Who do Regional Polls predict as the winner?

by Joel Sawat Selway

Image: NIDA's regional polls

Source: https://nidapoll.nida.ac.th/all-polls/

A less technical version of this article will appear on LatitudeTen tomorrow.

In the first article in this series, I introduced three simple ways of translating polling data into constituency seats. Using data from national polls, each method produced results that differed markedly from expert predictions, most notably, a People’s Party (PP) seat total well above the range suggested by expert commentary, alongside a sharp underperformance by Pheu Thai (PT) and no constituency seats for Kla Tham (KT).

A natural response is skepticism: national patterns obscure regional variation. We know, for example, that PT’s stronghold is the Northeast and has never been popular in the South; PP did well in Bangkok last time round and made in-roads in the North. In short, if we can get better data on regional patterns our models should improve in accuracy. In November-December of last year, NIDA conducted six polls in the North, Northeast, East, Central, South, and Bangkok. What does inputting this data into our model do to the results? Do seat projections get closer to the expert predictions?

The short answer is no. Using regional polling data makes PP’s seat advantage even larger, thus further away from expert predictions.

Dealing with the Undecideds

As with most polls 2-3 months before the election, there were a high number of “undecideds” in the regional polls—ranging from 24.36% in the Northeast to 40.20% in Bangkok—I incorporated them into the projections using two different assumptions: first, as in the national imputation method, I assumed the undecideds were distributed proportionally to the “decideds” (the proportional assumption). I did this by subtracting the undecided vote share between December and January (32.36%) from 100% (=67.64%) and dividing each party’s vote share by the new denominator. For PP this produced a predicted vote share of 36.28% (column 2). Second, I used information on how the undecideds in NIDA’s national poll in early December changed the overall distribution by the time of NIDA’s poll on January 30. The undecided rate between these two national polls decreased from 32.36% to just 2.92% (the retrospective assumption).

To calculate the retrospective assumption, we first ask: "how did the party shares change between the two polls?" Column 3 of Table 4 shows how December’s undecided voters showed up in the late January poll. We can see that for PP, that vote share was 33.56%. This means the proportional assumption overestimated PP’s share by 2.72 percentage points (column 4). In order to get to the 33.56% we observed in late January, 30.3% of the early December undecideds must have been PP-in-waiting voters (column 5) rather than the ~36% that the proportional assumption assumed. In contrast, the proportional assumption vastly underestimated BJT’s share of the undecideds. In early December, 9.92% of respondents said they supported BJT, but by late January it had risen to 22.76%. In order to get the ~13% bump in its vote, 43.1% of the December undecideds must have been BJT supporters-in-waiting (0.431*29.44%=12.69%).

Similarly, 17.5% of December undecideds must have been PT supporters-in-waiting, 4.8% DP, 4.2% KT, etc. UTN, TST, and PPRP must have had nobody in the December undecideds, and in fact some of the December respondents who chose those parties changed their mind by late January, hence their negative numbers in the last column.

In short, the primary benefactor of the Undecideds revealing their preferences is BJT. Alternatively, the December poll contained a whole lot of “Shy BJT voters”. Ultimately, it means that assuming the undecideds had the same distribution of party preferences as the decideds was wrong at the national level.

Party / Category	Dec 4–12 NIDA poll (Raw)	Dec 4–12 (Proportionally adjusted share)1	Jan 30 NIDA poll	Difference (Jan − Adj. Dec)	“Actual” Share of Dec undecideds (%)
People’s Party (PP)	25.28	36.28	33.56	−2.72	30.3
Bhumjaithai (BJT)	9.92	14.24	22.76	+8.52	43.1
Pheu Thai (PT)	11.04	15.84	16.92	+1.08	17.5
Democrat Party (DP)	11.80	16.94	12.76	−4.18	4.8
Economic Party (EP)	2.76	3.96	3.44	−0.52	2.3
United Thai Nation (UTN)	2.32	3.33	1.84	−1.49	−0.4**
Kla Tham (KT)	0.17*(imputed)	0.24	1.40	+1.16	4.2
Thai Sang Thai (TST)	2.00	2.87	1.08	−1.79	−3.0**
Palang Pracharath (PPRP)	1.12	1.61	0.27 (imputed)	−1.34	−2.9**
Other parties (ex-KT / ex-PPRP)	1.19	1.71	3.01	+1.30	4.4
Undecided	32.36	2.92	2.92	0.00

Table 4. What happened to Undecided voters between December and January?

