This document contains an epidemiological description of the *transmission risk level* used in the German [Corona-Warn-App](https://github.com/corona-warn-app/cwa-documentation) (CWA). As its name suggests, the transmission risk is an essential part when estimating the overall risk of a person to get infected in an exposure incident. Usage of the transmission risk level is specified in the [ExposureNotification API](https://developer.apple.com/documentation/exposurenotification) and in the [CWA Architecture](https://github.com/corona-warn-app/cwa-documentation/blob/master/solution_architecture.md#risk-score-calculation). In particular we use epidemiological information about COVID-19 from the literature to motivate the choice of levels for this parameter. To enhance transparency and reproducibility of the computations, we provide the mathematical derivations and the computations in one [Rmarkdown](https://rmarkdown.rstudio.com/) document. The methods sketched below are likely to be subject to change, once additional information about the characteristics of COVID-19 is obtained or as feedback from the use of the app arrives.
We are interested in the situation where a person **A** (potential infector) at time $t_0$ uploads information about being a laboratory confirmed SARS-CoV-2 case. The upload happens in terms of *A*'s *diagnosis keys* [see @ExposureNotificationCrypto]. Each *diagnosis key* is associated with a particular day in A's history (current and past 14 days) and also has an optional *transmission risk level* from I--VIII [see @AndroidExposureNotificationsApi].
Users can periodically download *diagnosis keys* from the *diagnosis server*. For each app user **B** (potential infectee), who downloads the list of valid *diagnosis keys* and discovers he or she has been in contact with *A*, a risk assessment will be made.[^a2b] This risk assessment is operationalized by the *total risk score*, of which one component is the *transmission risk level* $\lambda_A$ computed by *A*. The *transmission risk level* (provided by *A*) and its associated *transmission risk value* (set by *B*) are app-defined and should be based on the probability of transmission between the two persons being in close contact. In the present document we interpret this probability as a function of epidemiological information about *A* and the time of contact. Information about the closeness and the duration of the contact are not considered part of the transmission risk component, because they are handled separately in the computation of the *total risk score*. For more information on how to calculate the *total risk score* see the [Exposure Notification API](https://developer.apple.com/documentation/exposurenotification).
As the *transmission risk level* is computed on *A*'s device, additional information about *A* such as *being symptomatic or not*, *date of onset of symptoms*, *date of sampling* or *date of test result* could be used to estimate the **infectiousness of A** more precisely. We will use the currently known characteristics of COVID-19, especially its [infectiousness profile due to viral shedding](#viral-shedding) and the [operational delays](#delays) of its handling to estimate the infectiousness at certain times. We can obtain the information on how many days ago from **now** (i.e. $t_0$) the contact between *A* and *B* was: Let $t_C$ be the time of contact between *A* and *B*, then the contact was $d=t_0-t_C$ days ago. The aim of this document is thus to parametrize the time-dependent infectiousness of *A* as a function of $d$.
The better we can assess the probability of a transmission from *A* to *B*, the more accurate is the *combined risk score*[^risk] that is used to warn the user to take further action, e.g., to contact a local health authority. Having digital support for this type of contact tracing appears helpful in order to obtain a more complete coverage of contact tracing and to do this much quicker.
The present document is structured as follows. In section [Scenarios](#scenarios) we distinguish between four possible information states about *A* at the time of upload[^upload] depending on whether an onset of COVID-19 symptoms has occurred or not and whether this information can be used or not. We calculate a transmission risk for each of the four cases. However, since the initial version of the app did not allow a distinction between cases 1--4, and since the *transmission risk level* is only one of 4 components for the *total risk score*, we thus normalize the *transmission risk level* in a so called [base case](#base_case), which was used by the initial version of the app and is still used if no information on symptoms is provided. A section [Discussion](#discussion) summarizes the results and points out important limitations.
[^a2b]: Throughout we will assume that the contact between *A* and *B* was such that *A* infected *B*. At this point we explicitly ignore the possibility that *B* actually infected *A*.
[^risk]: For a more detailed description of the connection between the *total risk score* and the *combined risk score* see the [risk score calculation section](https://github.com/corona-warn-app/cwa-documentation/blob/master/solution_architecture.md#risk-score-calculation) of the [solution architecture document](https://github.com/corona-warn-app/cwa-documentation/blob/master/solution_architecture.md).
[^upload]: Due to technical restrictions, the upload of the diagnosis keys may not be a single-point-in-time event, but may happen over two consecutive days, without further interaction with the user after his or her initial consent. Given that the time point of the user's consent for upload may also be the last opportunity for providing additional information such as date of onset of symptoms, we use "consent for upload" when calculating delays and refer to it as simply "upload" throughout this document.
