1# RECORD 6edd1ccf2791c69eb6ab0393c5323dfff21ae6202# TITLE: The Impact of the Youth Housing Loan Subsidy Policy on Housing Prices: Evidence from New Taipei City3# YEAR: 2026.0 VENUE: Business and Economic Research DOI: 10.5296/ber.v16i4.240244# AUTHORS: Ching-Hsing Chang; Yi Lin; Wen-Tsung Wu; Chun-Hsien Wang5# PAGES: 26 CANDIDATE TABLE PAGES (1-based): [12, 13, 15, 16, 17, 18, 19]6# ABSTRACT: In August 2023, the Taiwanese government launched the New Youth Housing Loan Subsidy Policy (hereafter the New Youth Policy), which lowers the threshold for young people to buy homes through interest subsidies, higher loan ceilings, longer maturities, and extended grace periods. However, when housing supply is difficult to increase in the short run, such a demand-side subsidy may push up housing prices through a price capitalization mechanism. Taking New Taipei City as the study area, this paper uses more than one hundred thousand transaction records from the Ministry of the Interior’s actual price registration system between January 2021 and September 2024, and applies a hedonic price model to construct a multiple regression. After controlling for building characteristics, location, and annual trends, the study examines the effect of the policy on the unit price of housing. The pooled specification shows that the unit price of housing was about 4.4% higher in the post-implementation period, a differential equivalent to roughly one-fifth of the conditional appreciation of the market over the same window. To test whether this differential can be attributed to the policy rather than to a market-wide trend, we exploit the NT$10 million loan ceiling to construct a difference-in-differences design in which small dwellings, whose transaction prices fall within the ceiling, are treated as the high-exposure group and large dwellings, whose prices lie far above it, serve as the low-exposure comparison group. Once the control variables are allowed to differ across the two segments, the difference-in-differences estimate is economically and statistically indistinguishable from zero, and this null result is stable across four alternative size thresholds. An event-study specification reveals a significant pre-policy divergence between the two groups, and a placebo cut-off placed in the first half of 2023 yields a significant coefficient, so the parallel-trends assumption is not supported. A split-sample estimation by unit size likewise finds no significant difference in the policy coefficient between small and large dwellings. Building age and floor area have negative effects while floor level has a positive effect, reflecting the phenomenon of higher unit prices for smaller units. We therefore conclude that housing prices in New Taipei City rose markedly over the observation window, but that the data do not provide robust evidence that this increase is attributable to789===== PAGE 1 (score 3, numeric-paren 0, table-words 0, var-words 3) =====1011Business and Economic Research 12ISSN 2162-4860 132026, Vol. 16, No. 4 14http://ber.macrothink.org 157416The Impact of the Youth Housing Loan Subsidy Policy 17on Housing Prices: Evidence from New Taipei City 18Ching-Hsing Chang 19Department of Technology Management, National Chiayi University, Chiayi City, Taiwan 20E-mail: chc@mail.ncyu.edu.tw 21 22Yi Lin 23EMBA Program, College of Management, National Chiayi University, Chiayi City, Taiwan 24E-mail: irene173956@gmail.com 25 26Wen-Tsung Wu (Corresponding author) 27Chinese Sustainable Innovation Development Association (CSIDA), Taipei, Taiwan 28 E-mail: roywu070679@gmail.com 29 30Chun-Hsien Wang 31Department of Business Administration, National Chiayi University, Chiayi City, Taiwan 32E-mail: chwang@mail.ncyu.edu.tw 33 34Received: July 26, 2026 Accepted: September 9, 2026 Published: September 10, 2026 35doi:10.5296/ber.v16i4.24024 URL: https://doi.org/10.5296/ber.v16i4.24024 36 37Abstract 38In August 2023, the Taiwanese government launched the New Youth Housing Loan Subsidy 39Policy (hereafter the New Youth Policy), which lowers the threshold for young people to buy 40homes through interest subsidies, higher loan ceilings, longer maturities, and extended grace 41periods. However, when housing supply is difficult to increase in the short run, such a 42demand-side subsidy may push up housing prices through a price capitalization mechanism. 