Monitoring and moderating extreme indoor temperatures in low-income urban communities

Climate change presents significant threats to human health, especially for low-income urban communities in the Global South. Despite numerous studies of heat stress, surprisingly little is known about the temperatures actually encountered by people in their homes, or the benefits of affordable adaptations. This paper examines indoor air temperature measurements gathered from 47 living rooms within eight low-income communities of Accra and Tamale, Ghana. Using multiple temperature indices and a tiered analysis, we evaluate indoor temperature variations linked to roof type, ceiling insulation, presence of fans, and tree shade, for different housing types and locations. Our data reveal indoor temperatures in the range 22.4 °C to 45.9 °C for Accra, and 22.2 °C to 43.0 °C in Tamale. Using dummy regression analysis, we find that tree shade reduces the number of very hot days (>40 °C) and nights (>30 °C) by about 12 and 15 d per year, respectively. Building materials also strongly moderate indoor temperatures but in opposing ways: rooms with traditional mud walls and thatch roofs are on average 4.5 °C cooler than rooms in concrete block houses with uninsulated metal roofs during the day but are 1.5 °C warmer at night; rooms with ceiling insulation are on average 6.9 °C cooler in the day but 1.4 °C warmer at night. We conclude that sub-daily data are necessary for reporting extreme indoor temperatures, and that trade-offs between minimum and maximum temperatures require interventions to be assessed carefully before attempting to counter extreme heat inside homes.


Introduction
Millions of people are already exposed to deadly heat but this hazard is expected to become more frequent and severe as global temperatures rise (Zhao et al 2015, Nangombe et al 2018, Lie et al 2020. Populations in the tropics and subtropics are most likely to encounter the greatest rise in heatwaves with further warming (Dong et al 2015, Matthews et al 2017, Mora et al 2017, Rohat et al 2019, Raymond et al 2020. Moreover, high temperatures are exacerbated within dense cities by the urban heat island (UHI), which elevates heat stress compared to surrounding rural areas (Wilby 2007, Oleson et al 2015, Coffel et al 2017. Low-income communities are especially vulnerable given their relatively high exposure to extreme weather events and dependence on fragile water/energy infrastructure (Makaka and Meyer 2006, Maller and Strengers 2011, Ahmadalipour et al 2019, Gough et al 2019, Jagarnath et al 2020, Kayaga et al 2020. Hence, the global frontline for deadly heat is likely to be found in the densely populated informal settlements of the tropics in general, and sub-Saharan Africa in particular. Despite such concerns, surprisingly little is known about extreme indoor temperatures in urban areas (Scott et al 2017) and even less about affordable measures for lowering temperatures in homes within informal settlements of tropical Africa (Amos-Abanyie et al 2013, Dauda and Gao 2013). Exposure to high indoor temperatures matters because of the adverse impacts on human health, comfort, and productivity (Dapi et al 2010, White-Newsome et al 2012, Dunne et al 2013, Codjoe et al 2020. With more extreme temperatures and rapidly growing urban populations, there is mounting recognition that cities are priority places for climate action . More specifically, affordable/low-energy solutions are urgently needed to manage deadly heat, especially within low-income urban settlements in sub-Saharan Africa (Parkes et al 2019).
The effectiveness of building modifications to manage extreme temperatures depends on context. Many factors affect indoor temperatures, including: wall, roof and floor materials; roof pitch and reflectance, ventilation, size, and age of the structure; number of occupants and their behaviour; building use; artificial heat sources; building, window, and door size/orientation; site elevation, shade, local breezes, proximity to water bodies, and the urban heat island. Previous research has evaluated building age, window size/type, thermal mass or roof type using field measurements and/or building simulations (Taylor et al 2000, Amos-Abanyie et al 2013, Morakinyo et al 2016, Naicker et al 2017, Ayanlade et al 2019. Many studies use daily mean temperatures even though averages can obscure critical variations within the diel (24 h) thermal regime. Other research draws on qualitative evidence (e.g. Nematchoua et al 2014Nematchoua et al , 2017 or multi-variate indices of thermal comfort and heat stress (e.g. Holmes et al 2016). Metrics of human thermoregulation combine permutations of environmental data (such as air temperature, humidity, air speed and radiant energy) with information about occupants' metabolic rate and clothing. Dry-bulb air temperature features in the majority of heat warning systems and building energy models.
In this paper, we advance understanding by statistically relating building types and modifications to sub-hourly air temperature variations monitored in diverse homes within low-income urban communities. Temperatures were studied in two cities in Ghana, namely: Accra (5 • 33 ′ N, 0 • 12 ′ W, population 4.2 million, hot semi-arid/tropical wet climate) and Tamale (9 • 24 ′ N, 0 • 51 ′ W, population 0.4 million, tropical wet and dry climate). These communities and cities were chosen because they regularly endure adverse impacts from extreme heat (Gough et al 2019, Codjoe et al 2020-events that have become more frequent and severe in recent decades (Ringard et al 2016). The following analysis provides evidence of the efficacy of affordable measures intended to moderate extreme indoor temperatures.

