Modelling voter behaviours by geographic information technology: a case of Hong Kong in 2004

Abstract

Maps have been used widely to communicate election results and voter turnout. This article presents a practical approach to extracting and assembling the profiles of Hong Kong voters in the 2004 Legislative Council Election. It describes the use of modelling functions in a geographic information system to augment the examination of demographic and socio-economic constructs of voters and their voting behaviours. Voter profiles were used as predictors in forecasting election results. The predictive power of our model was assessed through visual and qualitative comparison with actual results of the election. The comparison showed that both maps exhibited similar patterns.

Keywords:

• election mapping
• geographic information systems
• predictive modelling
• voter behaviours

Research background

Maps are a popular graphic device to communicate election results and recount voting behaviours. They are fundamental to exploratory analysis that unravels the interplay among factors such as precincts, political parties and demographics to uncover hidden patterns, supply new insights and trigger speculations. The introduction of Geographic Information Systems (GIS) heightens our ability to use maps in exploratory analyses. The use of a GIS in redistricting work, such as locating polling stations or delimitating electoral or school districts, is not new (Agutu 2003). Novotny and Jacobs (1997) reviewed the use of GIS for political campaigns in the United States in the 1990s and concluded that such application had lasting impacts in the country. Dyck and Gimpel (2005) applied GIS to calculate distances between residential addresses of voters and their respective voting sites. They found that distance was the most important factor affecting the turnout rate and voting patterns. Similar usage of GIS was also found in a study by Gimpel et al. (2008) on the distance-decay effects of friends and neighbours in the voting patterns for local candidates.

GIS also provides a platform to correlate electoral results to factors such as weather, population and socio-economic variables (e.g. income, occupation, education attainment) for associative analyses. For example, Gomez et al. (2007) utilised a GIS to examine the relationships between voter turnout rate and bad weather in the United States. One interesting finding was that the Republicans seemed to benefit from bad weather in presidential elections. Burke (2003) demonstrated the significance of correlating election data with socio-economic indicators to identify voting behaviours. He argued that different cultures would influence each other and affect the voting propensity of residents in a multicultural community. In addition to social factors, Lay et al. (2007) found by means of indicators of spatial autocorrelation that spatial dependency did affect voting behaviours, using data from the 2004 Taiwan presidential election. However, map-based analyses of elections were not all well received. Fabrikant (2000) criticised the conventional use of the choropleth maps in displaying election results. She contended that the shading method would give a false impression of the overall distribution as a result of unequal sizes of the enumeration areas. Instead, she purported the use of a cartogram to stretch or shrink an enumeration area in proportion to the magnitude of electoral votes to give a faithful representation of the voters' support.

Previous studies on elections in Hong Kong, such as the reviews of Legislative Council (LegCo) elections by Lam (1995), Kuan (2006), Ngok (2001), Wong (1998, 2006) and Yu (2006), often involved the use of statistical and qualitative assessments. More recent studies by Lai and Loh (2007) and Lam et al. (2007) focussed on changes and development in the constitutional process of Hong Kong in the post-1997 era. There have been relatively few studies investigating the political patterns using the spatial dimension (Lai 2005). Studies that examine the correlation between socio-economic factors and patterns of election results in the Hong Kong context are also lacking.

This article demonstrates a means to predict the 2004 LegCo Election results by characterising the demographic and socio-economic profiles of supporters of each political camp. The propensity of voters to favour a political camp in each geographic constituency was constructed based on the identified demographic and socio-economic attributes. A comparative assessment between the predicted and the actual results summarises the power and deficiencies of our modelling approach. Our effort is made possible because of the coming together of the following constituents: (i) detailed election counts by 485 electoral areas, (ii) census data for each geographic constituency and (iii) a GIS platform to integrate these data and undertake various associative analyses.

