Flu seasons in residential aged care 2016-2022

Flu seasons in the general population

The 2020 and 2021 flu seasons in Australia did not arrive. In fact, in Figure 1 it is hard to even make out the lines representing the weekly number of confirmed flu cases for 2020 and 2021 (green and light blue) since they are so flat and close to zero. It is also clear from the graph that the 2022 flu-season started earlier than usual and had a higher peak.

Figure 1: Number of lab confirmed cases of influenza in Australia grouped by week, coloured by year. Figure from Department of Health and Aged Care: Australian Influenza Surveillance Report No. 10, 2022. The data is from NNDSS. (Department of Health and Aged Care 2022)

Flu seasons in residential aged care

The relevant quality indicator collected by MOA Benchmarking is called new respiratory tract infections (RTI) and it includes various infections: common cold, influenza-like illness, lower-respiratory tract infection, pneumonia. COVID-19 is now excluded from the definition since MOA has other tools for that. However, in this data there are some COVID-19 cases counted under RTI, but the mean profile for the aged care services included in this analysis are consistent with low COVID-19 cases. This was deduced by noting that the spike in January 2022 was small, despite being the peak for new COVID-19 cases in the general population.

Data was collected as the count of new respiratory infections in a month at an aged care service and the number of beds in that service at that time. The outcome of interest is the rate of new respiratory tract infections per 1000 bed days within each month computed on all aged care services included in the analysis. The calculation for the rate of new respiratory tract infections (RTI) in month M per 1000 bed days can be calculated from

$\text{Rate in month M} = \frac{\text{Sum of new RTI in M}}{\text{Sum of beds during M} \times \text{Days in month M}} \times 1000.$

To ensure that similar aged care services are analysed at the end of the period and the beginning I imposed the inclusion criteria that an aged care service had to have data for January 2018 and January 2022. This limited the sample size a lot compared to what would be available for benchmarking, but the data still tracked an average of 11,309 aged care consumers every month, which for a month with 30 days is 339,276 bed days.

Forecast of 2020 based on data available in 2019

The rate of respiratory tract infections per 1000 bed days for 2018 and 2019 is visualised in Figure 2. I observe that:

Infections are seasonal.
Infections peak around July and are at the lowest around February.

Figure 2: Monthly rate of RTI per 1000 bed days plotted up to the end of 2019.

We allowed the Seasonal ARIMA (SARIMA) model to be selected automatically using an information criteria (AICc) and the “auto.arima” function in the R package forecast (Hyndman et al. 2022). Using this method I arrive at the model that can be written ARIMA(0,0,1)(1,1,0)[12].

The first part, (0,0,1), indicates that current rates are affected by shocks in the previous month, i.e., shocks to the rate of RTI tend to linger for one month. The second part, (1,1,0)[12], indicates that current rate is a function of the rate 12 months ago and the difference between the rates 12 and 24 months ago.

Residuals pass the Ljung-Box test with a lag of 11 (p=0.824). The residuals are visualised in Figure 3. The correlelogram doesn’t show any serial dependence in residuals. The distribution of residuals appears to be somewhat gaussian, but with a high peak and left skew. However, on the whole, I conclude that residuals appear reasonable.

Figure 3: The residuals of the fitted seasonal ARIMA model.

The model can be written using the back-shift operator, B, as

$\underbrace{(1-\Phi_1 B^{12})}_{\text{Seasonal}\atop \text{AR(1)}} \times\underbrace{(1-B^{12})}_{\text{Seasonal}\atop\text{differencing}} \times y_t = \underbrace{(1+\theta_1B)}_{\text{Non-seasonal}\atop\text{MA(1)}}\times e_t,$

and expanding the model, I get the form

$y_t = y_{t-12} + \Phi_1 (y_{t-12} - y_{t-24}) + \theta_1 e_{t-1} + e_t.$

The parameters for this model estimated based on the training data are given in Table 1. The sign of the $\hat\theta$ indicates that a positive (negative) shock in the previous month tends to be associated with a positive (negative) shock in the current month. The $\hat\Phi$ indicates that the current rate is a combination of the rate a year ago and to a lesser extent the rate two years ago.

\hat\Phi — Table 1: The estimated parameters for the chosen SARIMA model.