Field data
Four communities were chosen in Acca and four more in Tamale based on their mix of building types and density, informality, and vulnerability to extreme heat (table 1). Tinytag Transit 2 thermistors were installed within homes to monitor air temperatures encountered by occupants during 2018. According to the manufacturer, the sensors have a precision of 0.01 • C and accuracy of 0.4 • C in the range 0 • C to 50 • C (Gemini Data Loggers 2015). The logging interval was set to 10 min.
Homes were selected to represent different structures, building materials, sizes, and locations (figure 1). This process was guided by the local knowledge of 'community champions' (residents nominated by their communities to liaise between households and the research team). A standard form captured information about building location (coordinates, elevation), number of floors and building height (as a proxy for size/thermal mass), alongside other characteristics (wall, and roof materials, number of doors and windows, amount of shade, and orientation). Any modifications to Figure 1. Example dwellings with: (a) wood walls and sheet metal-roof (Accra); (b) traditional mud walls and thatch roof (Tamale); (c) single-storey concrete block walls and sheet metal roof (Tamale); (d) two-storey/large concrete block walls (Accra); (e) some shade (Tamale); and (f) no shade (Accra).
counter high temperatures, such as ceiling insulation (plywood, plastic sheeting, Polyvinyl chloride, a synthetic plastic polymer tongue and groove panels, plaster), fans (ceiling, free-standing), tree shade or air conditioning, were also recorded. Room occupancy, usage, and artificial heat sources were noted too.
At each home, one thermistor was fixed outdoors on a self-shaded north-facing wall to measure ambient air temperature, whilst up to four were sited inside rooms with various ceiling types, usage, occupancy, and dimensions (figure S1 (available online at stacks.iop.org/ERL/16/024033/mmedia)). All thermistors were mounted ∼2 m above ground level, away from direct sunlight and artificial heat sources. These sites had to be as unobtrusive as possible and always beyond the reach of children. Spot measurements were taken within several homes to assess any height-dependency of air temperatures using a Testo 610 handheld sensor (Testo 2020). Variations in temperature were always smaller than instrument accuracy (±0.5 • C).
Thermistors were also installed within Stevenson screens at ∼1.25 m above ground level alongside thermometers operated by the Ghana Meteorological Agency (GMA) at three sites: (A) their compound on the University of Ghana (Legon, Accra) campus, (B) Accra airport, and (C) Tamale airport (figure S1). Indoor temperatures are expected to differ from these records because a key purpose of a dwelling is to regulate the thermal environment for occupants. Nonetheless, previous studies show strong positive linear associations between indoor and outdoor temperatures on a home-by-home basis (e.g. Quinn et al 2014, Nguyen andDockery 2016, Naicker et al 2017). Here, our thermistor records at GMA stations (A and C above) are used as points of reference for our outdoor and indoor temperature measurements. Sensors housed within these official meteorological stations are expected to have more stable environments than outdoor sites within communities. Initially, a network of 96 thermistors were deployed; after 1 year, 81 were still in place and operating (47 inside homes, 24 outside homes, plus 7 verandas, and 3 at meteorological stations). Data were downloaded every other month and quality assured. Day-to-day changes and basic range checks were performed, with records from GMA stations and homes cross-compared to eliminate outliers/suspect values associated with artificial heat exposure. Percentilebased indicators, rather than absolute extremes, were used for statistical analysis (section 2.3) to reduce possible contamination by erroneous outliers. Abrupt changes in the daily range helped to detect any relocation or tampering with a sensor. The cleaned archive yielded more than 3.5 million data values for analysis. Importantly, the high temporal resolution enabled evaluation of temperature variations within a day, so the effects of location and building factors can be discerned from a range of temperature metrics.
According to multi-decadal thermometer-based records provided by GMA (for sites B and C above), annual mean temperatures in Tamale and Accra for 2018 were 9th and 11th highest since 1960, respectively. Since the 1960s, annual mean temperatures rose by 0.26 • C decade −1 and 0.18 • C decade −1 at the same sites. Hence, we regard our study period as representative of the generally warmer conditions experienced in recent decades. Hereafter, we present results from thermistor data gathered during the 365 d period covering 22 January 2018-21 January 2019.

