Using social media to measure intentions to move house

Is it possible to measure the intentions of people to move house without actually asking them about their plans? Statistics Netherlands (CBS) has researched this question at the Center for Big Data Statistics. On the instructions of the Ministry of the Interior and Kingdom Relations, CBS already asks about the intention to move in the WoON survey, which forms part of the ongoing housing research. Every three years, this survey uses a questionnaire to gather information on respondents’ current living situation, including satisfaction with their homes and living environments, and their housing needs.

The ministry has commissioned CBS to investigate the use of alternative sources, such as social media, to measure the moving intentions. This CBS study shows that it is possible to automatically distil the moving intentions from information in social media posts. This is not to say that these people actually will move house in the near future, or that these people’s intention of moving reflects the intention of moving among the population of the Netherlands as a whole. Because the results deviate from the trends established in surveys and registers, they are not considered suitable to replace this data.

Working method

To test whether it is possible to measure the intention of moving using social media, the Center for Big Data Statistics has studied a large number of public social media posts written in Dutch. These items were posted between 1 January 2011 and the end of May 2018. As far as possible, posts written by people from Flanders and from countries other than the Netherlands were filtered out of the results.

The starting point for the study was a randomly selected set of posts containing the partial words ‘verhuis’ or ‘verhuiz’ (the first part of the Dutch words for ‘move house’). All the selected posts include at least one instance of a word containing this combination of letters. Two independent coders then manually classified this set of posts into two categories: ‘includes intention of moving’ and ‘includes no intention of moving’. This classification was uncontested in 95 per cent of cases. Cluster analysis was used to identify even more posts that included an intention of moving. These were classified manually. In total, 12,353 posts containing ‘verhuis’ and/or ‘verhuiz’ were studied, of which 4,324 (35 per cent) were identified as showing a moving intention. The coding instructions followed were that:
1. The post must originate from a private individual and not from a company;
2. The intention of moving must relate to the person who posted the message or to this person’s family;
3. The intention of moving must refer to a physical relocation to another country, municipality or house, rather than for instance the relocation of a website to another URL;
4. The move has not yet taken place;
5. The post must represent a serious intention of moving and not be meant sarcastically or as a joke.

This classified set of posts was then used to develop a model capable of identifying, with the greatest possible accuracy, social media posts that contain an intention of moving. This model uses logistical regression, which enables the model to choose the words, combinations of words or post characteristics that offer the best likelihood of detecting a moving intention. The classified data set was used as a training and test set. The model was compared to several alternative methods using machine learning. When it became clear that these methods were as reliable as or less reliable than the model, the decision was made to use the simplest model based on regression analysis. This model correctly identified whether or not a post contained an intention of moving in 87 per cent of cases. The first tests of the models were conducted using all social media platforms. To comply with applicable privacy legislation, however, it was decided that these analyses would be based exclusively on Twitter, excluding other platforms such as Facebook.

Findings

All public tweets containing the words ‘verhuis’ and/or ‘verhuiz’ that were available between 1 January 2011 and the end of May 2018 were included in the analyses: over 1.9 million tweets. In total, more than 4.8 billion Dutch-language tweets were available during this period. These tweets were used to assess the development of a moving intention in that time. The model used to classify tweets as ‘includes intention of moving’ or ‘includes no intention of moving’ is considered sufficiently reliable to be able to identify the intention of moving.

Figure 1 displays the findings from the relative number of tweets containing an intention of moving in the period from January 2011 to May 2018 inclusive. The series clearly shows two periods in which there is a decrease in the number of tweets containing an intention of moving: from January 2011 to December 2013 inclusive, and from August 2014 to May 2018 inclusive. The intervening period shows an increase in tweets regarding the intention of moving. As well as conducting an analysis in time, it is also interesting to see whether there is a seasonal pattern. The analysis showed that there is no strongly visible seasonal pattern, although the number of tweets containing a moving intention is lower in the period around December.

On average, 133 tweets including a moving intention were found each day during the period analysed, which translates to around 930 per week. In total more than 1.7 million tweets were posted every day: around 12 million a week. This puts the percentage of tweets including an intention of moving at 0.008 per cent after rounding. The series of tweets including an intention of moving reveals a different trend from that of the total number of tweets. This means that in some way the data reflects something meaningful. The daily and weekly series show the same patterns over time. The question, then, is whether an unusual occurrence such as an intention to relocate can be picked up with sufficient accuracy from social media. For validation purposes, 1,000 randomly-selected, anonymised social media accounts were followed for two years. Half the people who tweeted during this period indicated at a certain point that they had actually moved house.

Figure 1. Relative number of tweets containing an intention of moving posted each day from 1 January 2011 to 31 May 2018 inclusive (the inset shows the data before applying the spline filter)



The trend in social media posts containing a intention of moving deviates from the trend in the housing research, which only publishes annual figures on this issue. The intention of moving extrapolated from the WoON survey, with 40,000 participants, rose between 2009 and 2015. In 2009, around 2.1 million households indicated a potential intention to move house within two years; the same figure for 2015 was 3.2 million (Ministry of the Interior and Kingdom Relations, 2016).
If the hypothesis is correct that the intention of moving identified from social media posts can be used to indicate the intention of moving among the population of the Netherlands, the social media trend must be at least somewhat in advance of the actual number of relocations within the Netherlands. However, the entire period from 2011 to 2018 reveals an increase in the number of relocations. The long dip between 2011 and 2013, with a subsequent small peak in the intention of moving identified on social media, does not correlate with the actual number of relocations.



There may be a variety of reasons for the deviation from the trend based on social media data, compared with the trend in the WoON survey and the changes of address listed in the national register. The population on Twitter is not the same as the population of the Netherlands, and young people are overrepresented on social networks such as Twitter and Facebook (CBS, 2017). Previous research has also warned of the risks of using this kind of source: algorithms can change over time, thereby influencing trends (Lazer et al., 2014); public tweets are not necessarily a good reflection of all tweets; Twitter users are not necessarily representative of the population as a whole; and a Twitter account is not synonymous with a Twitter user, as people can have multiple accounts (Boyd and Crawford, 2012). Methods are currently being developed to resolve problems such as these. The findings show that the partial words ‘verhuis’ and ‘verhuiz’ appear less frequently in tweets over time, indicating the changing content of posts to this platform. Further research is needed to investigate whether the method researched can be used to monitor other phenomena, or whether there is greater correlation with other sources when more than one social medial platform is included. This study’s provisional conclusion is that a intention of moving can be automatically recognised from social media posts, but that this intention does not appear sufficiently frequently to produce a trend that can be used for statistical purposes.

Privacy

The study exclusively used publicly available tweets. Neither the content of the messages nor the associated accounts will be published. CBS will only publish the relative number of posts containing an intention of moving.

Application

The findings are not currently considered to be useful for monitoring developments in the intention of moving. This method may however be suitable for other, more common events such as buying a car or taking a holiday. CBS has no concrete plans for such a new application at this time.

Related items

You are using an unsupported browser. Please upgrade your browser. You provided an inconsistent user agent while solving the challenge. You may have browser extensions or settings enabled to spoof the user agent and should disable them to proceed. Some portion of Turnstile was accidentally cached. Please clear your cache. The time on the clock is incorrect. Please set your clock to the correct time. An unspecified error occurred.