An Algorithm to Identify Generic Drugs in the FDA Adverse Event Reporting System

Abstract

Introduction

Although generic drugs constitute approximately 88% of drugs prescribed in the US, there are no reliable methods to identify generic drugs in the US FDA Adverse Event Reporting System (FAERS).

Objective

The aim of this study was to develop an algorithm for identifying generic drugs in the FAERS.

Data Source

We used 1237 adverse event reports for tamsulosin, levothyroxine, and amphetamine/dextroamphetamine from the publicly available FAERS from 2011–2013, and 277 source case narratives obtained from the FDA.

Methods

Two reviewers independently and in duplicate used a three-item algorithm including the following criteria: manufacturer name, New Drug Application (NDA) number/abbreviated NDA (ANDA), and specific use of the term ‘generic’ or ‘brand’ to classify the focal drug of each case report as definitely generic (two of three criteria), probably generic (one of three criteria), brand, and cannot be assessed. Inter-rater reliability was estimated using kappa coefficients, and internal consistency was estimated using Cronbach’s alpha. We compared the classification of the drugs as generic versus non-generic in publicly available FAERS compared with the original case reports (reference).

Results

The focal drug was classified as generic (definite or probable) in 15.8% (39/234), 9% (67/742), and 16.7% (42/261) of tamsulosin, levothyroxine and amphetamine/dextroamphetamine cases, respectively (overall kappa 0.89, 95% confidence interval 0.85–0.93), while 37% of reports could not be classified due to incomplete information. Among the drugs classified as generics using the publicly available FAERS, we categorized 95.3% as generic drugs using the original case reports. Among those drugs that did not meet the algorithm-based definition of generic in the publicly available data, 20.9% were reclassified as generics using the original case reports.

Conclusions

The algorithm demonstrated high inter-rater reliability with moderate internal consistency for identifying generic drugs in the FAERS, in our sample. Future efforts should focus on improving the reliability and validity of identifying generics through improving the completeness of reporting in the FAERS.

1. Introduction

Utilization of generic drugs in the US has increased substantially since the 1980s, with approximately 88% of all prescriptions dispensed in the US in 2015 filled using generic drugs [1]. Despite widespread use, many patients still have concerns about generic drugs. A systematic review of stakeholder perceptions of generic drugs reported a lack of confidence in generic drugs, especially a belief that cheaper equals inferior, which was a unifying theme across studies [2]. In a survey assessing patient attitudes towards prescription drugs, 42% and 29% of participants reported that generics ‘are not the real thing’ and brand-name medications are safer than generics, respectively [3]. A national survey of commercially insured adults in the US reported that while 56% of respondents agreed that patients should use more generics, only 37.6% preferred to take generics themselves [4]. Perceptions that generics are less potent, require higher doses, have more side effects, and are for minor but not serious illnesses have also been reported [5]. Such patient-relevant concerns regarding generic drugs have been a long-standing issue for patients, physicians, and policy makers.

In the last 5 years, awareness and research regarding patient-relevant and patient-reported concerns regarding the safety of drugs has increased [6, 7]. The US FDA’s well-established voluntary reporting system (MedWatch) constitutes an unexplored resource to investigate patient-relevant concerns with generics. This system collects data on suspected adverse events due to drugs, devices, and medication errors reported by prescribes, consumers, and other healthcare professionals [8]. Several studies regarding the safety of drugs have been conducted using these data, providing useful signals [9, 10]; however, a sensitive and specific method of identification of generics in the FDA Adverse Event Reporting System (FAERS) has yet to be established. We aimed to develop an algorithm that will facilitate identifying generic drugs in the FAERS that will be useful for later investigations of adverse events attributable to generic products.

2. Materials and Methods

2.1. Data Resource

The FAERS is a database maintained to aid postmarketing drug safety surveillance programs, and contains adverse event and medication error reports. These reports can be voluntary, as submitted by physicians, healthcare professionals and consumers, or mandatory by regulation, as submitted by manufacturers [8]. The FAERS contains over 9 million reports on adverse events and medication errors submitted to the FDA. Its structure adheres to the international safety reporting guidance issued by the International Conference on Harmonisation, ICH E2B [11].