Since the API allows for a customization of the transmission component $\lambda_A$ in the above, we shall study it in more detail here. Particular interest will be in four information scenarios about *A*, the potential infector, at the time of the upload.
For an infected person the following sequence of event times occurs (but not necessarily in the given order):
* $T_E=T_{\text{infection}}$: transmission of SARS-CoV-2 to an **e**xposed person *A* from some unknown source
* $T_I=T_{\text{infectious}}$: start of the infectious period in person *A*, i.e. *A* is able to infect others
* $T_S=T_{\text{symptoms}}$: onset of symptoms in person *A* (also referred to as **DSO**, day of symptom onset)
* $T_P=T_{\text{sampling}}$: time of sampling of person *A* [dt. **P**robenentnahme]
* $T_R=T_{\text{result}}$: time of *A* obtaining the positive test result
* $T_U=T_{\text{upload}}$: time where person *A* uploads the positive test result to the system
Note that before observation these times are to be considered as random quantities and, hence, are denoted by an uppercase letter. Once observed, a lower case letter shall be used.
Furthermore, note that for SARS-CoV-2 it is likely that $T_I$ occurs before $T_S$, but it cannot be ruled out that the order is reversed. For never-symptomatic cases $T_S$ will never occur, but $T_I$ might already have occurred or will occur in the future. For pre-symptomatic cases $T_S$ lies in the future, i.e. $T_S > t_0$, but $T_I$ might already have occurred or lie in the future. From our above description we would usually have $T_{\text{upload}}=t_0$. In a later section we will study the delays between the different event times. For this we will have to use the [convolution of random variables](#convolution) to get the distribution of their sum.
To derive the transmission component $\lambda_A$ we will distinguish persons who will eventually develop symptoms and those which are never-symptomatic. The former set will be denoted $\mathcal{Symp}$ and the later $\mathcal{N\!symp}$. Throughout the text we will use the shorthand notation $\mathcal{N\!symp}_A$ to denote the event that person *A* belongs to the set of never-symptomatic. Likewise for $\mathcal{Symp}_A$. We reserve the notion of *never-symptomatic* for those persons who never develop symptoms, whereas *pre-symptomatic* at a particular time $t$ are those who will eventually develop symptoms, but at a later point in time than $t$. If $T_S^A$ denotes the time of symptom onset in *A*, we will use the shorthand notation $\mathcal{Symp}_A(T^A_S > t)$ to denote the event that *A* belongs to the set of individuals who will eventually develop symptoms, but the onset of these symptoms has not yet occurred by time $t$, i.e.
Likewise we define the event $\mathcal{Symp}_A(T^A_S \leq t)$. Both never-symptomatic and pre-symptomatic are characterized as being *asymptomatic* at a given time of reference $t$. We will use the following shorthand to denote this event:
which is just a formal way of saying that in order for *A* to be asymptomatic at time t, *A* either belongs to the set of never-symptomatic cases or *A* is pre-symptomatic at time $t$.
3. asymptomatic at time $t_0$, i.e. the event $\mathcal{Asymp}_A(T^A_S>t_0)$,
4. no knowledge of symptom status at time $t_0$ (the *base case*), i.e. the event $\Omega = \mathcal{Symp}_A(T^A_S \leq t_0) \cup \mathcal{Asymp}_A (T^A_S > t_0)$.
Differentiation of these scenarios requires person *A* to provide additional (optional) information on potential symptom onset and respective date. The 2nd scenario occurs, if *A* accepts to reveal that COVID-19 relevant symptoms have been observed by the time of upload, but *A* does not want to reveal (or does not know) the day of symptom onset. If *A* despite a positive test result either has not yet developed or never will develop any symptoms, then we would be in scenario 3. If *A* does not provide any additional information, this would lead to scenario 4.
## Infectiousness Profile due to Viral Shedding {#viral-shedding}
Infectiousness of COVID-19, i.e. how much infectious material is being shed, varies as a function of time since infection, the development of symptoms and (if available) the DSO, see for example @he_etal2020. Assuming the amount of virus shed by a symptomatic case *A* is described by the function $v_A(d)$, where $d$ is the days since DSO in *A*. We expect the function to be positive even for negative $d$ values, due to pre-symptomatic transmission. Note that the scale of $v_A(d)$ is in principle arbitrary for our purposes as we are interested in the amount of virus shedding compared to the potential maximum value which informs the eventual classification into the risk levels I to VIII. Note also that symptomatic cases with unknown date of onset and never-symptomatic cases are not handled by the function $v_A(d)$. These will be addressed later in this section.
The study by @he_etal2020 examines the temporal dynamics of viral shedding and infectiousness of symptomatic COVID-19 cases. They provide results informing on the transmission risk for contacts of symptomatic cases.
