Overall Prediction Model Structure
To generate projections for 2017 to 2040, we used data from 1990 to 2016 and modeled three disease groups found to be associated with high-sodium diets by the GBD (CVD and CKD at level 2 of the GDB hierarchical casualty structure, and SC at level 3). The latest GBD study (GBD 2017) included 195 countries and territories (including Japan) and estimated DALYs and other health indicators for 359 diseases and injuries for each year from 1990 to 2017. The results of the GBD study are available on a county-by-county basis, and the estimates are widely used by scientific researchers, policymakers, and other stakeholders in discussions on decision-making, prioritization, and strategic resource allocation. [9,10,11,12,13,14,15,16]The GBD hierarchy of causes has levels 1 to 4. The three cause groups at level 1 are communicable diseases, maternal and neonatal conditions, nutritional deficiencies, non-communicable diseases, and injuries. These are further subdivided into 22 disease and injury categories as level 2 of causes, which are then further subdivided into level 3 and finally into level 4 of causes, which is the most detailed and contains 293 disease groups. For example, ischemic heart disease is classified into non-communicable diseases (level 1), cardiovascular diseases (level 2), cerebrovascular diseases (level 3), and ischemic heart disease (level 4).
According to the GBD prediction research methodology [17]we developed a three-component model of disease-specific DALYs for the three diseases associated with high salt intake. The model consists of a component for change in key behavioral and metabolic risk predictors with salt intake as the main risk predictor of interest in this study, a component for per capita income, educational attainment, and total fertility rate under age 25 integrated into a socio-demographic index (SDI) expressed on a scale of 0 to 1, and an autoregressive integrated moving average (ARIMA) model to capture unexplained components correlated over time. Further details, including data sources and model equations, are provided below.
DALY and SDI data, 1990–2016
We used estimates of DALY rates per 100,000 population for CVD, CKD, SC, and SDI in Japan from 1990 to 2016 published in GBD 2017. [18]Detailed methods for estimating DALYs and SDIs are provided in the GBD 2017 overview publication. [18, 19]Data extraction and analysis were performed by sex (male, female, combined) and age group (20-49 years, 50-69 years, 70 years or older, all ages). The 0-19 years age group was not considered due to lack of risk prediction data (see below).
Behavioral and metabolic risk prediction data, 1990-2016
Based on data availability, we considered population-level average salt intake (grams/day) and the prevalence of smokers, drinkers, and obesity by sex and age group. These were obtained from the Japanese NHNS from 1990 to 2016 by sex and age group. The NHNS is a nationally representative household survey conducted annually by the Ministry of Health, Labor, and Welfare to clarify eating habits, nutrient intakes, and lifestyles at the population level in Japan. [20]The NHNS consists of three parts: 1) a physical examination, including blood tests, administered by a physician at a community center; 2) a face-to-face survey of household daily food records (weighted); and 3) a self-report lifestyle questionnaire (including smoking status and alcohol consumption) accompanying the dietary survey. Urinary sodium was not measured in the NHNS. A detailed description of the NHNS survey procedures is available elsewhere. [20, 21]Briefly, dietary intake surveys were conducted on designated days, excluding Sundays and public holidays. Trained enumerators (mainly registered dietitians) instructed the household representative (usually the person responsible for cooking meals) how to measure the amount of food and beverages consumed by household members using an open-ended record form. The allocation of dishes shared by each household member, food waste, leftovers, and eating out were also recorded, as were the portion sizes and quantities of foods consumed when measuring was not possible. Trained enumerators visited each household to ensure study participants’ compliance with the study, confirmed portion sizes when necessary, and converted estimates of portion sizes or food quantities. Each food was coded according to the dietary records and the corresponding food composition table in the Standard Tables of Food Composition in Japan, 6th Edition. [22].
In this study, salt intake (grams/day) was calculated as sodium (mg) × 2.54/1000. Obesity was defined as a BMI of 25 kg/m2 or higher according to the Japan Society for the Study of Obesity. [23, 24]Because the lifestyle questionnaire was not administered to younger people under 20 years of age, only people aged 20 years or older were included in this study. Analytical sample sizes ranged from 6149 to 26,594 from 1990 to 2016. A total of 275,468 people (127,571 men and 147,897 women) aged 20 years or older who completed the salt intake assessment were used to calculate the mean daily salt intake and prevalence of other predictors from individual surveys.
