Agreement of Administrative Pharmacy Dispensing With Self‐Reported Use of Oral Prednisone in US Veterans With Rheumatoid Arthritis

Abstract

Objective

Administrative claims are used to evaluate oral glucocorticoid use in rheumatoid arthritis (RA), despite limited evidence to support accuracy. We aimed to evaluate the performance of claims‐based algorithms for glucocorticoid use compared to self‐report in an RA population.

Methods

Participants with RA enrolled at seven Veterans Affairs Rheumatoid Arthritis (VARA) Registry sites were asked six questions as part of clinical care assessing current prednisone use and dose, recent use, “stockpiling,” and receiving prednisone outside the Department of Veterans Affairs (VA). Algorithms using VA prescription claims operationalized current use (active prescription on date of self‐report assessment), current dose (that prescription's mean dose), and recent use (active course overlapping the prior 30 or 90 days). We assessed performance characteristics and agreement, benchmarked on self‐report.

Results

Of 284 participants, 13% reported current prednisone use and 20% reported 90‐day use. Sensitivity, specificity, positive predictive value, and negative predictive value were 0.70, 0.98, 0.84, and 0.96, respectively, for current use and 0.71, 0.92, 0.72, and 0.92, respectively, for 90‐day use. Cohen's κ was 0.68 for current use and 0.63 for 90‐day use. Among participants reporting ≤5 mg/day, agreement for dose was high (weighted κ 0.67). One in four participants reported a stockpile, and one in four reported receiving prednisone from a non‐VA provider.

Conclusion

Algorithms derived from VA claims detecting prednisone prescriptions have high validity compared to patient self‐report. The modest sensitivity of these algorithms may reflect stockpiling and non‐VA prescriptions. These findings form a basis for contextualizing real‐world studies of glucocorticoid use in RA and improve clinical estimation of glucocorticoid use not captured in claims.

INTRODUCTION

Up to 90% of patients with rheumatoid arthritis (RA) use glucocorticoids, with 25% to 50% taking them long‐term. ¹ , ² Although glucocorticoids are effective in treating RA, there is concern that their dose‐dependent toxicity outweighs their benefits. ³ Clinical trials cannot assess real‐world usage patterns and are underpowered to assess toxicity, especially at lower doses. ⁴ Observational studies using electronic health records or administrative algorithms provide the opportunity to evaluate real‐world glucocorticoid use and long‐term toxicities, as well as to identify high‐risk glucocorticoid use to inform clinical practice or quality improvement initiatives. However, algorithms that estimate glucocorticoid exposure from administrative claims may not adequately account for use that is different than prescribed or that changes over time (eg, a taper or burst). There is also no clear reference standard against which to benchmark these algorithms, and assessments of previously proposed glucocorticoid algorithms are limited and have substantial methodologic variability. ⁵ , ⁶ Self‐report is likely to be the most accurate reference standard for glucocorticoid use in RA because patient‐directed changes in use (eg, dose increases or reductions, “stockpiling”) and prescribing by providers other than the patient's rheumatologist are common. Despite this, prior studies benchmarking on self‐report are small and may not be generalizable to the majority of patients with RA. ⁶ , ⁷

SIGNIFICANCE & INNOVATIONS.

• In a cohort of participants with rheumatoid arthritis (RA) enrolled in a national Department of Veterans Affairs (VA)–based registry, algorithms derived from pharmacy claims can acceptably rule in current prednisone use and use within 30 and 90 days, compared to a reference standard of self‐report.

• In an exploratory analysis of participants reporting current prednisone use, we saw substantial agreement between self‐reported and claims‐based assessments of daily dose when limited to participants reporting a dose ≤5 mg/day, corresponding to a claims‐based mean dose threshold of ≤7.5 mg/day. We observed reduced agreement for higher doses, perhaps due to dose variability not adequately captured by mean daily dose (eg, a prescribed taper, use of stockpiled prednisone).

• Using patient self‐report as a reference standard, our algorithm underestimated prednisone dose, contextualized by our finding of high self‐reported stockpiling and prescribing by non‐VA providers.

