Statistical Analysis Plan

Describing, characterising and predicting winter respiratory accident and emergency attendances, hospital and intensive care unit admissions and deaths in Scotland (plain text)

Version history

Log of edits made to document, starting from most recent entries.

Date	Who	What (adjustments)
2023-05-15	TS	Treated ARI hospitalisation (broad definition) as the primary outcome. Treated ARI related length of hospital stay, ICU admission and death as the secondary outcome. Proposed a multivariable logistic regression model to estimate the association between risk factors and ARI related prolonged hospital stay (a hospital stay longer than 5 days – median length of stay for ARI hospitalisations) Calculated % for outcomes ICU admission or death among all ARI hospitalisation. Used the national laboratory dataset to estimate the number of laboratory-confirmed influenza and SARS-CoV-2 cases in comparison to those identified using ICD-10 codes Conducted sensitivity analyses adjusting for either influenza vaccination status or SARS-CoV-2 vaccination status when looking at outcomes of influenza hospitalisation or SARS-CoV-2 hospitalisation.
2023-04-03	TS	Changes broad definition as the main analysis and strict definition as the supplementary analysis. Updated the death definition (only focusing on those admitted to hospitals as the scope of this project is hospital pressure). Instead of doing cox modelling for death data, we proposed to look at in-hospital case fatality ratio. Removed other severe outcomes from the Cox modelling “ICU admission and mechanical ventilation” as we don’t have enough data in subgroups. Treated prolonged hospital stay as a continuous variable. Proposed for pathogen specific analysis, we only produced descriptive tables due to small number of events and inaccurate results.
2023-03-21	TS	Change “respiratory disease” to “respiratory infection” to reflect on the focus of this work – acute respiratory infections
2023-03-13	TS	Removed the number of distinct BNF chapters prescribed in the last 6 months. Removed BMI for children. For two doses of influenza vaccines, take the first vaccine data only. Agreed on presenting the cox modelling results separately for children and adults. Hospitalisation definition: strict definition as the main result, broader definition as the supplementary result. Death definition: strict definition (within the hospitalisation cohort) as the main result, within 28 days of hospital admission with ARIs definition as the comparison (within the hospitalisation cohort), strict definition (all population) and broad definition as the supplementary result.
2023-03-08	TS	Removed covariates known respiratory conditions, previous SARS-CoV-2 infection, social care residence, smoking status, chronic disease treatment in the last six months, polypharmacy as of 1st September 2022 in the last six months. Added the number of distinct BNF chapters prescribed in the last 6 months. Removed oxygen therapy from the secondary outcome as the data are not available.
2023-01-30	TS, TM, BS, AFF	Updated plan based on comments
2023-01-09	TS	Initial plan based on bid

Research aims

1. Describe and characterise children and adults experiencing severe respiratory health outcomes across Scotland during the wintertime period.
2. Identify independent demographic and clinical risk factors for respiratory-related  Accident & Emergency (A&E) services attendance, hospital admission, Intensive Care Units (ICUs) admission and death.
3. Identify modifiable risk factors that can potentially improve respiratory outcomes in high-risk individuals.

Cohort design

Study design: A national population-based observational, retrospective cohort study using routinely collected linked data from across Scotland

A cohort of the general population defined on 1^st September 2022:

• Alive and living on 1^st September 2022.
• Have primary care records prior to 1^st September 2022 (registered with a GP practice).
• All age groups.
• Resident in Scotland for the study period or outcomes of interest.

Follow-up: The study window starts on 1^st September 2022 and ends on 31^st January 2023.

Data sources

• GP data: patient characteristics, acute respiratory infections and clinical history (Read codes version 2), prescribing and vaccination status;
• Secondary care data (emergency and non-emergency hospital admissions, ICU, maternity): Scottish Morbidity Record 01 and 02 and Rapid Preliminary Inpatient Data (RAPID);
• Accident & Emergency (A&E) services visit: Emergency Departments (EDs) and Minor Injury Units (MIUs);
• Laboratory data: including RT-PCR SARS-CoV-2, RSV and influenza and other common respiratory infection (multiplex) test and other test data, available through Electronic Communication of Surveillance in Scotland database;
• Medication data: Prescribing Information System: Prescribing in primary care data and GP data;
• Mortality data: National Records of Scotland database and GP data.