Primary and derived temperature indices
Primary series of daily minimum (Tn), daily maximum (Tx), daily mean (Tm), and daily range of temperature (DTR; Tx minus Tn) were obtained from thermistor recordings at each site (table 2). These data were then used to derive 12 temperature indices, again for each site. These are the number of very hot nights [when Tn > 30 • C] (Tn30C); the number of very hot days [when Tx > 40 • C] (Tx40C); temperature minima summarised by the very lowest (Tnn), the 1st (Tn01p), 10th (Tn10p), and 50th (Tn50p) percentile, and very highest (Tnx) of the daily distribution of Tn; temperature maxima characterised by the very lowest (Txn), the 50th (Tx50p), 90th (Tx90p) and 99th (Tx99p) percentile, and the very highest (Txx) of the daily distribution of Tx. Note that the temperature thresholds for Tn30C and Tx40C are higher than used elsewhere because the conventional definition of a 'tropical night' (Tn > 20 • C) (Sillmann et al 2013) would be exceeded 100% of the time at all sites (except for the GMA station in Tamale).

Statistical analysis
We used the above indices to discern variations in temperature associated with building location, type, and modifications that favour lower indoor temperatures. Three tiers of statistical technique were applied: (a) community level averaging of temperature to evaluate local variations in the strength of the UHI; (b) category level averages to show differences in temperature between groups of similar buildings; and (c) building level dummy regression analysis of variables influencing selected indoor temperature indices.
For tier 1, mean daily indoor and outdoor temperatures were stratified by community then compared with thermistor data from GMA sites (A and C). Since these meteorological stations are in opengrassed sites outside the city centre, differences in temperature between them and communities are indicative of local UHI intensity. This reveals ambient temperature variations across each city reflecting local building density, artificial heat sources, proximity to open/green space, water bodies, and distance from the coast (for Accra).
For tier 2, building types were categorized by height/number of floors, wall, and roof material as well as by measures intended to reduce indoor temperatures (table S1). The sampled housing archetypes were as follows. Single-storey homes (⩽5 m high) were comprised of wood walls with metal or asbestos sheet roofs ( figure 1(a)), or traditional mud walls with thatch roofs ( figure 1(b)), or concrete walls with metal or asbestos roofs ( figure 1(c)). Two-storey or large (>5 m high) buildings had concrete walls and asbestos sheet or tiled roofs (figure 1(d)). Shade by vegetation was classified as either present (figure 1(e)) or absent (figure 1(f)). This binary approach was justified given that rapid visual assessment of shade is problematic due to variations by time of day and effect of neighbouring structures.
Building-specific indices (Tn10p, Tx90p and mean Tm) were stratified by city, then averaged by housing archetype, or cooling measure. Two-sample t-tests of differences in the mean (assuming unequal variance) were applied to the groups of data. For example, indoor temperatures may be compared for sub-samples of single-storey (n 1 ) and two-storey (n 2 ) housing in Accra, with the null hypothesis of no difference. The number of degrees of freedom (n 1 + n 2 − 2) is sometimes small, in which case the statistical power to identify differences between factors is limited (i.e. there is a higher chance of accepting a null hypothesis that is actually false). Indoor temperatures were also compared for rooms with different modifications within the same house or with a nearby property of similar size, age, and design. Resulting scatterplots reveal daily variations of indoor Tx and Tn depending on outdoor temperatures (at the reference site); roof type (metal versus thatch); ceiling type (insulated versus uninsulated); shade (tree cover versus unshaded); and fans (present or absent).
For tier 3, we apply dummy regression modelling to i temperature indices (Tn30C, Tx40C, Tn10p, Tm, and Tx90p) to quantify the benefit of modifications relative to the effects of building location and characteristics. The influence of n dummy variables (v j ) were estimated from the coefficients of the following model, fitted via ordinary least squares: where T i is the indoor temperature index, α i is the model intercept, β i,j are the model coefficients, and e i is the model error. Eight dummy variables were applied, each coded either 0 for the reference case or 1 for the 'treatment' (table S2). The variables were city (0 for Tamale); coastal (0 for sites > 1 km from the sea); thermal mass (0 for buildings ⩽ 5 m high); wall material (0 for wood); roof material (0 for sheet metal or asbestos); ceiling insulation (0 for none); tree shade (0 for none); and ceiling or free-standing fan (0 for none). The coefficients β i,j are directly comparable (in each temperature index). Some dummy variables could have been sub-classified, such as ceiling insulation (by material), tree shade (by extent), or sheet roofing (by material). However, there is a danger of over-fitting the model given the modest sample size (45 rooms after excluding two homes with air conditioning). The dummy variables allowed us to pool temperature indices reflecting very different building types and locations.

