Policy‐induced selection bias in pharmacoepidemiology: The example of coverage for Alzheimer's medications in British Columbia

Abstract Purposes To assess the impact of a government‐sponsored reimbursement policy for cholinesterase inhibitors (ChEIs) on trends in physician visits with a diagnosis of Alzheimer's disease (AD). Methods Longitudinal population‐based study using interrupted time series methods. British Columbia outpatient claims data for individuals aged 65 and older were used to compute monthly AD visit rates and examine the impact of the ChEI reimbursement policy on the coding of AD. We examined trends in the number of patients with AD visits, the number of AD visits per patient, and visits with “competing” diagnoses (mental, neurological, and cerebrovascular disorders and accidental falls). Finally, we described demographic and clinical features of diagnosed patients. Results We analyzed 1.9 million AD visits. Faster growth in recorded AD visits was observed after the policy was implemented, from monthly growth of 7.5 visits per 100 000 person‐months before the policy (95% confidence interval [CI], 6.1‐8.9) to monthly growth of 16.5 per 100 000 person‐months after the policy (95% CI, 14.8‐18.3). After the implementation of the policy, we observed increased growth in the number of patients with recorded AD visits and the number of AD visits per patient, as well as a shift in diagnoses away from mental diseases and accidental falls to AD (diagnosis substitution). Conclusions British Columbia's reimbursement policy for ChEIs was associated with a significant acceleration in Alzheimer's visits. Evaluations of health services utilization and clinical outcomes following drug policy changes need to consider policy‐induced influences on the reliability of the data used in the analysis.


| INTRODUCTION
Administrative health claims data are generated after encounters with the health care system and are collected for administrative or billing purposes. 1 These data have been used in a growing number of pharmacoepidemiology and health policy studies. 2,3 One element of claims data is diagnosis codes. Diagnostic information is typically recorded in claims for outpatient physician visits and is frequently used to define study populations and identify study outcomes. Inaccuracies in diagnostic coding have been discussed as an important source of bias 3,4 ; however, such inaccuracies were generally assumed to be constant over the study period, in spite of evidence that this might not be the case. Multiple situations have been shown to influence the utilization of specific diagnostic codes, including a change in codes or the coding system, 5,6 the introduction of new diagnostic criteria, 7,8 increased awareness by physicians and the public for a specific disease, 9 and the substitution of codes for one disease with codes for another (diagnostic shift). [10][11][12][13][14][15][16][17] In the Canadian Province of British Columbia (BC), we identified a unique opportunity to examine the susceptibility of diagnostic coding behavior to a change in drug reimbursement policy. In October 2007, the provincial drug plan began covering the cholinesterase inhibitor medications (ChEIs: donepezil, galantamine, and rivastigmine) for patients with Alzheimer's disease (AD) as part of the Alzheimer's Drug Therapy Initiative (ADTI). The policy was implemented as part of an initiative for "coverage with evidence development," [18][19][20] and its details are presented elsewhere. 21,22 We sought to assess the effect of the ADTI reimbursement policy (aka "the policy") on health services utilization and cost in AD patients. We undertook this analysis as a preliminary measure to understand the possible influence of the policy on diagnostic information captured in the database, prior to using the database to evaluate the policy.

| Study design and data source
We conducted a longitudinal population-based study using interrupted time series analysis methods. We obtained administrative claims data from the BC Ministry of Health for the period 1 January 2001 to 31 December 2013. The anonymized data included records of fee-for-service payments to physicians and alternative providers, patient registration information and demographics, pharmacy records (PharmaNet), and hospital discharge records.

| Physician visits for AD
Our study assessed the impact of the ADTI policy on diagnostic coding for AD in outpatient visits ("Alzheimer's visits"). Alzheimer's visits were defined as a physician fee-for-service visits with an International Classification Disease, version 9 (ICD-9) code of 331, 290, 294, or 797 in any of the diagnostic fields. This definition was the only published AD definition validated on administrative data when this study was conducted, 23 and its sensitivity and specificity were 86%. 24,25 Individuals under the age of 65 years were excluded because dementia is rare and often secondary to other diseases in that age group. 26,27 Crude and standardized Alzheimer's visit rates were computed on a monthly basis. Crude visit rates were calculated as the numbers of Alzheimer's visits per 100 000 person-months of enrollment in the provincial medical plan. Directly standardized visit rates were computed to correct for variations in age and sex over time, using the 2007 British Columbia enrolled population as the reference population. Standardized visit rates were also corrected to a month-length of 30 days.

| Statistical methods
Interrupted time series analysis is considered the strongest, quasiexperimental design to evaluate the longitudinal effects of an intervention, eg, health policy, particularly when the researcher does not manage the intervention. [28][29][30] Applying this methodology, we included the following variables in our regression model: time in months, time after the ADTI policy, and a dichotomous variable for baseline level. implementation of the policy. This dichotomous variable tests for a change in AD visits between the months immediately before and immediately after the policy, accounting for the prepolicy trend. 31 In addition, we adjusted for autocorrelation and seasonality by including lag terms for up to 12 preceding months based on statistical significance (stepwise autoregression using the Yule-Walker method, SAS BACKSTEP selection option). Finally, to allow for a delayed effect of the policy, we excluded data from the first three months after the pol-

| ChEI reimbursement policy and additional visits/patients' parameters
We further tested for the effect of the ChEI reimbursement policy on several monthly parameters for individuals aged 65 and older using interrupted time series analysis. Visit ratio was defined as the number of physician visits with an Alzheimer's diagnosis divided by 100 000 total physician visits. Visit density was defined as the number of physician visits with an Alzheimer's diagnosis per 100 individuals with such visits. Alzheimer's patients were the number of patients with a physician visit with an Alzheimer's diagnosis per 100 000 person-months of enrollment. Lastly, Alzheimer's administrative incidence was defined as the first physician visit or hospital discharge with Alzheimer's diagnosis in patients with at least 18 months of continuous enrollment and no ChEI prescription during this period. Incidence was calculated per 100 000 person-months of enrollment.