43Taking New Taipei City as the study area, this paper uses more than one hundred thousand 44transaction records from the Ministry of the Interior’s actual price registration system 454647===== PAGE 12 (score 15, numeric-paren 0, table-words 2, var-words 8) =====4849Business and Economic Research 50ISSN 2162-4860 512026, Vol. 16, No. 4 52http://ber.macrothink.org 5385544. Empirical Results and Analysis 55This chapter analyzes the impact of the New Youth Policy on housing prices in New Taipei 56City. It first uses descriptive statistics to understand the basic characteristics of the sample; 57then conducts correlation analysis and a multicollinearity test to ensure the reliability of the 58model estimates; next applies OLS for regression analysis; then verifies the stability of the 59results through a robustness check; and finally examines sub-market heterogeneity by 60housing unit size. The chapter comprises five parts: descriptive statistical analysis; correlation 61and multicollinearity tests; regression results; the robustness check; and the sub-market 62heterogeneity analysis. 634.1 Descriptive Statistical Analysis 64This study collected 108,671 housing transaction samples in New Taipei City. The dependent 65variable, the log of the housing unit price, has a mean of 3.61 and a median of 3.65; the price 66distribution is slightly skewed, and log transformation helps reduce skewness and improve 67model stability. The policy dummy has a mean of 0.31, indicating that about 31% of the 68samples fall in the post-implementation period, covering transactions before and after the 69policy. For the control variables, building age averages 22.59 years with a median of 24, 70indicating that the sample is dominated by second-hand housing; floor area averages 37.28 71ping with a median of 33.09, showing a right-skewed distribution; floor level averages 7.26, 72reflecting a sample that includes both low-rise apartments and high-rise towers; about 45% of 73the dwellings have parking spaces, 72% have management organizations, and 72% are 74elevator buildings; in terms of regional distribution, high-price areas account for 58% and 75mid-price areas for 37%, consistent with the concentration of New Taipei City housing 76transactions in densely populated and conveniently connected areas; and the annual sample 77proportions are evenly distributed, effectively reflecting market changes across years, as 78shown in Table 4. 79Table 4. Descriptive statistics 80Variable 81Mean 82SD 83Min 84Median 85Max 86Log of housing unit price 873.61 880.41 890.39 903.65 915.49 92New Youth Policy 930.31 940.464 950 960 971 98Building age (years) 9922.59 10014.64 1011 10224 10373 104Floor area (ping) 10537.28 10620.13 1070.33 10833.09 1091172.84 110Floor level 1117.26 1125.67 1131 1145 11543 116Has parking 1170.45 1180.50 1190 1200 1211 122Has management 1230.72 1240.45 1250 1261 1271 128Elevator building 1290.72 1300.45 1310 1321 1331 134High-price area 1350.58 1360.49 1370 1381 1391 140Mid-price area 1410.37 1420.48 1430 1440 1451 1462022 1470.24 1480.43 1490 1500 1511 1522023 1530.26 1540.44 1550 1560 1571 1582024 1590.20 1600.40 1610 1620 1631 164Note: N = 108,671; the data period is January 2021 to September 2024. 165 166167168===== PAGE 13 (score 25, numeric-paren 0, table-words 6, var-words 10) =====169170Business and Economic Research 171ISSN 2162-4860 1722026, Vol. 16, No. 4 173http://ber.macrothink.org 174861754.2 Correlation and Multicollinearity Tests 176This study first uses correlation analysis to examine the degree of association among the 177variables; the results show that the direction of correlation between each explanatory variable 178and the log of the housing unit price is broadly consistent with theoretical expectations, and 179no abnormally high correlation appears among the explanatory variables. A further 180multicollinearity test using the VIF, shown in Table 5, indicates that the VIF values of the 181variables range from 1.317 to 5.601, all below the threshold of 10, showing no high 182collinearity among the explanatory variables and confirming the reliability of the regression 183estimates. The slightly higher VIF values for elevator buildings, high-price areas, and 184management organizations reflect a certain degree of association among building type, 185management mechanism, and location, but remain within an acceptable range. 