Variations in extreme temperature between sites
Our thermistor data reveal considerable variations of indoor temperature linked to site. In Accra, the lowest recorded indoor temperature (Tn) was 22.4 • C for a small asbestos-roofed, wood-walled room near the coast; the highest indoor temperature (Tx) was 45.9 • C for a small metal-roofed, wood-walled room in a densely built-up area. As expected, equivalent values (Tn = 20.7 • C, Tx = 37.0 • C) at Accra meteorological station were lower. The average number of very hot nights (Tn30C) within homes ranged between 7 (single-storey concrete houses) and 92 (two-storey or large concrete houses) per year, whereas none were recorded at the meteorological station (table 3). Likewise, the number of very hot days (Tx40C) was on average 40 inside single-storey wood homes but zero at the nearest meteorological station. In Tamale, the lowest indoor temperature (Tn) was 22.2 • C and the highest (Tx) was 43.0 • C (both for metal-roofed homes). Temperatures at Tamale airport ranged between 17.4 • C and 41.4 • C. Average Tn30C within metal and thatch roof homes were 96 and 102 nights per year respectively, compared with none at the GMA station (table 3). In contrast, average Tx40C was zero in thatch roof rooms, 7 in metal roof rooms, and 10 at the GMA station.
We also stratified our temperature data by community to elucidate variations across the two cities. In Accra, average indoor Tn were higher for all communities than at GMA stations (by up to +3.3 • C in Alajo) (figure 2). Likewise, indoor Tx was higher in all communities (by up to +3.1 • C in Alajo) except Odawna (−0.7 • C), an area with some two-storey houses ( figure 1(d)). In Tamale, Tn were higher than the GMA station in all communities, whereas Tx were lower on average than the meteorological station (by as much as −1.3 • C in Ward K).
Outdoor temperatures in communities were consistently higher than those at GMA stations (figure 2), as official sites (A and C) are located away from the city centre, in open, well-ventilated areas. Hence, by comparing outdoor temperature differences between communities and reference stations, we estimated Table 3. Arithmetic mean temperature indices based on the same from each site, pooled by building category and city. Single-storey houses are distinguished by wall material (in parentheses). Mean temperature indices are also given for outdoor (self-shaded) sites within communities and at reference meteorological stations. Refer to table 2 for index definitions.
Single-storey (wood)  Table 4. Differences in indoor temperature indices associated with building characteristics and modifications, based on two-sample t-tests assuming unequal variance. The number of degrees of freedom (df) are also given. Rooms with air conditioning were excluded.

