Prior Authorization: A Critical Pain Point
Prior authorization — the process by which insurers require physicians to obtain approval before delivering certain treatments — has become one of the most contentious issues in American healthcare. Physicians consistently report that prior authorization requirements delay necessary care, increase administrative burden, and sometimes force patients to abandon recommended treatments entirely. The American Medical Association has tracked these dynamics for years through its Prior Authorization Survey, one of the most widely cited data sources on the topic.
For Simsurveys, the AMA survey presented an important validation opportunity. Physician surveys are expensive and difficult to field. Response rates among practicing physicians have declined steadily, and the costs of reaching specialists who provide 20 or more hours of direct patient care per week are substantial. If a synthetic HCP model can reliably reproduce the patterns found in authoritative physician research, it opens the door to faster, more affordable insights on healthcare policy and practice management questions.
The Benchmark
We compared Simsurveys synthetic physician responses against the AMA Prior Authorization Survey, which sampled 1,000 practicing physicians across the United States. Eligibility required that respondents provide at least 20 hours of direct patient care per week and personally complete prior authorization requests. The survey covered 15 questions organized around care delays, treatment abandonment, clinical outcomes, peer review processes, resource utilization, and payer-specific administrative burden.
We generated a matched synthetic sample using the Simsurveys Healthcare (HCP) model and measured distributional alignment for each question. For the 14 single-select questions we computed Kullback-Leibler (KL) divergence against the AMA distribution, with Laplace smoothing applied to handle zero-probability responses. The one multi-select question — the ways prior authorization drives higher resource utilization — was scored with Rank-Biased Overlap (RBO) similarity, which is better suited to comparing rankings of independent response categories.
Where It Worked Well
The synthetic model performed strongly on the questions most central to the prior authorization debate. On Q1, which asked physicians about the frequency of care delays attributable to prior authorization, the KL divergence was just 0.048 — indicating near-identical response distributions between the AMA sample and our synthetic respondents. Q4, on how the number of prior authorization denials has changed over the last five years, was among the very best in the study at 0.017, and Q2, covering treatment abandonment due to prior authorization barriers, came in at 0.093.
The model also excelled on process-oriented questions. Q5, which asked how the frequency of peer-to-peer reviews has changed over the last five years, scored 0.008 — the lowest divergence in the entire study. These results suggest the model has internalized the professional norms and frustrations that practicing physicians express around the mechanics of authorization processes.
Perhaps most notable was the performance on payer-specific burden questions. When physicians were asked to rate the administrative burden imposed by individual insurers, the synthetic model closely matched real-world sentiment. Anthem/Elevance scored a KL divergence of just 0.004 — the tightest alignment of any single-select question — while Blue Cross Blue Shield came in at 0.048, Cigna at 0.051, and Aetna at 0.057. These are the kinds of granular, brand-level questions where synthetic data is often expected to struggle, and the model held up well.
Strongest results: Anthem/Elevance burden (KL=0.004), peer-to-peer review frequency (0.008), denial trends (0.017), care delay (0.048), and BCBS burden (0.048). The model reliably captured the core dynamics of the prior authorization experience as reported by practicing physicians.
Where It Struggled
Not every question produced tight alignment. Q3, which asked physicians about their perception of the overall impact of prior authorization on patient clinical outcomes, showed a KL divergence of 0.257 — the highest in the study. Q9, on how often a delay or denial leads to a patient paying out of pocket for a prescribed medication, scored 0.232. Q6, on whether the health plan's peer reviewer has the appropriate qualifications, came in at 0.140. These questions tend to be more subjective and involve longer-term professional judgment rather than concrete, observable events.
On the payer-specific side, Humana burden scored 0.198 and UnitedHealthcare 0.174 — the two loosest of the six insurer questions. These larger divergences may reflect the fact that physician sentiment toward specific insurers varies significantly by region, specialty, and practice type — variables that are difficult to fully capture in a general-purpose HCP model without additional segmentation. The single multi-select question (Q8), on the specific ways prior authorization drives higher resource utilization, was scored with Rank-Biased Overlap rather than KL and reached an RBO similarity of 0.22, indicating the synthetic ranking of contributing factors only partially matched the AMA ordering.
The pattern is instructive: the model performs best on factual, behavioral, and process-oriented questions and shows more variance on subjective perception and trend-judgment items. This is consistent with what we see across other validation studies and reflects a known characteristic of synthetic survey data.
What This Means
The AMA validation demonstrates that synthetic physician data can reliably replicate the response distributions of real physicians on a majority of prior authorization questions — particularly those related to care delays, treatment abandonment, insurer interactions, and payer-specific burden. For research teams studying prior authorization policy, practice management, or payer relations, the Simsurveys HCP model offers a viable path to fast, scalable physician insights without the cost and timeline of traditional physician panels.
Where the model showed weakness — subjective clinical outcome perceptions and denial trend judgments — researchers should exercise caution and consider supplementing synthetic data with qualitative interviews or targeted traditional fielding. This is not a limitation unique to synthetic data; even traditional physician surveys struggle with the reliability of subjective trend questions.
The full validation report, including question-level distribution tables and KL divergence scores, is available for download on our validation studies page. To explore the Healthcare (HCP) model further, visit the model page or create a free account to run your own physician study.
Frequently Asked Questions
What was validated in the AMA Prior Authorization study?
Simsurveys synthetic physician responses were compared against the American Medical Association's Prior Authorization Survey, which sampled 1,000 practicing physicians across 15 questions covering care delays, treatment abandonment, peer review processes, and payer-specific administrative burden.
How did synthetic physician data perform on care delay questions?
The synthetic model achieved a KL divergence of just 0.048 on the care delay question and 0.093 on treatment abandonment, indicating near-identical response distributions between the AMA sample and synthetic respondents. Payer-specific burden questions performed especially well, with Anthem/Elevance at 0.004, Blue Cross Blue Shield at 0.048, and Cigna at 0.051.
Where did the synthetic physician model show limitations?
The model showed higher divergence on subjective perception questions such as clinical outcome perceptions (KL 0.257, the highest in the study) and out-of-pocket cost frequency (KL 0.232). The model performs best on factual, behavioral, and process-oriented questions and shows more variance on subjective judgment items.