What predicts patient experience?

The three outcomes, from the GP Patient Survey 2026 (fieldwork January–April 2026; practices with at least 30 responses):

Overall experience — Q32: "Overall, how would you describe your experience of your GP practice?" (% very or fairly good).

Most recent contact — Q16: "Overall, how would you describe your experience of contacting your GP practice on this occasion?" (% very or fairly good). This is about the last attempt to get in touch, not access in general.

Easy to contact on the phone — Q1: "Generally, how easy or difficult is it to contact your GP practice on the phone?" (% very or fairly easy). The phone-specific report, and the outcome the operational data should predict best.

How to read the tables. Every predictor except training-practice status is standardised, so each figure is the difference in the outcome (percentage points) associated with being one standard deviation higher on that predictor; training-practice status is the difference associated with hosting at least a quarter of a full-time GP registrar. GP staffing is split by contract type: partners, salaried (including retainers), regular locums, and registrars in training. "Alone" looks at one predictor at a time. "Together" puts the operational measures and the practice's circumstances in one model. "+ patients report" adds the three measures patients themselves report: seeing a preferred clinician (Q7, Q6) and being told to contact the practice again (Q12). Figures in grey are not statistically distinguishable from zero (p≥0.05).

The largest coefficient in every model belongs to patients reporting they were told to contact the practice again another day (Q12). This measure is strongly tied to capacity — deprivation, unanswered calls and GP numbers predict a fifth of it (final section) — so a high figure partly reflects demand exceeding the appointments available, which is not quickly fixable. It is not fully determined by capacity, though:

Practices under the same measured pressure differ several-fold: a tenth of even the most pressured keep the figure at or below 3.4% — beneath the national median — while a tenth exceed 19%. What separates them is not measurable in these data; the survey does not record what patients are offered when no slot is left. The largest positive coefficient belongs to patients getting to see the clinician they prefer (Q7), which is substantially a matter of how appointments are allocated. Among the purely operational measures, answering the phone reliably matters most, and calls ending in the phone menus carry a separate unfavourable association. The Monday–Wednesday answering gap and contacts per appointment look important alone but stop mattering once the overall answer rate and calls per patient are in the model.

Deprivation works largely through being turned away. Alone, deprivation carries one of the biggest disadvantages in the table (−2.27 / −2.44). With the patient-reported measures in the model it shrinks to a fraction (−0.54 / −0.77): practices in deprived areas score worse substantially because more of their patients report being told to come back another day.

The size disadvantage decomposes almost entirely. Alone, larger practices score much lower (−2.33 / −3.35). With continuity, deflection and the phone measures accounted for, a small fraction remains (−0.46 / −0.97). Most of what looked like a cost of being large is the loss of continuity, more patients turned away, and fewer calls answered — all of which are things practices do, not things they are.

Who the GPs are matters, not just how many — and the types relate to experience in different ways. Partners carry the largest staffing association with every outcome when looked at alone, but most of it operates through continuity: control for whether patients get their preferred clinician and the partner coefficient falls from +1.68 to +0.55 on overall experience, while the salaried-GP coefficient changes much less (+1.42 to +0.96). The composition model of continuity itself makes the mechanism explicit:

Partners have the strongest positive association with continuity of any staff group. Trainees, who rotate, are associated with lower continuity — yet training practices as organisations are continuity-positive (+2.16) despite that, and score 1.9–3.4 points higher than their staffing and circumstances predict on every experience outcome. The likeliest reading is that training accreditation marks organisational quality the other variables cannot see; the same association was reported from the 2012 survey by Ahluwalia and colleagues (BJGP 2014), when 29% of practices trained. Our designation is cruder: hosting at least a quarter of a full-time registrar in the March 2025 census (59% of practices), which misses accredited practices between placements.

Regular locum GPs show no relationship with any outcome, favourable or unfavourable, in any specification — with two caveats that limit what this can mean: only one practice in ten carries more than a quarter-FTE of regular locum cover, and ad-hoc locums are excluded from practice-level workforce data entirely (NHS England reports them only in national annexes), so heavy sessional locum use is invisible to every practice-level analysis, ours included.

The same-day share might seem counterintuitive. Alone it looks slightly unfavourable; with the patient-reported measures held constant it turns slightly favourable (+0.52 / +0.61). Booking models built around same-day access tend to cost continuity and turn more people away when the day fills; once those costs are accounted for separately, the configuration itself is not the problem.

Admin staffing might seem counterintuitive. Alone, more admin staff goes with higher scores (+1.00). The combined model asks a narrower question: of two practices the same on everything else measured — size, deprivation, GPs, nurses, appointments — does the one with more admin staff score higher? It scores slightly lower (−1.11, shrinking to −0.45 with the patient-reported measures in). We tested the obvious explanations. It is not phone workload: adding calls per patient to the model leaves the result unchanged. It is not reception quality: patients at practices with more admin staff rate their reception teams as more helpful, not less (Q4), and allowing for that barely moves it. So we do not know what this remaining association reflects — possibly workload we don't measure, differences in which roles get counted as admin, or the difficulty of dividing credit among staffing numbers that rise and fall together. It should not be read as admin staff lowering scores — and equally, the table gives no support for expecting more admin staff alone to raise them. Nor does admin staffing predict how well the phones are answered:

The pattern is flat, and stays flat comparing practices of similar size. Admin headcount includes many roles besides answering phones, so total admin staffing is a poor guide to how well a practice's phones are answered.