City
Factors

Variations in temperature extremes between and within building types
Home type and community-level averages mask considerable temperature variations between and within building archetypes, especially during very hot days and nights. Such differences are generally obscured by daily mean statistics (table 4). In Accra, significant (p < 0.01) variations were found in high indoor temperatures (Tx90p) between rooms with and without ceiling insulation (−6.9 • C). Rooms in two-story residences were also significantly cooler on average than those in single-storey homes (−4.4 • C). However, night temperatures (Tn10p) were higher in insulated versus uninsulated rooms (+1.4 • C) and in two-versus single-storey homes (+1.5 • C). The effect of shade (−0.2 • C) or fans (+0.4 • C) was insignificant (p > 0.1). In Tamale, the largest difference in Tx90p was between homes with insulated thatch roofs compared with uninsulated metal roofs (−4.5 • C) (table 4). Similarly, Tx90p was lower in rooms with insulated ceilings (−2.8 • C). Overall, rooms with thatch roofs had lower Tx90p than those with metal roofs (−2.2 • C) but only marginally lower Tm (−0.6 • C). There were weakly significant (p < 0.1) differences in Tn10p (−0.8 • C) for shaded rooms and in Tx90p (+1.3 • C) for rooms with fans. Differences also emerge when comparing daily Tx (−4.9 • C) and Tn (+1.3 • C) for rooms with thatch or metal roofs, in this case, within the same house (figures 3(a) and (b)). During the hottest day sampled in Tamale, the smallest daily temperature range was recorded in a room with thatch underlain by plastic sheeting (installed to protect occupants from dust and insects rather than for insulation) ( figure 4(a)). Variations within the diel regime were likewise recorded amongst rooms with metal roofs (figure 4(b)). The room with the smallest daily variation was in a large, single-storey house inhabited by one family, with partial shade to the rear and PVC cladding on high ceilings. Conversely, the metal-roofed room with greatest daily temperature range had no shade and a low, uninsulated ceiling.
Similarly, diel temperature profiles for the hottest day in Accra were different for single-storey houses with insulated ceilings compared to those with uninsulated ceilings (figure 4(c)). The sample size is not large enough to discern the effect of floor-level (in two-storey buildings) or roof type; however, the data reveal a narrower indoor daily temperature range for insulated than uninsulated rooms, with a diel regime similar to thatch ( figure 4(a)). Although Tx was lower, temperatures in rooms with insulated metal/asbestos and thatch roofs were consistently higher at night than uninsulated metal-roofed rooms (see below).

Effect of individual modifications on indoor temperatures
Modifications to homes, such as ceiling insulation, shade, and fans are intended to improve thermal comfort. However, consequences for indoor temperatures were mixed, depending on time of day and building archetype. Rooms with insulated ceilings were up to 7 • C cooler in the day than ones without insulated ceilings within the same home (figures 3(c) and (d)) but during one night, the insulated room did not fall below 32.2 • C. Hence, insulation reduces Tx but at the penalty of higher Tn.
Tree shade reduces indoor temperatures most during the day (Tx) in the dry season ( figure S2(a)) and, as expected, there is no difference between shaded and unshaded rooms at night ( figure S2(b)). Although lower indoor temperatures were recorded in rooms with some shade (table 4), the difference in Tx peaked during April-July (2.2 • C) for the two exemplar rooms. On some days, Tx in a shaded room was at least 8 • C cooler than the meteorological station ( figure 3(e)).
Discerning the effect of (ceiling) fans on indoor temperatures is problematic as their presence or absence tends to coincide with that of insulation. Moreover, even when present, fans may not be used. Overall, fans are not expected to reduce indoor temperatures unless by promoting ventilation of cooler air from outside. This was confirmed by the absence of any statistically significant temperature differences in our pairwise analysis of all rooms with, and without fans (table 4). Likewise, there was no differences in temperature due to fans in the subset of rooms with no ceiling insulation (figure S3). Although Tn was higher in some rooms with PVC ceiling cladding and a fan (figures 3(g) and (h)), this may be explained by a greater number of electrical appliances (i.e. artificial heat sources) and fewer doors.