| ChEI reimbursement policy and diagnosis substitution
We considered that the ChEI reimbursement policy might have been associated with diagnosis substitution, ie, the substitution of codes for one diagnosis by codes for another disease. Specifically, for patients with the same clinical presentation, we considered whether physicians increased the use of AD coding after the policy over other codes that had been previously used. We tested for diagnosis substitution from four "competing" disease categories that may be relevant: mental disorders (ICD-9 codes 290-319), neurological disorders (ICD-9 codes 320-359 or 430-438), cerebrovascular disorders (ICD-9 codes 430-438), and accidental falls (including orthopedic trauma, ICD-9 codes E880-E888, 800-849). In the absence of accepted methodology to identify diagnosis substitution in epidemiology studies, we a priori defined diagnosis substitution as a combination of two criteria: the first is a decrease in the coding for the "competing" diseases, and the second is an increase in the ratio of Alzheimer's visits to the "competing" diseases. The analysis included a series of two interrupted time series analyses for each "competing" disease categories. For the first criterion, we analyzed visit rates with "competing" disease diagnoses and required significantly fewer visits with "competing" diseases after the new policy, ie, smaller slope or lower baseline level after policy initiation. For the second criterion, we analyzed the visit rate ratio: the product of rates of physician visits with an Alzheimer's diagnosis divided by the rates of visits with a diagnosis of a "competing" disease.
For this criterion, we required a larger slope or higher baseline level after policy initiation. in the standardized rates of visits was 6.2% (5.0%) and 8.7% (2.7%) in the periods before and following the ChEI reimbursement policy.
In the interrupted times series analysis, we observed a significant association of the ChEI reimbursement policy with trends of physician visits with an Alzheimer's diagnosis ( Figure 1, Table 1). Before the policy was implemented, the monthly growth in standardized Alzheimer's visits (the slope) was +7.5 visits per 100 000 person-months (95% confidence interval [CI], 6.1-8.9). After the policy was implemented, the slope increased by 9.0 (95% CI, 6.6-11.5), to a monthly growth of +16.5 (95% CI, 14.8-18.3) without a significant change in the baseline level of visits.  (Table 1 and Figure 3). We also compared Alzheimer's administrative incidence before and after the policy. We found that the new ChEI reimbursement policy was associated with a level increase of +19.5 (95% CI, 10.9-28.2) incident cases per month.

| Additional analysis
It was also associated with a change in direction of the slope, from a decreasing trend of −0.8 (95% CI, −0.9 to −0.6) incidence cases per 100 000 person-months before the policy, to a constant trend of  Alzheimer's visits, physician visits with Alzheimer's diagnosis codes; Visit ratio, the number of physician visits with an Alzheimer's diagnosis divided by 100 000 total physician visits; Visit density, the number of physician visits with an Alzheimer's diagnosis per 100 individuals with such visit; Alzheimer's patients, the number of patients with a physician visit with an Alzheimer's diagnosis code. Alzheimer's administrative incidence is based on the first physician visit or hospital discharge with an Alzheimer's diagnosis in patients with at least 18 months of continuous enrollment and no ChEI prescription during this period. Results are presented as estimated regression parameters (95% confidence interval). † Per 100 000 patient-months; *Significant at the .05 probability level.  (Table 1 and Figure 3).
Finally, we checked whether physicians increased their use of Alzheimer's diagnosis coding over coding of other "competing" diseases after the new policy, ie, diagnosis substitution (Table 2 and Figure 4). On the basis of the two predefined criteria, we detected diagnosis substitution from mental disorders and accidental falls (Table 2). In both categories, the new policy was associated with a significant decrease in the baseline level of visits with the "competing" diseases and an increase in the baseline ratio of Alzheimer's visits to visits with "competing" diseases ratio.

| DISCUSSION
In this study, we tested for the effect of a new drug reimbursement  Our results are different from a previous BC study that examined the effect of drug coverage policy on visits. 39 This published study estimated no effect of drug cost sharing on trends of visits with depression.
The main difference between the studies is the direction of effect on copayment; in our study, it was lower after the policy, and in the previous study, it was higher. In addition, the previous study examined a more general policy that included a few medication groups. Trends in absolute numbers of visits with AD or dementia diagnosis codes have only been studied in a single study. 40,41 The researchers of that study This study has several strengths. The Canadian health care system is based on the principles of fairness and equity, comprehensiveness, accessibility, and universality; hence, it is well suited to study the effects of new policies. We analyzed data from a population-based databases in which data were collected prospectively in a systematic manner, and examined data had been collected over a long period of  Alzheimer's visits, we find the validation secondary to the main purpose of the study. Regardless of the accuracy of the AD codes, we demonstrated a change in trend related to the new drug reimbursement policy.

| CONCLUSIONS
The observed increase in the number of physician visits with an AD diagnosis after the implementation of a government-sponsored reimbursement policy for ChEI could present a challenge when studying other aspects of the new drug coverage policy. Policy-induced influences on the selection of a study population could bias assessment of health services utilization and clinical outcomes in before-after designs even when they include historical or concurrent control groups. We encourage researchers to critically evaluate the accuracy of diagnostic coding and trends and consider describing the effect of the policy on the cohort studied as part of the policy effect.

ETHICS STATEMENT
The study protocol was approved by the Clinical Research Ethics Board of the University of British Columbia (H09-01696) and the Human Research Ethics Board of the University of Victoria (08-164).