186Table 5. Multicollinearity test 187Variable 188VIF 189Tolerance 190New Youth Policy 1913.404 1920.294 193Building age 1943.112 1950.321 196Floor area 1971.562 1980.640 199Floor level 2001.317 2010.759 202Has parking 2032.222 2040.450 205Has management 2064.750 2070.211 208Elevator building 2095.601 2100.179 211High-price area 2125.156 2130.194 214Mid-price area 2155.138 2160.195 2172022 2181.379 2190.725 2202023 2211.979 2220.505 2232024 2243.894 2250.257 226 2274.3 Regression Results 228The regression results are shown in Table 6. The model’s R² and adjusted R² are both 0.533, 229indicating that the policy, building characteristic, regional, and annual variables included in 230this study explain about 53.3% of the variation in the log of the housing unit price; the F 231value is 10346.121, significant at the 1% level, indicating that the overall model is 232statistically significant. 233For the main explanatory variable, the coefficient of the New Youth Policy is 0.044 with a t 234value of 12.829, significant at the 1% level, indicating that after controlling for building age, 235floor area, floor level, parking, management organization, building type, region, and year, the 236housing unit price in the post-implementation period is still significantly higher than in the 237pre-implementation period; since the dependent variable is the log of the unit price, the policy 238coefficient can be interpreted as an increase of about 4.4% in the housing unit price after 239implementation. This result suggests that the New Youth Policy may, by lowering the burden 240of home-purchase loans, raising loan ceilings, and improving financing conditions, bring part 241of the purchase demand into the market earlier, thereby exerting upward pressure on housing 242prices in New Taipei City—consistent with the expectation of this study and with the 243inference of price capitalization theory. 244245246===== PAGE 14 (score 24, numeric-paren 0, table-words 6, var-words 12) =====247248Business and Economic Research 249ISSN 2162-4860 2502026, Vol. 16, No. 4 251http://ber.macrothink.org 25287253The economic magnitude of this coefficient should be assessed against the contemporaneous 254appreciation of the market rather than in isolation. The year dummies in Table 6 indicate that, 255relative to 2021, the conditional unit price of housing in New Taipei City was 9.5% higher in 2562022, 14.4% higher in 2023, and 21.3% higher in 2024, implying an average conditional 257appreciation of roughly 7.1% per year over the observation window. Measured against this 258benchmark, the 4.4% policy differential is equivalent to approximately one-fifth of the 259cumulative conditional appreciation recorded between 2021 and 2024, or about seven to eight 260months of typical market appreciation. Measured against the raw index, the Ministry of the 261Interior housing price index for New Taipei City rose by about 31% between the first quarter 262of 2021 and the third quarter of 2024, so the policy differential corresponds to roughly 263one-seventh of the unconditional price growth of the same period. The interpretation is 264therefore two-sided. The differential is not trivial: for a dwelling transacting at the sample 265median it is comparable in order of magnitude to several years of the interest saving that the 266subsidy itself confers, which is the pattern that price capitalization theory predicts when 267short-run supply is inelastic. At the same time, the differential is clearly a minority share of 268the observed price increase, so the New Youth Policy cannot be regarded as the dominant 269driver of housing price growth in New Taipei City over this period; the great majority of the 270increase is attributable to the broader market trend captured by the year effects and to 271locational and structural attributes. Readers should also bear in mind that, for the 272identification reasons set out in Section 3.4, this 4.4% figure represents an upper bound on 273the policy contribution to the extent that any unobserved time-varying demand shock 274coincides with the August 2023 cut-off. The interpretation is therefore two-sided. The 275differential is not trivial: for a dwelling transacting at the sample median it is comparable in 276order of magnitude to several years of the interest saving that the subsidy itself confers, 277which is the pattern that price capitalization theory predicts when short-run supply is inelastic. 