The legislative council and the elected body

The LegCo is an elected governmental body in Hong Kong. It comprises 60 council members elected through a combination of corporate and individual voting according to the Basic Law of Hong Kong. The LegCo's major functions are to enact laws, examine and approve budgetary matters, monitor the Government's performance and debate issues of public interest. It is one of the bodies that restrain the power of the Chief Executive of Hong Kong (Wong 1997). LegCo members in the past comprised a selection of appointed official and unofficial members. Functional constituency elections and direct elections of some seats in the LegCo were introduced respectively in 1985 and 1991 (Lai and Loh 2007, Lam et al. 2007). Further reforms in 1995 dramatically increased 20 seats to be elected directly through geographical elections, 30 seats through functional constituency elections and 10 seats by an Election Committee made up of directly elected District Board members. A Provisional LegCo was established in 1996 to replace the LegCo after the transfer of the sovereignty. The LegCo was suspended from 1997 until 1998 and temporarily replaced by the Provisional LegCo, which resumed its operation in 1998. The first post-1997 LegCo Election in 1998 overturned the 1995 electoral arrangements. The 20 single-member constituencies through geographical elections were replaced by five multi-member constituencies, and the size of electorates of the functional constituency was significantly reduced, whereas a less democratic system of corporate voting was introduced. In 2000, 24 seats were directly elected, 6 seats elected by the Election Committee of Hong Kong comprising 800 members and 30 seats through the functional constituency.

The 2004 LegCo Election held on 12 September had a total of 60 seats, with 30 directly elected members from five geographic constituencies (Figure 1 and Table 1) and one member each from the 30 functional constituencies representing occupational and professional groups (HKSAR Legislative Council 2004). The present election adopts the list system of proportional-representation method by which the required votes to win a seat within a geographic constituency is equal to the number of ballots cast divided by the number of available seats. Each voter is allowed to cast a single vote for a candidate list, which may be represented by one or more candidates. A candidate list receiving more than the number of required votes will win a seat for the first candidate on the list. This method ensures that all parties with sufficient voter support have a representation in the LegCo. Candidate lists with the highest remaining votes (having deducted the first cut of required votes from the total votes) get the remaining seats. This accounting mechanism also makes clear that not all candidates on the list receiving the largest number of votes within a constituency will win the election. The remaining votes must meet the set criteria to claim a seat and so forth. LaMoshi (2004) commented that this election system averts monopolisation of one political party in the election outcome and thus presents a fairer chance to independent candidates.

Table 1. The 2004 LegCo Election – various statistics by geographic constituency

			Political camps^a
Geographic constituency	Number of seats	Number of candidate list	(A) Pro-Democracy	(B) Pro-Government	(C) Moderate	(D) Independent	Required votes to win a seat
Hong Kong Island	6	6	3	2	0	1	16.7% of total votes
Kowloon East	5	5	3	2	0	0	20% of total votes
Kowloon West	4	6	4	1	0	1	25% of total votes
New Territories East	7	6	3	2	1	0	14.3% of total votes
New Territories West	8	12	6	2	1	3	12.5% of total votes
Total	30	35	19	9	2	5
Note: ^aPolitical camps were classified on the basis of shared political ideology among candidates.
(A) Pro-Democracy: (1) Democratic Party, (2) The Frontier, (3) Hong Kong Association for Democracy and People's Livelihood, (4) Hong Kong Confederation of Trade Union, (5) Neighbourhood and Workers Service Centre, (6) United Front, (7) Independent Democrats.
(B) Pro-Government: (1) The Hong Kong Progressive Alliance, (2) Democratic Alliance for the Betterment of Hong Kong, (3) Hong Kong Federation of Trade Unions, (4) New Century Forum, (5) Independent.
(C) Moderate: (1) Liberal Party.
(D) Independent: (1) Independent.

Figure 1. A reference map of Hong Kong (an interactive map of Hong Kong is available from Centamap at http://www.centamap.com or YPmap at http://www.ypmap.com).

