How survivorship bias can affect reports of two-year outcomes
When we quote a two-year success rate for an operation, the figure typically describes the outcomes for the patients who came back to be assessed at two or more years after surgery.
Those who did not return are not a random sample of all the patients having the surgery. Patients lost to follow-up tend to have fared worse than those who complete follow-up, because the reasons they drop out are often themselves adverse — death, revision, or disappointment with an early result that leads them to transfer their care elsewhere [1, 2, 3].
Excluding patients who have done poorly prior to the two-year mark — considering only those who remain for the two-year analysis — makes the operation look better than it actually was, a phenomenon known as survivorship bias [4].
A model of survivorship bias
Here is an illustrative hypothetical example of a thousand patients having an RSA for cuff-intact osteoarthritis with two-year follow-up.
In the first six months patients were lost as follows: 20 shoulders revised, 5 patients dead, 15 transferred to another practice, and 40 who simply stopped responding — 80 in all. Of note, certain problems occur early in the postoperative period: acute infection, instability, acromial fracture, and life-limiting frailty.
During the second six months there were 7 more revisions and 8 deaths, 25 transfers, and 110 who did not respond for unknown reasons — 150 in the interval.
The second-year losses are almost entirely loss of contact rather than loss of the shoulder: 3 revised and 12 more dead, 35 gone to other practices, and 200 more who stopped answering — 250 in that interval.
By the time of the two-year analysis, 30 of the 1,000 have been revised, 25 have died, and 75 have transferred their care; 350 more have gone quiet without a recorded reason. That is 480 lost from view, none of whom are included in the two-year analysis of results; only 520 remain available for study.
Figure 1. A hypothetical cohort of 1,000 patients followed over two years. At each interval, patients drop out of view for four kinds of reasons, and the mix changes as time passes. Of the 520 analyzed at two years, 9 out of 10 report success; but out of all 1,000 operated on, a successful outcome is documented for only about 4.7 out of every 10.
The four reasons carry different weight, and their proportions change over the two years. Revisions come early: RSA’s dominant early failures — acute infection, instability, and acromial fracture — appear mostly in the first months, so revisions decline from 20 in the first half-year to 3 in the second year. Deaths run the other way, accumulating with time in an elderly group, from 5 to 8 to 12 across the intervals. Transfers of care, and above all patients who simply stop responding, grow steadily as contact is lost — from 40 silent patients in the first six months to 200 in the second year. By the time of the two-year analysis, loss of contact, not loss of the shoulder, accounts for most of the missing.
Applying the model of survivorship bias to actual data on two-year outcomes for RSA for osteoarthritis
The model becomes meaningful when anchored to what the literature reports. Two published figures set the scale.
The first is the rate of two-year follow-up. In a multicenter shoulder arthroplasty registry, only about half of patients — about 5 out of 10 — provided two-year patient-reported outcomes [5]; registries that send repeated reminders might improve the return to 8 out of 10.
The second is the success rate among those who do return: in high-volume single-surgeon series of RSA for this diagnosis, about 9 out of 10 report being better and satisfied [6, 7].
Consider the combined effect of these two rates. Of the 520 with a known two-year result, 9 out of 10 — 468 patients — report success; but across the full 1,000 who had the surgery, those 468 provide an overall documented success rate of only about 4.7 out of 10.
Had the follow-up rate reached 8 out of 10 rather than 5, about 800 would have a known result, and the same 9 out of 10 would give about 720 documented successes — roughly 7 out of 10.
The span from 4.7 to 7 out of 10 is set entirely by the follow-up rate.
The 4.7 out of 10 is the floor. It counts every patient without a documented success as a non-success, so that the true whole-cohort success rate (if it could be known) might be higher than 4.7 out of 10.
The missing patients are not a random subset of the cohort. Those lost to revision, death, or a transfer of care each had reason to fare worse than the returners. The larger unresponsive group has, on average, done less well than those who answer — though that association is weaker and less certain than for the other types of losses [1, 2, 3].
Reporting the returners’ 9 out of 10 as the rate for the overall cohort makes the operation look better than the data support [4].
However, a successful outcome is only documented for about 5 of every 10 patients considering all those having the surgery, not 9 out of 10.
A fuller accounting would mean following the patients who leave — the revised, the transferred, and above all the ones who quietly stop answering — well enough to know how they actually did. Until a series does that, the appropriate two-year estimate for successful outcome for RSA in osteoarthritis is somewhere above the documented 4.7 out of 10 rate for the entire cohort and below the rate of 9 out of 10 considering only the patients returning for two year analysis.
And two years is the favorable case. The effect of survivorship bias grows as follow-up lengthens: over five, ten, and fifteen years, follow-up falls further and the number of missing patients grows, so the distance between the returners’ success rate and the whole-cohort success rate only widens. The longer the follow-up a reported success rate claims, the greater the effect of survivorship bias.
It's about survivorship
Follow on twitter/X: https://x.com/RickMatsen
Follow on facebook: https://www.facebook.com/shoulder.arthritis
References
[1] Solberg TK, Sørlie A, Sjaavik K, Nygaard ØP, Ingebrigtsen T. Would loss to follow-up bias the outcome evaluation of patients operated for degenerative disorders of the lumbar spine? A study of responding and non-responding cohort participants from a clinical spine surgery registry. Acta Orthop. 2011;82(1):56–63.
[2] Murnaghan ML, Buckley RE. Lost but not forgotten: patients lost to follow-up in a trauma database. Can J Surg. 2002;45(3):191–195.
[3] Torrens C, Martínez R, Santana F. Patients lost to follow-up in shoulder arthroplasty: descriptive characteristics and reasons. Clin Orthop Surg. 2022;14(1):112–118.
[4] Elston DM. Survivorship bias. J Am Acad Dermatol. Published online June 18, 2021.
[5] Patel M, Sekar MG, McDaniel L, Kisana HM, Sykes JB, Amini MH. Changes from baseline in patient-reported outcomes and patient satisfaction do not vary significantly between 1 and 2 years postoperatively after shoulder arthroplasty: a multicenter analysis of 2580 patients. Semin Arthroplasty JSES. 2025;35(2):235–245.
[6] Puzzitiello RN, Moverman MA, Glass EA, Swanson DP, Bowler AR, Le K, Kirsch JM, Lohre R, Jawa A. Clinically significant outcome thresholds and rates of achievement by shoulder arthroplasty type and preoperative diagnosis. J Shoulder Elbow Surg. 2024;33(7):1448–1456.
[7] Ahmed AF, Glass EA, Swanson DP, et al. Predictors of poor and excellent outcomes following reverse shoulder arthroplasty for glenohumeral osteoarthritis with an intact rotator cuff. J Shoulder Elbow Surg. 2024;33(6S):S55–S63.