278At the same time, it is clearly a minority share of the observed price increase, so the New 279Youth Policy cannot be regarded as the dominant driver of housing price growth in New 280Taipei City over this period. Whether even this minority share can be attributed to the policy 281is a separate question, and it is the one addressed in Section 4.5, where a 282difference-in-differences design based on differential exposure to the loan ceiling is used to 283test whether the differential is concentrated in the segment the policy actually reaches. 284For the control variables: the coefficient of building age is −0.009 and significant, indicating 285that each additional year of age reduces the unit price by about 0.9%, reflecting depreciation; 286the coefficient of floor area is −0.001 and significant, confirming the “small unit, high unit 287price” phenomenon in the New Taipei City market; the coefficient of floor level is 0.006 and 288significant, indicating that each additional floor raises the unit price by about 0.6%. The 289coefficients of parking (−0.080) and management organization (−0.116) differ from the 290expected direction: the parking variable may be diluted because parking area or ancillary 291space is included in the per-ping unit price calculation, while management organization is 292highly correlated with elevator buildings, so its positive effect may have been absorbed by 293the building type variable; these two results should not be read directly as parking or 294management reducing housing value, but rather as overlapping explanatory effects and 295product-structure differences under the per-ping unit price model. The coefficient of elevator 296297298===== PAGE 15 (score 44, numeric-paren 16, table-words 8, var-words 7) =====299300Business and Economic Research 301ISSN 2162-4860 3022026, Vol. 16, No. 4 303http://ber.macrothink.org 30488305buildings is 0.150 and significant, indicating that the unit price of elevator buildings is about 30615% to 16% higher than that of apartments. For the regional variables, the coefficients of 307high-price areas (0.964) and mid-price areas (0.511) are both significant at the 1% level, 308indicating that location remains one of the most important factors affecting housing prices. 309The coefficient of the year variable rises from 0.095 in 2022 to 0.213 in 2024, reflecting the 310overall continued rise of the New Taipei City housing market during the study period. 311Table 6. Regression analysis 312Variable 313Coefficient 314t value 315p value 316New Youth Policy 3170.044 31812.829 319<0.001*** 320Building age 321−0.009 322−87.388 323<0.001*** 324Floor area 325−0.001 326−19.028 327<0.001*** 328Floor level 3290.006 33032.799 331<0.001*** 332Has parking 333−0.080 334−30.929 335<0.001*** 336Has management 337−0.116 338−27.854 339<0.001*** 340Elevator building 3410.150 34233.041 343<0.001*** 344High-price area 3450.964 346243.965 347<0.001*** 348Mid-price area 3490.511 350126.682 351<0.001*** 3522022 3530.095 35440.181 355<0.001*** 3562023 3570.144 35852.217 359<0.001*** 3602024 3610.213 36250.526 363<0.001*** 364Intercept 3652.959 366477.348 367<0.001*** 368F value 36910346.121 370 371<0.001*** 372R² 3730.533 374 375 376Adjusted R² 3770.533 378 379 380Note 1: N = 108,671; the dependent variable is the log of the housing unit price. Note 2: *** significant at 1%, 381** at 5%, * at 10%. 382 3834.4 Robustness Check 384To confirm the reliability of the baseline estimates, this study, based on the standardized 385residuals of the baseline model, removes observations whose absolute standardized residual 386exceeds 3, so as to reduce the interference of abnormal transaction prices; after this treatment, 387the valid sample decreases from 108,671 to 107,518, removing 1,153 extreme observations, 388about 1.06% of the total. As shown in Table 7, the model after removing extreme values 389(Model 2) is highly consistent with the baseline model (Model 1) in the direction and 390statistical significance of the coefficients: the coefficient of the core explanatory variable, the 391New Youth Policy, is slightly