Information from individual case reports, as submitted to the FDA by patients, healthcare providers or manufacturers, is extracted into the FAERS data files available for public download from the FDA website. The publicly available data are included in seven data tables: patient demographics, drug or biologic reported, adverse events, patient outcomes, report sources, drug therapy start and end dates, and indications for use/diagnosis for the reported drugs. The drugs are tagged as primary suspect, secondary suspect, concomitant, or interacting drugs. Drugs names are recorded by their validated trade name, i.e. the standardized name based on the National Library of Medicine RxNorm ingredient names, or verbatim as submitted in the case report. Adverse events and medication errors are coded using Medical Dictionary for Regulatory Activities (MedDRA®) terminology [12]. Details about the source of the report and occupation of the reporter are also available in the dataset.

We accessed the publicly available FDA Adverse Drug Reaction data files for 12 quarters from 2011 to 2013. We selected three representative drugs for testing the developed algorithm, i.e. tamsulosin, levothyroxine and amphetamine/dextroamphetamine, as these were either drugs with concerns about generic substitution or had periods of brand-name-only availability followed by periods when both brand names and generic drugs were concurrently available.

2.2. Selection Criteria for Adverse Event Reports

We limited our evaluation to reports from the US, and included case reports that met a minimum measure of completeness: age and sex of the patient, name of the suspected drugs, and the adverse event. We limited the case reports to those describing events with one of the three selected drugs as the primary suspect. A single case report may exist in multiple versions in the FAERS since the reports are updated as more information is obtained on follow-up. These reports have the same ‘case id/case’ identifying number but unique ‘isr/primaryid’ number. We selected the latest report as it usually has the most complete information, and removed duplicates by identifying reports that had the same event date, age, sex, and adverse event, regardless of their case identifying number. Furthermore, we created a list of all brand-name drugs and their generic equivalents, along with their manufacturers and respective New Drug Application (NDA)/abbreviated NDA (ANDA) numbers from the FDA Orange Book [13], Truven Red Book [14], and National Drug Code Directory [15]. This list was used to categorize whether the focal drug of interest was generic or brand. Our unit of analysis was at the individual case report, which could be associated with multiple adverse events.

2.3. Grading Algorithm

We used a combination of the following three criteria to classify the focal drug as either a brand version or generic equivalent (Fig. 1).

Fig. 1.

Grading algorithm to identify whether the drug in the FAERS is generic or brand. FAERS FDA Adverse Event Reporting System, NDA New Drug Application, ANDA Abbreviated New Drug Application

1. Name of the manufacturer: The name of the manufacturer in the report was matched against the reference list of manufacturer names for drugs available in the market to identify the focal drug as generic or brand.

2. NDA or ANDA number: The NDA number for brands, and ANDA number for generics, was used to identify whether the focal drug was a generic or brand, using the reference list.

3. Name of the drug: The information from this variable was used if the name of the drug, as translated verbatim from the original narrative into the publicly available datasets, contained specific terms such as ‘generic’ or ‘brand’, e.g. ‘generic Synthroid’.

We used the above criterion and categorized the focal drug of each case report as either definite generic, probable generic, or brand. If the focal drug satisfied at least two of the three criteria specific to generics, for example generic manufacturer and explicit mention of the word ‘generic’ in the drug name, it was classified as definitely generic, and those that met one of the three criteria were classified as probable generic. Drugs were classified as ‘brand’ if they satisfied criteria for brand-name drugs as described above. We classified drugs with missing or insufficient information on all three criteria as cannot be assessed.

Two investigators (GI and SPM) assessed each report independently and classified the report into the categories above. A third investigator (SS) adjudicated any discrepancies. We pilot tested the grading algorithm prior to conducting the classification.

2.4. Reliability

We measured the inter-rater reliability to assess concordance of the two investigators using the Cohen’s kappa (κ) for the two raters [16]. A κ statistic of 0.01–0.2 was rated as slight, 0.2–0.4 as fair, 0.4–0.6 as moderate, 0.6–0.8 as excellent, and 0.8–1 as almost perfect. We deemed a moderate level of reliability as acceptable. Internal consistency of the grading algorithm was measured using Cronbach’s alpha [16].

2.5. Validity

There is no reliable ‘gold standard’ approach for identifying generic drugs in the FAERS. We considered the original case reports with narratives as an acceptable reference [17]. The original case reports are submitted by reporters to the FDA and contain both structured and unstructured text. The narrative text description includes information about the development of the adverse event, including details about the suspected drug, relevant diagnostic tests, and concomitant medications. Via the Freedom of Information Act, we requested the source case narratives for a sample of the case reports. We obtained the source narratives of all generic drugs, classified as definite and probable, and a random sample of those that had insufficient information and brand-name drugs, from the FDA.