Based on 77 identified transmission pairs @he_etal2020 inferred the infector's infectiousness profile with respect to symptom onset by fitting a left-shifted gamma-distribution to the empirical frequency of observed transmissions occurring at $d$ days before or after symptom onset of the infector. The left-shifting of the gamma-distribution allows for potential pre-symptomatic transmission. The resulting infectious profile is shown in Fig. 1c (middle) in @he_etal2020 and is subsequently
corrected in @he_etal2020correction, suggesting that infectiousness starts as early as 5-6 days prior to DSO and ends 8 days after DSO, with peak infectiousness at one day before DSO.
Thus, for the profile function $v_A(d)$ we implement a discretized version of the infectiousness profile as inferred by the corrected version of @he_etal2020. The discretized profile for each day $d$ computes the mean infectiousness within $[d,d+1)$, for instance the value $v_A(-2)$ refers to the mean infectiousness within the interval $[-2,-1)$ days, i.e. with respect to time of symptom onset. The profile $v_A(d)$ is normalized such that the maximum infectiousness by day equals one, which occurs at $d = - 1$.
xlab("Time relative to symptom onset (Days) of infector") +
ylab("relative infectiousness") +
coord_cartesian(ylim = c(0, 1))
```
It should be noted that the inferred infectious profile by @he_etal2020 has some bias, since the infectiousness profile represents the conditional probability density of the time point of transmission relative to the infector's date of symptom onset, given that a transmission from A to B occurs. Thus, for transmission pairs with frequent contacts within a wide exposure window, this observed time of transmission merely represents the time of 'first successful transmission'. However, this eliminates the possibility for further transmissions (to be observed) between these persons, although the actual infectiousness of the infector might persist (and even further increase) after the observed transmission. The infectiousness was estimated to have completely vanished by 8 days after symptom onset, which might be just a consequence of this bias.
An indicator for this bias is the amount of virus shedding as a function of time since onset of symptoms, which is also provided within @he_etal2020. Virus shedding very likely corresponds well to the actual infectiousness of a symptomatic person and it was found that viral load only gradually decays until the end of the testing window (21 days after symptom onset), which suggests that infectiousness persists even beyond 8 days after symptom onset as inferred through the transmission pairs. However, a study by @woelfel_etal2020 suggests, that persistent viral shedding beyond 8 days after symptom onset does not yield viable viruses that could be amplified in cell culture, and @arons_etal2020 observed in a 'Skilled Nursing Facility' setting, that viable viruses were recovered more often in the days prior to symptom onset than in the subsequent days. Thus, the exact connection between the amount of shedding and the corresponding risk of transmitting the virus is still unclear.
Still, a more realistic infectiousness profile could be informed by also accounting for the duration of virus shedding, resulting in a longer duration of persisting infectiousness. Combining both of these data sources from @he_etal2020 with available and future data on virus viability in cell culture within a holistic statistical framework might yield such an adjusted infectiousness profile curve, which should be utilized in a future adoption of the transmission risk calculation.
Note that the chosen probabilities in these two distributions are no more than educated expert guesses. They have to be adjusted based on real data once the system is running.
### Sampling to Upload
The convolution of the time from (sampling to getting the result) and (getting the result to uploading) leads to the following distribution for the time between sampling and upload:
```{r d_samp2upload, fig.cap = "PMF of the duration between sampling and upload."}
To compute the delay between onset of symptoms and and upload we need a distribution for the time between symptom onset and time of sampling. Note that for this distribution we assume that the infected person gets tested *because* of the developed symptoms and thus this delay cannot be negative. For the alternative scenario, in which a person gets tested for other reasons, the onset of symptom might occur after taking a sample. This case is treated further below.
We assume $T_{\text{sampling}} - T_{\text{symptoms}}$ for symptomatic individuals is a discrete distribution with the following PMF:
ggplot_pmf(d_symp2upload) + xlab("Duration Symptom Onset until Upload (Days)")
```
### Upload to Symptoms for Pre-Syptomatic
Here, we cover the scenario in which people get tested independent of having developed any symptoms beforehand and in case of a positive test result, upload their diagnosis keys. This could, e.g., be the case for individuals which are tested as part of contact tracing, because they were identified as a contact at risk after having had contact with an infected person. For these cases the time from symptom onset to upload may be in fact negative, since symptoms might begin after receiving the test result and upload. Note that we here condition on DSO actually occurring, which is not the case for truly asymptomatic cases.