The 2004 LegCo Election was represented by 35 registered candidate lists for the 30 available seats (Table 1). The political scene of Hong Kong has been dominated by two political camps: Pro-Democracy and Pro-Government. Such polarisation in the political scene of Hong Kong was in part a result of fear of the communist regime by some citizens and in part caused by the Sino-British dispute over political reforms made by Christopher Patten, then Governor of Hong Kong [see also Wong (1998) and Ngok (2001)]. Other competitors include the Moderate camp and independent candidates.

The Pro-Democracy camp comprises representatives of the Democratic Party (DP), the Frontier, the Hong Kong Association for Democracy and People's Livelihood, Article 45 Concern Group, a few other front-line groups and individuals sharing similar political beliefs. The camp, particularly the DP, strives to speed up the process of democracy (such as encouraging universal suffrage for LegCo members and the Chief Executive) while developing closer economic ties with the Mainland (Wong 1998). The younger and better educated generation also seems to share the same stance and is more supportive of these values (Ngok 2001).

The Pro-Government camp includes the Democratic Alliance for the Betterment of Hong Kong (DAB), Hong Kong Federation of Trade Unions and groups or individuals having a patriotic view towards China and its policy makers (which is why the camp is sometimes labelled Pro-China). The camp strives to maintain a strong tie with China through adamant support of many reforms put forth by the Hong Kong Government. This camp, especially the DAB, seems to draw support from the more conservative voters and migrants from the Mainland. Members of the Pro-Government camp are known allegedly to have an extensive people network and good ability to mobilise their supporters (Ngok 2001).

The Moderate camp is represented by the Liberal Party known for its conservative, business-friendly and pro-Beijing policies. The party appeals mostly to businessmen and proprietors yearning for a stable environment for their investments. The Independent camp is represented by individuals whose political stance varies from those described above. Each of the political camps has a distinctive political ideology and agenda that appeal to supporters from different sectors. Such is the basis of our approach to devise social profiles of voters to discriminate characteristic supporters for the various political camps.

Although some researchers were uncertain about the role between social class difference in voters and their choice preference, there was general consensus that voters tended to pick candidates using information gained from flyers, mailings and campaign activities. The accessibility to information, however, is somewhat related to socio-economic and demographic factors. Kuan and Lau (2002) found that younger and better educated voters were often better informed and therefore more inclined towards democratic parties. The less-informed supporters of the democratic parties were a result of their personal attachments (e.g. influenced by family members, friends or current events). Lui (2003) reviewed the role of the middle class in the development of socio-economic and political scenes of Hong Kong in the past few decades. His findings indicated that the middle class was frustrated about adverse effects arising from the changing socio-economic and political environments of Hong Kong after the 1997 handover. The 1st of July rally in 2003 and active participation in the district council election that year showed that the middle class were moving from what Lui (2003) described as the ‘withdrawn’ mode to have more say in politics. The turnout rate in the 2003 district council election of Hong Kong was the highest ever (HKSAR Electoral Affairs Commission 2003). Such a high turnout rate was perhaps due to factors other than network mobilisation, especially in middle-class residential districts (Cheng 2004, 2005). The findings also showed that public perception and political inclinations of certain social groups can impact and shape the political development of Hong Kong.

Methodology

The study comprised four stages facilitated by GIS processing: (1) profiling of voters by political camps, (2) profiling of census population by voting districts, (3) predicting election results by matching profiles of voters against census population and (4) assessing the predicted and actual results.

Stage 1 attempted to establish characteristics of potential supporters for each political camp. Our basis for the generalised social profile of those more likely to render their support to the three major political camps were based on descriptive statistics of exit polls of previous District and LegCo elections (HKUPOP 1988–1999, 2000, Chung 2000, 2003, 2004). Then we tried to profile census population of the voting districts in Stage 2 using demographic and socio-economic characteristics identified in Stage 1. Our study used the 2001 census data to determine the dominant demographic and socio-economic groups in each District Council Constituency Area (DCCA). ¹ Given the census population data, our intent is to uncover possible spatial concentration of selected characteristics of potential supporters of the political groups in various voting districts.