Sample selection and weighting

• Keep all those who had the outcome (respiratory related A&E attendance, hospital admission, requirement for oxygen therapy, ICU admission or death) during follow-up and are given a weight of 1.
• Randomly select a sample of all those who did not have the outcome during follow-up. Sample without replacement at a ratio of 10:1. For every person who had the outcome, we want 10 who did not.
• Weight for those without the outcome is:
  • Total number who did not have the outcome / 10 x total number who had the outcome

Effectively, we are down sampling the number of people without the outcome and then weighting them to represent the number of those we dropped. We do this to save computation time when model fitting, as we do not gain much extra information from using the full cohort (millions) versus a large random sample (100,000s). We pick a large ratio of 10:1 to overcome any potential covariate imbalance by randomly sampling.

Outcome

• Primary outcome is respiratory infection related emergency hospital admission and death.
• Secondary outcome is respiratory infection admissions related length of hospital stay or percentage of deaths among hospitalised individuals (in-hospital case fatality ratio).
• The plan is to use broad definitions for respiratory-related severe health outcomes, followed-by a sensitivity analysis using strict definitions.
• Broader definitions:
  • Respiratory-related hospitalisation is defined as a hospital emergency admission with an ICD-10 code for respiratory infection in any position of the first episode (i.e. admitted due to or with respiratory infection)
  • Respiratory-related death is defined as a death with respiratory infection being listed in any position relating to the death (i.e. primary, secondary cause, or underlying factor) among people who were hospitalised with respiratory infection.
• Strict definitions:
  • Respiratory related hospitalisation is defined as a hospital emergency admission for one or more days, with an ICD-10 code for respiratory infection in the first position of the first episode, after removing ICD-10 codes starting with R or Z
  • Respiratory-related death is defined as death with respiratory infection being listed as the primary cause of death among people who were hospitalised with respiratory infection.
• Observations are censored before the end of the study window if the person:
  • Moved out of the respective nation.
  • Died due to other causes.

Exposure

SARS-CoV-2, RSV, influenza, Streptococcus A or other acute respiratory tract infections (ARTIs) including co-infection (via RT-PCR test, rapid antigen test (RAT) and clinically coded diagnosis in GPs, and hospitalisation records); dates of these events will be the index date. Each of these is being treated as a separate exposure. There will also be a composite exposure for any of these.

Covariates – demographics and clinical characteristics

• Age (5-year bins): 0-4 (also 0-2 if possible), 5-9, 10-17, 18-49, 50-54, 55-59, …, 75-79, 80-110 for use as a covariate in the modelling. Also consider a continuous polynomial variable or similar.
• Age (5 categories): 0-17, 18-49, 50-64, 65-79, 80-110 for coarser analysis such as the exploratory work.
• Sex: Male, Female.
• Urban/rural classification of residence on 1^st September 2022. Categories collapsed:
  • Urban = (Urban city and town, Urban city and town in a sparse setting).
  • Rural = (Rural town and fringe, Rural town and fringe in a sparse setting, Rural village and dispersed, Rural village and dispersed in a sparse setting).
• Area deprivation quintile on 1^st September 2022.
  • SIMD (or respective national indicator), 1 is Most deprived, 5 is least deprived.
• Ethnic groups are categorised with the ONS's five (plus unknown) categories:
  • Asian, Black, Mixed, White, Others.
  • Code missing as Unknown.
• QCOVID measures as of 1^st September 2022, excluding BMI, converted to a sum score:
  • 0, 1, 2, 3, 4, 5+.
• Number of previous emergency hospitalisation in the last six months before 1^st September 2022.
• NHS appropriate geography (e.g. Health Board).