Combined effects of modifications on indoor temperatures
We used dummy regression to evaluate the relative effects of geographic location (city, distance from sea), building archetype (number of floors/ thermal mass, wall, and roof materials), and modifications (ceiling insulation, tree shade, and fans) on indoor temperature indices (figure 5 and table S2). The amount of variance explained by these factors ranged from 17% (Tx40C) to 73% (Tx90p). Statistically significant (p = 0.05) model coefficients (with standard errors) emerge for rooms with: Dummy variable coefficients can be summed to give combined effects of several measures. For example, traditional homes in Tamale are always ⩽5 m high, have thatch roofs and mud walls but are not located by the coast. Hence, to reduce Tx90p in such homes, the modelled measures (i.e. ceiling insulation −2.6 • C, tree shade −0.1 • C, and a fan −0.2 • C) would yield a benefit of −2.9 • C relative to the reference building (table S2). However, the model also indicates that this set of modifications would increase the number of very hot nights by 28 per year. Adding ceiling insulation, tree shade and a fan to a concrete walled and metal roofed home would change Tx90p by +0.8 • C and Tn30C by −7 nights per year. Day/night trade-offs are also predicted by the models for modifications to low-income homes in Accra. For example, upgrading a reference building  (⩽5 m high, wooden wall, metal roof, with no insulation) to one with concrete walls, insulation, and fan, could change Tx90p by −5.2 • C but Tn30C by +57 nights per year. Overall, choice of roof material and ceiling insulation are the most heavily weighted variables for reducing Tx90p. Addition of tree shade reduces the frequency of hot nights by 15 per year.