Two independently trained investigators (GI and SPM), blinded to the status of the records, extracted data from the source narratives using a similar grading scheme as the reliability study. We evaluated the text of the narratives, especially the development of adverse events, and paid special attention to any language indicating whether the drug was generic or brand, such as “patient was taking generic Synthroid”, any mention of switching from brand to the generic version, name of the manufacturer, etc. A third expert (SS), blinded to the results of the earlier evaluation, adjudicated in case of discordance between the two investigators. Among the sample of cases for which we had the original case reports, we compared the classification of the drugs as generics versus non-generics in the public FAERS versus those in the original case reports.

2.6. Sample Size

The number of subjects required, in a two-rater study, to detect a statistically significant κ (p ≤ 0.05) on a dichotomous variable, with 80% power at two proportions of positive diagnoses (30 and 50%), assuming the null hypothesis value of κ to be 0.4 (below 0.4 is unacceptable), was 762 and 660, respectively. We assumed that half of the reports would be positive and planned to examine at least 660 reports [18].

3. Results

Between January 2011 and December 2013, approximately 2 million unique adverse events were reported in the publicly available FAERS data (Fig. 2). Of these, 654,424 adverse events were reported from the US and had complete information regarding the age and sex of the patient, suspected drug, and MedDRA preferred term for the event [12]. After limiting the adverse event reports to those with levothyroxine, amphetamine/dextroamphetamine, and tamsulosin as primary suspects, a total of 1237 case reports, comprising 234 reports for tamsulosin, 742 for levothyroxine, and 261 for amphetamine/dextroamphetamine, were included in the analysis. The demographic characteristics and occupation of the reporters are described in Table 1. We observed that information on our three variables of interest was missing in 25–71% of case reports.

Fig. 2.

Selection process of the case reports (N = 1237). FAERS FDA Adverse Event Reporting System, AEs adverse events, CRs case reports

Table 1.

Descriptive characteristics of the selected case reports in the FDA Adverse Event Reporting System from 2011 to 2013

	Tamsulosin [N = 234 reports]	Stimulantsa [N = 261 reports]	Levothyroxine [N = 742 reports]
Age, years [median (range)]	71 (2 months–101 years)	27 (8 months–77 years)	57 (1–98 years)
Males (%)	94.4	48.7	16.2
Reporter (%)
Consumer	59.4	39.1	62.9
Physician	11.1	18.8	15.6
Pharmacist	12.8	7.3	5.1
Other healthcare professional	9	15.7	8.9
Missing	7.2	17.6	7.4

^aStimulants include amphetamine and/or dextroamphetamine

3.1. Reliability

The inter-rater reliability was consistently above 0.8 for tamsulosin, levothyroxine, amphetamine/dextroamphetamine, and overall, as shown in Tables 2, 3, 4, and 5, respectively. The Cronbach’s alpha for internal consistency for tamsulosin was 0.52, 0.47 for amphetamine/dextroamphetamine, 0.34 for levothyroxine, and 0.40 overall (Table 6).

Table 2.

Concordance of two raters in identifying generic drugs in the FDA Adverse Event Reporting System from 2011 to 2013 for tamsulosin (Kappa statistic = 0.87, 95% confidence interval 0.77–0.96)

I-1	I-2
	Generic definite	Generic probable	Brand	Not assessable	Total (I-1)
Generic definite	29	0	0	0	29
Generic probable	7	0	0	1	8
Brand	10	0	94	0	104
Not assessable	2	0	0	91	93
Total (I-2)	48	0	94	92	234

I-1 investigator 1, I-2 investigator 2

Table 3.

Concordance of two raters in identifying generic drugs in the FDA Adverse Event Reporting System from 2011 to 2013 for levothyroxine (kappa statistic = 0.92, 95% confidence interval 0.86–0.973)

I-1	I-2
	Generic definite	Generic probable	Not assessable	Brand	Total
Generic definite	54	0	0	0	54
Generic probable	9	0	1	2	12
Not assessable	0	0	479	1	480
Brand	2	0	16	178	196
Total	65	0	496	181	742

I-1 investigator 1, I-2 investigator 2

Table 4.