• These findings may be used to estimate the performance of claims‐based algorithms for glucocorticoid use in research and clinical contexts and to improve clinical estimation of glucocorticoid use not captured in claims.

This study aimed to evaluate the accuracy of claims‐based algorithms for current prednisone use in a population of veterans with RA using patient‐reported use as a reference standard. Although self‐reported medication use has limitations, primarily related to recall and desirability bias, it shows good agreement with use derived from other sources (eg, medical records, prescription data), is easier to obtain than provider report, and is more sensitive than provider report in capturing use occurring without that provider's knowledge (eg, patient‐initiated changes, prescribing by another provider). Thus, demonstrating comparable performance between self‐reported and provider‐reported glucocorticoid use would help facilitate future research in this area.

PATIENTS AND METHODS

Participants

As part of clinical care, 284 participants enrolled in the multicenter prospective Veterans Affairs Rheumatoid Arthritis (VARA) Registry ⁸ were asked semistructured questions about their glucocorticoid use at the time of a routine in‐person rheumatology clinic visit between October 2023 and August 2024 (Supplemental Table 1). Data reported were obtained from the following VARA sites: Omaha Department of Veterans Affairs (VA) Medical Center, Rocky Mountain Regional VA Medical Center, VA Ann Arbor Healthcare System, VA Durham Health Care System, VA Philadelphia Healthcare System, VA Puget Sound Health Care System, and VA Salt Lake City Healthcare System. Each site obtained institutional review board approval, and participants provided informed consent to participate in the VARA Registry.

Reference standard: Self‐reported glucocorticoid use

We limited our analysis to oral prednisone‐dispensing episodes to enable accurate dose estimation by participants. In a prior evaluation of prescription data, we found that prescriptions for prednisone represented 95.9% of oral glucocorticoids prescribed to VARA‐enrolled patients (Supplemental Table 2). To assess self‐reported current prednisone use, participants were asked, “Did you take prednisone yesterday?” with response options “yes” or “no.” This served as the reference standard for current use. Participants who responded “yes” were asked to select the dose they took from the following categories: <5 mg, 5 mg, 10 mg, 15 mg, >15 mg, or a different dose. Participants were also asked about “recent use” in the past 30 and 90 days using the following categories: none, less than half the days (≤14 or ≤44 days in a 30‐ or 90‐day period, respectively), more than half the days (≥15 or ≥45 days in a 30‐ or 90‐day period, respectively), or “I took prednisone in the past 30 [or 90] days, but don't remember how often.” Two barriers to accurate claims‐based evaluation of prednisone use were also assessed: prednisone “stockpiling” and obtaining prednisone from non‐VA providers.

Administrative algorithms

We generated three administrative algorithms to evaluate current prednisone use using only prescription information captured directly in the VA Corporate Data Warehouse (CDW). These included the following: (1) an algorithm for current use, defined as an active prednisone‐dispensing episode overlapping the self‐report assessment date; (2) a 30‐day algorithm for recent use, defined as an active prednisone course overlapping the 30 days before the self‐report assessment date; and (3) a 90‐day algorithm for recent use, defined as an active prednisone course overlapping the 90 days before the self‐report assessment. ⁹

For the current use algorithm (algorithm 1), we recorded the dispensing episode start date, the number of tablets dispensed, the days’ supply of medication, and the unit dose for each prednisone‐dispensing episode. We calculated the dispensing episode end date by adding the days’ supply to the dispensing episode start date, and we calculated the cumulative prednisone dose by multiplying the number of tablets dispensed by the unit dose for the dispensing episode. To meet algorithm 1, the day before the self‐report assessment was required to fall within an active prednisone‐dispensing episode. ⁹