ARIMA Models for Forecasting
An ARIMA model was used to forecast future DALY rates by adjusting for multiple risk predictors. The model uses shifts and lags in historical information to generate forecasts based on its own past values in the time series (autoregressive, or AR term) and the error from previous forecasts (moving average, or MA term). The integral (I term) in the ARIMA model represents differencing the raw observations to make the time series stationary, i.e., data values are replaced by their difference from the previous value.
For ARIMA models, the standard notation is ARIMA with p, d, q, where integer values are used instead of parameters to represent the type of model written as ARIMA(p, d, q). The parameters are defined as follows: p is the order of autoregression, d is the order of differencing included, and q is the order of moving average. Zero values can be used as parameters to indicate that a particular component is not used in the model. For example, ARIMA(1, 0, 2) indicates that there is no differencing and the model has one AR term and two MA terms. As is commonly known, the ARIMA model is expressed as follows:
$$ \left(1-\sum \limits_{i=1}^p{\alpha}_i{L}^i\right){\left(1-L\right)}^d{y}_t=\left(1+\sum \limits_{i=1}^q{\beta}_i{L}^d\right){\varepsilon}_t,\kern0.5em $$
(1)
where yt is the outcome of interest. εt is the (white noise) error term, the residual defined as the time series of the difference between observed and forecasted values at time t. L is the time lag operator defined as Lkyt = yt − k. αi and βi are the ith coefficient parameters of p (AR part) and q (MA part). [25]The important point is that the time series model should have serial correlation in the observed data, and therefore the residuals themselves are independent and identically distributed with zero mean and covariance. If the left-hand side of equation (1) contains difference values, appropriate adjustments are also applied to the right-hand side. Before fitting the model, we used the Dickey-Fuller test to check whether the observed data of the time series are stationary, i.e., whether the data are constant over time. [25]If assumed to be likely non-stationary, the data were transformed into a stationary time series by taking the appropriate difference at order d. Autocorrelation and partial autocorrelation functions were used to identify stationary states and to determine the extent of the grid search for the model order. Model parameters were estimated using maximum likelihood methods. The Akaike Information Criterion (AIC) was calculated to select the best model by order.
To project future DALY rates, a two-stage approach was used. In the first stage, the values of each predictor were projected separately at the population level from 2017 to 2040 (i.e., SDI, mean salt intake (grams/day), obesity prevalence (%), current smokers (%), and current alcohol drinkers (%)) using equation (1). In the second stage, after substituting the projected values of the above predictors into xtj, the log-scaled DALY rate yt was projected using the following equation (2):
$$ \left(1-\sum \limits_{i=1}^p{\alpha}_i{L}^i\right){\left(1-L\right)}^d{y}_t=\sum \limits_{j=1}^4{\gamma}_j{L}^d{x}_{tj}+\left(1+\sum \limits_{i=1}^q{\beta}_i{L}^d\right){\varepsilon}_t, $$
(2)
where xtj is the value of the jth predictor variable at time t, and γj is the coefficient parameter of the jth predictor variable. Equation (2) is a general form of the so-called ARIMAX model (ARIMA with exogenous inputs), which captures the influence of external factors. [26]Widely adopted in epidemiological time series studies [27,28,29]ARIMAX has the ability to generate forecasts while identifying underlying patterns of change of both internal and external nature. All analyses were conducted using R version 3.6.1. The parameter sets in equation (2) were estimated separately for each age and sex category.
Future scenarios for salt intake
To evaluate the impact of salt intake on the DALY rates of three diseases (CVD, CKD, and SC) from 2017 to 2040, several future scenarios were assumed: a baseline forecast and three alternative scenarios (best, intermediate, and worse). In the baseline forecast, it was assumed that the current trend would be maintained; that is, future salt intake from 2017 to 2040 was projected using the ARIMA model defined in equation (1). In the best scenario, the daily salt intake target (8 g/day), which is the target of Health Japan 21 (Phase 2), will be achieved in 2023. [5] This will continue to decline until it reaches 5g per day in 2040, in line with WHO guidelines. [6]In this scenario, a constant monotonic decreasing function is assumed from 2017 to 2023 when 8 g/day is achieved, and a further monotonic decreasing function from 2024 to 2040 when 5 g/day is achieved. In the medium scenario, the intake below 8.0 g/day set in the best scenario is achieved in 2040 instead of 2023, and a monotonic decreasing function is assumed. The worst scenario is when the latest salt intake (i.e., the 2016 value) remains constant from 2017 to 2040. By inputting these assumed scenario values as predictors into equation (2), we can obtain the final projected DALY rates to 2040 for these alternative scenarios. Note that the salt intake in 2040 is by definition the same in the best and medium scenarios, and the projected DALY rates converge mathematically to the same value in 2040.