Concordance of two raters in identifying generic drugs in the FDA Adverse Event Reporting System from 2011 to 2013 for amphetamine/dextroamphetamine (kappa statistic = 0.79, 95% confidence interval 0.70–0.87)

I-1	I-2
	Generic definite	Generic probable	Brand	Not assessable	Total
Generic definite	19	0	8	2	29
Generic probable	6	0	4	3	13
Brand	0	0	49	0	49
Not assessable	0	0	6	164	170
Total	25	0	67	169	261

I-1 investigator 1, I-2 investigator 2

Table 5.

Concordance of two raters in identifying generic drugs in the FDA Adverse Event Reporting System from 2011 to 2013 for all three drugs (kappa statistic = 0.89, 95% confidence interval 0.85– 0.93)

I-1	I-2
	Generic definite	Generic probable	Brand	Not assessable	Total
Generic definite	102	0	8	2	112
Generic probable	22	0	5	6	33
Brand	10	0	622	1	633
Not assessable	4	0	22	433	459
Total	138	0	657	442	1237

I-1 investigator 1, I-2 investigator 2

Table 6.

Internal consistency of the grading algorithm using Cronbach’s alpha

Drug	No. of observations	Cronbach’s alpha	ITC—Item 1/Item 2	ITC—Item 2/Item 3	ITC—Item 3/Item 1
Tamsulosin	234	0.5212	0.9239	−0.0195	−0.1056
Stimulants	261	0.4711	0.8006	−0.0632	−0.0506
Levothyroxine	742	0.3362	0.5685	−0.0862	−0.049
Total	1237	0.3981	0.6685	−0.0662	−0.0603

Item 1 NDA number, Item 2 – manufacturer name, Item 3 drug name, ITC inter-item correlation, NDA New Drug Application

The inter-item correlation between manufacturer name and NDA number was acceptable for tamsulosin and amphetamine/dextroamphetamine (0.93 and 0.80, respectively), but poor for levothyroxine (0.57); however, the inter-item correlation between NDA number/manufacturer name and drug name was negative with all drugs. We observed that information from the variable drug name was useful in categorizing drugs as generic or brand in only 6 of 1237 cases, when the other two variables (NDA number and manufacturer name) were unavailable.

3.2. Proportion of Generics

After adjudication of case reports with discordant categorization, the suspected drug was classified as a generic product (definite or probable) in 15.8, 9, and 16.7% of case reports for tamsulosin, levothyroxine, and amphetamine/dextroamphetamine, respectively (Table 7). The suspected drug was classified as the brand-name drug in 44, 64.5, and 18.8% of case reports for tamsulosin, levothyroxine, and amphetamine/dextroamphetamine, respectively. Overall, 37% of the reports did not have sufficient information to be classified.

Table 7.

Classification of suspected drugs using grading algorithm after adjudication [N = 1237]

Categories	Tamsulosin [n = 234]	Stimulants [n = 261]	Levothyroxine [n = 742]	Total [n = 1237]
Generic definite	30 (12.8)	29 (11.1)	55 (7.4)	114 (9.2)
Generic probable	9 (3.8)	13 (5)	12 (1.6)	34 (2.7)
Brand	103 (44)	49 (18.8)	479 (64.6)	631 (51)
Not assessable	92 (39.3)	170 (65.1)	196 (26.4)	458 (37)
Total	234	261	742	1237

Data are expressed as n (%)

3.3. Validation

We requested the source case narratives of a sample of 277 case reports because time and resource constraints limited access to the full sample. It was not deemed pragmatic to manually redact the entire sample for identifiable information before the source case narratives could be released by the FDA. Among the 277 case reports, 148 had the focal drug categorized as generic using the algorithm, 54 were categorized as cannot be assessed, and 75 were categorized as brand. Levothyroxine was the primary suspect in 119 case reports, tamsulosin in 77 case reports, and amphetamine/dextroamphetamine in 81 case reports. The Cohen’s κ statistic for inter-rater reliability was 0.96 for tamsulosin, 0.86 for amphetamine/dextroamphetamine, and 0.87 for levothyroxine (Table 8). The overall κ statistic was 0.89 for all case reports.

Table 8.