Discussion
Relatively little is known about indoor temperatures in tropical Africa; even less for those living and working within high-density, informal communities. In this paper we have shown that building materials and modifications can have significant impacts on the indoor temperatures endured by residents of low-income communities. Here, we consider the wider implications of our research for other parts of the Global South.
We show that conventional dry-bulb air temperature measurements at meteorological stations significantly under-report the extraordinary indoor temperatures faced by some households; yet such station data (or gridded derivatives) are routinely used as reference points for heatwave forecasts and climate change studies (e.g. Alexander et al 2006, Donat et al 2013, Raymond et al 2020. This adds to the sense that heatwave impacts are being overlooked for sub-Saharan Africa (Harrington and Otto 2020). Subhourly, indoor temperature data reveal that residents may be living much closer to a 'thermal precipice' than conventional ambient air temperature measurements suggest (Kenney et al 2004). We suspect that this situation is not unique to Ghana.
Although often used, we show that mean indoor temperature data can hide important trade-offs. For example, thatch with insulated ceilings has relatively low indoor Tx but comparatively high Tn. If the aim is to bring respite from high indoor Tx, preferred options are ceiling insulation, tree shade, and/or thatch roof materials. If improved sleeping conditions at night are the priority (i.e. lower Tn), then uninsulated metal roofs and/or tree shade are desirable. Such measures can reduce the frequency of very hot nights by 30 per year. This means that the diel temperature regime must be considered when evaluating heatwave impacts and adaptations.
There are also important considerations about the cultural and practical aspects of some modifications in high-density urban contexts (Satterthwaite et al 2018). For instance, concerns about security, pests, or poor air quality may discourage use of natural ventilation. Even though tree shade reduces insolation, this option is not always feasible within the dense urban fabric of informal settlements. Saplings may also struggle to establish in extreme hot-arid conditions. Furthermore, there are taboos against tree-planting amongst some communities in northern Ghana. Although such beliefs are waning (Hansen et al 2012), the matter was raised by several households in our study and may act as a deterrent.
The benefits of some building modifications appear modest when expressed as mean temperature changes. For example, tree shade reduces Tm by just 0.5 • C, yet the average frequency of very hot days (Tx40C) (as an indicator of potentially dangerous conditions) is reduced by 12 d per year (table S2). Benefits may also be framed as the amount of climate change that is countered, or the equivalent time delay for a given rise in temperature. For example, insulating a metal roof in Tamale (figure 1(e)) could reduce indoor Tx90p by 2.6 • C, effectively negating an outdoor temperature rise of 4.2 • C ( figure S4(c)). This is the median warming by 2100 projected by the CMIP5 ensemble for the hot season in Ghana under RCP8.5 emissions (IPCC 2013). Hence, we advise the use of multiple temperature indices (as in table 2) when evaluating the impact of adaptation(s) to buildings.
We demonstrate how empirical data and statistical models can be used to evaluate building features and measures that are widely deployed by households in Ghana. However, other options, such as cool paint, more passive ventilation, increased ceiling height and increased wall thickness need to be tested (Santamouris 2014, Akbari and Kolokotsa 2016, Kolokotroni et al 2018. The influence of floor level on temperatures (inside multi-storey homes) and long-term trends in preference for certain materials are also worthy of further exploration. In our sample, ceiling insulation was present in 74% of the rooms. Amongst these, 69% used plywood, 20% PVC cladding or plastic sheeting, and 11% plaster. More widespread use of PVC tongue and groove panels in place of plywood would favour higher night-time indoor temperatures over the long-term (Naicker et al 2017).
We acknowledge that the benefits of fans are difficult to quantify without accompanying information about their use. Nonetheless, it is accepted that the protective value of fans is via increased airflow and evaporative cooling from sweating. Previous research shows that indoor fan use is beneficial up to ∼45 • C and 10% relative humidity or ∼40 • C and 60% relative humidity; beyond these limits there is increased risk of dehydration (Jay et al 2015). Some recommend against use of fans in very hot (⩾45 • C) and arid (<10% relative humidity) regions (Morris et al 2019)-conditions that are already approached during heatwaves in some parts of sub-Saharan Africa. Moreover, electricity supplies, and hence access to fan cooling, can be curtailed during heatwaves (Gough et al 2019, Kayaga et al 2020 while power networks feeding cooling systems may be damaged by cyclones, increasing the risk of deadly heat during blackouts (Matthews et al 2019). Hence, solutions should be both energy efficient and resilient to extreme weather.
Measures taken by many individual households to moderate indoor temperatures could also affect the intensity of the local urban heat island, when aggregated at community scales. For instance, cool paints and more reflective materials reduce the overall heat load to the urban environment (Kyriakodis and Santamouris 2018). Experiments with a house in Ghana showed that cool paint applied to a poorly insulated roof reduced the surface temperature by 7.8 • C and ceiling temperature by 6.2 • C during the hottest month (Kolokotroni et al 2018). Other studies show that green spaces, roofs, facades and walls also contribute to reduced intensity of the daytime UHI (for a review see: Akbari and Kolokotsa 2016). In turn, lower outdoor temperatures may also have indirect benefits by reducing energy consumption for space cooling (Mastrucci et al 2019).
Finally, we quantified the benefits of selected modifications but the socio-ecological-technological dimensions of adapting to extreme heat should not be overlooked (Markolf et al 2018). These include the social acceptability of measures, such as treeplanting, or preference for modern materials over traditional. Heat avoidance behaviours, including outdoor sleeping and other night-time activities, such as socializing, night school, household chores, and small-scale economic activity significantly increase malaria risk (Monroe et al 2015). Consequently, there is a strong imperative to moderate indoor temperature extremes at night as well as during the day. Since most home-based livelihood activities take place outside the physical structure of the home, in a kiosk or under an adjacent veranda (Gough 2010, Amankwaa et al 2017, measures are also needed for reducing ambient heat in outdoor spaces.

Conclusions
This study exposes the level to which observations at meteorological stations and use of daily means under-report the extraordinary indoor temperatures encountered by some households. Our data also show the relative effects of building location, size, wall/roof materials, insulation, shade, and cooling systems on extreme indoor temperatures. We used statistical modelling to quantify the benefits of affordable adaptation measures (traditional roof materials, ceiling insulation, shading by vegetation, and electric fans). Critical trade-offs emerge when evaluating consequences of these measures using more nuanced temperature indices. Such evidence is urgently needed to properly monitor and then guide modifications to low-income housing. In this way, vulnerable people in tropical cities will have an opportunity to become more resilient to climate change.

Data availability statement
The data that support the findings of this study are available upon reasonable request from the authors.

Acknowledgments
This work was funded by the British Academy under the Cities and Infrastructure Programme CI170211. The VEWEC team thank the communities, in particular the Community Champions, for facilitating the fieldwork. We are especially grateful to all the participants who repeatedly welcomed us into their homes and workplaces to record the temperatures they experience. Ben Roberts kindly supplied data on temperature variations within rooms linked to measurement heights above ground level.