Effect of the Lookback Period’s Length Used to Identify Incident Acute Myocardial Infarction on the Observed Trends on Incidence Rates and Survival
Cardiovascular Disease in Norway Project
Background—In studies using patient administrative data, the identification of the first (incident) acute myocardial infarction (AMI) in an individual is based on retrospectively excluding previous hospitalizations for the same condition during a fixed time period (lookback period [LP]). Our aim was to investigate whether the length of the LP used to identify the first AMI had an effect on trends in AMI incidence and subsequent survival in a nationwide study.
Methods and Results—All AMI events during 1994 to 2009 were retrieved from the Cardiovascular Disease in Norway project. Incident AMIs during 2004 to 2009 were identified using LPs of 10, 8, 7, 5, and 3 years. For each LP, we calculated time trends in incident AMI and subsequent 28-day and 1-year mortality rates. Results obtained from analyses using the LP of 10 years were compared with those obtained using shorter LPs. In men, AMI incidence rates declined by 4.2% during 2004 to 2009 (incidence rate ratio, 0.958; 95% confidence interval, 0.935–0.982). The use of other LPs produced similar results, not significantly different from the LP of 10 years. In women, AMI incidence rates declined by 7.3% (incidence rate ratio, 0.927; 95% confidence interval, 0.901–0.955) when an LP of 10 years was used. The decline was statistically significantly smaller for the LP of 5 years (6.2% versus 7.3%; P=0.02) and 3 years (5.9% versus 7.3%; P=0.03). The choice of LP did not influence trends in 28-day and 1-year mortality rates.
Conclusions—The length of LP may influence the observed time trends in incident AMIs. This effect is more evident in older women.
WHAT IS KNOWN
Applying a lookback period (LP) to identify incident (first) events is a commonly used method in trend analyses using registry-based data.
Although different LPs are used in different studies, the potential effect of the LP length on the observed trends has not been studied previously.
WHAT THE STUDY ADDS
The length of the LP can influence observed trends of disease occurrence.
In the case of acute myocardial infarction, LPs between 7 and 10 years yield similar results, whereas shorter LPs may yield significantly different results.
Acute myocardial infarction (AMI) is often used as an end point in medical research. The distinction between incident (first) and recurrent events is important as they can differ in the pathogenesis,1,2 prognosis,3,4 and time trends.5
Information on an individual’s lifelong medical history is the best way to identify the incident event. However, such information is not available in register-based studies using data from hospital patient administration systems. To overcome this, a simple method is often used. Any time an AMI is identified in the data set, a retrospective search for previous AMI hospitalizations for the same individual is performed within the same register data. If no previous hospitalizations for AMI are found, the identified AMI is defined as incident. The time period used to check for previous hospitalizations is called lookback period (LP). To avoid bias in time trends, it is important to use a fixed LP for all individuals in a study.6
If an LP of n years is selected, the first n observation years will serve as a wash-out period and, hence, cannot be included in the analyses. Thus, the choice of the LP length is a tradeoff between the level of accuracy in identifying the incident event and the number of observation years included in the data set. Long LPs identify more accurately incident AMIs compared with short LPs but with the cost of excluding valuable years of information on disease occurrence to search for previous hospitalizations. Short LPs on the other hand carry the risk of misclassifying a recurrent event as incident. Various studies have used different LPs, typically 3 to 10 years.5,7–13
The effect of different LPs used to identify incident AMIs on time trends in incident AMI rates or short- and long-term mortality rates after an incident AMI has not been previously explored. Thus, the purpose of this study was to investigate whether identifying incident events using different LPs would have an influence on such trends.
Cardiovascular Disease in Norway Project
Cardiovascular Disease in Norway project (https://cvdnor.b.uib.no/) is a research project established in collaboration between the University of Bergen and the Norwegian Knowledge Center for the Health Services.14 Information on all hospital stays with International Classification of Disease (ICD)-9 codes 390 to 459 or ICD-10 codes I00-I99 was retrieved from the electronic patient administrative systems from all somatic hospitals in Norway during 1994 to 2009. Information includes patient’s age and sex, date of admission and discharge, main and secondary (≤20) discharge diagnoses, and information on diagnostic and treatment procedures performed during each hospitalization. In addition, information on date and cause of death for all Norwegian residents who died during the same time period was retrieved from the Norwegian Cause of Death Registry.
