Persistence in regional voting patterns in Turkey during a period of major political realignment

Using province-level data from five nationwide elections held during the past decade, we examine the main voting patterns in Turkey. By means of cluster analysis, we classify the 81 provinces according to vote shares of the major parties and independent candidates, and repeat this exercise for each election held between 1999 and 2009. We find that three-way and five-way partitions of the country adequately capture the main political cleavages in Turkey. Although the conservative right-wing parties receive a plurality of votes in all regions of the three-way partition, they receive significant challenge from left-wing and Turkish nationalist parties in the west and from the Kurdish nationalist parties in the east. In addition to these patterns, the five-way partition brings out shifts in the relative strength of the parties within each main division. Our results also show that, despite the major political realignment that occurred during the period under examination, the groupings of provinces remain mainly unchanged. Therefore, we construct ‘composite clusters’ by classifying provinces in the group in which they appear the majority of the time. The distinct socioeconomic and demographic characteristics of the composite clusters suggest that differences in social and economic structures lie at the root of differing regional political tendencies and their persistence.


Introduction
Turkey went through an unprecedented political transformation between 1999 and 2009. After experiencing rampant corruption, poor economic performance and constant infighting under various coalition governments during the preceding decade, voters in the November 2002 election ousted all of the parties that had entered the parliament in 1999. Included among these were the Motherland Party (ANAP), which held the premiership during 1983-91 and 1997-9, the True Path Party (DYP) and the  Democratic Left Party (DSP), which led governments during the 1991-6 and 1999-2002 periods, respectively, and the Nationalist Action Party (MHP), which was part of the ruling coalition between 1999 and 2002 together with the DSP and the ANAP. None of them was able to exceed the 10 percent nationwide vote share threshold required to be represented in the Turkish Grand National Assembly. In addition, the Virtue Party (FP), the successor to the Welfare Party (RP) -which was the only other party to lead a government during the 1990s -had already been banned in 2001 by the Constitutional Court, just like its predecessor. The combined share of votes for the ANAP, DYP, DSP, MHP and FP was 81 percent in 1999 but only 24 percent in 2002 (see Table 1). Of these parties, so far only the MHP has been able to engineer a come-back. The party's vote share of 16 percent in 2009 was almost the same as it was in 1999. The rest of the parties continued to lose ground In hindsight, the building of voter frustration was quite visible. In each of the parliamentary elections since 1987, a different party finished first. Since 1991, each election produced a coalition government involving different combinations of parties. During this post-cold war period, rapid urbanization and the market-oriented economic restructuring initiated earlier and greater integration with the global economy, leading to a major transformation of Turkish society, which many of the existing parties failed to recognize and adjust to. 1 Based on a 2001 survey, Çarkoğlu and Hinich (2006) find that rising disenchantment with existing parties at the time was the most pronounced among the large segment of voters positioned at the centre of the secularist vs. pro-Islamist and the Turkish vs. Kurdish nationalist cleavages that -according to the authors -dominate the ideological competition in the Turkish party system. Given their finding, it would have been easy to anticipate eventual replacement of the existing centrist parties by a new movement that would capture much electoral support by pursuing moderate conservative policies.
In 2002, observing no change in party leaderships and policies despite diminishing vote shares, voters finally lost faith in most of the established parties and cast their votes overwhelmingly for three parties: the Justice and Development Party (AKP), the Young Party (GP) and the Republican People's Party (CHP). The first two parties were formed shortly before the election, and the last one had been out of the parliament during the previous legislative term, having received less than 10 percent of the votes in the 1999 election. The AKP, which emerged from the ashes of the banned FP, received the lion's share of the votes under the leadership of Recep Tayyip Erdoğan, a former mayor of Istanbul. The party's disavowal of its Islamist roots, its embrace of the political and economic reforms necessary for Turkey's accession to the European Union (EU), the non-corrupt and populist image of its mayors at the local level and the likelihood of its forming a singleparty government all appealed to the electorate, which deserted the right-wing ANAP, DYP and MHP. 2 Interestingly, this support did not come at the expense of the traditional pro-Islamist voter base, which also transformed like the rest of society. The Felicity Party (SP), the other party rooted in the FP, toed the old party line and received only 2 percent of the votes, in contrast to the AKP, which captured just over 34 percent. The AKP's vote share expanded to 42 percent in 2004 and to 47 percent in 2007, but declined to 38 percent in 2009. The CHP vote share has hovered around 20 percent, and exceeded 23 percent in 2009. The GP, however, turned out to be a flash in the pan. Its 7 percent share in 2002 fell to 3 percent in 2007 and to zero in 2009, when it did not even participate in the election.