Results of Stage 2 became inputs for our prediction model in Stage 3. Our approach considered one political group at a time and made prediction of its likelihood to win support from the population of various voting districts. Our model would rate each voting district according to the profile of its census population against the voter profile for each political group. A corresponding match between the two profiles would flag a greater success rate for a political camp to win support in the respective voting districts. A map revealing the potential of various political camps to win in the voting districts would result from the predictive modelling. Our predicted result in Stage 3 was compared against the actual election results for similarities and discrepancies in the final stage.

Results

Stage 1: Profiling of voters by political camps

The ultimate success of our predictive model depends largely on our ability to extract characteristics representative of potential supporters of various political camps (Burke 2003). Descriptive statistics from exit polls of the previous District and LegCo elections of Hong Kong (HKUPOP 1988–1999, 2000, Chung 2000, 2003, 2004) were used to derive characteristics of the voters. These results allow us to sketch out a fairly consistent profile of supporters for each political camp, and the selected variables also exist in census data, which makes it possible to test our voter profiles against the population at large.

Analyses of exit poll data from 1998 to 2003 yielded the following profiles of supporters for the three political camps (Table 2):

Table 2. Characteristic profiles^a of potential supporters by political camps

Political camp	Age (years)	Monthly household income (HK$)	Occupation	Education attainment	Marital status	Duration of residence
Pro-Democracy	Below 40	20K–40K	Clerk, Professional, Elementary	Upper secondary and tertiary education	Single	Over 10 years
Pro-Government	Over 40	Below 19K	Worker, Service Worker, Craftsman, Clerk	No schooling, primary or lower secondary school	Married	Less than 10 years
Moderate	Relatively senior (40–60)	Over 40K	Manager, Proprietor, Administrator	Above upper secondary school	Married	Over 10 years
Note: ^aDerived from exit polls of various elections held between 1998 and 2003 in Hong Kong. The exit polls were conducted by the Population Opinion Polls of The University of Hong Kong (HKUPOP).

1. The Pro-Democracy camp appealed to the young and middle-aged professionals, generally with a higher educational attainment and earning median to high income. Ngok (2001) and Lui (2003) indicated that this group tended to be more receptive of views towards democracy but had reservations about the present political construct.

2. The Pro-Government camp seemed to gain support mostly from the lower income groups, generally of lower educational attainment (Leung 1995, Cheung 2005). Migrants with less than 10 years of residency in Hong Kong were also likely to support this camp because of their mobilisation power and extensive connections with the local networks (such as clan associations or organisations).

3. The Moderate camp was backed by the relatively senior (around 40–60 years old) managerial or administrative professionals and businessmen with a high income (Leung 1995, Cheung 2005).

Stage 2: Profiling of census population by voting districts

The Census and Statistics Department of Hong Kong provides demographic and socio-economic data of Hong Kong by different regional breakdowns (such as Tertiary Planning Units or Street Blocks or DCCA) in their 2001 Census and Statistics Survey. To model results of the 2004 election, we used the 2001 census data, which was the only reliable and up-to-date source of demographic and socio-economic information of Hong Kong. Based on our findings from Stage 1, we made an attempt to profile the population of each DCCA according to age, income, education level and occupation. For example, the population of each DCCA would be classified into four age groups of 20–29, 30–39, 40–49 and over 50. The age group with the highest population number (or percentage) would be the dominant group. Each age group was assigned a colour/pattern, and the DCCAs would inherit the corresponding colour/pattern of the dominant age group, as shown in Figure 2a. The process was repeated for each of the variables of interest, and their spatial variations were portrayed in Figure 2b–d.

Figure 2. Distribution maps of demographic and socio-economic attributes.