To compute the delay distribution for this case we define the delay distribution from sampling to DSO for cases which were tested within their pre-symptomatic phase and will develop symptoms afterwards. Since a PCR-test result is most likely to be positive in a phase of considerable viral shedding, we assume that only tests within the pre-symptomatic phase will come out positive, which means that only tests from samples taken within the 7 days prior to DSO will be positive. Thus, as we have no further information when suspected cases get tested during their pre-symptomatic stage, we assume that it takes between 1 and 7 days from sampling to symptom onset, each with probability $1/7$. In other words, the sampling occurs 1 to 7 days _prior_ to DSO for pre-symptomatically tested. Thus the time from DSO to sampling is negative. We denote the distribution of this delay by `d_symp2samp_presymptomatic`.
By convoluting the (negative) time delay from symptom onset to sampling `d_symp2samp_presymptomatic` with the delay from sampling to upload `d_samp2upload` we obtain the overall time delay from upload to symptom onset, denoted by `d_upload2symp`. Note that this delay may be negative or positive, since some cases, although tested in a pre-symptomatic stage, might have already developed symptoms at the time of upload.
```{r d_upload2symp, fig.cap = "PMF of the duration between DSO and upload."}
Assume that we know that *A* got a positive test result, which is uploaded to the server on day $t_0$. All contacts that occurred $d = t_0 - t_C$ days ago will be assigned the same *transmission risk level* $\lambda_A(d)$, containing information on the infectiousness of a generic contact of *A* on that day. The infectiousness and therefore also the risk level $\lambda_A(d)$ might depend on further information provided by *A*, in particular whether *A* developed symptoms by the time of upload and in that case the day of symptom onset $T^A_S=t^A_S$, which would imply that $t^A_S \leq t_0$.
The *transmission risk level* $\lambda_A(d)$ is derived in two steps:
1. First, we compute the _raw relative infectiousness_ (also referred to as _transmission risk_) $\lambda^{(raw)}_A(d)$ of *A* given the provided information, which serves as a continuous infectiousness value on the scale $[0,1]$. For this step we distinguish between the four possible scenarios of available information introduced in section [Scenarios and Events](#scenarios).
2. In a second step, this raw infectiousness is translated into a *transmission risk level* $\lambda_A(d)$, which is denoted by roman numerals from $\text{\Romannum{1}}$ to $\text{\Romannum{8}}$ and which will be further translated into a transmission risk value to be used within the *total risk score* calculation. This classification will be explained in section [Transmission Risk Levels](#trl).
If the date of symptom onset is available, i.e. in the event $\mathcal{Symp}_A(T^A_S = t^A_S)$, then at time $t_0$, the onset of symptoms in the primary case *A* happened $t_0 - t^A_S \geq 0$ days ago. So the relative infectiousness of case *A* relative to $t_0$ is just $v_A(-d)$ shifted $t_0 - t^A_S$ days to the right, i.e $v_A(-d + (t_0 - t^A_S)) = v_A(t_C- t^A_S)$.
```{r d_infprofile_t0, fig.height=5, warning=FALSE, fig.cap = "Plot of the infectiousness profile, if symptom onsets happened 1 (a) or 2 (b) days before upload."}
where $d = t_0 - t_C$ refers to the duration in days between the time of exposure and the time of upload. Again, here we only provide the raw value $\lambda^{(raw)}_A$, i.e. the relative infectiousness on a $[0,1]$-scale. These raw infectiousness values as a function of DSO and time since exposure are shown below.
```{r ggplot_rel_inf, echo=FALSE, message=FALSE, fig.width=8, fig.height=8, fig.cap="Raw transmission risk for Case 1: A is symptomatic with DSO provided."}
### Case 2: Symptomatic, but no Availability of $t_S^A$, only Day of Upload
Here we consider the case, in which we know that *A* is symptomatic at time of upload, but do not know the exact DSO, i.e. the event $\mathcal{Symp}_A(T^A_S \leq t_0)$. In this case, we can infer a probability distribution for the DSO $T_S^A$ from our knowledge about the duration between DSO and upload given by `d_symp2upload` and knowing the date of upload $t_0$. We therefore define the raw transmission risk as the conditional expectation of $\lambda_A^{(raw)}$ given that *A* was symptomatic somewhere before $t_0$. We obtain this expectation by marginalizing over the possible DSO $T_S^A$, i.e.
Due to marginalization, the resulting raw transmission value (as a function of $d = t_0 - t_C$) is flatter and wider compared to a raw transmission value curve for a known date of symptom onset $t_S^A$ (see case 1). In particular, the maximum raw infectiousness value is reached at `r round(max(df_M_case2$M), 2)` compared to 1 in the case of a known DSO. This difference shows the additional gain in risk assessment within the case of knowing the DSO.