For the purpose of this study, we identified all AMI events in individuals ≥25 years during 1994 to 2009, either hospitalizations with an AMI as a discharge diagnosis (ICD-9 code 410 or ICD-10 codes I21 and I22) or out-of-hospital deaths (ICD-9 codes 410–414 or ICD-10 codes I20–I25).13
Definition of the Incident AMI Using Different LPs
Using an LP of 10 years, we defined an AMI event as incident if not preceded by an AMI hospitalization (for the same individual) during the previous 10 years. Similarly, we defined AMI events as incident by using LPs of 8, 7, 5, and 3 years. This resulted in 5 different sets (cohorts) of incident AMIs (including time periods 2004–2009 for the LP of 10 years; 2002–2009 for the LP of 8 years; 2001–2009 for the LP of 7 years; 1999–2009 for the LP of 5 years; and 1997–2009 for the LP of 3 years). The time period 2004 to 2009 was included in all sets (cohorts) and was therefore used to compare the effect of different LPs on incidence and mortality trends.
Trends in Incident AMI Rates During 2004 to 2009
Changes in age-adjusted AMI rates were calculated from Poisson regression analysis with the count of incident AMIs as the outcome, calendar year as a continuous independent variable, and the logarithm of stratum-specific population as the offset. Based on the regression coefficient (b), we derived the incidence rate ratio (IRR) for year 2009 versus 2004 (5-year change) as IRR=exp(5b).
Trends in 28-Day and 1-Year Mortality Rates After an Incident AMI
Patients hospitalized for an incident AMI were followed up to 1 year from the hospitalization date. The survival status (alive or dead) was assessed at day 28 (for the analyses of trends in 28-day mortality) and at day 365 (for analyses of trends in 1-year mortality) linking data from the Cause of Death Registry with hospitalization data using the personal identification number, unique to each Norwegian resident. In analyses of 1-year mortality, only patients who survived the first 28 days were included.
To obtain a standard score matrix to be used in postestimation comparison tests, we applied parametric survival-time regression models when modeling 28-day and 1-year mortality instead of Cox regression. Because the risk of dying after an AMI is higher in the beginning of follow-up and decreases with time, we chose to use the Weibull distribution.15
For each LP, we calculated separately the average annual change in 28-day and 1-year mortality rates after an incident AMI by including the calendar year of the incident AMI as a continuous variable in the regression model. Models were adjusted for age, educational level, comorbidities (diabetes mellitus, hypertension, chronic pulmonary obstructive disease, and heart failure), and indicators of severity (atrioventricular blocks, cardiogenic shock, and ventricular fibrillation) reported as being present during the incident AMI hospitalization.
On the basis of the regression coefficients obtained from the models (b), we again calculated changes in mortality rates for year 2009 versus 2004 (5-year change) as hazard ratio [HR]=exp(5b).
The obtained HRs were similar to those obtained when applying Cox regression.
Comparing Results Obtained From Analyses Using Different LPs
We calculated the IRRs and HRs for each of the first AMI cohorts as defined by applying different LPs (from 10 to 3 years). The estimates obtained from the analyses using an LP of 10 years were considered as gold standard because of the high accuracy in identifying incident AMIs. The regression coefficients obtained using the gold standard were compared with regression coefficients obtained from analyses using each of the 4 LPs (8, 7, 5, and 3 years) by applying postestimation tests based on seemingly unrelated estimation (suest command in STATA 13). This method combines the parameter estimates and the corresponding covariance matrices for 2 separate models into 1 parameter vector and 1 simultaneous covariance matrix of the sandwich/robust-type, which is appropriate even if the estimates are obtained on the exact same or overlapping data.16 A P value of <0.05 indicates that there is a statistically significant difference between the regression coefficients obtained from a certain LP and the results obtained from the analyses using an LP of 10 years.
Analyses were conducted for men and women separately overall and by age group (25–64 years and 65+ years), using STATA 13. The study protocol was approved by the Regional Committee for Medical and Health Research Ethics, Health Region West.
Influence of Different LPs on the Number of Incident AMIs
Using an LP of 10 years, we identified 54 487 men and 37 741 women experiencing an incident AMI during 2004 to 2009. Compared with this, the application of LPs of 8, 7, 5, and 3 years resulted in an overestimation of the number of incident events by 2.5%, 4.0%, 7.2%, and 11.8%, respectively, in men and 1.9%, 3.0%, 5.7%, and 10.1%, respectively in women (Table 1). An overestimation in the number of incident AMIs when using LPs of 8, 7, 5, and 3 years when compared with an LP of 10 years was observed in both younger men (1.6%, 2.6%, 4.9%, and 7.9%, respectively) and women (0.9%, 1.3%, 2.7%, and 4.3%, respectively) and in older men (3.0%, 4.7%, 8.6%, and 14.0%, respectively) and women (2.1%, 3.2%, 6.2%, and 11.0%, respectively).
