Multi-Site Analysis for Independent Data Sources

The multi-site analyses included in this suite are intended to be executed against data that are all stored in the same place. However, there may be some instances where the data associated with each site is stored in independent locations. This vignette outlines how the multi-site analysis can be executed in these instances.

After following the instructions to reproduce the analysis, you will also need to change the output_function column to tell the prc_output function which check you executed. Reference the table below for the labels that are associated with each check:

Check Type	output_function
Multi Site, Exploratory, Cross-Sectional	prc_ms_exp_cs
Multi Site, Exploratory, Longitudinal	prc_ms_exp_la
Multi Site, Anomaly Detection, Cross-Sectional	prc_ms_anom_cs
Multi Site, Anomaly Detection, Longitudinal	prc_ms_anom_la

Multi-Site Exploratory Analysis

The process for the exploratory analysis is the same for both the cross-sectional and longitudinal configurations.

First, execute either of the Single Site, Exploratory analyses, configured appropriately for your study, against each data source.

library(patientrecordconsistency)

my_table <- prc_process(cohort = my_cohort,
                        multi_or_single_site = 'single',
                        anomaly_or_exploratory = 'exploratory',
                        time = T / F,
                        ...)

Then, combine these results into a single table with the different sites delineated in the site column.

my_final_results <- my_table1 %>% dplyr::union(my_table2) ... %>%
  dplyr::union(my_table_n) %>%
  dplyr::mutate(output_function = '{see table above}')

Multi-Site Anomaly Detection Analysis

Cross-Sectional

First, execute the Single Site, Anomaly Detection, Cross-Sectional analysis, configured appropriately for your study, against each data source.

library(patientrecordconsistency)

my_table <- prc_process(cohort = my_cohort,
                        multi_or_single_site = 'single',
                        anomaly_or_exploratory = 'anomaly',
                        time = F,
                        ...)

Then, combine these results into a single table with the different sites delineated in the site column.

my_final_results <- my_table1 %>% dplyr::union(my_table2) ... %>%
  dplyr::union(my_table_n)

Finally, use this combined table as input to the compute_dist_anomalies and detect_outliers functions, both available through the squba.gen package. The p_value can be selected by the user.

# First, execute the compute_dist_anomalies function
df_start <- compute_dist_anomalies(df_tbl = my_table,
                                   grp_vars = c('fu_bin'),
                                   var_col = 'jaccard_index',
                                   denom_cols = c('fu_bin'))

# Finally, use that output as input for the detect_outliers function
df_final <- detect_outliers(df_tbl = df_start,
                            tail_input = 'both',
                            p_input = p_value,
                            column_analysis = 'jaccard_index',
                            column_variable = 'fu_bin') %>%
  dplyr::mutate(output_function = '{see table above}')

Longitudinal

Start by executing the same steps as the exploratory analysis. Then, apply some additional formatting (below) to the final combined table and pass it into the ms_anom_euclidean function, available through the squba.gen package.

## Additional formatting for the table
event_categorization <- prc_tbl %>%
  uncount(pt_ct) %>%
  mutate(stat_type = case_when(event_a_num == 0 & event_b_num == 0 ~ 'Neither Event',
                               event_a_num == 0 & event_b_num != 0 ~ 'Event B Only',
                               event_a_num != 0 & event_b_num == 0 ~ 'Event A Only',
                               event_a_num != 0 & event_b_num != 0 ~ 'Both Events')) %>%
  group_by(site, time_start, time_increment, 
           event_a_name, event_b_name, total_pts, stat_type) %>%
  summarise(stat_ct = n(),
            prop_event = stat_ct / total_pts) %>% ungroup()

## Apply Euclidean distance computation
df <- ms_anom_euclidean(fot_input_tbl = event_categorization,
                        grp_vars = c('site', 'stat_type'),
                        var_col = 'prop_event')