In short, a massive shift in votes occurred between 1999 and 2009, which was far more than could be attributed to the usual depreciation in the political capital of ruling parties or voter dissatisfaction with poor economic performance. What occurred during this period would be better described as a realignment of the electorate, in that a large number of voters changed the party that they identified as representing their interests and ideological leanings. 3 In this period, support for the centre-right parties ANAP and DYP, which dominated Turkish politics during the 1980s and 1990s, melted away. After capturing former supporters of the FP, ANAP and DYP, the AKP now represents much of the right wing, except its Turkish nationalist branch, which is still represented mainly by the MHP. The centre-left parties, which used to capture collectively about onethird of the votes, even reaching 40 percent on occasion, are now able to capture no more than onequarter. With the virtual disappearance of the DSP, the CHP is left as the only centre-left party. Despite having leftist elements in its programme, the Democratic Society Party (DTP) should be viewed primarily as a Kurdish nationalist party. 4 The fact that the combined vote share of the ANAP, DYP and DSP was quite small in 2009, and has stabilized in the last two elections, can be taken as an indication that the realignment in question has essentially come to an end. 5 The main purpose of this paper is to study the extent of the change in the political map of Turkey during the realignment process by identifying regions that are distinct from the rest of the country in terms of voting patterns and discussing specific patterns within them. This will be accomplished in the next section with the application of cluster analysis to cross-provincial data from five elections held between 1999 and 2009. Owing to its size, geography and history, Turkey exhibits a great deal of regional diversity in terms of demographic, economic, social and cultural characteristics, and, consequently, in terms of political tendencies. It is both a European and an Asian country. It can be simultaneously characterized as being in the Middle East, in the Caucasus and in the Balkans. It is one of the countries on the shores of the Black Sea and one of those on the Mediterranean as well. Its size is more than twice that of Germany, and more than three times that of the United Kingdom. It borders EU countries Greece and Bulgaria at one end and Iran, Iraq and Syria at the other. Turkey was at the centre of the Ottoman Empire and is its successor state. Consequently it houses people with a variety of ethnic roots. For example, its south-eastern region is home to a majority of Kurds and some Arabs. The Marmara region is where those who migrated from the Balkans and Crimea during the decline of the Ottoman Empire are concentrated, whereas the eastern Black Sea region is where those arriving from the Caucasus predominantly settled. Although the religion of the overwhelming majority of Turkish people is Islam, not all belong to the same sect, and the degree of their religiosity and the emphasis they place on religious freedoms and secularism vary. Turkey incorporates many different types of climate and thus different types of economic activity. For example, the main crops produced along the Black Sea are hazelnuts and tea, whereas they are cotton and olives in the Aegean and Mediterranean regions. Central Anatolia, which is sealed from the climactic effects of the Black Sea in the north and the Mediterranean in the south by high mountain ranges, grows mainly wheat, other types of cereals and sugar beets. The provinces around the northern, eastern and southern shores of Marmara are heavily industrialized, whereas those along the Aegean and the Mediterranean rely greatly on tourism and on light industry based on industrial crops. In the mountainous east, animal husbandry is important.
With this much diversity, it is natural to expect people in different parts of the country to have different economic interests and ideologies and, consequently, to vote differently. 6 Furthermore, in light of the much-discussed link between culture, economic development and political outcomes in the works of scholars such as Ronald Inglehart and Robert Putnam, it is likely that the electorates within each region alter their votes in tandem when changing economic and social conditions dictate it. 7 For this reason, there seems to be a lot of merit in carrying out an empirical analysis of regional voting patterns during a period of political realignment. In an attempt to observe whether political outcomes are indeed related to other characteristics of society, the current study also briefly examines whether the 'political' clusters obtained are also useful representations of underlying social and economic disparities.

Cluster analysis of voting patterns in Turkey
Before discussing the provincial clusters obtained, it will be useful to explain in greater detail the methodology employed in achieving them. To avoid arbitrariness in partitioning a population or sample, researchers often employ cluster analysis. This technique is designed to arrange the units in a population or sample into groups (or clusters) of relative similarity with respect to a number of characteristics. In our case, the 81 provinces are classified according to the vote shares of the major parties and independent candidates, and this exercise is repeated for each election held between 1999 and 2009. Three of these elections (1999, 2002 and 2007) were held to elect members of the parliament and two (2004 and 2009) were for local administrations. In the case of the latter, we took the results of contests for provincial councils because the electorate for these is the same as for the parliamentary elections. 8 There are two types of algorithms one can use in cluster analysis: hierarchical clustering and k-means clustering. In the former, initially each unit in the population is treated as a separate cluster. Then clusters that are closest to each other are merged successively, until all units form a single cluster. The practitioner then decides, depending on the context, which level of agglomeration is the most appropriate to work with. In the latter approach, first the number of clusters to be formed (k) is decided and the units in the population are assigned randomly to one of the k clusters. The units are then moved between the clusters so as to minimize variability within clusters and to maximize it between clusters. We selected the k-means procedure for our analysis. As West (2005) explains, when making comparisons between partitions involving different numbers of clusters, as we intend to do, the partitions obtained with the k-means method are more meaningful. In hierarchical clustering, each step in the process of aggregation is built upon the organization of all the previous steps. This means that the composition of both of the clusters to be compared is embedded in the initial pairings. With the k-means clustering, in contrast, each cluster pattern begins from an independent organization of the units in the population. The hierarchical method also tends to produce unevenly sized clusters, which are less likely to reveal the main political patterns.