Influence of Different LPs on Trends in Incident AMI Rates During 2004 to 2009
In men, we observed a decline in incident AMI rates of 4.2% (IRR, 0.958; 95% confidence interval [CI], 0.935–0.982) from 2004 to 2009 when using an LP of 10 years (Figure 1; Table I in the Data Supplement). Using other LPs resulted in slightly different (but not statistically significant) reductions in incident AMI rates.
In women, incident AMI rates declined by 7.3% (IRR, 0.927; 95% CI, 0.901–0.955) from 2004 to 2009 when using an LP of 10 years. When using an LP of 5 years, the reduction in rates during the same time period was 6.2% (IRR, 0.938; 95% CI, 0.911–0.965), statistically significantly lower compared with the result obtained from the LP of 10 years (P for comparison=0.02). Applying an LP of 3 years yielded a reduction of 5.7% (IRR, 0.943; 95% CI, 0.917–0.970) in AMI incidence rates from 2004 to 2009. Again this was statistically different from the result obtained by applying an LP of 10 years (P for comparison=0.01; Figure 1; Table I in the Data Supplement).
Age Group–Specific Analyses
We found no significant differences in trends of incident AMI rates among younger men and women, or among older men, when using the different LPs to define incident AMIs. Older women experienced a decline of 8.6% (IRR, 0.914; 95% CI, 0.885–0.943) in incident AMI rates from 2004 to 2009 when using an LP of 10 years. Compared with this result, the decline was significantly attenuated to 7.5% (IRR, 0.925; 95% CI, 0.897–0.954) when using an LPs of 5 years and to 7.0% (IRR, 0.930; 95% CI, 0.903–0.959) when using an LP of 3 years (both P for comparisons=0.01; Figure 1; Table I in the Data Supplement).
Influence of Different LPs on 28-Day and 1-Year Crude Mortality Rates After an Incident AMI
Twenty-Eight–Day Mortality Rates
Gender-specific 28-day and 1-year crude death rates after an incident AMI are shown in Table 2 for each LP for the total study population and by age group. Using an LP of 10 years, we identified 44 932 men and 29 085 women hospitalized for an incident AMI during 2004 to 2009. Of these, 12.9% of men and 18.5% of women died within the first 28 days. The use of LPs of 8, 7, 5, and 3 years resulted in 28-day mortality rates of 13.0%, 13.0%, 13.1%, and 13.2% in men and 18.5%, 18.5%, 18.5%, and 18.6% in women. Age group–specific analyses showed no differences in 28-day mortality rates when an LP of 10 years were compared with other LPs in either men or women (Table 2).
One-Year Mortality Rates
Among hospitalized patients with AMI who survived the first 28 days, 4111 (10.5%) men and 4029 (17.0%) women died within the first year of hospitalization when an LP of 10 years was used. This proportion was different from results obtained when using an LP of 3 years (10.5% versus 11.2%; P<0.001 for men and 17.0% versus 17.7%; P<0.001 for women; Table 2).
In age group–specific analyses, the length of LP did not have an effect on the proportion of younger men and women dying after 28 days and ≤1 year after an incident AMI. In older men, 1-year mortality rates using an LP of 10 years was lower compared with results obtained from analyses using an LP of 3 years (17.0% versus 17.8%; P<0.001), whereas no differences in the mortality rates between different LPs were observed for older women (Table 2).
Influence of Different LPs on Trends in 28-Day and 1-Year Mortality Rates After an Incident AMI Hospitalization During 2004 to 2009
During 2004 to 2009, 28-day mortality rates after an incident AMI declined overall by 23.5% (HR, 0.765; 95% CI, 0.709–0.825) in men and 17.6% (HR, 0.824, 95% CI; 0.762–0.891) in women when an LP of 10 years was used. The decline was similar when other LPs were used (Figure 2; Table II in the Data Supplement).
During the same time period, 28-day mortality rates declined by 43.4% (HR, 0.566; 95% CI, 0.450–0.712) in younger men, 31.5% (HR, 0.685; 95% CI, 0.475–0.987) in younger women, by 21.2% (HR, 0.788; 95% CI, 0.727–0.854) in older men, and by 17.4% (HR, 0.826; 95% CI, 0.763–0.895) in older women when an LP of 10 years was used. Similar results were observed for other LPs (Figure 2; Table II in the Data Supplement).
When an LP of 10 years was used, overall 1-year mortality rates did not change significantly in either men (HR, 0.931; 95% CI, 0.851–1.017) or women (HR, 0.927; 95% CI, 0.847–1.013). The results were similar for analyses using other LPs (Figure 3; Table III in the Data Supplement).