Figure 2a showed that the predominant age group of Hong Kong in 2001 was 30–39. The 40–49 middle-aged group was found to prevail in the northern parts of the Hong Kong Island. The younger age group in the 20s was prominent in several regions, including West Kowloon, Ngau Chi Wan and Sai Kung. Figure 2b portrayed the distribution of dominant income groups by DCCAs. Persons with monthly income over HK$20,000 concentrated in most parts of the Hong Kong Island and the Kowloon Peninsula. Lower-income persons were found in southern Lantau and the northeast part of New Territories. The distribution map of educational attainment (Figure 2c) showed that the prevalent population of Hong Kong completed secondary school education. The majority of residents of the Discovery Bay in Lantau and those in the central parts of the Hong Kong Island and Kowloon had higher educational achievement. Interestingly, the dominant occupation of population in these DCCAs seemed to be professionals, managers and administrators although the population of Hong Kong was predominantly unskilful workers (Figure 2d). DCCAs dominated by professionals, managers and administrators (darker areas in Figure 2d) seemed to be more affluent (darker areas in Figure 2b).

The distribution maps of marital status and duration of residence were also produced but the patterns were rather homogenous and therefore not shown here. In general, the married category was rather widespread with only a few pockets of villages in the rural areas dominated by unmarried individuals. Most residents had lived in Hong Kong for more than 10 years. These variables were not used as predictors in our study.

Stage 3: Predicting election results by matching profiles of voters and population

We made use of the GIS technology to match profiles of the population from Stage 2 against those of the voters from Stage 1. Our intent was to determine whether it was possible to forecast political camps more likely to win in each DCCA based on historical profiles of their respective supporters. Several underlying assumptions are essential for a proper forecast: (1) the characteristics of voters for each political camp are distinctively different (as evident from studies by Cheung (2005) and Leung (1995)), (2) the characteristics of supporters of political camps remain invariable in the short-term transition period, (3) the exit polls are representative of the population at large (surveyed at about 1% of voters who turned out to vote or 0.4% of all eligible voters; see also the 2000 LegCo Hong Kong Island By-election exit poll, HKUPOP 2000) and (4) the generalisation based on DCCA is indicative of voting behaviours in the regions.

We first formalised the criteria of supporters for each political camp derived in Stage 1 and applied these criteria to extract results of the four demographic and socio-economic maps presented in Figure 2. Two stages of extraction were carried out using the overlay method of analysis: the first-level extracted DCCAs with dominant groups that matched the criteria of supporters of a select political camp; the second-level extract was less stringent to involve a much wider range of subgroups. The extracted results for each political camp were presented in two different colour/shading tones (Figure 3a–c). In general, a political camp would likely capture the majority of electoral votes in the shaded DCCAs; the darker shades showed geographic units of first-level agreement whereas the lighter shades were indicative of the secondary concurrence. Figure 3a predicted that potential regions of Pro-Democracy supporters would include eastern parts of the New Territories and Kowloon, as well as northern and eastern stretches of the Hong Kong Island. Areas of potential support for the Pro-Government group, as highlighted in Figure 3b, covered areas complementary to those for the Pro-Democracy camp to include northwestern New Territories, western Kowloon and eastern part of the Hong Kong Island. The geographically polarised distribution of the two political camps was not unexpected as their supporters originated from different sectors that had longstanding social differences. On the contrary, the Moderate camp featured in Figure 3c exhibited an almost similar distributional pattern as the Pro-Democracy camp because supporters of both camps appeared to possess rather similar socio-economic background. Potential supporters of the Moderate camp were more likely found in the Peak area and western part of the Hong Kong Island.

Figure 3. Potential regions of support for various political camps.

We combined Figure 3a–c to yield a composite map in Figure 4 that represented a prediction map of the 2004 LegCo Election. Figure 4 revealed the power distribution of the three political camps in the then forthcoming election. The map highlighted three types of power struggle: (1) unilateral, (2) bilateral or bipartisan and (3) indeterminate. Support for the Pro-Government or Pro-China camp was rather clear-cut in most instances, which extended over the north and northwest regions of the New Territories, West and East Kowloon (including Kwun Tong and Tseung Kwan O) and the eastern part of the Hong Kong Island (North Point, Chai Wan and Shau Kei Wan). We predicted that the Pro-Government camp would not face much competition from their political counterparts in these unilateral areas.