As to the value of k, since our ultimate aim is to compare and contrast voter bases, we tried different values until a meaningful picture emerged. It turned out that, for each of the five elections between 1999 and 2009, a three-way partition of the provinces captures much of the variation in voter behaviour across the country, but a five-way partition allows a little finer separation. There are two reasons why we decided not to work with even larger number of clusters. First of all, the two stopping rules available in the STATA program, namely the Calinski & Harabasz and the Duda & Hart indices, were inconclusive. Secondly, when the value of k was increased beyond 5, it resulted in new clusters that were quite similar to the existing ones in terms of vote shares, or contained very few provinces that stood out as having atypical voting patterns (usually having to do with the candidacy of a locally popular person). Consequently we reached the conclusion that the five-way partition yields maps that adequately capture the main political divisions in Turkey.
Cluster analysis has previously been applied in several studies to group Turkey's provinces based on voting patterns. Çarkoğlu (2000, 2002 and 2010) does this for the 1999elections, respectively, West (2005 for the 1999 and 2002 elections, and Şekercioğlu and Arıkan (2008) for the 2007 election. Three-way partitions of the provinces emerge as useful representations of the main voting patterns in those studies as well. 9 However, comparisons between the findings of these studies are not entirely appropriate, because they differ in the clustering method and the algorithm used. Our aim is to minimize such concerns by applying exactly the same methodology to each data set at hand. 10 The three-way and five-way groupings of the provinces We now turn to the spatial characteristics of the clusters obtained with the methodology discussed in the previous section, and discuss how they have evolved over time. The three-way partitions obtained for each election are given in Figure 1. It is remarkable that, in a period of major political turmoil, so many of the provinces remain in the same cluster election after election. Going from west to east, cluster 1 (black), typically follows the Mediterranean, Aegean and Marmara coasts and juts out from the eastern Marmara Sea inland all the way to Kırşehir. It also includes provinces adjacent to the coastal ones along the Aegean. In 1999, the cluster includes almost the entire Black Sea region as well, but after 2002 only a few provinces scattered along that coast remain in it. Actually, some of the provinces along the Marmara-Kırşehir axis leave the region in 2002, but their departure is only temporary. The number of provinces in this cluster ranges from 25 to 33, and, with the exception of 2002, these provinces account for more than half of the electorate. Cluster 2 (dark grey) covers much of the rest of the nation apart from mid-eastern and south-eastern Anatolia. It accounts for just over one-third of the electorate, with the exception of the 2002 elections, in which it accounts for nearly 60 percent of voters. Cluster 3 (light grey) covers a triangular region made up of south-eastern and mid-eastern Anatolian provinces. It comprises 10-15 provinces and fewer than 10 percent of the voters. This region is populated heavily by ethnic Kurds.
Going from west to east, the clusters in the fiveway partition are numbered 1 to 5, and they are shown as white-dotted black, black, dark grey, white-dotted dark grey and light grey, respectively, on the maps in Figure 2. Comparing the k = 3 and k = 5 maps for each election, one can see that the provinces that make up the new black cluster in each election come, almost exclusively, from the black and dark grey clusters of the three-way partition.  white-dotted dark grey in the five-way partition. No provinces change from black to white-dotted dark grey in the 2002 and 2007 elections. In light of these figures, we expect to find the newly formed clusters displaying a mixture of the main patterns captured through the three-way partition of the country.