¹ Computed by distributing parties proportional to their raw share (minus undecideds) amongst the 32.4% undecideds minus the 2.92% who were still undecided in late January (29.44%)

*KT’s December vote share was imputed at 1/8 the Other category, given eight parties were listed with no exact vote share.

** Negative values indicate parties that lost existing supporters over the period, offsetting any gains from undecided voters; the figures therefore technically represent net changes relative to the undecided pool, not the gross distribution of undecided voters.

Translating the regional polls into seats

With the two ways of dealing with undecideds explained, we can now apply those assumptions to the regional polling data. Table 1 shows the results from the regional imputation method. This is identical to the national imputation method (described in full in the longer version of my first article), except the model uses both regional vote shares from the 2023 election and the results from NIDA’s regional poll results. I show predictions using both the multiplier and absolute change method introduced in the first article (two assumptions x 2 methods = 4 projections)

Overall, the regional imputation models predict that PP will get more seats than it did using the national polling data, and maybe even a majority in parliament. BJT’s tally is fairly consistent at the high end, but one of the models dropped them to just 106 seats, well below the 140-seat floor of expert predictions. DP’s range is slightly higher than its range under the national imputation method, but it is PT that is the biggest loser of using regional polling data, with the lowest prediction down to just 36 seats and maxing out at just 53 seats. Its lost seats go to primarily to PP and EP.

Party	Constituency Seats (multiplier)	Constituency Seats (absolute change)	Party List Seats	Total Seats	Percent (%)
People’s Party (PP)	209-220	208-215	33	241-253	48.2% - 50.6%
Bhumjaithai (BJT)	89-122	82-106	24	106-146	21.2% - 29.2%
Democrat Party (DP)	42-55	28-43	13	41-68	8.2% - 13.6%
Economic Party (EP)	7	0	5	5-12	1.0% - 2.4%
Pheu Thai (PT)	19-23	34-36	17	36-53	7.2% - 10.6%
Thai Sang Thai (TST)	1-3	5	2	3-7	0.6% - 1.4%
United Thai Nation (UTN)	0-2	6-7	2	2-9	0.4% - 1.8%
Palang Pracharat (PPRP)	0	0-12	1	1-13	0.2% - 2.6%
Other	0	0	3	3	0.6%
Total	400	400	100	500

Table 1. Seat outcomes implied by regional polling (Undecideds-adjusted multiplier)

Where the seats changes are coming from

How do parties perform in each region? Under the regional imputation method PP no longer sweeps Bangkok. BJT now gets 4-6 seats there. PP does show the possibility of slight improvement in the Central region (up to 6 seats), and even more in the South gaining up to 12 more seats there than under the national imputation method. However, PP’s biggest gains come in the Northeast—a whopping 30-32 seat increase is predicted when using these regional data. Meanwhile, BJT is fairly consistent across all regions at the high end, but has lower predictions in one model in the Central region (-13), the South (-6), the North (-6), and the Northeast (-10), though it does gain a few in Bangkok in both models. DP losses come, as expected, in the South where it could lose as many as 10 compared to the national imputation method. PT’s seat losses are almost entirely in the Northeast where at worse it could win as few as 13 seats.

Region	PP	BJT	DP	PT	EP	TST	UTN	PPRP	Total
Bangkok	27-28	4-6	0-1	0	0	0	0	0	33
Central	44-51	11-23	8	1-3	2-3	0	0-2	0	78
East	17	4	2-3	1	2	0	0	0	26
North	38-44	13-20	3	3-4	0	0	0	0	64
Northeast	65-67	45-53	7-9	13-15	0	1-3	0	0	139
South	13-18	12-16	22-31	1	2-3	0	0	0-1	60
Total	209-220	89-122	42-55	19-23	7	1-3	0-2	0-1	400

Table 2. Constituency Seats by Region (Regional Imputation Method)

Region	PP	BJT	DP	PT	TST	Other	Total
Bangkok	33	0	0	0	0	0	33
Central	45	24	7	2	0	0	78
East	18	4	2	2	0	0	26
North	39	19	3	3	0	0	64
Northeast	35	55	6	36	4	3	139
South	6	18	32	0	0	4	60
Total	176	120	50	43	4	8	400

Table 3. Constituency Seats by Region (National Imputation Method)

Are regional polls less accurate?

Are regional polls less accurate?