Note that for intermediate scenarios in which some additional information, like the date of sampling, is provided by *A*, one can compute the raw transmission value similarly by marginalizing out the date of symptom onset with respect to the corresponding delay distribution. This again could yield a more informative value compared to relying on the upload date alone.
It might be, that a person does not remember the DSO precisely, but can only constrain the DSO to lie within a certain period of time, e.g. only the week it was in ("in the last 7 days", "1-2 weeks ago", "more than 2 weeks ago"). In this case, we can assume an uniformly distributed DSO within that period or with a slightly skewer (geometric) distribution (motivated by the [Forgetting Curve](https://en.wikipedia.org/wiki/Forgetting_curve)). In either case we we can argue similarly as before:
where $[t_{\text{min}}, t_{\text{max}}]$ is the interval in which we know the DSO to be located in (e.g. the week) and $d_{\texttt{missingDSO}}(t_S^A)$ is the probability mass function representing the belief in where the true DSO value is located in the interval $\{t_{\text{min}}, \ldots, t_{\text{max}}\}$.
The difference in the geometric and the uniform approach is marginal. In the current week the later days are slightly rated higher, since we can assume that more recent days are remembered more often correctly and thus given as a precise date, not as a period. For the previous and penultimate week the differences to the uniform case almost disappear:
```{r m_missing_dso_dists, fig.height=2, fig.cap="The probability mass function representing the likelihood of the DSO.", include=TRUE, echo=FALSE}
rem <- .1
dist0 <- Reduce(
function(x, y, ...) merge(x, y, all = TRUE, ...),
list(
tibble(
i = 0:6,
y = d_missing_dso_unif(7),
v = "u"),
tibble(
i = 0:6,
y = d_missing_dso_geom(0,6,rem),
v = "0"),
tibble(
i = 0:6,
y = d_missing_dso_geom(7,13,rem),
v = "1"),
tibble(
i = 0:6,
y = d_missing_dso_geom(14,20,rem),
v = "2")))
ggplot(dist0, aes(i,y)) +
geom_point(aes(colour = v), size = 2) +
geom_line(aes(colour = v), size = 1) +
labs(y = "Probability", x="Day", color="", title = "Probability Mass Function") +
scale_colour_brewer(palette = "PiYG",
labels = c("in the last 7 days", "1-2 weeks ago", "more than 2 weeks ago", "uniform")) +
```{r m_case2_and_plt_w0, fig.height=2, fig.cap="Raw transmission risk for Case 2: A is symptomatic with DSO only known to be in the current week (in the last 7 days)."}
```{r m_case2_and_plt_w1, fig.height=2, fig.cap="Raw transmission risk for Case 2: A is symptomatic with DSO only known to be in the previous week (1-2 weeks ago)."}
```{r m_case2_and_plt_w2, fig.height=2, fig.cap="Raw transmission risk for Case 2: A is symptomatic with DSO only known to be in the penultimate week (more than 2 weeks ago)."}
For this case we consider the scenario in which *A* at the time of upload provides the information that there was no onset of symptoms so far, i.e. the event $\mathcal{Asymp}_A(T_S^A > t_0)$. This case implies the two mutually exclusive situations:
At the time of upload we will not know which of these two situations applies. For the purpose of computing a transmission risk for case 3 we thus require the probabilities for each of the two scenarios (and for the pre-symptomatic scenario for each sub-scenario $T_S^A = t_0 + n$, for $n \geq 1$) which will then be appropriately weighted together. Thus, let
denote the probabilities for the different sub-cases of the pre-symptomatic scenario, i.e. $w_n$ is the conditional probability for the event that *A* develops symptoms $n$ days after the upload date, where $n \in \{1, \ldots, N\}$ with $N$ being some plausible maximum time delay between upload date and DSO. Analogously, we define
In order to compute $w_n$ we thus require the probability that symptom onset occurs $n$ days after the upload date given that onset of symptoms will happen, i.e. $P(\mathcal{Symp}_A(T_S^A = t_0 + n) \>|\> \mathcal{Symp}_A)$. This is covered by the delay distribution `d_upload2symp` introduced further above. Furthermore we need the probability for any infected person to develop symptoms at all, i.e. $P(\mathcal{Symp}_A)$. This probability was in @mizumoto_etal2020 estimated to be $\hat{P}(\mathcal{Symp}_A) = `r p_symp`$. For $P(\mathcal{Symp}_A(T_S^A > t_0) \>|\> \mathcal{Symp}_A)$ we rely on the random time from data upload to symptom onset for suspected cases, who were tested in a pre-symptomatic stage.