For the recent use algorithms (algorithms 2 and 3), each prednisone course was defined as a group of all consecutive dispensing episodes that had a ≤90‐day gap between the end date of one dispensing episode and the start date for the next dispensing episode. We defined course duration as the period from the start date of the first dispensing episode in the course through the end date of the last dispensing episode of the course (ie, the end of the last dose before a ≥90‐day gap). We calculated the cumulative prednisone dose for each course by summing the total doses for each dispensing episode in that course, and we calculated the average daily dose by dividing the cumulative dose of prednisone for a course by the course duration (course end date minus course start date). To meet algorithm 2, a patient was required to have any part of an active prednisone course overlapping the period between the self‐report assessment date and 30 days before the self‐report assessment date. To meet algorithm 3, a patient was required to have any part of an active prednisone course overlapping the period between the self‐report assessment date and 90 days before the self‐report assessment date ⁹ (Figure 1). When assessing agreement for recent use, self‐reported prednisone use during the past 30 and 90 days was dichotomized as any use versus no use.

Figure 1.

Schematic representation of claims‐based algorithms for assessing current and recent prednisone use. Blue bars represent prednisone‐dispensing episodes. Orange bars represent prednisone courses, defined as a group of all consecutive dispensing episodes with a ≤90‐day gap between the end date of one dispensing episode and the start date for the next dispensing episode. Course duration was defined as the period from the start date of the first dispensing episode in the course through the end date of the last dispensing episode of the course (ie, the end of the last dose before a ≥90‐day gap).

Performance of algorithms against the reference standard

We calculated the sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of each administrative algorithm using prednisone‐dispensing data from the VA CDW against the reference standard of self‐reported current use. Because there is no validated benchmark for assessing glucocorticoid use, we also calculated Cohen's kappa to assess agreement between administrative algorithms and self‐reported use. Agreement based on Cohen's κ was classified as near perfect (0.8–1.0), substantial (0.6–0.8), moderate (0.4–0.6), fair (0.2–0.4), or slight (0.0–0.2). ¹⁰ We used Youden's J, equivalent to sensitivity + specificity − 1, to summarize algorithm performance; it ranges from −1 to 1, with 0 being equivalent to random chance and 1 representing a perfect test. ¹¹

For participants reporting current use, we performed an exploratory analysis assessing agreement between the following: (1) self‐reported current use and an active dispensing episode in claims (Figure 2) and (2) self‐reported daily dose (<5 mg, 5 mg, 10 mg, 15 mg, >15 mg, or other) and mean observed daily dose in claims for any active dispensing episode (Figure 3). To allow for error in our claims‐based assessment of prednisone dose (eg, the fact that prescriptions intended to be taken as a burst and taper will be represented in our analysis as a single mean daily dose), we categorized claims‐based mean prednisone dose thresholds as follows: no active course, ≤4 mg, 4.1 to 7.5 mg, 7.6 to 12.5 mg, 12.6 to 17.5 mg, and ≥17.6 mg. To assess agreement, we calculated the weighted Cohen's kappa for (1) all participants reporting current use, (2) participants with both self‐reported current use and an active dispensing episode, and (3) participants with self‐reported current use ≤5 mg and an active dispensing episode. For participants reporting never use, we identified the year of their last prescription‐dispensing episode by examining all prednisone‐dispensing episodes in the CDW since October 1, 1998 (Figure 4). Analyses were completed using Stata version 18. Data are reported in accordance with the policies of the VA Office of Research and Development.

Figure 2.

Error matrix: current prednisone use. CDW, Corporate Data Warehouse; GC, glucocorticoid; NPV, negative predictive value; PPV, positive predictive value. Color figure can be viewed in the online issue, which is available at http://onlinelibrary.wiley.com/doi/10.1002/acr.25580/abstract.

Figure 3.

Daily dose from claims for participants reporting current prednisone use. Weighted κ for 26 participants with an active dispensing episode: 0.55 (95% CI 0.43–0.67, P < 0.0001). Weighted κ for 21 participants with an active dispensing episode reporting ≤5 mg/day: 0.67 (95% CI 0.49–0.85, P = 0.0001). * Two participants reported current use but did not report a verifiable dose. CDW, Corporate Data Warehouse; CI, confidence interval. Color figure can be viewed in the online issue, which is available at http://onlinelibrary.wiley.com/doi/10.1002/acr.25580/abstract.

Figure 4.