Modifiable covariates

• Most recently recorded BMI in the last 5 years up to 1^st September 2022. Imputed at 45% missing:

  • Six categories: <18.5, 18.5-24.9, 25-29.9, 30-34.9, 34-39.0, 40+.
  • Single imputation using log(BMI) as the outcome and the following as covariates:
    • Simplified outcomes: ever had respiratory related hospital admission or death.
    • Simplified exposures: SARS-CoV-2, RSV, influenza, Streptococcus A or other respiratory pathogens.
    • All covariates used for analysis, including individual QCovid measures
  • Remember to exponentiate after imputing.
• Vaccination for COVID-19 (any time periods), influenza (at the time of hospitalisation looking back to 1^st September 2022) and other vaccinations.

Exploratory analysis

For the full population cohort:

• Provide an epidemiological description of those individuals who have contributed to NHS compound winter pressures (including prevalence by characteristics).
• Plot weekly counts for (or a smoothed time series):
  • SARS-CoV-2, RSV, influenza, Streptococcus A or other ARTIs.
  • Severe respiratory health outcomes.
  • Move out of the nation, death due to other causes.
• Empirical cumulative incidence of SARS-CoV-2, RSV, influenza, Streptococcus A or other ARTIs by age groups: 0-4, 5-17, 18-64, 65-79, 80-110.
• Univariate summary of characteristics plus collinearity plots of outcomes/covariates.
• Empirical cumulative incidence of severe respiratory health outcomes (respiratory related A&E attendance, emergency hospital admission, length of stay, requirement for oxygen therapy, ICU admission, need for mechanical ventilation and death) by vaccination status and age groups: 0-4, 5-17, 18-64, 65-79, 80-110.

Modelling analysis

• Associations between risk factors and the outcomes of interest will be evaluated using cohort approaches i.e., logistic regression, Poisson regression and Cox proportional hazards models whereby patients will be followed up from index date until the earliest date of outcomes of interest, death or end of the study period.
• Specifically, we will look at whether combinations of risk factors place individuals more at risk of poor health outcomes and if the risk factors for COVID-19, RSV and influenza are the same.
• Conduct a subgroup analysis focusing on healthcare usage in individuals with multimorbidity (as individuals with multimorbidity are amongst the highest users of healthcare).

Describe population characteristics

Using the proposed cohort, make four tables of descriptive characteristic summaries for:

• For those who had SARS-CoV-2, RSV, influenza, Streptococcus A or other ARTIs, total counts and column percentages (Table 1):

 

• Number of severe respiratory health events and incidence rate per 1,000 persons per year stratified by respiratory pathogen (Table 2)

Using the full cohort, make two figures:

• Weekly number of severe respiratory health outcomes over time.
• Cumulative incidence of severe respiratory health outcomes from 1^st September 2022, stratified by respiratory pathogen.

Modelling of risk factors for severe respiratory health outcomes

• Cut follow-up time into intervals based on:

  • Calendar weeks 
• Create person-years summary data set given formula:
  • Start_time, stop_time, event_flg ~ [all the covariates].
  • Use sample weights.
• Fit one overall model, and four subset models, one for each respiratory pathogen. Each model is a Poisson GLM with:
  • Outcome is number of events.
  • Main effects for all covariates Offset with log(person years).
  • Covariate selection.
• Report estimates with 95% Wald confidence intervals.

Acute respiratory infection code lists

Acute respiratory infection code lists

Condition	ICD10
Acute upper respiratory tract infection (URTI)	J00, J02-06
Lower respiratory tract infection (LRTI)
Pneumonia & influenza	J09-18
Bronchiolitis and bronchitis	J20-21, J40
Unspecified LRTI	J22
COVID-19	U07.1, U07.2, U08-10
RSV	J12.1, J20.5, J21.0
Group A Streptococcus	B95. 0

Other code lists

Other code lists

Respiratory	All J codes
Heart	I00-I52
Chronic Obstructive Pulmonary Disease	J40-J44, J47
Asthma	J45-J46
Wheeze	R06.2