As in the previous maps, it is noticeable in the five-way partition maps that some drastic changes occurred in 2002. In 1999, the provinces in the Aegean and Marmara regions are white-dotted black, but the coastal provinces of both the Mediterranean and Black Sea are black. The dark grey cluster is land-locked and is in two separate groups. In 2002, the Mediterranean coast turns white-dotted black and remains that way until 2009. On the other hand, Istanbul, Turkey's largest metropolis and lying on two continents along the northern Marmara coast, changes from white-dotted black to black and remains that way in the next three elections as well. The dark grey cluster becomes contiguous and expands mainly at the expense of the black and white-dotted dark grey clusters. The eastern half of the Black Sea coastline and parts of the western Black Sea and southern Marmara Sea areas turn dark grey. Although the black colour returns to these areas again in 2004, it is only until the next election. In 2007 and 2009, most of the Black Sea coast, except for a few provinces, is in cluster 3. The five-way maps also show that there is considerable mobility between white-dotted black and black provinces. That is probably because the voting patterns in those two regions are quite similar, as we will explain below. The white-dotted dark grey cluster mostly contains the provinces located along the southeastern border of k = 3's dark grey cluster and the eastern border of k = 3's light grey cluster. Some other provinces scattered around the country also become white-dotted dark grey in one election or another but not persistently.

Composite clusters
As discussed in the previous section, in both the three-way and five-way partitions, provinces that change clusters between elections are few in number, and those with multiple changes are even fewer. The stability of regional groupings brings to mind the possibility of obtaining a summative picture of the political divisions in the country throughout the decade under examination. In other words, it seems reasonable to construct 'composite' three-way and five-way maps by placing provinces in the cluster colour that they take on the majority of the time. The clusters we have constructed in that manner will be presented in this subsection and utilized in ensuing analyses as the basis of interregional comparisons.
In obtaining the composite clusters, it was found that ties occur between two colours in the cases of one province in the k = 3 maps and nine provinces in the k = 5 maps. In those cases, the tie was broken in favour of the cluster that is numbered closer to the one in which the province appears once (assuming that cluster numberings order these regions in terms of the similarity in voting patterns as well). There were also two provinces (Kars and Bartın) that take on four different colours. For the sake of completeness, an 'average' cluster number was computed to classify them by taking into account the voting patterns in all clusters involved.
The resulting maps are presented in Figure 3. It is interesting that these groupings reflect the current situation closely. For example, the three-way composite map differs from the 2009 election's three-way map in only four provinces. This finding may be interpreted as another indication that, after a long realignment, Turkey's political map may be stabilizing near to what might be considered a longterm equilibrium. The maps also reveal an advantage of working with composite clusters, which is that they eliminate outlying observations at the provincial level. In many cases it is observed that provinces that change their colour return to their previous colour after one election. By construction, the composite clusters eliminate such anomalies from the picture.
Comparing the three-way and five-way composite maps in Figure 3, one can see that the black cluster of Figure 3(a) splits almost evenly into two in Figure 3(b). Of the 29 black provinces, 15 along the Mediterranean, Aegean and western Marmara coasts turn white-dotted black. However, the 14 along the very east Mediterranean, the Black Sea and the eastern, southern and northern Marmara coasts, together with 4 previously dark grey provinces, make up the new black region. Of the 38 dark grey provinces in the three-way composite map, 29 stay dark grey in the five-way map. Another 5, together with 2 light grey provinces from the three-way division, compose the white-dotted dark grey cluster in the fiveway composite map. Finally, the 12 light grey provinces in the five-way partition are all among the 14 light grey provinces in the three-way partition.

Socioeconomic characteristics of the composite clusters
Before discussing the differences in voting patterns between the composite clusters, it will be useful to examine the differences in their socioeconomic and demographic characteristics first, which we turn to next. Since the most recent population census was carried out in 2000, the figures we make use of, in this regard, will pertain to that year.
The population-weighted averages of provincelevel data, given in Table 2, reveal that, of the three clusters in the three-way partition, cluster 1 is the most urbanized, most densely populated, richest, most educated and most modern (as indicated by the median age, the proportion of non-agricultural employment and women's share in it). It has a Notes: The two composite maps are obtained from the maps in Figures 1 and 2, respectively, by placing each province in the cluster in which it appears three or more times. For a few provinces, which appear twice each in two regions and once in another, the one of the former that is adjacent to the latter is chosen. In the case of five-way partitions, Bartın and Kars take on four different colours. A weighted average is used to place these provinces, assuming that the regions rank from 1 to 5 in their closeness also. The clusters in Figure 3 positive net migration rate, as opposed to negative rates for the other two. It also incorporates more mobile and more cosmopolitan segments of the population, with close to half of its population born in another province. Cluster 3 lies at the other extreme and cluster 2 lies somewhere in between, but is closer to the third cluster than to the first. Cluster 2 is not that different from the first in median age and education level but differs significantly in all other areas. On the other hand, cluster 2 is similar to cluster 3 in urbanization and the proportion of non-agricultural employment, but differs from it considerably in income, net migration, education, median age and share of females in non-agricultural employment. Although no reliable recent data exist on the ethnic composition of provinces, the projections made by Mutlu (1996Mutlu ( ) from 1965Mutlu ( to 1990 show that all except one of the provinces with a Kurdish share of more than 40 percent are in composite cluster 3. The provinces in cluster 2 all have less than 20 percent, and those in cluster 1 all have less than 10 percent Kurdish population shares. Thus the third region differs from the other two in terms of ethnic composition as well as the other aspects presented here.
