Electrical appliances moderate households’ water demand response to heat

Analysis of potentially interconnected residential water and energy demand is sparse. In a 1-in-10 random sample of Singapore households living in apartments, water use per capita declines over the socioeconomic distribution, whereas electricity use rises. Here I show that in this leading Asian city and tropical climate, water and electricity demand respond differentially to heat across different socioeconomic groups. When temperatures rise, water demand increases among lower-income households but remains unchanged among higher-income households. In sharp contrast, heat induces larger shifts in electricity demand among higher-income households. With air-conditioner penetration ranging from 14 to 99% across different socioeconomic groups, my interpretation is that water provides heat relief for households that have yet to adopt air conditioning. How Singaporeans’ resource demands respond to heat at different income levels can inform the future responses of a vast urban population on rising incomes living in the water-stressed tropics, in similar and warming climates.

. Ground-level wind patterns in Singapore a., Average wind speed (km h -1 ), b., Prevalence of wind from north (thick black line) and east (thin red line), c., Prevalence of wind from south (thick black line) and west (thin red line). The sample period is August 1, 2012 to December 31, 2015. For each day in the sample, "wind direction from north" is assigned: (i) a value of 1 when recorded wind direction (at 8 am, local time) lies between 330 degrees (from north, clockwise, direction from which the wind blows) and 30 degrees; (ii) a value of 0.5 when wind direction lies between 300 and 330 degrees or between 30 and 60 degrees; and (iii) a value of 0 otherwise. The other three wind direction controls are defined analogously. For better visualization, the plots show means over days within year by month. Source: NOAA Integrated Global Radiosonde Archive (IGRA Version 2).

. Allowing for a nonlinear response in the household demand model
95% confidence intervals for the coefficients on average temperature bins, converted to percent increase in use, obtained from two separate OLS regressions: a., log water use (m 3 month -1 ), and b., log electricity use (kWh month -1 ). The regression models and samples follow that in specification 4 of Table 1 in the main text, with one exception: the variable of interest, average temperature over the usage period, enters flexibly via bins (e.g., between 26.9 °C and 27.3 °C), Temp. ≥ 28.5 ∆ electricity use (% increase) vs. <26.9 °C Electricity demand response to heat (reference<26.9 °C) a b rather than linearly. The reference category is average temperature < 26.9 °C. The empirical support for average temperature is 26.1 to 29.1 °C. Regressions include controls for household, time (month, year, day-of-the-week, public holidays, school holidays), weather (relative humidity, dew point depression, wind speed, precipitation) and PM2.5.

Supplementary Fig. 5. Further robustness tests
Robustness to: a., Temperature variable, based on the proportion of days in the usage period with daily mean temperature above 28.5 °C, as in Supplementary Table 4, rather than the average temperature in the usage period. b., Dependent variables in levels, namely water use (converted to L day -1 ) and electricity use (kWh day -1 ), rather than the natural logarithm of resource use. c., Estimation by apartment type subsample: 2 resources × 6 apartment types = 12 separate OLS price was constant in 2012-15). f., Include interactions between apartment-type indicators and PM2.5, allowing for households' avoidance behavior to PM2.5 to vary by apartment type. Each panel reports 95% confidence intervals for the percent increase in water demand against the percent increase in electricity demand, by apartment type, for a +1 °C variation.
Taking Fig. 5a's specification (in the main text) as the point of departure: a., PM2.5 is instrumented with Southeast Asia fire activity variables and Singapore thermal gradients and wind direction (as in specification 6 of Table 1 in the main text). b., Similar specification as in a now including interactions between apartment-type indicators and PM2.5 (similarly, PM2.5 instruments), thus allowing for households' avoidance behavior to PM2.5 to vary by apartment type. c., Similar specification as in a now instrumenting for temperature (as measured by MSS) with temperature Electricity meter reading date in the sample a b Supplementary Fig. 11. Distribution of duration of actual use observations between successive meter readings in the sample.
An observation is a household by usage period. The density at 59-62 days, characterized by a meter reading frequency of once every 2 months by an SP Services meter reader, is 94% for both a., water use observations, and b., electricity use observations. The density at 28-31 days, typically due to some customers calling in to report actual readings within one month of a meter reader's visit, is 5% for both water use observations and electricity use observations. The modal duration in each distribution is 61 days, with a 63% density for both water use observations and electricity use observations. Duration is shown up to 200 days for better visualization. Source: SP Services microdata. Notes: An observation is a household by usage period among active accounts in the billing period from September 2012 to December 2015. SP Services randomized-a 1-in-10 random sampleamong all accounts that were active when the extraction code went into production, in August 2015. For each household in the random sample, SP Services provided usage by monthly billing cycle going back to September 2012, and subsequently tracked these accounts to December 2015. A string variable in the microdata informs apartment type, taking on the "values": HDB01 = 1room apartment, HDB02 = 2-room apartment, HDB03 = 3-room apartment, HDB04 = 4-room apartment, HDB05 = 5-room apartment, HDB06 = 6-room apartment, HDBEX = executive apartment, PTEAP = condominium apartment. HDB06 apartments are rare (only 51 identifiers in the microdata). The labels "HDB" and "PTE" denote the lead contractor for the development of apartments that are sold to individuals, whether the Housing Development Board or a private condominium developer. Source: SP Services microdata.