In age group–specific analyses, 1-year mortality rates declined only in younger women (HR, 0.564; 95% CI; 0.358–0.888) but not in younger men, or in older men and women. Similar to the overall analyses, the choice of LP did not affect the results (Figure 3; Table III in the Data Supplement).
In this study, we focused on the potential effect of the LP length on number and time trends of AMI incidence rates and survival.
The application of an LP of 3 years overestimated the number (and hence rates) of incident AMIs by 11.8% and 10.1% (in men and women, respectively) compared with the LP of 10 years. This overestimation was greater in older compared with younger patients.
We also found that the length of the LP has an effect on the observed trends in AMI incidence rates, especially among older women. Although LPs between 7 and 10 years produce similar results, shorter LPs underestimate the magnitude of changes in AMI rates among older adults.
However, we observed no differences in the trends of survival (either short or long-term) after an incident AMI when different LPs were used to identify the incident AMI. This observation applied to overall and age group–specific analyses for both men and women.
Application of the LP Concept in Previous Studies
The LP (alternatively named retrospective search) technique has been used within administrative data to capture comorbidities in an individual hospitalized for a certain medical condition. In this context, studies have tried to identify an optimal length of the LP to search for comorbidities to improve the predictive performance of comorbidity scores17–19 and to correctly assess the prevalence of those comorbidities in the various populations.20
In epidemiological studies based on patient administrative data, retrospective searches are used to exclude previous AMI (or coronary heart disease) hospitalizations before defining an event as incident.5,7–11,13 Such studies have used different LPs, but the potential effect of differences in LPs’ length on time trends of disease occurrence (either incidence or survival) has not been previously evaluated. The only study addressing this issue was conducted in Denmark and found that AMI incident rates changed by 11% to 13% in men and 10% to 11% in women during 1987 to 1992 when different LPs were applied.6 However, in the analyses, the incident event was defined using an LP of 3 years and no age group–specific analyses were conducted. On the basis of data presented in the Danish study,6 we have calculated that compared with an LP of 10 years, the use of an LP of 3 years may overestimate AMI incident rates by 13.5% in men and 8.2% in women.
We could not identify any study focusing on the potential effect of the LP length on trends in survival after an incident AMI.
Determinants of the LP Length
The most important factor determining the length of the LP period is the number of years with available data. If the data come from newly established registries, there is no possibility to apply long LPs. When data allow choosing the LP length, other important factors should be considered. Some of these factors include characteristics and timing of the disease occurrence, case fatality rates, age of the study population, and time trends in survival and recurrence rates.
Distribution in Time of Disease Occurrence
In diseases with a high fatality rate, short LPs could be safely used to identify the incident event. In the case of chronic diseases, such as ischemic heart disease, patients often survive the incident AMI. Therefore, short LPs cannot capture previous hospitalizations occurring long time ago. This could lead to misclassification of some late recurrences as incident events. Thus, short LPs overestimate the number of incident AMIs (and rates) compared with longer LPs. In such cases, longer LPs should be preferred. Previous studies have shown that the risk of recurrences is higher the first year after the incident event.21,22 Thereafter, such risk declines, but remains present during the first 5 to 7 years.21,22
Age of the Study Population
Age is a strong predictor of morbidity. The younger the patient, the lower the probability that he/she has had a previous manifestation of the disease. Therefore, the length of the LP has a minor effect among younger adults. In elderly patients, the likelihood of having had a previous event is higher and therefore the length of the LP might have a greater role in identifying previous events.
Trends in Survival After the First Event
Many studies9,11,23–26 (including a study from our group27) have reported improvements in short- and long-term survival after the first AMI. The widespread and effective use of revascularization procedures and other treatment modalities, as well as secondary prevention strategies, has led to an increased number of incident AMI survivors, exposing them to more years at risk of experiencing a recurrent event. Therefore, the effect of defining the first event by using different LPs is greater now compared with decades ago.
A recent study focusing on sex differences in outcomes among patients having an AMI found that the gap between women and men with regard to long-term survival is narrowing.28 Women, who decades ago had higher mortality rates, are now experiencing more pronounced improvements in survival compared with men. Such pronounced improvement in survival exposes more women to the risk for recurrences compared with decades ago. As a consequence, the proportion of recurrences being misclassified as incident events is expected to rise among women when short LPs are applied. Because the decline in AMI incidence rates is more pronounced than the decline in recurrence rates,5,13,29 the increasing proportion of recurrences being misclassified as incident events among women (present with the use of short LPs) may attenuate the decline in true AMI incident cases (as defined by an LP of 10 years). This phenomenon can be more evident in studies including long-time periods as such differences in trends caused by using different LPs could be larger compared with what we found in a short-time period (2004–2009).
