covid19br / nowcaster Goto Github PK

Through a simple linear model adjusted to the trajectories, we can infer the tendency of the epidemic curve, to do so we can construct functions that take the posteriors output from the main function nowcasting_inla and adjust the linear model and return the slope of the trajectories, to give a more quantitative tendency inference:

There is a bug on the use of parameter K. When the last week is incomplete and K = 0 the last week to correct delay is the last complete week. But when K>0, it should forecast the current (and incomplete) week and the following K-1 weeks, but it seems to forecast K weeks beyond the current. As as example, when K=1, it should forecast the current (and incomplete) week but it forecasts two weeks ahead, the current one + 1.

data(sragBH)

sragBH20 <- sragBH %>% 
  filter(
    # Excluding typing date before symptoms onset
    DT_SIN_PRI <= DT_DIGITA, 
    # Excluding no typing date
    !is.na(DT_DIGITA),
    # cases reported in 2020
    year(DT_DIGITA) == 2020
  ) 

max(sragBH20$DT_DIGITA)
weekdays(max(sragBH20$DT_DIGITA))

epiweek(max(sragBH20$DT_DIGITA))
# Incomplete week (recording ends on Thursday)

# Nowcasting ignores an incomplete week (OK)
nowk0 = nowcasting_inla(dataset = sragBH20,
                        date_onset = DT_SIN_PRI,
                        date_report = DT_DIGITA,
                        silent = T)
nowk0$total %>% 
  mutate(
    epiweek = epiweek(dt_event)
  )  %>% tail()



# Week to forecast: week 53 only, but it forecast weeks 53/20 and 1/21
nowk1 = nowcasting_inla(dataset = sragBH20,
                        date_onset = DT_SIN_PRI,
                        date_report = DT_DIGITA,
                        K = 1,
                        silent = T)
nowk1$total %>% 
  mutate(
    epiweek = epiweek(dt_event)
  )  %>% tail()


# Weeks to forecast: week 53 and 1, but it forecast weeks 53/20, 1 and 2/21
nowk2 = nowcasting_inla(dataset = sragBH20,
                        date_onset = DT_SIN_PRI,
                        date_report = DT_DIGITA,
                        K = 2,
                        silent = T)
nowk2$total %>% 
  mutate(
    epiweek = epiweek(dt_event)
  ) %>% tail()

We need on this line to verify which are date formats, the actual problem it is the onset dates are moved to the last complete day of the week .

Maybe a solution:

turn back the dates to its original format, which it is the inverser of the formula in dados_w()

Create new internal functions for nowcasting with and without spatial structured data, nowcasting_spatial for spatialy structured data and the existing nowcasting_age & nowcasting_no_age for non-spatial structured data

• nowcasting_spatial created and implemented
• stable and tested for the README

Hessian test is constructed to avoid problems when the INLA solver stack in negative eigenvalues, to avoid it we can work on the length iteration steps. The test can be implemented through a while test until it reach a certain amount of tries. Follows a schematic test to be put on the nowcasting_age and nowcasting_no_age functions:

hess.min <- -1
h.value <- 0.01
h.trials <- 0.001
trials <- 0
while (hess.min <= 0 & trials < 50){
    
    # Running the Negative Binomial model in INLA
    output0 <- inla(model, family = "nbinomial", data = dados.age,
                    control.predictor = list(link = 1, compute = T),
                    control.compute = list(config = T, waic=F, dic=F, openmp.strategy='huge'),
                    control.family = list(
                      hyper = list("theta" = list(prior = "loggamma", 
                                                  param = c(0.001, 0.001))
                      )
                    ),
                    num.threads = 6,
                    control.inla = list(h = h.value),
                    ...
                    
    )
    hess.start <- which(output0$logfile == 'Eigenvalues of the Hessian')
    hess.min <- min(as.numeric(output0$logfile[(hess.start+1):(hess.start+3)]))
    h.value <- h.trials + 0.001
    h.trials <- h.value
    trials <- trials + 1
    
  }

Correct the formulas for the nowcasting model on README to a more formal version:

Here for the non-structured model

Here for the structured model

@lsbastos

@rafalopespx nowcasting_age is the workhorse of the package, needs to be more friendly user

• Silent parameter is not working properly needs fixing
• Return.age is obsolete, needs fixing

Implement a trajectory flag on nowcasting_inla to give back the option trajectory devolution

Posterior sampling parameter implementation, through the parameter n on inla.posterior

Require INLA with the NAMESPACE and Warning if it is not installed:

• INLA load with the NAMESPACE
• Warning if it is not installed

Need to be changed to have all distinct dates

To guarantee a well written documentation and understandable articles, we have to review the articles presented in the site of the package:

• Get started!
• Forecasting
• Nowcasting for decision-making
• Structured data

Manage to make pkgdown site version working

To build more cleaverly the package, without problems when calling nowcasting_inla(), create another functions that will by called by nowcasting_inla to make nowcasting without stratifying and make nowcasting with clustering specified data.

Tentative names, nowcasting_no_stratum() and nowcasting_cluster(), add the parameters to call those functions on nowcasting_inla(), these new functions will have a structure similar to nowcasting_age()

• #6
• #19
• implement above functions on nowcasting_inla()
• test stability of any kind of data passed to nowcasting_inla()

👋🏽 I maintain the cran checks badges. Please change to the new cran checks badge URL (e.g., https://badges.cranchecks.info/worst/dplyr.svg). Old badges at (e.g. https://cranchecks.info/badges/worst/dplyr) will be unavailable as of Jan 1st 2023.

• Blog post from June 2022 announcing the change: https://recology.info/2022/06/cran-checks-badges/
• New badges repo: https://github.com/r-hub/cchecksbadges
• Old badges repo: https://github.com/sckott/cchecksapi

Create an article performing a forecasting using the parameter K > 0

• create a rmarkdown performing the forecasting
• build article
• build site

Stabilize the Imports for the package, now is import by the export specification on the NAMESPACE:

• Verify Imports needed
• NAMESPACE update
• Warning to certify if INLA package is installed

Add a column with the observed number of cases in the nowcaster outputs.

> aaa$total
# A tibble: 15 × 7
    Time dt_event   Median    LI    LS   LIb   LSb
   <int> <date>      <dbl> <dbl> <dbl> <dbl> <dbl>
 1    17 2019-10-26    154  154   154    154  154 
 2    18 2019-11-02    173  173   173    173  173 

Need to check the last date in the dataset and the last date in the nowcast estimates. It seems they don't match properly.

Use of native pipe into pipelines |>

Implement and test a model with covariate of tweets counts and google trends searches on the main function nowcasting_inla:

• creates new shadow function nowcasting_covariates()
• flag on the main function nowcasting_inla
• all together