Ethics statement
The RAND Corporation Institutional Review Board approved this analysis. The requirement for informed consent was waived because this was a retrospective analysis of existing health care data in which the researchers did not have access to identifiable patient information that would have allowed patients to be contacted.
Study design
This was a retrospective difference in differences analysis of de-identified records from e-prescribing adopter and non-adopter cohorts before and after FDS implementation.
Setting
In late 2004, Horizon Blue Cross Blue Shield of New Jersey (BCBSNJ) led an initiative to offer subsidized iScribe standalone electronic prescribing (e-prescribing) software to high volume prescribers. In a prior publication, we described levels of e-prescribing use among 297 primary care physicians (PCPs) who participated in this initiative by adopting iScribe during 2005[12]. They were compared with 1892 PCPs who were also offered the e-prescribing system during this time period, but did not adopt it. We found that solo practitioners, pediatricians, and physicians with more patients from predominantly African American zip codes were less likely to adopt e-prescribing. In the current study, we compare the pharmaceutical claims (claims) of these PCPs’ assigned primary care patients before and after implementation of FDS.
Isolating and classifying pharmaceutical claims
We obtained a dataset containing all claims for medications dispensed between June 3, 2003 and July 21, 2006 and submitted to Horizon BCBSNJ for the assigned primary care patients of the 2189 PCPs. Because there are other ways of increasing brand to generic switches (e.g. state laws that require generic medications be dispensed when available), we selected two medication classes without generic medications available during the time period considered. We thus isolated all claims for angiotensin receptor blocker (ARB) and inhaled steroid (IS) medications (Figure 1). For the latter class, we recognize that adherence calculated from claims is often lower than for pills. Nonetheless, prior studies show that proportion of days covered (PDC) can be reliably measured within this lower range[13–16].
Then, we excluded claims with less than 15 days supplied to eliminate trial starts and a small number of claims that appeared erroneous. We further restricted the dataset to first, new claims (‘index’ claims) because the decision to select a given medication within a class is considered most when a medication is started. This restriction was accomplished by excluding any claims preceded by another same-class (ARB or IS) claim during the prior six months. Finally, for our adherence analysis, we also excluded index claims from 2006. Since all of the patients were continuously enrolled through June 30, 2006, this ensured that we had six further months of claims to calculate adherence subsequent to each index claim.
Different pharmaceutical benefit plans, and their effect on patient copayments
Among studied patients, pharmaceutical benefit plan coverage was heterogeneous. The most common benefit plan used a three tiered formulary with ascending copayments for generic, preferred, and non-preferred brand medications. However, some patients had two tiered plans with identical copayments for all branded medications. Others had percentage coinsurance requirements that did not differ by brand status. A small proportion of patients were required to pay all costs at the pharmacy and later submit for reimbursement, in which case copayments could not be deduced from claims. Finally, even among patients with similar plan structures, there were differences in terms of actual copayment amounts, coinsurance percentages, deductibles, out-of-pocket maximums, flex spend plans, and “gap” insurance that would affect patients’ actual out-of-pocket costs. As with nearly every study using pharmaceutical claims as a data source, we did not have access to all of this cost information, but we nonetheless used the copayment amounts listed in the claims to model the overall relationship between copayment and tier. For example, if a patient’s plan required 10% coinsurance for a $200 claim, the patient responsibility on the claim would show as $20, so we would use $20 as the ‘copayment’ for that claim.
Intervention – initial non-interruptive FDS changed to combined non-interruptive and interruptive FDS
For the aforementioned e-prescribing initiative, Caremark began activating participating physicians’ e-prescribing software in January of 2005, and continued to do so on a rolling basis throughout 2005. We isolated the claims of patients attributed to the 297 PCPs studied in our prior manuscript, each of whom activated their iScribe e-prescribing software during calendar year 2005. For each study PCP, Caremark provided us with an e-prescribing activation date.During the study period, the software initially used only non-interruptive FDS, but later added interruptive FDS. Thus, e-prescribers were initially only exposed to non-interruptive FDS, which consisted of automatic display of the medication tier at the time of e-prescribing (Figure 2). Beginning September 16, 2005, the software exposed e-prescribers to both non-interruptive and interruptive FDS, which included the following: For the ARB medication class, physicians selecting candesartan, eprosartan, losartan, olmesartan, or telmisartan were advised to “Consider preferred brands Avapro, Diovan”. Atacand was added on February 17, 2006 after it also became a preferred brand. For the IS medication class, physicians prescribing inhaled beclomethasone, flunisolide, or triamcinolone were advised to “Consider preferred brands Flovent HFA, Pulmicort Turbuhaler”. Thus, for the purposes of our analysis, each study claim was linked to the appropriate PCP e-prescribing adoption date and subsequently classified as belonging to one of these three time periods: Pre-FDS, e-prescribing with non-interruptive FDS only, and e-prescribing with (both) interruptive and non-interruptive FDS.