That the predictions from regional polls are even further from the experts’ predictions leads us to three important lessons about polling data:

First, is question wording. The NIDA regional polls do not ask the same question as in the national polls. The national polls ask who the respondent intends to vote for on the constituency ballot. In contrast, the regional polls ask: “Which party do you support today?” This question may elicit responses such as: “I like PP the best, even though I will not vote for them in the constituency race”. In other words, it is likely to elicit responses closer to party list voting intentions.

Second, is timing. These polls were conducted primarily in November, with more than two months until the election. History has shown us that voting intentions change significantly amongst the Thai electorate in this short window. Even if all the undecideds intended to vote in a similar pattern as the rest of the region, the entire region’s vote is not fixed this far in advance.

This shows up particularly in the “Undecided” category: all the polls have undecided shares over 24%, with Bangkok’s as high as 40.2%! However, the two assumptions used to deal with undecided voters mask a lot of unknowns. Thus, the third lesson is that all predictions deal with numerous unknowns--and have to make decisions about how to infer information from other sources to the polling data. In statistical analysis, we call this making ecological inferences. In all the prediction models I have presented thus far, I have made ecological inferences: for example, inferring national or patterns are the same in constituencies across Thailand, or inferring that how undecideds resolved their party choice nationally is similar to how they do at the regional level. The key is to make those inferences transparent.

Party / Category	South (multiplier, national poll change assumption)	South (multiplier, proportional assumption)	NIDA National Poll (Jan 30)	Nakhon Si Thammarat	Songkhla
Democrat Party (DP)	29.96	40.01	12.76	51.08	44.42
People’s Party (PP)	26.42	24.90	33.56	15.18	13.96
Bhumjaithai (BJT)	23.91	16.30	22.76	16.87	10.87
United Thai Nation (UTN / RTSC)	3.79	5.46	1.84	~0.33	1.87
Pheu Thai (PT)	7.43	3.43	16.92	2.25	1.87
Prachachat Party	0	2.73	~0.273	~0.33	~0.28
Economic Party (EP)	0.13	2.31	3.44	1.41	1.22
Palang Pracharath (PPRP)	0.72	2.17	~0.273	~0.33	~0.28
Kla Tham (KT)	1.44	~0.33	1.40	1.31	~0.28
Other parties	—	~1.98	~2.80	~2.00	~1.14
Undecided	(absorbed)	(absorbed)	2.92	8.43	23.81
Constituency seats for DP, BJT, and PP in the South	22, 16, 18	31, 12, 13	32, 18, 6	39, 11, 8	39, 9, 10

Table 4. Poll results in the South

A Focus on the South

To understand why regional polling does not resolve the gap between model-based projections and expert expectations, it is useful to look more closely at the South. Table 4 (columns 1 and 2) compares the two methods just explained using regional data with the January 30^th national poll and two provincial polls. The bottom row shows how three parties—DP, BJT, and PP—perform under each model.

Table 4 illustrates how different—but all reasonable—ways of predicting voting intentions in the South produce a wide range of seats predictions, from 22-39 seats for DP, 9-16 for BJT, and 6-18 for PP.

Which one is the most accurate? The two provincial polls in the final columns, which show raw percentages unadjusted for undecideds, reinforce the idea that purely mathematical formulas to transfer polling numbers into seat predictions are fraught with all sorts of uncertainty. Both provinces show a much higher preference for the Democrat Party than the regional poll results. Could this mean the regional polls have it wrong for all Southern provinces, and thus we should apply these numbers to our predictions (which is how we get the 39 seats for DP in the bottom row)? Or should we assume that other provinces have much lower vote shares for DP to balance them out? If we apply these polls’ proportions to the entire South, we get the highest projections for DP at 39 seats. But notice that both Nakhon Si Thammarat and Songkhla still contain sizable undecided shares, even though they were conducted much closer to the election than the regional polls—the Nakhon Sri Thammarat poll was conducted on January 27^th.

Thus even these provincial polls mask underlying constituency patterns. They are not able to reveal whether DP’s gains in Nakhon Sri Thammarat are concentrated in constituencies it already holds or spread into marginal districts it previously lost. If the former, it could stay at the same six seat victories; if the latter it is possible DP could sweep all ten seats. In other words, we are still left making ecological inferences, even with granular provincial polling data. The lesson from Table 4, then, is not that one method is correct, but that plausible assumptions can all be justified and yet lead to very different seat outcomes.

In closing, transparency is good, but accuracy along with transparency is even better. Thus, the next step is to improve the accuracy of our predictions by incorporating more constituency-level information.