Supplementary Table 2. Household characteristics by apartment type
Mean annual household income per person 1000 US$ Living in 1-or 2-room apartments 9.3 Living in 3-room apartments 20.9 Living in 4-room apartments 23.9 Living in 5-room and executive apartments 29.7 Living in condominium apartments 68.9 Mean household size (occupancy) Persons Living in 1-or 2-room apartments 2.1 Living in 3-room apartments 2.7 Living in 4-room apartments 3.6 Living in 5-room and executive apartments 3.9 Living in condominium apartments 3.4 Three weights capture proximity to Singapore. First, the inverse distance from the hotspot's location to Singapore's centroid. Second, the cosine of the difference between wind direction in Singapore and the initial bearing from Singapore's centroid to the hotspot's location, where this wind direction-initial bearing difference is not to exceed 90 degrees in absolute value (the cosine weight is bounded from below by zero). Third, the interaction between the first two weights. Source data are provided as a Source Data file.  Notes: The table reports estimates for 24 water or electricity use regressions. An observation is a household by usage period in the 2012-2015 utility usage microdata. The dependent variable is log water use (m 3 month -1 ) in panels A and C, and log electricity use (kWh month -1 ) in panels B and D. The key regressor is the proportion of days with daily mean temperature above 28.5 °C in panels A and B (empirical support 0 to 83%), or above 29 °C in panels C and D (empirical support 0 to 77%). Proportions are taken over the same days that are concurrent to each usage observation. Other notes to Table 1 in the main text apply here exactly. OLS regressions in columns 1 to 5, 2SLS regressions in column 6. Standard errors (se), in parentheses, clustered by household. ***, **, * denote significance at the 0.01, 0.05 and 0.1 levels, respectively. Notes: The table reports estimates for four water use regressions. An observation is a household (HH) by usage period in the 2012-2015 utility usage microdata. The dependent variable is log water use (or water use, m 3 month -1 , in specification 4). The regressors of interest are average daily mean temperature (°C) and its interaction with average electricity use for the household's geographic area and apartment type, computed as mean electricity use over all household and time varying observations within postal code by apartment type in the sample. This time-invariant variable has empirical support 6 to 1,104 kWh month -1 and shifts by apartment type (six) within postal code (73 in total). I divide the interaction variable by 1000 in order to scale up its coefficient by this factor. In specification 3, average electricity use is adjusted for the mean household size within apartment type. As in Table 1 in the main text, all regressions include controls for household, time (month, year, day-of-the-week, public holidays, school holidays), weather (relative humidity, dew point depression, wind speed, precipitation) and PM2.5. Average market-level electricity use is included but subsumed in the household fixed effects. OLS regressions. Standard errors, in parentheses, clustered by household. ***, **, * denote significance at the 0.01, 0.05 and 0.1 levels, respectively.