In the five-way partition, cluster 2 emerges as the most urbanized, most densely populated, most industrialized, most educated, most cosmopolitan and richest region. It ranks second to cluster 1 only in median age and net migration rate. However, the latter is partly owing to two major earthquakes that hit a number of the provinces in cluster 2 in 1999. It should also be noted that clusters 1 and 2 do not differ much from each other with respect to education level, per capita income, median age, female share in non-agricultural employment and net migration rate. Moving in the eastern direction, all of the reported indicators decline, except the ratio of nonagricultural employment to total employment, which is slightly higher in cluster 5 than in cluster 4. The net migration rates in clusters 3-5 are negative, in contrast to the first two clusters, where the rate is positive. As a consequence of this, the proportion of people residing in a province other than the one they were born in.
In the absence of data from nationwide surveys on attitudes and beliefs, the variation in cultural norms that lead to different political outcomes in different parts of the country is not easy to demonstrate. However, in view of the well-documented link between cultural characteristics and economic development and education, the socioeconomic evidence provided in this section can be taken as support for the presence of a similar amount of cross-regional variation in cultural characteristics. The specific regional voting patterns to be discussed in the next section, involving major interregional discrepancies in conservative and/or nationalistic tendencies, are indeed suggestive of a society with multiple political cultures.

Voting patterns in the composite clusters
Up to this point, we have identified the regions that consist of provinces that exhibited similar voting patterns during the past decade, and compared them in terms of various socioeconomic characteristics. However, we still have not discussed what those 'voting patterns' are. We now do this by summarizing the vote shares of the major parties in each of our composite clusters in the five elections being studied. Interested readers can find the vote share figures in Akarca and Başlevent (2010a). For brevity and to facilitate our explanations, we will discuss voting patterns in the five-way partition first.
The AKP has a significant presence in all regions. However there is a tendency for the party's vote share to increase from west to east, reaching a maximum in either cluster 3 or cluster 4, and then declining a little. The FP and the SP exhibit a similar pattern and reach their peak in cluster 4. The AKP vote share peaks in 2007 in all clusters except the first, where it reaches its peak in 2004. The vote losses experienced by the AKP between 2007 and 2009 also show a tendency to rise from west to east, peaking in cluster 4. Contrary to the impression created by the maps presented in the Turkish media immediately after the March 2009 election, the decline in the AKP's vote share is lowest in cluster 1. As a matter of fact, of the 15 provinces in this cluster, the party's vote share increased in 3 and declined by fewer than five percentage points in 5. The SP is almost non-existent in cluster 1. In the remaining clusters, its vote share tends to rise in local elections (2004 and 2009) and to fall in parliamentary ones, which is consistent with the concept of 'strategic voting'. In other words, the religious-right voters seem to be using this party to send a warning to the ruling party in local elections, without any risk of toppling the government. At the same time, they do not risk wasting their vote either, because the 10 percent national threshold does not apply in local administration elections.
