In the first in a new NatCen, European Social Survey (ESS) European Research Infrastructure Centre (ERIC), City University methodology seminar series, Professor Nigel Gilbert talked about the HomeSense project, which is exploring what role digital sensor data could play in social science research. HomeSense aims to catalogue the technologies and methodology, explore the ethical issues and provide guidance to social researchers who are thinking about using sensors. The project involves around 20-30 households having sensors placed in their homes and individuals wearing a sensor wrist band. Sensors will collect data on location, movement, noise levels, air quality, temperature, and energy use within the home and individuals' activity.
His talk raised two important issues: how can we make sense of the vast amounts of data produced by sensors; and how might sensor data more generally be useful to social scientists?
Making sense of sensor data
Sensors continuously collect observations such as motion. However, as social scientists, our interest is not in motion per se but in what that motion represents; is the individual exercising, travelling, or something else? We have to make inferences from the sensor data – to interpret what the observations mean. There are several challenges: the sheer volume of data sensors produce, how to identify patterns and how to identify activities. Professor Gilbert and his team are employing methods more commonly used in the physical and computing sciences to help tackle these challenges. For example, Change Point Detection is a statistical method for simplifying continuous data, which identifies the times when there is a significant change in the value of the observation. The application of Hidden Markov Models (HMM), a statistical method used to identify (hidden) activities from the observations collected by the sensors, can be employed in machine learning algorithms. Having identified likely patterns of activity within the household using HMM these are compared with household time use diary data, to see how good the model is at identifying the activity. Early data from the HomeSense pilot comparing HMM and diary activity data show good levels of agreement for cooking (see Example 1).
Example 1: HMM example output and time use diary data for cooking
Source: Jie Jiang, Riccardo Pozza, Kristrún Gunnarsdóttir, Nigel Gilbert and Klaus Moessner. Recognising Activities at Home: Digital and Human Sensors. Forthcoming at the 2017 International Conference on Future Networks and Distributed Systems (ICFNDS '17), 2017
How might sensor data be useful to social scientists?
There is considerable commercial interest in sensors: from smart homes to cheaper car insurance and possibly social care, the internet of things is opening up many possibilities. But what do sensors offer social scientists? They offer us an alternative way to measure behaviour to survey self-reports, and potentially a means of collecting more data more cheaply (depending on how the technology develops and how wide the uptake is of it by the general population). One example is the measurement of physical activity in surveys. We know that survey respondents can over-estimate the amount of exercise they do. Wearing an activity monitor could provide more accurate information on actual body movements and the 2008 Health Survey for England (HSE) experimented with their use among the general population. There are two important messages in this experiment that are as relevant today as they were back in 2008 .
First, the report’s authors concluded that the accuracy of the sensor data depends on how well the sensor measures physical activity and how willing survey participants are to wear it. Whilst the technology has moved on since 2008, as researchers we still need to know how sensitive activity monitors are to movement and to what level of accuracy measurements are made. We also need to consider how willing people are to wear a sensor or install an App on their smartphone, and whether some groups may be less inclined to wear one than others.
Secondly, the HSE experiment demonstrates that sensors are not a replacement for survey data. To make sense of the activity monitor data the HSE researchers needed to know what activities participants were engaging in at what times of day. Combining survey data with the sensor data is tricky. A recent CLOSER workshop exploring the use of sensors in longitudinal studies for non-health data collection concluded that more research is needed to identify valid methods for combining self-report and objective measures from sensors. The HomeSense project is part of this effort.
Follow me on Twitter: @109deb