Figure 4. A prediction map of the 2004 LegCo Election of Hong Kong. The map was made before the election to reveal our speculation of potential areas of support for each political camp. (Available in colour online.)

Some areas were not so clear-cut in terms of their political supporters. Support for the Moderate and Pro-Democracy camps overlapped a great deal because of a significant degree of similarity in the demographic and socio-economic profiles of their potential supporters. These political camps seemed to appeal to the more affluent voters residing in Central Kowloon (including Mongkok and Yaumatei), southern regions of the New Territories (Tsuen Wan and Kwai Chung) and various districts of the Hong Kong Island (Central, the Peak, Wanchai and Causeway Bay). There were also a few areas of overlap between the Pro-Democracy and the Pro-Government camps, likely because of a crossover of dominance in demographic and socio-economic profiles of the population. Supporters of both the Pro-Democracy and the Pro-Government camps appeared to dominate equally in isolated areas of Shatin, Ma On Shan and the Chek Lap Kok Airport. Moderate and Pro-Government camps did not appear to compete with one another because of the polarised nature of the demographic and socio-economic profiles of their respective supporters. We predicted that areas exhibiting bipartisan support would give the concerned political camps an equal chance in claiming victory in these areas.

Less than one-fifth of DCCAs in Hong Kong was indeterminate. These areas remained unclassified because they were either sparsely populated or inhabited by people of varied socio-economic backgrounds with no dominant subgroups. We predicted these areas to be free-for-all and all political camps would have an equal chance in capturing the electoral votes.

Stage 4: Assessing the predicted and actual results

The actual performance of each political camp at the 2004 LegCo Election was summarised in Figure 5. The map of the election results clearly highlighted an overwhelming support for the Pro-Democracy camp shown in various green tones (in which darker tones indicated stronger support). Strong support was also evident from the over 50% of votes within the majority of voting areas secured by the Pro-Democracy camp, which included heavily populated areas in Kowloon, the Hong Kong Island as well as southern and eastern parts of the New Territories. Pro-Government support was endorsed by residents in suburban areas of Hong Kong including Tuen Mun, Yuen Long, southern parts of the Lantau Island and some outlying islands. The Moderate camp managed to capture the majority votes in northern parts of the New Territories bordering China (Figure 5).

Figure 5. Results of the 2004 LegCo Election. The map shows the majority vote counts of political camps by voting areas. (Available in colour online.)

A visual comparison between maps of predicted and actual election results (Figures 4 and 5) suggested almost similar voting patterns. The Pro-Democracy camp was predicted and also won support from urban and populated areas in Eastern Kowloon and the Hong Kong Island. The Pro-Government camp was predicted and earned votes as well in the New Territories and peripheral areas. The greatest disparity between the two maps appeared in Western Kowloon, which was predicted to favour the Pro-Government camp but eventually won by the Pro-Democrats. Furthermore, our prediction map was not able to single out areas of majority support for the Moderate camp, in all likelihood because of the following reasons: (i) the demographic and socio-economic profiles of the potential supporters were not characteristically clear and (ii) the comparatively small number of candidates for this camp.

The rendering of quantitative assessment was not straightforward. Our major problem arising from incompatible geographic units between maps of predicted and actual election results warranted further processing to resolve the issue of area mismatch. The common areas for quantitative comparison were obtained by intersecting both maps, which had 390 and 485 polygons, respectively. We adjusted for minor disagreement along area boundaries and removed small polygons. Although our adjustments were not entirely scientific, the approach would not affect the comparative assessment of area-based counts as shown in Table 3 in which the χ² analysis shows a statistically significant relationship between predicted and actual results for the two largest political camps. The greatest mismatch occurred in areas predicted to have Pro-Government support that did not materialise in the election. We would attribute such a misfit to the continual influx of new migrants with no voting rights into the new territories, which might have distorted the representation of voters through the census statistics. The predictive power of our model was not exceptional but the map was rather revealing. We were able to extract the major trend of political stands even though our model was based on voters' profiles extracted from exit polls and constructed from merely four characteristics.