Time (years) from most recent prednisone prescription to self‐report assessment for participants who report never using prednisone but have a pharmacy claim for it (n = 21). Color figure can be viewed in the online issue, which is available at http://onlinelibrary.wiley.com/doi/10.1002/acr.25580/abstract.

RESULTS

Two hundred eighty‐four participants provided data on prednisone use. Demographics reflect those of prior studies in the VARA Registry, with a mean age of 69.2 years, 84.2% of participants being male, 83.4% being White, and 68.7% being current or former smokers (Table 1). Eighty‐seven percent of those surveyed were taking disease‐modifying antirheumatic drugs (DMARDs), with 43.7% taking biologic or targeted synthetic DMARDs. The mean Clinical Disease Activity Assessment score in this population was 9.33, representing low‐moderate RA disease activity.

Table 1.

Characteristics of patients with RA asked about glucocorticoid use^*

Variable	Participants (N = 284)
Age, mean (SD), y	69.2 (10.9)
Male, n (%)	239 (84.2)
Site, n (%)
Rocky Mountain Regional VAMC	15 (5.3)
Omaha VAMC	70 (24.6)
VA Philadelphia Healthcare System	23 (8.1)
VA Salt Lake City Healthcare System	104 (36.6)
VA Puget Sound Health Care System	53 (18.7)
Other sites a	19 (6.7)
Race, n (%)
White	237 (83.4)
Black	27 (9.5)
Other or not reported	20 (7.0)
Ethnicity, n (%)
Non‐Hispanic	267 (94.01)
Hispanic or not reported	17 (6.0)
Smoking status, n (%)
Current	46 (16.20)
Former	149 (52.46)
Never	82 (28.87)
Unknown	7 (2.46)
RA duration, mean (SD), y	16.1 (12.84)
Seropositivity, n (%)
Anti‐CCP	185 (65.14)
RF	173 (60.92)
Active MTX course at time of survey, n (%)	115 (40.49)
Active csDMARD course at time of survey, n (%)	202 (71.13)
Active btDMARD course at time of survey, n (%)	124 (43.66)
Active btDMARD or csDMARD course at time of survey, n (%)	248 (87.32)
Most recent CDAI at time of survey, mean (SD)	9.33 (9.71)

^*btDMARD, biologic or targeted synthetic disease‐modifying antirheumatic drug; CCP, cyclic citrullinated peptide; CDAI, Clinical Disease Activity Index; csDMARD, conventional synthetic disease‐modifying antirheumatic drug; MTX, methotrexate; RA, rheumatoid arthritis; RF, rheumatoid factor; VA, Department of Veterans Affairs; VAMC, Department of Veterans Affairs Medical Center.

^a VA Ann Arbor Healthcare System and VA Durham Health Care System.

Thirty‐seven participants (13%) reported current prednisone use, 52 (18%) reported use in the past 30 days, and 59 (20%) reported use in the past 90 days. Of those reporting use in the past 30 and 90 days, 23 (44%) and 29 (49%) reported use on less than half the days and 29 (56%) and 29 (49%) reported use on more than half the days, respectively.

Table 2 shows the performance of each claims‐based algorithm for current prednisone use. Algorithm 1, assessing current use, had a sensitivity of 0.70, a specificity of 0.98, a PPV of 0.84, and an NPV of 0.96. Algorithms 2 and 3, estimating use in the past 30 and 90 days, had sensitivities of 0.63 and 0.71, specificities of 0.96 and 0.92, PPVs of 0.79 and 0.72, and NPVs of 0.92 and 0.92, respectively. Substantial agreement was observed between administrative algorithms and self‐reported use. Cohen's κ and Youden's J for each algorithm were 0.68 and 0.68 for current use, 0.64 and 0.59 for use in the past 30 days, and 0.63 and 0.57 for use in the past 90 days, respectively. Percent agreement was 0.94 for current use, 0.90 for use in the past 30 days, and 0.88 for use in the past 90 days.

Table 2.

