Adding care seeking rate

[OD1992]

@goldingn

I have modify static and dynamic covariates into .RData files in lines code: https://github.com/OD1992/epiwave.mapping/blob/master/NGA_example.Rmd#L237&242.
Currently, dynamic covariates are not included in the X matrix: https://github.com/OD1992/epiwave.mapping/blob/master/NGA_example.Rmd#L925

I want to include treatment seeking (a vector rows=553593) into the reporting fraction r to multiply by the clinical_case_incidence ( matrix 553593x36) in line code: https://github.com/OD1992/epiwave.mapping/blob/master/NGA_example.Rmd#L1052
I fail because of the dimensions.
**Note: this is fixed by

convert the vector to a greta array

s_ts <- as_data(all_s_ts) # greta array of shape (n)
s_ts_matrix <- s_ts %*% t(rep(1, 36)) # now 553593 × 36
reported_clinical_case_incidence <- s_ts_matrix * clinical_case_incidence # r * clinical_case_incidence**

I am still thinking how to create catchment weights for LGAs in line code: https://github.com/OD1992/epiwave.mapping/blob/master/NGA_example.Rmd#L1118.

In section: https://github.com/OD1992/epiwave.mapping/blob/master/NGA_example.Rmd:define_prevalence
I have error (Error in daily_new_infections[prev_index, ]: ! subscript out of bounds) when I run the code without doing this !is.na(prev_index).

I have modify in line code: https://github.com/OD1992/epiwave.mapping/blob/master/NGA_example.Rmd#L1149
prevalence_estimates <- prevalence_prev_locs[idx] (giving dim=3992 x 36) replaced by prevalence_estimates <- prevalence_prev_locs[idx[,2],] (giving dim= 1996 36).
But I still have this issue (Error in initialize(): ! incompatible dimensions: 1996x1, 1996x36) in line #L1155.

[goldingn]

Hi Ousmane. I think it will be easier for us to discuss and collaborate if you use a separate GitHub issue for each problem you need to resolve. For now, let's address these issues here. But for any new ones, please try to open a separate GitHub issue, with a clear description of the problem, and ideally a reprex. Then we can close the issue as each one is resolved.

For your first problem here, your solution works fine. Here's an alternative approach which might make the model run just a little faster (note that this is a reprex I made with a smaller fake dataset and rendered using the reprex package as Nick Tierney showed you):

library(greta)
#> 
#> Attaching package: 'greta'
#> The following objects are masked from 'package:stats':
#> 
#>     binomial, cov2cor, poisson
#> The following objects are masked from 'package:base':
#> 
#>     %*%, apply, backsolve, beta, chol2inv, colMeans, colSums, diag,
#>     eigen, forwardsolve, gamma, identity, rowMeans, rowSums, sweep,
#>     tapply

# example objects
all_s_ts <- runif(10)
clinical_case_incidence <- ones(10, 3)
#> ℹ Initialising python and checking dependencies, this may take a moment.
#> ✔ Initialising python and checking dependencies ... done!
#> 

# Ousmane's code
s_ts <- as_data(all_s_ts)
s_ts_matrix <- s_ts %*% t(rep(1, 3))
reported_clinical_case_incidence <- s_ts_matrix * clinical_case_incidence

# alternative code, probably more computationally efficient
reported_clinical_case_incidence2 <- sweep(clinical_case_incidence,
                                          1,
                                          all_s_ts,
                                          FUN = "*")

results <- calculate(reported_clinical_case_incidence,
                     reported_clinical_case_incidence2)

identical(
  results$reported_clinical_case_incidence,
  results$reported_clinical_case_incidence2
)
#> [1] TRUE

^{Created on 2025-04-22 with reprex v2.0.2}

[goldingn]

I am still thinking how to create catchment weights for LGAs in line code: OD1992/epiwave.mapping@master/NGA_example.Rmd#L1118.

From what I understand of your application, you don't want to use catchment fractions (weights) for each pixel. You are using data that is recorded at the LGA of residence. So for each pixel, instead of some fraction of cases from that pixel reporting in each of several 'clinics' (the fraction being the probability that a randomly selected case attends that clinic), you have the fraction of cases from that pixel that report in each LGA. Since you have the LGA of residence, you can just assume that fraction is 1 for the LGA that the pixel falls in, and 0 for all other LGAs. So the catchment weights matrix can just be 1s and 0s: 1 in matrix cells where the pixel corresponds to the LGA, and 0 where it doesn't. Then you can use the same matrix multiply to aggregate them.

Here is another reprex, showing how you can make the matrix from a vector of the LGAs that each pixel belongs to:

# make some fake data
n_pixels <- 20
n_lgas <- 9

# a fake vecctor of the LGAs to which each pixel belongs
pixel_lga_membership_vec <- sort(rep(seq_len(n_lgas), length.out = n_pixels))

# make an n_pixels x n_lgas matrix of the LGA membership
pixel_lga_membership_mat <- matrix(0, n_pixels, n_lgas)
membership_index <- cbind(seq_len(n_pixels),
                          pixel_lga_membership_vec)
pixel_lga_membership_mat[membership_index] <- 1

# check there is only a single 1 in each row (each pixel is only in one LGA)
all(rowSums(pixel_lga_membership_mat) == 1)
#> [1] TRUE

# test it on some fake data
n_times <- 5
reported_clinical_cases_shifted <- matrix(runif(n_pixels * n_times),
                                          n_pixels, n_times)

# use the membership as the weights
catchment_weights <- pixel_lga_membership_mat
reported_clinical_cases_hf <- t(catchment_weights) %*% reported_clinical_cases_shifted

# check this has the right dimensions
all(dim(reported_clinical_cases_hf) == c(n_lgas, n_times))
#> [1] TRUE

^{Created on 2025-04-22 with reprex v2.0.2}

[goldingn]

I don't understand your last issue, and I don't have a way to help you debug these sorts of dimension mismatches, since I don't have the objects (or time) to rerun all of your code. Could you try running it with some smaller example objects, like in my reprexes above, and running the code until you can reproduce the same error in a way that I can quickly reproduce it?