Concordance of two raters in identifying generic drugs in the validation subset [N = 277]

Drug	No. of case reports	Kappa statistic (95% CI)
Tamsulosin	77	0.96 (0.82–1.00)
Stimulants	81	0.86 (0.71–1.00)
Levothyroxine	119	0.87 (0.74–0.99)
All	277	0.89 (0.81–0.97)

CI confidence interval

We classified 95.3% (141/148) of case reports as generics in the original case reports among those reports meeting the algorithm-based definition of generic drugs (Table 9). Of the 129 case reports that did not meet the algorithm-based definition of generic drugs (brand or unable to assess), we classified 20.9% (27/129) as generics using the original case narratives. We did not estimate other accuracy parameters because of a lack of data on the full sample.

Table 9.

Number of case reports categorized as generic from the validation sample [N = 277] using the algorithm and source narratives

Categories	Generic in source narratives	Drugs not classified as generics in source narratives (brands or unable to assess	Total
Generic in public FAERS	141	7	148
Non-generic in public FAERS	27	102	129
Total	168	109	277

FAERS FDA Adverse Event Reporting System

4. Discussion

The inter-reliability of our algorithm was excellent, with κ coefficients consistently above 0.8, but the internal consistency of the items varied across the three drugs, ranging from poor to moderate. Information on the three criterion was missing in a large proportion of case reports. Approximately one-third of cases could not be classified as generics or brand-name drugs using the algorithm in the publicly available FAERS. From our validation subsample of 54 cases that were categorized as cannot be assessed, we were able to use additional information from the case narratives to reclassify 28 of 54 case reports as either generic equivalents or brand-name drugs. Although the majority of drugs we identified as generics were also classified as generics in the original case reports (≈95.3%), approximately one-fifth of those categorized as non-generics were reclassified as generics after examination of the original case reports. Thus, while the positive predictive value of the algorithm is high, estimation of other accuracy parameters was limited by lack of access to the full sample.

There are several potential explanations for the low Cronbach’s alpha estimates observed. Cronbach’s alpha provides an estimate of the lower bounds of reliability [19]. The measurement of alpha assumes unidimensionality of the underlying construct and is based on the tau-equivalence model which stipulates that each test item measures the same latent trait. The model may have also been violated because of poor inter-relatedness between items, as demonstrated by the low inter-item correlation for the drug-name variable with the other two variables. The inter-item correlation between the manufacturer’s name and NDA/ANDA variable was acceptable (>0.7) for tamsulosin and amphetamine, but was low for levothyroxine. The NDA number and manufacturer name were often reported together in the majority of reports for tamsulosin and amphetamine; however, only manufacturer name, without a corresponding NDA/ANDA number, was reported in a substantial number of levothyroxine reports. These reports associated with levothyroxine could not be classified as brands or generics because the same manufacturer produced both generic and brand-name drugs.

We included the variable ‘drug name’ in our algorithm because the coding of the publicly available FAERS data also includes the verbatim name of the drug if any deviation from the standard name is identified; however, this variable was only useful for classifying drugs as probable generics in only 0.5% of reports. In these reports, neither the manufacturer’s name nor the NDA/ANDA number were reported. The incremental value of including this variable in any algorithm for generic drug identification appears to be marginal. Regrettably, the high popularity of some of the brand names, such as Flomax, Synthroid, etc., preclude the use of the drug names as a variable for the algorithm.

4.1. Comparisons with Other Studies

A recent study by Bohn et al. [17] evaluated 2500 case narratives of antiepieptic drugs to identify the extent and quality of product identifying information in these case narratives. Those researchers used manufacturer name as well as specific mention of ‘generic’ or ‘brand’ to describe the drug in the source narratives to classify the drug as generic or brand. Furthermore, they observed that approximately 68% of the FAERS case narratives did not have any information about the product manufacturing fields, and that the name of the manufacturer may be unreliable for identifying whether the drug was generic or brand, especially as certain manufacturers may submit the case report under their own name if they cannot identify the actual suspected product manufacturer. In contrast, we examined several classes of drugs to formally evaluate a grading algorithm in both the publicly available FAERS and validation using their source narratives. Similar to their findings, we found that, overall, 37% of reports could not be assessed, which was variable across the three drugs (65% for amphetamine/dextroamphetamine and 26% for levothyroxine).