Table 3. Results of the quantitative comparison of correct matches between predicted and actual election results

	Predicted majority^a of votes
Actual majority^a of votes	Pro-Democracy	Pro-Government	Moderate^b	Indeterminate^b	Row total
Pro-Democracy	97	168	8	45	318
Pro-Government	33	26	5	8	72
Moderate^b	0	0	0	0	0
Column total	130	194	13	53	390
Note: ^aThe political camp receiving the highest numbers of votes in a particular District Council Constituency Area (DCCA) in the prediction map.
^bModerate and Indeterminate cases were excluded from the χ^2 test.
χ^2 test for 2 × 2 contingency table.
H o, there is no relationship between predicted and actual values; H A, there is a relationship between predicted and actual values; χ^2 (df = 1) = 7.503.
The critical values lie between 3.841 and 6.635 for probabilities 0.05 < ρ < 0.01.
Thus, the null hypothesis that there is no relationship between predicted and actual values of voting counts is rejected. The relationship between the predicted and the actual values is statistically significant.

Constraints and limitations of our modelling approach

Various survey methods, economic models and statistical approaches have been employed to predict election but the use of maps is a relatively novice approach. Mapping enables the use of census data to build a population profile for each geographic constituency. Although census or aggregated data may not be the best predictor, they do provide a means to make inferences about the background of the population or voters when reliable data at the individual level are not available. The approach also offers valuable clues on the voting behaviours of individuals of certain demographic and socio-economic attributes.

Our GIS modelling and mapping techniques are not without problems. The foremost concern involves the four underlying assumptions of the approach stated earlier in Stage 3 of the Results. Some of the conditions are difficult to realise at times. The first assumption presupposes that characteristics of the voters favouring each political camp are distinctively different. It may be possible to derive a discrete profile for supporters of a political party already in existence, but newly formed political groups will not have a discerning profile of their supporters. We encountered similar situations in the 2004 LegCo Election in which a few celebrities or well-known public figures decided to turn politicians. These individuals might not have their political ideology known but their names were well recognised to influence voters of all social classes. Incidentally, the star-effect was also observed in the 2004 Presidential Election in the Philippines between the incumbent Gloria Mapacagal Arroyo and the action film star Fernando Poe Jr., although the former was eventually the proclaimed winner.

Secondly, we assumed that socio-economic status and demographic characteristics were the only influential factors. This assumption did not take into account changes or disruptive events (such as wars or radical birth control policy) that are both variable and transitory. Past elections allowed us to witness how propaganda, campaign strategy, publicity and libel could overturn profoundly the election outcomes (Butler and Stokes 1974, Krassa 1995). Researchers have contended that the social environment has a role in shaping information, which, in turn, influence an individual's political perception and propensity to vote. A voter's electoral choice appeared to be associated not only with the individual's social characteristics but also with the social quality of the area of residence. For example, Butler and Stokes (1974) found that the working class living in working-class areas in England was more likely to vote for the Labour Party than the working class living in middle-class areas, whereas the middle class living among its own kind was more likely to vote for the Conservative Party than the middle class residing in working-class areas. Lay et al. (2008) also pointed out that partisan support in Taiwan exhibits significant regional differences and suggested targeting voters in the more unsettled regions to safeguard victory. We were not sure whether this ‘neighbourhood’ effect was widespread in Hong Kong but did observe that the Pro-Government candidates seemed more confident in securing votes from residents in the North Point district (along the northern stretch of the Hong Kong Island) than elsewhere in Hong Kong. Our profiling approach is deficient in handling such irregularities.