Performance characteristics for administrative algorithms vs self‐reported GC use^*

Concept	Question	CDW algorithm	Sensitivity	Specificity	PPV	NPV	κ	Youden's J
Current GC use	Did you take prednisone yesterday?	1. Active prednisone prescription overlapping day of assessment	0.70	0.98	0.84	0.96	0.68	0.68
Recent GC use	In the past 30 days, how often did you take prednisone?	2. Active prednisone course 9 overlapping the period from 30 days before self‐report assessment to date of assessment	0.63	0.96	0.79	0.92	0.64	0.59
	In the past 90 days, how often did you take prednisone?	3. Active prednisone course 9 overlapping the period from 90 days before self‐report assessment to date of assessment	0.71	0.92	0.72	0.92	0.63	0.63
Barriers to estimating GC use in claims	Do you have a “stockpile” or “reserve supply” of extra prednisone pills saved up?	–	–	–	–	–	–	–
	Have you ever gotten prednisone from a health care provider outside the VA?	–	–	–	–	–	–	–

^*Bold values indicate positive predictive value. CDW, Corporate Data Warehouse; GC, glucocorticoid; NPV, negative predictive value; PPV, positive predictive value; VA, Department of Veterans Affairs.

Figure 2 shows the error matrix for participants reporting current use compared to the claims‐based algorithm. Of 37 participants reporting current use, 26 (70%) had a concurrent overlapping prednisone prescription. Of 247 participants reporting no current prednisone use, 242 (98%) had no concurrent overlapping prednisone prescription. The most common discordance was self‐reported use without an overlapping prednisone prescription course (11 of 37 self‐reporting use).

Figure 3 shows the agreement between self‐reported prednisone dose and mean observed daily dose for the 35 participants who reported current use and provided a dose. Of these, 9 (26%) had no prednisone prescription overlapping the date of the self‐report assessment. Four (44%) of these nine reported having a prednisone stockpile, and two (22%) of these nine reported receiving a prednisone prescription from a non‐VA provider in the past year. We observed moderate agreement between self‐reported current dose and mean observed daily dose in this population, with a weighted Cohen's κ of 0.55 and agreement of 89%. Among the subset of participants with an active prednisone prescription who self‐reported a current dose ≤5 mg/day (n = 19, 51% of those reporting current use), we observed substantial agreement with a claims‐based dose of ≤7.5 mg/day, with a weighted Cohen's κ of 0.67 and agreement of 95%.

Thirty‐two participants (11%) reported that they had never taken prednisone. Of these, 21 (66%) had a previous prednisone prescription fill documented in the CDW. Figure 4 shows the year of the most recent prednisone prescription for these 21 participants. Of these, zero had a prednisone prescription filled in the 90 days before the self‐report assessment, three (9%) had their most recent prescription filled within the prior year, and four (19%) had their most recent prescription filled ≥10 years prior.

Sixty‐six participants (23%) reported having a stockpile of prednisone. Of these, 49 (74%) reported no current use, 40 (60%) reported no use in the past 30 days, and 35 (53%) reported no use in the past 90 days. One hundred forty‐five participants (51%) reported that they had never received prednisone from a non‐VA provider. Of the remainder, 13 (9%) reported receiving prednisone from a non‐VA provider in the past year, and 60 (43%) reported receiving prednisone from a non‐VA provider more than a year prior.

DISCUSSION

In a cohort of participants with RA enrolled in the VARA Registry, ⁸ claims‐based algorithms evaluating prednisone use in the past day, the past 30 days, and the past 90 days performed acceptably when benchmarked on self‐report. In particular, the NPV of these algorithms was excellent, exceeding 0.9 in all three cases, partly because the prevalence of use was generally low. In a cohort with a higher prevalence of use, we may expect a superior PPV using these algorithms but a somewhat worse NPV.

