Title: Using multiple imputation to account for missing stage and to generate a population-based late-stage incidence rate

Authors & affilliations: Shane Johnson^1,2,5, Kirstin Roberts^1,2, Sophia Lowes¹, Fahmina Fardus-Reid¹, Carolynn Gildea², Sean McPhail², Jason Poole³, David Foord³, Lorna Wills¹, Samantha Hinks³, Mark Rutherford⁴, Matthew Barclay⁵, Jon Shelton¹

¹Cancer Intelligence, Cancer Research UK, ²National Disease Registration Service, NHS England, ³Cancer Analysis and Insights Team, NHS England, ⁴Division of Public Health & Epidemiology, University of Leicester, ⁵Epidemiology of Cancer Healthcare & Outcomes, University College London

Background: 

• Reliable stage at diagnosis data are important for monitoring ambitions to diagnose cancers at an earlier stage. 

• Completeness of staging data is variable over time, making accurate population-based surveillance of stage at diagnosis difficult.

• As an alternative to stage proportion metrics¹, a rate-based metric of stage at diagnosis has been proposed to aid interpretation of trends, by taking into account population and age structure, and changes in cancer incidence over time.

• We aimed to develop a robust multiple imputation model for missing stage data, enabling calculation of incidence rates by stage at diagnosis.

Methods:

Data were extracted for cancer diagnoses in England from 2014 to 2022 from the National Cancer Registration Dataset (NCRD). Multiple imputation (MI) models were run to impute missing stage at diagnosis values.

• tumour characteristics (diagnosis year, stage at diagnosis, cancer site, route to diagnosis [screening, emergency presentation or other])

• patient characteristics (age, gender and deprivation quintile)

• treatment (fact of first systemic anti-cancer therapy, radiotherapy and surgery²)

• vital status (at 12 months post-diagnosis for cumulative hazard³)

• Stage distribution

• Cohen’s kappa – measure of agreement between the known and imputed stage values⁴

Results:

After imputing missing stage, late-stage incidence rates were more consistent and higher over time compared to incidence rates prior to imputation (Figure 1). Late-stage incidence rates in 2022 were 282 late-stage diagnoses per 100,000 persons where stage was imputed compared to 203 prior to imputation.

Around 19% of the original dataset had missing stage values (Figure 2). After stage imputation, stage 4 increased by 6.4 percentage points and stage 1 increased by 4.5 percentage points. This is because stage 4 was more common (34%) than stage 1 disease (24%) in the imputed missing stage patients.