4.2. Limitations

Our study has some limitations that largely reflect the quality of reported data. A large number of reports frequently lacked sufficient detail to allow for full evaluation. The current labels on prescription bottles do not provide sufficient information, such as the manufacturer’s name and the NDA/ANDA number, to allow consumers to distinguish between generics and brands, in cases of direct reports from consumers. In some cases, the name of the reporting manufacturer does not guarantee it is the actual product manufacturer. Studies have shown that even though branded drugs account for approximately 1–10% of the prescribed drugs, the brand manufacturers submit more than half of the reported adverse events [20]. Recently, the FDA has initiated a ‘Safety First Initiative’ that requires stricter postmarketing standards from manufacturers of generic drugs [21]. We also noted that approximately one-fifth of the drugs classified as brand name or unknown were reclassified as generics using the original reports, either because additional information in the original report was not included in the publicly available FAERS, or there was conflicting information on manufacturer and NDA/ANDA number from the case reports. A few case reports had an NDA/ANDA number that did not correspond to the manufacturer’s name.

We were only able to conduct validation on a subsample of reports because of pragmatic considerations. We requested the source case narratives of a sample of 277 case reports because time and resource constraints limited access to the full sample. Since our primary objective was measuring the overall performance of the algorithm rather than drug-specific performance, we requested a random subsample of the entire set for validation instead of a proportional set for each drug.

4.3. Strengths

Despite these limitations, there are several strengths to our study. We used a reliable method to identify generic drugs using both the publicly available FAERS and the source narratives. The rigorous data extraction process using dual and independent reviewers ensured the integrity of our results. We have outlined the feasibility and internal consistency of an algorithm to identify generic drugs in the FAERS, and demonstrated the performance of different variables in identifying generic drugs in the FAERS for more than one drug class.

4.4. Implications

There are several implications of our findings. The label of dispensed drugs should contain the name of the manufacturer and the NDA/ANDA number. These efforts to identify generic drugs in the FAERS can only be successful when there is adequate and complete reporting of adverse events. A separate field, which classifies the drug as generic or brand, should be added in future updates of the MedWatch forms. ‘Hard stops’ to ensure complete reporting when filling out web-based forms, with an override only possible in cases where data are unavailable, can be evaluated. Manufacturers should streamline their approach to reporting by including the above variables, including the NDA/ANDA number.

Further replication of this approach using other generic and brand-name drugs by other investigators is needed. Future iterations should consider factor analysis to identify the underlying constructs and dimensions. Additional reliability analyses including item–total correlation studies are needed. The performance of the algorithm may change depending on the specific drug, as well as the time since patent expiry. The algorithm may perform better when reporting is more complete during periods of stimulated reporting or shortly after generic availability. This algorithm could be potentially automated through more structured reporting in the FAERS, as well as natural language processing of data in the case narratives. However, for any of these methods to be reliable and valid, complete reporting in the FAERS is necessary.

Further development of approaches to identification of generic drugs, such as those outlined above, can inform future analysis of the FAERS. Currently performed disproportionality analyses do not usually distinguish between generic and brand-name drugs. In certain instances, where there may be differences in ADRs due to the excipients present in the generic drug, such separate analyses may be desirable; however, such analyses require complete data on the adverse event and suspected drug, while we used our algorithm on reports that also had complete data on age, sex, and country of reporting.

5. Conclusion

We have developed a novel algorithm comprising of the combination of manufacturer name and NDA/ANDA number to identify generic drugs in the FAERS. This approach is sensitive to the use of variables, the specific nature of the product, and the completeness of reporting. Future efforts should focus on improving the reliability and validity of identifying generics. Improving the completeness of reporting of generic identifiers will be needed for the FAERS to realize its full potential on proving reliable information on adverse events associated with generic drugs.

Key Points.

A three-item algorithm including variables from the publicly available US FDA Adverse Event Reporting System (FAERS) database, i.e. manufacturer name, New Drug Application (NDA) number/abbreviated NDA (ANDA), and specific use of the term ‘generic’ or ‘brand’ to identify the drug name, was used to classify the focal drug of each case report. Reliability, and validity using a sample of source case narratives of the algorithm was assessed.

The algorithm demonstrated high inter-rater reliability with moderate internal consistency for identifying generic drugs in the FAERS in our sample.

In the validation subsample, the focal drug for approximately 95% of case reports was correctly identified as generic; however, approximately one-fifth of the case reports in our validation sample that were identified as non-generic using the algorithm were reclassified as generic.

It is possible to identify generic drugs in the FAERS database, but complete information in the case reports is a prerequisite for adequate performance of the algorithm.