The CHP and the DSP votes are highest in cluster 1. They decline a little in cluster 2 and drop sharply in the eastern direction. In 2002, the combined vote share of these two parties drops to its lowest levels in all clusters. Although the vote share increases between 2002 and 2009 in clusters 1 and 2, it never reaches its 1999 level. In the other clusters, the already low vote share of the two parties either stays that way or declines further. Thus, it might be argued that these parties are increasingly becoming regional parties with a significant presence in only the western provinces of the country. Actually, the DSP has become negligible even in the west. As for the MHP, the party's vote share generally declines from cluster 1 to 2, but increases from 2 to 3, peaking in that region, and then declines continuously and sharply towards the east. In cluster 1, the party's vote share is higher in 2009 than in 1999. In the remaining clusters, however, the situation is the opposite. In other words, the centre of gravity for the party's support appears to be shifting towards the west. GP votes decline steadily from west to east, whereas the DYP/DP and ANAP votes appear to be distributed almost evenly across the country. However, all three parties seem to be moving towards extinction or marginalization over time in all of the clusters. Finally, the votes of the Kurdish nationalist parties (HADEP in 1999, DEHAP in 2002, SHP in 2004, and DTP in 2007 come mostly from cluster 5 and to some extent from cluster 4 (see the notes to Table 1  Looking at the overall voting patterns by clusters, we find that, compared with cluster 1, in cluster 2 the combined vote shares of left-wing parties (CHP and DSP) are slightly less and of right-wing parties (AKP, SP, MHP, DYP/DP, ANAP and GP) are slightly more. The main difference between the two clusters is in the distribution of votes among the right-wing parties. In cluster 2, the AKP and the SP vote shares are higher and those of all other rightwing parties are lower, whereas, in region 1, the proportion of Turkish-nationalist MHP voters is larger. If we consider the fact that the left-wing parties in Turkey also take a Turkish-nationalistic stance, we can say that cluster 1 shows a greater Turkish-nationalist bent than cluster 2. The rise in nationalism in cluster 1 can, at least in part, be attributed to the massive influx of internal migrants over the past two decades (as demonstrated by the net migration rate of 26 percent in Table 2), especially from provinces with large Kurdish populations. Cultural disparities and economic struggles between the natives and the migrants have inevitably caused ethnic tensions and given rise to nationalist feelings, as observed in parts of Western Europe where the influx of foreign migrants has led to the strengthening of ultra-nationalist movements. 11 Cluster 3 is overwhelmingly conservative, of all shades. Right-wing parties receive votes far above their national averages there, whereas left-wing parties are far below theirs. Cluster 4 differentiates itself from its two neighbouring clusters in that the Kurdish nationalist votes are higher than in cluster 3 but are much lower than in cluster 5. However, the Turkish nationalist and centre-left votes are lower than in cluster 3 and higher than in cluster 5. In addition, the religious-right and centre-right votes tend to be slightly higher in cluster 4 relative to cluster 3, but substantially lower relative to cluster 5. Finally, the domination by the Kurdish nationalist parties and the AKP, and the lack of a strong presence by other parties, characterize cluster 5.
The votes received by the Kurdish nationalist parties in cluster 5 (i.e. the south-east of the country), although substantial, are hardly sufficient to cross the 10 percent nationwide threshold applicable in national elections, because these parties receive very few votes in the other parts of the country. This leads some ethnic Kurds to vote for independent candidates or for a party that is not their first choice, in order not to waste their vote. Furthermore, some block voting occurs when tribal leaders run, very often as independent candidates. These no doubt play some role in the emergence of this region's distinct voting pattern. 12 Once the voting patterns in the five-way partition are uncovered, those in the three-way partition are easy to anticipate. As noted above, the voting patterns in clusters 1 and 2 of the five-way partition differ only in nuance, and these are essentially the two halves of k = 3's first cluster. Consequently, the vote shares in cluster 1 of the three-way partition are basically the averages of the first two clusters of the fiveway. There is also much overlap between the provincial make-up of the dark grey clusters of the three-way and five-way partitions. Thus, it is not surprising that their voting patterns are also alike. Likewise, the light grey clusters of both partitions are almost identical, and exhibit similar voting patterns. In short, the three-way partition identifies regions based on the relative strengths of main political divisions in the country. Right-wing parties receive the majority of votes in all regions of the three-way partition, but they receive a significant challenge from left-wing parties in cluster 1 and from the Kurdish nationalist parties in cluster 3. The five-way partition also picks out differences in the relative strength of the parties within each main division.

Conclusions
The purpose of our study was to explore the regional differences in voting patterns in Turkey and to present these differences along with some of the socioeconomic characteristics on which they may be built. Unlike earlier studies dealing with these themes, our findings are based upon an entire decade, covering five nationwide elections held between 1999 and 2009. We also analysed each one of these elections with exactly the same procedure in order to discern areas with differing voting patterns more clearly, without the hindrance of noise generated by temporary or local factors, and to check the persistence of observed patterns over time.
Applying the k-means version of cluster analysis to province-level data from the five elections, we found that three-way and five-way partitions of the country yield maps that adequately capture the main political divisions in Turkey. More remarkably, we found that the provincial make-up of the clusters remains essentially unchanged during a decade when Turkish politics was going through a major realignment. 13 It appears that, even when they change their party allegiances, voters in each partition behave in a similar manner. It seems that the cultural, ethnic, socioeconomic and historical ties that bind people together have lasting effects and are the main determinants of political patterns in the long run. This finding inspired us to construct summative groupings, which we termed 'composite clusters'. The fact that these clusters also exhibit distinct variation in socioeconomic and demographic characteristics indicates that they capture more than just voting patterns.