Among the subset of participants reporting current use, we saw moderate agreement between self‐reported and claims‐based assessments of current daily dose; agreement became substantial when we limited the analysis to the 51% of participants reporting a daily dose of ≤5 mg/day. For our current use algorithm, misclassification was low and predominantly underrepresented exposure. For our current dose algorithm, misclassification underrepresented exposure in the setting of one in four participants reporting they had a prednisone stockpile and one in four participants reporting they received prednisone from a provider not captured in the claims source used. Overall, our results suggest that the claims‐based algorithms we present here perform adequately to (1) rule out prednisone use during the past 90 days in the context of low rates of use (NPV > 90%), (2) rule in current use in a cohort with similar rates of use (PPV > 80%), and (3) estimate current daily dose for participants with claims‐based dose estimates ≤7.5 mg/day of the prednisone equivalent. These findings support the use of such algorithms, compared to a benchmark of self‐report, for estimating glucocorticoid exposure in both research and clinical contexts.

We observed reduced agreement with higher reported mean daily doses, which may be due to prescribing patterns not adequately represented by a mean daily dose, such as a taper, or to use not adequately captured in claims, such as stockpiled or externally prescribed prednisone. We observed some decline in claims‐based algorithm performance, as our algorithms assessed longer lookback periods, and reduced agreement for questions assessing ever use of prednisone. This may be a limitation of the reference standard due to inferior recall for remote use and suggests algorithms benchmarked on self‐report may not be ideal for validating administrative measures of long‐term exposure.

Prior work in this area includes a recent study of 494 patients with RA enrolled in the CorEvitas registry that compared 90‐day average daily glucocorticoid dose in linked Medicare claims to a gold standard of physician‐reported use at a registry visit during that period. ⁵ Although agreement benchmarking on physician report was similar (overall κ of 0.61 and percent agreement of 0.9), the prevalence of physician‐reported 90‐day use was 31% (vs our finding of 20% self‐reported use), sensitivity was higher at 0.88, and specificity was lower at 0.79. These differences may be due to different algorithm criteria and a distinct reference standard used.

A small UK‐based study of 78 patients with RA reported excellent algorithm performance (sensitivity 0.84, specificity 0.87, PPV 0.87, NPV 0.85) when comparing presence of an active glucocorticoid prescription in the Clinical Practice Research Datalink to self‐reported glucocorticoid use via mail‐in survey. However, an important distinction from the current study is that only patients with documented glucocorticoid use in the past two years were eligible, and the survey response rate was only 16%, raising some concern for nonparticipant bias. ⁶ A single‐center study of 91 participants in the Brigham Rheumatoid Arthritis Sequential Study registry found a moderate to high correlation between self‐reported cumulative glucocorticoid dose and cumulative dose derived from chart review (Pearson's r 0.59) and strong agreement between self‐report and chart‐derived quartiles of dose (weighted κ 0.67). Thus, claims‐based algorithms evaluating current and recent glucocorticoid use appear to have reasonable performance despite differences in population, reference standard, and algorithm criteria. As the largest and most recent evaluation to benchmark glucocorticoid use on self‐report as a reference standard, our current study supports its use as a reference standard for validation of claims‐based algorithms. This is especially relevant in settings where provider report may underestimate glucocorticoid use (eg high patient sharing, lower provider connectedness). ¹²

We chose self‐report as our reference standard for several reasons. First, self‐reported glucocorticoid use in RA has only been examined as a reference standard in Europe, where glucocorticoid prescribing patterns differ from those in the United States. ¹³ Second, self‐report not only is more likely to capture glucocorticoid use occurring without a provider's knowledge but also allows us to ask about and assess the impact of potential sources of such use, as we do here. Third, although self‐reported glucocorticoid use suffers from recall bias, it is not clear that provider report suffers less from such bias. In prior studies, provider‐reported use was either derived directly from the medical record or collected during a patient visit when the provider had access to both the medical record and the patient. ⁵ , ⁶ , ¹⁴ Provider‐reported glucocorticoid can thus be conceptualized as the provider's transcription of (1) patient‐reported use at a visit and (2) prescribing data and/or previous provider reports from the electronic medical record. Provider‐reported glucocorticoid use is highly variable in consistency and detail and does not capture use of glucocorticoids from outside sources (stockpiles, outside prescribers), patient‐initiated dose reduction or discontinuation, or nondocumented verbal instructions given by the provider. ¹⁵ Fourth, the literature comparing provider report and self‐report as benchmarks for glucocorticoid use is limited. Demonstrating comparable performance between benchmarks would help facilitate future research in this area. It is also reassuring that self‐reported medication use generally shows good agreement with use derived from other data sources, even for medications that are used episodically (eg, antacids, antibiotics, muscle relaxants). ¹⁶ , ¹⁷ , ¹⁸ Although we designed this study with self‐report use as the reference standard, it is likely more appropriate to consider prescription claims‐based algorithms and self‐reported use as complementary assessments. Each has its own strengths and limitations, and combining these two sources together would be ideal, when feasible.