Thirdly, we suggested that voters who turned up at the polling stations were faithful representation of the population in the respective constituencies. The corollary is that the election results are representative of the people's will at large. Although the methods of sampling and statistical adjustment were designed to improve the degree of representation (HKUPOP), voters appearing at polling stations in a constituency might be overwhelmingly biased to skew the results. This situation was evident in the 2004 LegCo Election in Hong Kong, where a group of elderly persons in a village were persuaded to vote for a bespoken candidate. This sudden emergence of a specific voter group at a polling station would disrupt the representation of the election results or exit polls that recorded data uncharacteristic of voters of that constituency and would lead to unreliable prediction in the modelling.

Fourthly, the political inclination of specific groups could have changed over time (Abramson 1983). The participation of the middle class in politics is a good example. This group of voters, traditionally considered withdrawn and politically inactive before the handover, took a more active attitude in 2004 (Cheng 2004, 2005). Besides, campaigns by various political camps and targets of these campaigns could have implications on their supporters. It is important for one to examine the changing political atmosphere and campaign targets of various political camps if one were to adopt our approach in modelling.

Finally, we employed census data to establish our predictors under the assumption that the generalisation based on DCCA is indicative of voting behaviours in the regions. There was the basic question on whether the census population (a sampled enumeration of the total population) was characteristic of registered voters that made up 48% of residents in Hong Kong in 2004. The dominant subgroups from each of the demographic and socio-economic attributes (i.e. age, income, education and occupation) might represent a fraction of the registered voters and therefore not be representative. Furthermore, the dominant subgroups of many geographic constituencies had percentage values slightly higher than the next group within the same precincts. As a result, some prominent subgroups of a specific attribute would have been inadvertently omitted in our modelling approach. We attempted to alleviate the problem by imposing a two-stage extraction: the first pass to include the dominant subgroup and the second pass the secondary dominant subgroups. This modification broadened our scope and extended the ability of our model to increase the rate of matching the voters' profiles against those of potential supporters of individual political camps.

Conclusion

Despite limitations described above, the study has demonstrated the use of GIS methodology in forecasting election results. The predictability of our profiling approach was limited largely because of unavailability of more individualistic statistics about the voters other than the aggregated census data. Besides, the robustness of a prediction model hinges on the ability of the modeller to identify with political development and the local situations. Although our model might not seem very accurate, the approach was sufficiently sound to warrant further examination with better voter-related statistics. In addition to demographic and social-economic traits, data on political preference of voters in the recent past and methods to account for effects of transient events significant enough to swing votes would be relevant.

Election events tend to draw widespread attention over a relatively short time span. Although the use of numerical reporting of election results involving sufficiently few electoral areas (e.g. 5 LegCo constituencies and 18 districts) was not challenged in the past, this tabular method of accounting was no longer effective nor acceptable given finer breakdown of data into 485 geographic units. The GIS technology seems a viable and favoured alternative given its mapping and modelling functions. In the case of Hong Kong, electoral and census enumeration areas were unfortunately a total misfit, which defeated the efficiency of the GIS techniques in undertaking electoral analysis. There is also a fundamental need for the Election Offices to institute more effective methods of counting, recording and dissemination of vote counts to the public.

Acknowledgements

We thank Ms. K.W. Chan and student helpers for their proficiency in data collection, verification, processing, modelling and cartographic presentation. Special thanks are extended to the Department of Geography for the use of GIS facilities and to the following news media for their sponsorship in the election mapping exercise: Apple Daily, Hong Kong Economics, Ming Pao and Radio Television Hong Kong.

Notes

1. DCCA is defined by the Electoral Affairs Office (EAO) for District Council election. A few regions may vary from year to year as the EAO will redefine the districts according to factors, like number of population. More details are available from the Report on the Recommended Constituency Boundaries for the 2003 District Councils Election. Available from: http://www.info.gov.hk/eac/en/distco/2003dc_boundary_f.htm [Accessed 19 February 2010].