With the help of the clusters obtained, we were able to study the support given to various parties in Turkey and the realignment that has taken place, both spatially and temporally. The vote share of the AKP, the conservative party in power, increases in all parts of the country until 2007, except in the very west, where it peaks in 2004. The strength of the party shows a tendency to increase going from west to east. However, the losses the party suffered in the 2009 election tend to rise going from west to east as well. Consequently, one can say that the relative strength of the party has evened out across various parts of the country. The combined vote share of the two left-wing parties (the CHP and the DSP), on the other hand, decreases going from west to east. Their support has been rising in the western and southern coastal provinces since bottoming out in 2002, but it has not yet reached its 1999 level. Their share in central and northern Anatolia does not change, but has been declining continuously towards negligible levels in the east. Therefore, in contrast to the AKP's evolution into a mass party, we can talk of a 'regionalization' of these parties. 14 The same can be said of the Turkish nationalist MHP. The party's already small presence in the central-eastern and southeastern provinces has dwindled to almost nothing, whereas its vote share in western, southern and central parts of the country increased substantially after bottoming out in 2002. One distinctive pattern regarding this party is that its 2009 vote share exceeds its 1999 level only in the coastal areas of the Aegean and the Mediterranean regions.
The challenge posed by the MHP to the left-wing parties, right in their strongholds, may explain why their platforms have become increasingly nationalistic. All of these parties exhibit a lack of enthusiasm for the reforms and compromises necessary for EU accession and the settlement of the Cyprus and Kurdish minority issues, as well as the implementation of liberal economic policies that would promote Turkey's integration with the global economy. This strategy, however, cedes the ethnic Kurdish votes to the AKP and to the Kurdish nationalist BDP, which explains the disappearance of the CHP and the DSP in the east. It also leaves the AKP as the sole party in the field to champion democracy (as witnessed in the constitutional reform debates of 2010) and globalization, and, thus, the only alternative for those who support these reforms. Viewed from this perspective, the evolution of electoral support for the AKP is also in line with the idea that growing pro-EU tendencies in Turkish society have to do with people's belief that EU membership will facilitate the consolidation of liberal democracy in Turkey (McLaren, 2000;Öniş, 2003). Büthe et al. (2009) argue that there are strong reasons to expect that EU membership will hold the same promise of democratic consolidation for Turkey as it previously did for Greece, Spain, Portugal and the ex-communist states in Eastern Europe. It seems safe to assume that such aspirations are one reason the current government continues to receive the support of a major portion of the society.
In a rapidly transforming but highly conservative society such as Turkey, opposition parties are not likely to increase their support unless they drop their advocacy of the status quo, or the AKP drops its proreform stand. Therefore, barring the emergence of such new rivals in the centre-right and in the centreleft, it seems unlikely that the AKP will lose much more ground than it already has between 2007 and 2009. 15 However, the recent change in the leadership of the CHP may pose a challenge for the AKP, especially if it leads to significant changes in the party's views on critical issues such as the democratic rights of the ethnic Kurdish population. The new leader of the party, Kemal Kılıçdaroğlu, a popular politician of eastern origins, could have a considerable effect on CHP votes in the east, where it has performed poorly during the past decade. On a related point, another threat to the AKP could come from an escalation of ethnic tensions. The rising Turkish nationalistic tendencies in the western provinces (as demonstrated here by the expansion of the cluster of provinces with a high Turkish-nationalistic tendency), coupled with the existing Kurdish nationalistic ones in the east, pose a major challenge for the party: solving issues related to the Kurdish minority in a manner satisfactory and palatable to both groups. Without concrete steps by the government toward a permanent solution, increased terrorist attacks by the separatist PKK and the resulting response by the Turkish military are bound to cost the AKP support in both the west and the east. 16 From the perspective of temporal trends in electoral support during the past decade, the results of the current study can also shed some light on the ongoing debate about whether the rising support for the AKP during the 2000s was due to the party's pro-EU, pro-democratization agenda or to rising religiosity. Previous econometric work on inter-party vote movements in Turkey during the past decade based on micro data ) and province-level data (Akarca, 2010) have already shown that the AKP's rise mainly had to do with its ability to appeal to larger segments of Turkish society by its determination and ability to carry out economic and democratic reforms. The findings of the current study corroborate this view by revealing that the party's substantial rise between 2002 and 2007 was the result of a nationwide success rather than being mainly limited to areas where conservative and nationalistic tendencies are the strongest. 17 The AKP's decline from 2007 to 2009 also signifies the importance of economic factors, because this reversal can hardly be explained as resulting from a decline in the degree of religiosity in the society. 