Because there was no date of e-prescribing activation for non-participating PCPs, we assigned each non-participating PCP a ‘synthetic’ activationdate by random sampling with replacement from study PCPs’ actual activation dates. This assignment was done to make the distribution of activationdates similar in both groups, such that our analyses would be robust to secular trends. We then used these dates to separate control PCPs’ claims into pre and post-FDS claims. Because interruptive FDS was added on a specific date, we used this date to further classify post-FDS claims into the period of non-interruptive FDS only versus the period of both interruptive and non-interruptive FDS.
We could not definitively determine whether individual claims had been electronically prescribed, but we were able to associate each claim with a PCP’s level of e-prescribing usage. Because of our prior finding that levels of e-prescribing usage varied greatly but were generally stable, we classified e-prescribing users as high versus low users based on whether they used it more or less than 30% of the time. Based on our previous analysis of usage levels, this represented the 70th percentile, and the mean usage rate among this subgroup was 61% (61 e-prescriptions per 100 claims)[12].
Outcome variables – medication tier, patient copayment, and adherence
Medication tier was found in Horizon BCBSNJ formularies. Claims contained patient copayment data. Because cost variables often contain a skewed distribution with many outlying data points[17, 18], a Box-Cox transformation was used to determine the best way of transforming the patient copayment variable to minimize root mean square error. Adherence was quantified using the proportion of days covered (PDC): after a patient filled a new medication, the PDC was the percentage of the subsequent 180 days when any medication within the index class (ARB or IS) was available to them, based on the days of medication supplied according to claims data[19].
Covariates and intervention variables
Caremark provided physician specialty and practice size information. Horizon BCBSNJ provided de-identified demographics for each of the PCPs’ assigned primary care patients. As described in prior work, we used patients’ zip codes to estimate their household income, race (black vs white) and ethnicity (Hispanic vs non-Hispanic)[12]. Dosing frequency was calculated from claims data.
Data analysis
We first compared characteristics of the three groups of studied PCPs and their patients, including across the three time periods studied. We made bivariate comparisons between FDS use and medication tier, between tier and patient copayments, between patient copayments and adherence, and finally directly between FDS use and adherence. We then constructed four multiple regression models to control for possible confounders.
Because our prior work found that usage of the FDS intervention varied widely[12], the FDS:tier model includes the interaction between the extent of FDS usage and the type of FDS present. We used claims from non-participating PCPs in corresponding times periods (before e-prescribing activation, after activation of e-prescribing with non-interruptive FDS, and after the addition of interruptive FDS) to control for secular trends. A difference in differences approach was used to compare the temporal differences within like PCPs across groups of PCPs. The estimated effects in this model were obtained from generalized estimating equations (GEEs) with a logit link function.
The tier:copayment model used multiple linear regression, and assumed that insurers only consider tier and medication class in determining copayments. These covariates were therefore the only ones tested. Linear mixed effects models were used to examine copayment:adherence and FDS:adherence associations. In the copayment:adherence and FDS: adherence regression models, a one-dimensional random effect was used to control for clustering of patients within PCPs.
Because these three underlying models required irreconcilable specification differences, the final model that directly analyzed the relationship between FDS and adherence was not just an identical, overarching model, but rather a separate analysis. The regression models were generally constructed by beginning with all available and theoretically tenable predictor variables included, and then using a backward variable selection procedure to eliminate covariates determined not to be associated. A p-value threshold of 0.05 and model fit criteria were jointly used to make this determination. Model fit was assessed using the quasi-likelihood information criterion for the FDS:Tier model and the Akaike's information criterion for the copayment:adherence model. All analyses were performed using SAS, release 9.2 (SAS Institute, Inc; Cary, NC).
After developing these three models, we used the FDS:tier and tier:copayment model estimates to project the effect of FDS on patient copayments. We also used the tier:copayment and copayment:adherence model estimates to project the effect of tier on adherence, and we combined all three model estimates to project the effect of FDS on adherence. Finally, because the FDS:tier model was the most important new knowledge generated in our analyses, and because there is extensive prior evidence regarding tier:copayment and copayment:adherence relationships, we combined our FDS:tier model estimates with this prior evidence. Specifically, we used annual survey results from the Kaiser Family Foundation that included copayments for different medication tiers to summarize existing knowledge of tier:copayment relationships, and we use a landmark meta-analysis of cost-sharing studies to understand copayment:adherence relationships[20, 21]. In doing so, we generated FDS:copayment and FDS:adherence illustrative projections that were independent of the tier:copayment and copayment:adherence relationships we found in the studied setting. Adherence projections assumed a baseline PDC of 60%.