0.85) 0.87 (0.85, 0.89)1998 0.83 (0.81, 0.85) 0.87 (0.84, 0.89)1999 0.83 (0.81, 0.85) 0.87 (0.85, 0.89)2000 0.88 (0.86, 0.90) 0.92 (0.90, 0.94)2001 0.95 (0.93, 0.97) 0.99 (0.98, 1.01)2002 0.98 (0.96, 1.00) 1.03 (1.01, 1.05)2003 1.01 (0.99, 1.03) 1.05 (1.03, 1.07)2004 1.08 (1.06, 1.10) 1.12 (1.10, 1.14)2005 1.04 (1.02, 1.06) 1.08 (1.06, 1.10)2006 1.04 (1.02, 1.06) 1.08 (1.06, 1.10)2007 1.05 (1.03, 1.07) 1.08 (1.06, 1.10)


in a single model, both factors significantly affectedreporting rates. However, Vision software experienceappeared to explain some of the variability in reportingrates initially attributed to year (year 0.00003,p¼ 0.001; Vision experience 0.00007, p< 0.001). Inthe fully adjusted model, year retained statisticalsignificance, with each year resulting in an increase of0.00006 (p< 0.001) (Table 2).With respect to Vision software experience, each 1-

year increase in Vision software experience within apractice resulted in an increase in the adjustedincidence rate of all cancers of 0.0001 (p< 0.001)(Figure 3). In comparison to no experience, 1 year ofexperience with Vision software was associated with asmall insignificant increase in reported incidence ratesfor all cancers (0.000116, p¼ 0.10) while 2 or moreyears of experience was associated with significantlyhigher reported rates (0.0006, p< 0.001). In the fullyadjusted model, the strength of the association with 2or more years of Vision experience was less strong(0.0001, p¼ 0.04).Constituent country was also associated with

incidence rates for all cancers combined. If the practiceis located in Scotland, the rates were 0.0007 higher andif the practice is located in Wales the rates were 0.0004higher than if the practice is located in England.

DISCUSSION

This study found the recording of all cancer in theTHIN database is consistent with that reported incancer registries for the incidence of all cancers

All cancer excludingNMSC and codesfor neoplasm of

uncertain significanceand codes for history

of and treatment for cancer

All cancer excludingNMSC, patients 2 yearsafter the final medical

encounter orprescription record censored

1.01 (0.98, 1.04) 0.92 (0.90, 0.95)1.01 (0.98, 1.04) 0.95 (0.92, 0.98)0.95 (0.93, 0.98) 0.89 (0.86, 0.91)0.96 (0.94, 0.99) 0.9 (0.87, 0.92)0.93 (0.91, 0.96) 0.87 (0.84, 0.89)0.91 (0.88, 0.93) 0.84 (0.81, 0.86)0.90 (0.88, 0.92) 0.84 (0.82, 0.86)0.91 (0.89, 0.93) 0.84 (0.82, 0.86)0.96 (0.94, 0.98) 0.89 (0.87, 0.91)1.04 (1.02, 1.06) 0.96 (0.94, 0.98)1.07 (1.05, 1.09) 0.99 (0.97, 1.01)1.1 (1.08, 1.12) 1.02 (1.00, 1.04)1.19 (1.17, 1.21) 1.08 (1.06, 1.10)1.16 (1.14, 1.18) 1.05 (1.03, 1.07)1.22 (1.20, 1.25) 1.04 (1.02, 1.06)1.15 (1.13, 1.17) 1.06 (1.04, 1.08)


0.5

11.

5

SIR

1990 1995 2000 2005 2010Year

0.5

11.

5S

IR

1990 1995 2000 2005 2010Year

0.5

11.

5

SIR

1990 1995 2000 2005 2010Year

0.5

11.