18 In addition to facilitating the examination of the spatial aspects of political tendencies, we believe that the provincial groupings obtained in this paper may also serve as a guide for those studying regional differences, for those conducting polls to capture the mood of the electorate, and for those analysing the above-mentioned realignment process in greater detail. They may also provide a reference point against which past and future election results can be put in perspective. Indeed, the three-way partition of the provinces suggested here has already been utilized by Köksal et al. (2010), and Akarca (2010) in studying regional variations in inter-party vote movements between the 2007 and 2009 elections. Finally, the clusters presented here and the regional political tendencies they imply may provide input for academicians and experts, who may choose to interpret them in a completely different light. Özbudun (2006a, 2006b). Özbudun finds significant continuities between the worldview of the AKP and the former centre-right parties that dominated Turkish politics from the beginning of the multi-party period, and notes that the party's understanding of conservatism is open to changing the existing institutions and relationships while preserving traditional values (including religious ones). 3. In an in-depth analysis of the November 2002 elections, Tosun (2003) argues that the outcome of this election alone constitutes a major realignment in Turkish politics. Akarca and Tansel (2006) Başlevent and Akarca (2009) and Akarca and Başlevent (2009) and province-level data in Akarca (2010) and Akarca and Tansel (2009). 6. Among the studies from a recently growing literature on the examination of various social and economic patterns and outcomes pertaining to Turkey from a regional perspective are Dulupçu (2005), Özdemir andDarby (2009), Gedikli (2009) and Lagendijk et al. (2009). 7. See, for example, Putnam (1993), Inglehart (1997), Inglehart and Baker (2000). 8. In Turkey, elections for all local administrations (mayors, city councils and provincial councils) are held simultaneously and nationwide. Consequently, they are contested by all major parties, take place in a similar atmosphere to that of the parliamentary elections, and have similar participation rates. For example, the participation rates for the 1999, 2002 and 2007 parliamentary elections were 87.1, 79.1 and 84.2 percent, respectively. The corresponding figures for the 2004 and 2009 elections for provincial councils were 76.3 and 85.2 percent, respectively. With the exception of the three most populous provinces, each province constitutes a single election district in both parliamentary and provincial council elections, implying that the provinces are the ideal unit of analysis for our work. 9. Güvenç and Kirmanoğlu (2009) use cluster analysis jointly with correspondence analysis, which is an alternative data reduction technique, to examine geographical voting patterns in Turkey since the 1950 elections. As in cluster analysis, provinces with similar political preferences are classified into the same group. Allowing for the possibility of a different number of groups in each election, the technique ends up dividing the country into 7-9 non-contiguous regions depicting the strongholds of different parties. 10. Aside from pre-setting the number of clusters for all elections at three and five, we use data from the same provinces (those with licence plate numbers 1 to k) to serve as our 'initial cluster centres'. The purpose of this approach is to ensure that the ultimate cluster memberships do not depend on the initial centres, which -if chosen randomly -might lead to different outcomes in each run of the k-means algorithm. To measure the distances between the data points, we use the Euclidean distance, which is the default option in STATA. 11. The voting behaviour of Turkey's internal migrants is studied by Akarca and Başlevent (2010b), who show that migrants vote quite differently from the natives. 12. The 10 percent threshold leads to some distortion in the party choices of Turkish voters, in that it causes some supporters of the minor parties to cast their ballot for parties or independent candidates that are not their first choices. However, since the threshold applies to nationwide, rather than provincial, vote shares, its impact can be assumed to be more or less uniform across the provinces and, therefore, too small to affect the provincial make-up of the clusters obtained from observed voting patterns. As we mention here, the south-east of the country constitutes an exception in this regard, but it is almost certain that the same provinces would have formed a cluster with or without the threshold. Furthermore, the similarity of the clusters obtained from the three parliamentary and the two local election results supports this argument, since there is no threshold in local elections. 13. Only some provinces on the Black Sea and eastern Mediterranean coasts appear to have changed clusters permanently after 1999. 14. Çarkoğlu (2010) also uses the concept of 'regionalization' to describe this situation. 15. This argument is in line with the finding of Başlevent (2009) that the voter bases of the AKP and the CHP are quite distinct from each other. The study suggests that there can be little vote movement between the two parties also because the majority of the supporters of one party cite the other party as the one they would never vote for. 16. According to Çarkoğlu (2010), part of the erosion in the AKP vote between 2007 and 2009 among ethnic Kurdish voters is attributable to the military operations. Based on province-level data from the 1990s, Kıbrıs (2010) finds a significant link between the number of casualties resulting from terrorism and support for certain groups of parties. 17. Based on a 2002 survey, Çarkoğlu (2003) reports that support for EU membership is weakest in the cluster of provinces where conservative and nationalist votes are the highest. Apparently, the AKP has managed to push its pro-EU agenda without alienating its more conservative voters. 18. Akarca (2010) estimates that 4.2 of the AKP's 8.1 percentage point loss in 2009 can be attributed to the poor economic conditions prevailing in the wake of the global economic crisis of 2008-9, and another 3.0 to the 'strategic voting' typically observed in local elections.