Since according to `d_upload2symp` the maximum delay between upload and DSO is three days we set $N=3$. By plugging in the individual numbers we can compute the scenario probabilities $w_n$ for $n \in \{1,2,3\}$. For instance, for $w_1$ we obtain
Knowing the probabilities of the possible scenarios and sub-scenarios in the case of a person being non-symptomatic at the time of uploading a positive test result, we can compute the transmission risk for case 3 by applying the [law of total probability](https://en.wikipedia.org/wiki/Law_of_total_probability), i.e.
Thus, the first term shows that we require the transmission risk function for never-symptomatic people. Here we assume, that never-symptomatic cases have the same infectiousness profile (as a function around the date of upload) like symptomatic people, but with their infectiousness reduced by a factor of `r factor_asymp_reduce_infectiousness` [see @mizumoto_etal2020]. However, note that in this case we cannot rely on the computations from case 2 (symptomatic with unknown DSO), since here the DSO does not follow the distribution `d_symp2upload`, which applies for people who were known to be symptomatic at the time of upload. For this case we instead focus on people who were tested in an non-symptomatic stage, such that the DSO (for people who develop symptoms) is distributed according to `d_upload2symp` subject to the upload date $t_0$. Utilizing the transmission risk from case 1 with known DSO, this yields
If person *A* has not developed any symptoms up to the day of upload, the transmission risk is highest at the day of upload itself, and gradually decays for each day that a risk contact lays further in the past.
### Case 4: Unknown Symptom Status {#base_case}
The final case represents the scenario in which no further information is available beyond the fact that person *A* uploaded a positive test results at time $t_0$. This case covers several possible scenarios from above and summarizes them into a condensed transmission risk.
We decompose the case into the following possible mutually exclusive scenarios and apply corresponding probabilities for each subscenario:
(i) $\mathcal{Symp}_A(T^A_S \leq T^A_{\text{sampling}})$, i.e. person *A* got tested because they developed COVID-19 related symptoms,
(ii) $\mathcal{Asymp}_A(T^A_S > T^A_{\text{sampling}})$, i.e. person *A* got tested as a suspected case, e.g. because of a risk contact, and was non-symptomatic at the time of sampling. This can be decomposed into:
(b) $\mathcal{N\!symp}_A \cap \mathcal{Asymp}_A(T^A_S > T^A_{\text{sampling}})$, i.e. person *A* is a never-symptomatic case and will thus never develop symptoms.
The transmission risk as a function of time difference $d$ between day of upload and day of contact for each of these scenarios can be derived based on results and considerations from the above cases 1 to 3.
For (i) we can apply the risk value function from case 2 (symptomatic with unknown DSO), i.e.
For (ii.a) we can rely on the same considerations as made for case 3 (non-symptomatic at time of upload). Again, note that for the here reflected scenario we assume that person *A* was pre-symptomatic at time of testing. For considering the potential DSOs, we therefore again rely on the delay distribution between upload and DSO used for pre-symptomatically tested people, i.e. `d_upload2symp`. Thus, we compute the expected risk value by marginalizing over the DSO which follows `d_upload2symp` subject to the given date of upload $t_0$, i.e.
For (ii.b) we follow the same arguments as in case 3 for never-symptomatic cases, which effectively yields $\lambda_A^{(raw)}(d, \mathcal{Symp}_A(T^A_S > T^A_{\text{sampling}}))$ times a factor for the reduced relative infectiousness, i.e.
To define the overall transmission risk for case 4, we require probabilities for how likely each of the distinct scenarios is to occur. In order to weigh scenarios (ii.a) and (ii.b) we again rely on the probability that a proportion of `r scales::percent(p_symp)` of infected cases develop symptoms (@mizumoto_etal2020). This leaves the question on what proportion of cases, who upload their positive test result, were already tested while being asymptomatic. For now, we can only assume that this proportion is rather low, at $p_{\texttt{suspect}} = `r p_suspect`$. However, the proportion $p_{\texttt{suspect}}$ is likely subject to change over time, in particular due to the impact of the _Corona-Warn-App_ itself, as it has the primary purpose of identifying potentially infected people and initiate corresponding tests in an early stage of their course of infection. Altogether, this leads to the transmission risk for case 4 being defined as follows.
Note that case 4 represents the baseline case that is applied if the app does not or cannot collect any additional information about person *A*, which will apply at least for the initial public version of the _Corona-Warn-App_.
## Transmission Risk Level {#trl}
The final step for generating the *transmission risk level* is to translate the raw transmission risk from the continuous scale $[0,1]$ into a discrete scale (categories from _I_ to _VIII_ in roman numbers).