5S

IR

1990 1995 2000 2005 2010Year

(a) (b)

(d)(c)

Figure 2. (a) SIRs for pancreatic cancer in THIN 1992–2007. (b) SIRs for colorectal cancer in THIN 1992–2007. (c) SIRs for lung cancer in THIN 1992–2007.(d) SIRs for lymphoma cancer in THIN 1992–2007


combined. This was true as well for recording oflymphoma. However for pancreatic, colorectal andlung cancer, the reporting was lower than that reportedin cancer registries. Of note, recording of these cancers

Table 2. Results of mixed effects models examining the association ofpractice factors and calendar year with reported rates of any cancer

Univariate analysis(beta coefficient, p-value)

Fully adjusted analysis(beta coefficient, p-value)

Calendar year 0.000071, <0.001 0.000057, <0.001Vision experience0 year Reference Reference1 year 0.000116, 0.10 �0.000117, 0.122 or more years 0.000589, <0.001 0.000146, 0.04

CountryEngland Reference ReferenceIreland 0.000196, 0.28 0.000149, 0.41Scotland 0.000694, <0.001 0.000713, <0.001Wales 0.000367, 0.03 0.000291, 0.08

Practice size<5000 Reference Reference5000–10 000 �0.000065, 0.37 �0.000016, 0.82>100 000 �0.000015, 0.89 0.000019, 0.85


approached the expected rates in the later years, withSIRs greater than 0.8. Our regression models demon-strated a significant increase in recording of cancerincidence over time and with a practice’s duration ofexperience in using Vision software. While constituentcountry was also associated with recorded incidencerates, practice size was not.Recorded incidence rates of all cancers and

lymphoma are consistent with the generalizability ofTHIN to the UK population. However, an importantquestion raised by these results is why the recording ofthe three solid cancers was lower than that expected bycancer registry data and why these appeared to increaseover time. There are several potential explanations forthe increased cancer recording rates over time. First,while some of the practices in THIN have alsoparticipated in the General Practice Research Data-base1, for others participation in THIN is their firsttime contributing data to a research database. Thesepractices have likely never received recording instruc-tions with research in mind prior to joining THIN;


0.5

11.

5S

IR

0 5 10 15Years of Vision Experience

Figure 3. Unadjusted SIRs for all cancer by years of Vision softwareexperience


recruitment to THIN started in 2003. Additionally, acancer quality improvement measure was instituted inthe UK in 2003 by the National Health Service, whichis consistent with the notable improvement following2003 in our study. Importantly, THIN providesfeedback to the participating practices regarding theirperformance on UK quality metrics.We hypothesized that experience using electronic

medical records and specifically Vision software mayhave an impact on the recording. We did in fact find anincrease in reporting with years of experience withVision software. From Figure 3 and the adjusted multi-variable mixed effects mode, 2–3 years of experienceresulted in more complete recording of cancer incidencerates. Of note, nearly 89% of the THIN practiceswere using electronic medical records of some form by1997 and half were using Vision software by 2000.Thus, while we hypothesize that enrollment in THINand receipt of training in data entry may havecontributed to the observed secular trends in cancerrecording, increased experience using the Visionsoftware may also have contributed to these trends.Although we did not identify many specific diag-

noses of colorectal, pancreas, or lung cancer bysearching the free text, the free text field may still be animportant source of data to explain the difference in ourlymphoma and all cancer results compared to the otherthree site-specific cancer results. We hypothesize thissince the non-specific cancer codes, such as adeno-carcinoma, are more likely to be used for solid cancersthan for lymphoma. Most of the free text that we hopedto search had not yet been depersonalized for researchuse. As more of these data are depersonalized for use inresearch, it will be interesting to retest this hypothesis.


Our study has several limitations. We only studiedfour selected cancers in addition to all cancers.We selectedindividual cancers of importance to pharmacoepide-miology research and that are of major public healthimportance, each being common in men and womenalike. Lung and pancreatic cancer are highly lethal, asis advanced stage colorectal cancer. Colorectal cancerhas also been a focus of prevention and early detectioninterventions with screening, thereby creating thepossibility that the pattern of recording of this cancerwould differ from lung and pancreas cancers. However,this was not observed. There is a possibility that diffe-rent results would be observed with other individualcancers, but the consistent results across these threesolid cancers suggest that this is less likely.Another way to assess data quality is through review

of consultant notes, pathology reports, and hospitaldischarge summaries to assess the positive predictivevalue of a recorded diagnosis. We did not conduct sucha validation for the cancer diagnosis codes. Thus, somemisclassification is possible which would likely resultin over estimate of recording rates. Of course, failure torecord any diagnosis or recording of an unrelateddiagnosis rather than cancer would lead to an under-estimate of recording rates. Additionally, because weused a constant set of diagnosis codes, changes in ratesover time should not be biased by our choice of studycodes. Finally, it should be recognized that incidencerates reported in cancer registries may not be perfectlyaccurate due to missed cancer cases or misclassifieddiagnoses. Nonetheless, the registry data are generallyaccepted to be relatively close to the true rate.We used cancer registry data for England and Wales