These considerations may help explain the sex differences in the effect of the LP length on AMI incidence trends observed in this study.
Strength and Limitations
A major strength of this study is the inclusion of all AMI events occurring in Norway, both those hospitalized and those who died outside the hospital in both men and women without any age restriction. This allowed us to report on overall as well as sex and age group–specific analyses free of selection bias. Data available in our project (covering 16 calendar years) allowed us to make comparisons between long LPs (offering a high accuracy in identifying the incident event) and shorter ones.
Our study carries some limitations. One is the inability to distinguish between ST-segment–elevation MI and non–ST-segment–elevation MI as such information is not available in our register.
Another limitation of our study is the consideration of the LP of 10 years as gold standard for identifying incident events. However, better alternatives for correctly identifying previous events would include either self-reported information or lifelong medical history of patients, both unavailable in studies using data from patient administrative systems covering a specific time period.
Our study is based on the homogenous white population of Norway and covers a relatively short-time period. We do not know whether our findings are generalizable to other countries with different ethnic structure or in another time period. Therefore, other studies are needed to validate our findings, preferably studying populations experiencing different time trends than ours and covering different time periods.
LPs of 7 to 10 years are reliable for identifying the incident AMI in a given individual. Shorter LPs may overestimate the true number (and hence rates) of incident events by creating a mixture of incident and recurrent cases and might also distort the real trends in the incidence, especially among elderly women. In the case of trends in survival after an incident AMI, the choice of the LP length does not seem to influence the results.
We thank Tomislav Dimoski at The Norwegian Knowledge Centre for the Health Services, Oslo, Norway for his contribution by developing the software necessary for obtaining data from Norwegian hospitals, conducting the data collection, and quality assurance of data in this project.
The Data Supplement is available at http://circoutcomes.ahajournals.org/lookup/suppl/doi:10.1161/CIRCOUTCOMES.114.001703/-/DC1.
- Received October 15, 2014.
- Accepted May 11, 2015.
- © 2015 American Heart Association, Inc.
- Yusuf S,
- Hawken S,
- Ounpuu S,
- Dans T,
- Avezum A,
- Lanas F,
- McQueen M,
- Budaj A,
- Pais P,
- Varigos J,
- Lisheng L
- Chan WC,
- Wright C,
- Tobias M,
- Mann S,
- Jackson R.
- Yang D,
- Dzayee DA,
- Beiki O,
- de Faire U,
- Alfredsson L,
- Moradi T.
- Schmidt M,
- Jacobsen JB,
- Lash TL,
- Bøtker HE,
- Sørensen HT.
- Nedkoff LJ,
- Briffa TG,
- Preen DB,
- Sanfilippo FM,
- Hung J,
- Ridout SC,
- Knuiman M,
- Hobbs M.
- Koopman C,
- Bots ML,
- van Oeffelen AA,
- van Dis I,
- Verschuren WM,
- Engelfriet PM,
- Capewell S,
- Vaartjes I.
- Sulo G,
- Igland J,
- Nygård O,
- Vollset SE,
- Ebbing M,
- Tell GS.
- Sulo G,
- Igland J,
- Vollset SE,
- Nygård O,
- Øyen N,
- Tell GS.
- Cleves M,
- Gould W,
- Gutierres R,
- Marchenko Y.
- Weesie J.
- Smolina K,
- Wright FL,
- Rayner M,
- Goldacre MJ.
- Gulliksson M,
- Wedel H,
- Köster M,
- Svärdsudd K.
- Kostis WJ,
- Deng Y,
- Pantazopoulos JS,
- Moreyra AE,
- Kostis JB
- Roger VL,
- Weston SA,
- Gerber Y,
- Killian JM,
- Dunlay SM,
- Jaffe AS,
- Bell MR,
- Kors J,
- Yawn BP,
- Jacobsen SJ.
- Hardoon SL,
- Whincup PH,
- Petersen I,
- Capewell S,
- Morris RW.
- Sulo E,
- Vollset SE,
- Nygård O,
- Sulo G,
- Igland J,
- Egeland GM,
- Ebbing M,
- Tell GS.
- Bucholz EM,
- Butala NM,
- Rathore SS,
- Dreyer RP,
- Lansky AJ,
- Krumholz HM.
- Sulo G,
- Vollset SE,
- Nygård O,
- Igland J,
- Egeland GM,
- Ebbing M,
- Tell GS.