Our data also provide valuable input into glucocorticoid stockpiling, external glucocorticoid sources, and patient‐directed glucocorticoid consumption. Although outside the scope of this study, future work investigating ways to reduce stockpiling could be beneficial in limiting avoidable glucocorticoid exposure in RA populations.

Our study is among the first and largest to evaluate the performance of claims‐based algorithms for glucocorticoid use compared to self‐report. It is also, to our knowledge, the first to directly assess potential barriers to developing accurate algorithms, such as stockpiling and prescribing not captured in claims. Limitations include that we assessed prednisone use only among patients with RA who were both enrolled in a prospective registry and had a rheumatology visit during the assessment interval. Thus, our results may not be generalizable to patients who do not receive regular rheumatologic care or who receive care at lower‐resourced sites. Relatedly, we found a relatively low prevalence of prednisone use in general and high‐dose use (≥10 mg/day) in particular. Although this finding is consistent with American College of Rheumatology recommendations to taper glucocorticoids promptly and avoid high‐dose use, ³ it also limits our ability to validate claims‐based algorithms assessing high‐dose use. Additionally, although self‐reported current prednisone dose assessed the day before the questionnaire reasonably approximates the mean dose for an active dispensing episode in our sample, these measures do not reflect the fact that prednisone dose can vary from day to day within a single dispensing episode, for example, when a taper is prescribed. We were not able to capture prednisone prescribed by non‐VA providers, though prior work suggests veterans with RA are relatively unlikely to use medical care. ¹⁹ To maximize algorithm performance, we did not assess prednisone use for RA specifically, such that our results likely capture prescriptions given by nonrheumatologists for other indications (eg, chronic obstructive pulmonary disease flares). Our analysis was limited to veterans and used VA claims; however, our results largely agree with and support prior evidence in nonveteran populations. ⁵ , ⁶ , ¹⁴ We cannot exclude the possibility of survey bias in our questionnaire responses, for example, recall bias (which may be increased among older participants) or demand bias from underreporting use unrelated to RA treatment or reporting a prescribed dose when taking differently. We attempted to mitigate this by limiting our evaluation to prednisone use, which constitutes >95% of oral glucocorticoids prescribed to this population. We also note that prior work evaluating the validity of self‐reported medication use in the past year, including evaluations focused on older adults, shows excellent agreement with pharmacy claims. ²⁰ , ²¹

In conclusion, we found that algorithms derived from VA claims can acceptably assess ongoing prednisone use in patients with RA compared to a reference standard of patient self‐report. Such algorithms may also be useful for estimating mean daily doses ≤7.5 mg/day of the prednisone equivalent, though we had limited ability to assess their utility for higher doses due to low prevalence. Our algorithm somewhat underestimated prednisone dose, contextualized by our finding of high self‐reported glucocorticoid stockpiling and prescribing by non‐VA providers. These findings may be used to estimate the performance of claims‐based algorithms for glucocorticoid use in research and clinical contexts and to estimate the impact of use that cannot be captured in claims.

AUTHOR CONTRIBUTIONS

All authors contributed to at least one of the following manuscript preparation roles: conceptualization AND/OR methodology, software, investigation, formal analysis, data curation, visualization, and validation AND drafting or reviewing/editing the final draft. As corresponding author, Dr Wallace confirms that all authors have provided the final approval of the version to be published and takes responsibility for the affirmations regarding article submission (eg, not under consideration by another journal), the integrity of the data presented, and the statements regarding compliance with institutional review board/Declaration of Helsinki requirements.

Supporting information

Disclosure form.

Appendix S1: Supplementary Information