as our reference rates. In the regression models, theconstituent country of origin of the practice was asso-ciated with the cancer rates. Although there are no majordifferences in cancer rates between the four constituentcountries represented in THIN, the small differencesamong constituent countries observed in the regressionmodels are entirely consistent with true differences incancer rates between these regions.9 These differences,with Scotland having the highest cancer rates, would alsoresult in the observed SIRs for all THIN practicescombined being very slight over-estimates.We used acceptable record flags and AMR as data

quality indicators in this study. Future research shouldfocus on developing additional quality improvementindicators. Completeness of cancer recording is one ofseveral possible quality indicators. Although incidencerates of specific types of cancer are too low to allow forreliable estimates, a standardized incidence rate or SIRfor all cancers might be a useful measure to examinecompleteness of recording within a practice.


KEY POINTS

� Cancer rates in THIN are comparable to those reportedin cancer registry data for the years 2001–2007.

� Increased Vision software experience may contributeto more complete recording of cancer diagnoses.


CONCLUSION

The recording of all-cancer excluding NMSC inci-dence rate in THIN appears to be relatively completeover time, particularly since 2001. Recording ofspecific diagnoses of solid cancers appears lesscomplete, but this may be due to the use of non-specific cancer codes. The specific recording of thesesolid cancers has increased over time and withincreasing expertise with Vision software and is nowclose to that expected based on cancer registry data.This suggests that investigators wishing to use THIN toestimate cancer incidence rates in specific populationsshould limit studies to more recent years and topractices with at least 2 years of Vision softwareexperience, should consider chart validation of thediagnosis codes, and should perform similar calcu-lations of SIRs for a cancer in question for the overallTHIN population to assess the completeness ofrecording.

CONFLICTS OF INTEREST

None of the other authors have any other potentialconflicts of interest to report.


ACKNOWLEDGEMENTS

This study was supported in part by an Agency for Health-care Research and Quality (AHRQ) Centers for Educationand Research on Therapeutics cooperative agreement (grant#HS10399), a Clinical and Translational Science Award(grant #UL1-RR024134), an unrestricted educational grantfrom Merck (K. H.) and a grant from Cegedim EPIC inLondon, United Kingdom. The authors designed and imple-mented the study. Cegedim EPIC in London, United King-dom was allowed to provide non-binding comments on thestudy design and manuscript. The final opinions expressedare those of the authors. Dr. Lewis has served as an unpaidmember of an advisory board for THIN.

REFERENCES

1. Lewis JD, Schinnar R, Bilker WB, Wang X, Strom BL. Validationstudies of The Health Improvement Network (THIN) database forpharmacoepidemiology research. Pharmacoepidemiol Drug Saf 2007;16(4): 393–401.

2. Hall GC. Validation of death and suicide recording on the THIN UKprimary care database. Pharmacoepidemiol Drug Saf 2009; 18(2): 120–131.

3. Maguire A, Blak BT, Thompson M. The importance of defining periodsof complete mortality reporting for research using automated data fromprimary care. Pharmacoepidemiol Drug Saf 2009; 18(1): 76–83.

4. Lewis JD, Bilker WB, Weinstein RB, Strom BL. The relationshipbetween time since registration and measured incidence rates in theGeneral Practice Research Database. Pharmacoepidemiol Drug Saf2005; 14(7): 443–451.

5. Great Britain. Office for National Statistics. Cancer statistics regis-trations: registrations of cancer diagnosed in 1992, England and WalesStationery Office, 1998. Downloaded on 01 January 2009. http://www.statistics.gov.uk/StatBase/Product.asp?vlnk=8843.

6. Ashworth M, Jenkins M, Burgess K, et al. Which general practices havehigher list inflation? An exploratory study.Fam Pract 2005; 22(5): 529–531.

7. Begg MD, Parides MK. Separation of individual-level and cluster-levelcovariate effects in regression analysis of correlated data. Stat Med 2003;22(16): 2591–2602.

8. Fitzmaurice GM, Laird NM,Ware JH. Applied Longitudinal Analysis. WileySeries in Probability and Statistics. Wiley-Interscience: Hoboken, NJ, 2004.

9. Westlake S, Cooper N. Cancer incidence and mortality: trends in theUnited Kingdom and constituent countries, 1993 to 2004. Health Stat Q2008; 38: 33–46.


cancer incidence in the health improvement network

Documents