To do so, we set equidistant thresholds within the $[0,1]$-scale, such that the scale is decomposed into 8 equal-sized intervals. A raw transmission risk $\lambda_A^{(raw)}(d, \cdot)$ within such an interval is then mapped onto the respective level $\lambda_A(d, \cdot)$, where the $\cdot$ stands for one of the four case scenarios, i.e.
where $l \>\in \{1, \ldots, 8\}$. Note: we define that raw values with $\lambda^{(raw)}_A(d, \cdot) = 1$ are also classified as *transmission risk level* VIII.
### Transmission Risk Levels for Cases 1--4
This classification rule leads to the following risk levels for the raw risks obtained from each of the above considered case scenarios.
```{r plot_case1, echo=FALSE, message=FALSE, fig.cap="Transmission risk level for Case 1: A is symptomatic with DSO provided.", fig.width=5.5, fig.height=8}
df_M_case1_discrete <- discretize_matrix(M_case1)
plot_risk_levels(
df_M_case1_discrete,
breaks_y = 0:21,
ylab_text = "Days since symptoms in A"
)
```
```{r plot_case2, echo=FALSE, message=FALSE, fig.height=1.3, fig.width=5, fig.cap="Transmission risk level for Case 2: A is symptomatic but with unknown DSO."}
```{r plot_case3, echo=FALSE, message=FALSE, fig.height=1.3, fig.width=5, fig.cap="Transmission risk level for Case 3: A is non-symptomatic at day of upload."}
```{r plot_case4, echo=FALSE, message=FALSE, fig.height=1.3, fig.width=5, fig.cap="Transmission risk level for Case 4: no information about A at upload."}
For the DSO only known by the week it was in, we get:
```{r plot_case2_w0, echo=FALSE, message=FALSE, fig.height=1.3, fig.width=5, fig.cap="Transmission risk level for Case 2: A is symptomatic but with DSO only known to be in the current week (in the last 7 days)."}
```{r plot_case2_w1, echo=FALSE, message=FALSE, fig.height=1.3, fig.width=5, fig.cap="Transmission risk level for Case 2: A is symptomatic but with DSO only known to be in the previous week (1-2 weeks ago)."}
```{r plot_case2_w2, echo=FALSE, message=FALSE, fig.height=1.3, fig.width=5, fig.cap="Transmission risk level for Case 2: A is symptomatic but with DSO only known to be in the penultimate week (more than 2 weeks ago)."}
The initial version of the _Corona-Warn-App_ does not offer the possibility to provide additional information on symptom status when uploading a positive test result. Thus, the transmission risk is always calculated as in case 4, where no further information is available, i.e. the event $\Omega$.
For the risk level calculation, the transmission risks are classified into levels based on the highest achievable value from case 1, in which the DSO is known. This leads to the highest possible *transmission risk level* in case 4 being at `r df_M_case4_discrete$M %>% max %>% as.roman`. However, since no additional information is provided for the initial version of the app, this case does never actually apply. Due to averaging over all possible scenarios in case 4, selecting the transmission level from case 4 in the above calculations implies that the *transmission risk level* has an upper level of `r df_M_case4_discrete$M %>% max %>% as.roman`. Therefore the app does not make use of the full level scale for the transmission risk, in particular levels VII and VIII are effectively never used. This could mean that the transmission risk weights less in the subsequent *total risk score* computations.
To address this issue we re-adjust the level classification by accounting for the maximum possible transmission risk achieved within case 4. For this purpose we will also refer to case 4 as the _base case_ since it represents the only possible case within the initial app version. Thus we obtain
where $l \>\in \{1, \ldots, 8\}$ and where again raw values with $\lambda^{(raw)}_A(d, \Omega) = \max_{d \in\{0\ldots,`r max_dse`\}} \lambda^{(raw)}_A(d, \Omega)$ are also classified as level 8. This leads to the following risk levels for this single base case:
```{r plot_level_case_4, echo=FALSE, message=FALSE, fig.width=5, fig.height=1.3, fig.cap="Transmission risk level for the Base Case."}
plot_risk_levels(df_M_isolated_case4_discrete, title = "Base Case")
```
These are the *transmission risk levels* to be used for the initial app version. Once the app offers the possibility to provide additional information, which then enables the cases 1 to 3, we advise to fall back to the case-specific risk levels $\lambda_A(d, \cdot)$ as defined above (and maybe adjust alarm thresholds within the app accordingly).
# Discussion {#discussion}
The present document reflects an epidemiological motivation of the *transmission risk level* used as part of the Corona-Warn-App. In particular we show how to inform this parameter by epidemiological information about COVID-19 and the processes leading to a confirmed diagnosis in order to identify periods of increased infectiousness. A stochastic framework was used to characterize the sequence of events. However, any such stochastic model is subject to assumptions and parameter choices, which need to be carefully evaluated. Given the present state of information about COVID-19 and the current simultaneous demands on resources, such an extensive evaluation and sensitivity analysis was not possible. As a consequence the present document reflects our best knowledge at the time of publication, but we strongly urge that relevant areas are identified for further evaluation. In particular this could consist in quantifying some of the operational delays from real-world data, once the app is in use.
