Identifying care-home residents in routine healthcare datasets: a diagnostic test accuracy study of five methods

Background: There is no established method to identify care-home residents in routine healthcare datasets. Methods matching patient’s addresses to known care-home addresses have been proposed in the UK, but few have been formally evaluated.

Study design: Prospective diagnostic test accuracy study Methods: Four independent samples of 5,000 addresses from Community Health Index (CHI) population registers were sampled for two NHS Scotland Health Boards on 1st April 2017, with one sample of adults aged ≥65 years and one of all residents. To derive the reference standard, all 20,000 addresses were manually adjudicated as ‘care-home address’ or not. The performance of five methods (NHS Scotland assigned CHI Institution Flag, exact address matching, postcode matching, Phonics and Markov) was evaluated compared to the reference standard.

Results: The CHI Institution Flag had a high PPV 97-99% in all four test sets, but poorer sensitivity 55-89%. Exact address matching failed in every case. Postcode matching had higher sensitivity than the CHI flag 78-90%, but worse PPV 77-85%. Area under the receiver operating curve values for Phonics and Markov scores were 0.86-0.95 and 0.93-0.98 respectively. Phonics score with cut-off 13 had PPV 92-97% with sensitivity 72-87%. Markov PPVs were 90-95% with sensitivity 69-90% with cut-off 29.6.

Conclusions: More complex address matching methods greatly improve identification compared to the existing NHS Scotland flag or postcode matching, although no method achieved both sensitivity and positive predictive value (PPV) >95%. Choice of method and cut-offs will be determined by the specific needs of researchers and practitioners.