$\hat\theta$	$\hat\Phi$
0.6	-0.358

The RTI forecast using this SARIMA model over the next two years using data available at the end of 2019 is visualised in Figure 4 with 95% confidence intervals. The point forecast looks quite reasonable, if a season tends to be similar to the previous season.

Figure 4: Forecasted rate of new respiratory tract infections based on the SARIMA model with a 95% confidence interval.

What actually happened

It is probably obvious that the above forecast would not have captured the rate of new respiratory tract infections in 2020 since the flu season did not arrive, due to lockdowns and social distancing measures. The actual observed rate of new respiratory tract infections in aged care services over the analysis period is shown in Figure 5. In 2020 and 2021 not only was the seasonal effect much weaker, but the overall level was lower. In Figure 6 that the difference between the rate during flu-seasons (winter) and out of season (summer) disappeared in 2020.

Figure 5: The observed rate of new respiratory tract infections plotted over time. The introduction of COVID-19 restrictions in Australia is indicated for context.

Figure 6: The observed rate of new respiratory tract infections in each complete year (bar), the rate computed only for winter months (black diamond), and the rate computed only for summer months (red square).

Actual rate compared with forecast

The first border restrictions in Australia in response to COVID-19 were introduced March of 2020 (Parliment of Australia 2020). In Figure 7 the forecast looks reasonable up to and including March 2020. At March 2020 there appears to be a structural break and the sinusoidal wave that was predicted to continue did not appear in 2020 and 2021. The forecast error plotted in Figure 8 makes the missing flu seasons even more apparent.

While the flu season didn’t arrive in 2020 and 2021 as per the data in the overall population (Department of Health and Aged Care 2022), there was still a level of non-flu respiratory tract infections observed in the population of aged care consumers.

In more recent months, i.e., May-July 2022, there was a high level of respiratory tract infections. In these months, compared to the forecast based on pre-2020 data, the rate of RTI was high and peaked earlier than usual. The same features of the 2022 flu season were present in the general population as well, which can be observed in Figure 1.

Figure 7: The observed rate of new respiratory tract infections plotted over time along with forecast.

Figure 8: The forecast error plotted over time, i.e., actual rate of RTI minus the forecasted rate.

The total cumulative deviation from the forecast at the end of 2021 was 0.372 per person. According to GEN Aged Care Data, the number of people using residential aged care as of 30th of June 2021 was 371,000 (Australian Institute of Health and Welfare 2022). Based on this, I estimate that about 138,000 respiratory tract infections were averted over 2020 and 2021.

Comparing directly with previous years

In addition to forecasting the rate of respiratory tract infections, I also compared 2020 and 2021 with previous years directly as in Figure 9. This leaves all the random variation of 2016-2019, but allows for an intuitive comparison with previous observed rates. The structural break after March 2020 can also be observed in this figure.

Figure 9: The observed rate of new respiratory tract infections plotted over time of year coloured by year.

Note

Different restrictions were present across states in response to COVID-19, so there are some differences in how RTI incidence changed by state. However, these differences do not alter the overall interpretation of this analysis.

Conclusion

The lack of 2020 and 2021 flu seasons in Australia, due to lockdowns and social distancing, were also observed in the aged care setting based on data on respiratory tract infections. However, a non-negligible number of new respiratory tract infections were still present throughout 2020 and 2021. Comparing forecasts to actual observations suggest there is a structural break in the rate of new respiratory tract infections after March 2020. In April 2022 the lower level of RTI incidence appears to have ended with the arrival of an earlier than usual flu season.

References

Australian Institute of Health and Welfare. 2022. “GEN Aged Care Data.” https://www.gen-agedcaredata.gov.au/Topics/People-using-aged-care#Aged%20care%20use%20in%20Australia.

Department of Health and Aged Care. 2022. Australian Influenza Surveillance Report No. 10, 2022.

Hyndman, Rob, George Athanasopoulos, Christoph Bergmeir, Gabriel Caceres, Leanne Chhay, Mitchell O’Hara-Wild, Fotios Petropoulos, Slava Razbash, Earo Wang, and Farah Yasmeen. 2022. “Forecast: Forecasting Functions for Time Series and Linear Models.” https://pkg.robjhyndman.com/forecast/.

Parliment of Australia. 2020. “COVID-19: A Chronology of State and Territory Government Announcements (up Until 30 June 2020).”