With regards to some of the other parameters, a special challenge lies in the decentralized structure of the app. This was a non-negotiable design decision in the run-up of developing the app. However, in order to improve the app, epidemiological studies with a sample of voluntary users could be designed, in order to evaluate the appropriateness of not just the *transmission risk level*, but the overall classificational approach of the app.
One component to evaluate continuously should be the infectiousness profile utilized in this model. Our current estimate inspired by @he_etal2020 contains several time- and context-specific quantities such as the contact structure and the resulting serial interval. As an example, the serial interval is a time-dependent quantity of an outbreak, e.g., because susceptible persons are depleted (@svensson2007), or because increased contact tracing efforts, reduce the period of infectiousness. Finally, the introduction and increased and successful use of the app may alter the very parameters that we estimated for its calibration.
Altogether, it would be important to continuously evaluate and update the infectiousness profile, e.g., by being better able to connect information about viral shedding with the risk of infection.
Another important limitation is that throughout the document we assumed that *A* infected *B*. However, the likelihood and the consequences that it was *B* who infected *A* should be further investigated - this seems particularly important in a setting with many asymptomatic or pre-symptomatic transmissions.
One important point of this document is, that a better transmission risk assessment would be possible, if information about day of symptom onset of the person uploading his or her *diagnosis keys* would be available. The specific gain, for example in terms of sensitivity and specificity, is currently not provided, but could be quantified using simulation.
However, besides the epidemiological aspect two other aspects are important: Firstly, using additional information about *A* and mapping that to levels from I-VIII can be a privacy risk, because a potential attacker could reveal some information about *A* from level and time of upload. In the specific case, if provided, the DSO of *A* can probably be inferred in some situations. Secondly, the information about DSO will be self-reported. In our computations we assume DSO to be reported correctly, but misreports are likely to occur, which could also lead to serious bias in the *transmission risk level*. A malicious user can also falsely report his or her DSO to cause false alarms. Hence, one has to carefully balance these two aspects, e.g., by evaluating different formats of DSO specification in the App. For a thoroughly discussion of potential risk of a proximity tracing app see [Privacy and Security Attacks on Digital Proximity Tracing Systems](https://github.com/DP-3T/documents/blob/master/Security%20analysis/Privacy%20and%20Security%20Attacks%20on%20Digital%20Proximity%20Tracing%20Systems.pdf).
The epidemiological motivated *transmission risk level* has value in contact tracing beyond the App. Close work together with local health authorities and those actually performing the contact tracing in the field is, however, necessary to make this added-value more specific.
## Quality Assurance
The calculations in the present document reflect one of two implementations, which were made as part of a [two-mindset approach](https://staff.math.su.se/hoehle/blog/2017/09/02/pairprogramming.html) for quality assurance of the *transmission risk level* computations.
## Acknowledgements
We thank Eric H. Y. Lau for his quick and friendly response to some additional questions regarding underlying data of the @he_etal2020 publication.
# Appendix
## Convolution of Discrete Random Variables {#convolution}
In what follows we will use the [fact](https://en.wikipedia.org/wiki/Convolution_of_probability_distributions) that for two *independent* integer random variables $X$ and $Y$ with PMFs $f_X$ and $f_Y$, respectively, we will have for their sum $Z=X+Y$ that
$$
f_Z(z) = \sum_{x} f_X(x) f_Y(z-x).
$$
This can be handled in `R` using the following function, where we assume that the PMF is formulated as a vector where the elements denote the individual probabilities and the names of the vector contain the support of the distribution.
```{r convolute, purl=TRUE}
#' Function to convolute two discrete probability distributions with
#' support on 0, 1, ... I.e. we compute the PMF of Z = X + Y.
#'
#' @param fX - the PMF of X given as a named vector, where the names
#' represent the support X_min, ..., 0, 1, ..., X_max
#' @param fY - the PMF of Y given as a named vector, where the names
#' represent the support Y_min, ..., 0, 1, ..., Y_max
#'
#' @return A names vector containing the PMF of Z = X+Y.
```{r plot_case2_wX, include=FALSE, echo=FALSE, message=FALSE, fig.height=1.3, fig.width=5, fig.cap="Transmission risk level for Case 2: A is symptomatic but with DSO only known to be in a certain period."}
# Plot TRL table for all periods (2 til 18)
# (change max_dso to a higher value if you want the full range til 21)