adjusted from 0.044 to 0.043, remaining positive and significant 392at the 1% level; the coefficients and significance of the building characteristic variables are 393almost identical; and although the coefficients of high-price and mid-price areas decline 394slightly, their direction and significance are unchanged. The R² of Model 2 rises from 0.533 395to 0.557 and its F value to 11273.002, indicating an improvement in the overall explanatory 396power after removing extreme values. In sum, the significant positive impact of the New 397Youth Policy on housing prices in New Taipei City does not change with the minor 398adjustment of the sample composition, and the empirical findings of this study exhibit good 399robustness. 400 401402403===== PAGE 16 (score 53, numeric-paren 26, table-words 7, var-words 9) =====404405Business and Economic Research 406ISSN 2162-4860 4072026, Vol. 16, No. 4 408http://ber.macrothink.org 40989410Table 7. Estimation results of the baseline model and the robustness check model 411Variable 412Baseline (Model 1) 413After removing outliers (Model 2) 414New Youth Policy 4150.044*** 4160.043*** 417Building age 418−0.009*** 419−0.009*** 420Floor area 421−0.001*** 422−0.001*** 423Floor level 4240.006*** 4250.006*** 426Has parking 427−0.080*** 428−0.078*** 429Has management 430−0.116*** 431−0.115*** 432Elevator building 4330.150*** 4340.149*** 435High-price area 4360.964*** 4370.897*** 438Mid-price area 4390.511*** 4400.451*** 4412022 4420.095*** 4430.097*** 4442023 4450.144*** 4460.149*** 4472024 4480.213*** 4490.217*** 450Intercept 4512.959*** 4523.018*** 453F value 45410346.121 45511273.002 456R² 4570.533 4580.557 459Adjusted R² 4600.533 4610.557 462N 463108,671 464107,518 465Note 1: The dependent variable is the log of the housing unit price. Note 2: *** significant at 1%, ** at 5%, * at 46610%. 467 4684.5 Sub-market Heterogeneity and a Difference-in-Differences Assessment 469The New Youth Policy is not directed at the housing market as a whole. Eligibility is 470restricted to first-time owner-occupier buyers with no other self-use residence, and the loan is 471capped at NT$10 million, so the scheme reaches dwellings with a low total-price threshold 472far more than it reaches large residences. If the 4.4% differential estimated in Section 4.3 473reflects capitalization of the subsidy rather than a market-wide trend, it should be 474concentrated in the segment the policy actually touches. This section tests that proposition in 475two ways: first by re-estimating the baseline model on size-split sub-samples, and then by 476using the exposure contrast described in Section 3.4 to construct a difference-in-differences 477estimator with an event study and placebo tests. 4784.5.1 Split-sample Estimation by Housing Unit Size 479Table 8 reports the baseline specification estimated separately for dwellings below and at or 480above 25 ping, with heteroskedasticity-robust standard errors. The policy coefficient is 0.040 481in the small-unit sub-sample and 0.045 in the large-unit sub-sample. Both are significant at 482the 1% level, but they are almost identical in magnitude, and the difference between them is 483−0.005 with a standard error of 0.008 (z = −0.67, p = 0.503). The difference is therefore 484statistically indistinguishable from zero, and its sign runs opposite to the direction that price 485capitalization would predict. The remaining coefficients differ substantially across the two 486sub-samples — the high-price-area coefficient is 1.241 among small units against 0.845 487among large units, and the floor-area coefficient is an order of magnitude larger in absolute 488value among small units — which indicates that the two segments are governed by materially 489different price structures. This observation matters for the specification choices reported 490below. 491492493===== PAGE 17 (score 82, numeric-paren 53, table-words 7, var-words 9) =====494495Business and Economic Research 496ISSN 2162-4860 4972026, Vol. 16, No. 4 498http://ber.macrothink.org 49990500Table 8. Split-sample estimation by housing unit size 501Variable 502Small units (< 25 ping) 503Large units (≥ 25 ping) 504New Youth Policy 5050.040*** (0.007) 5060.045*** (0.004) 507Building age 508−0.008*** (0.000) 509−0.010*** (0.000) 510Floor area 511−0.003*** (0.000) 512−0.000*** (0.000) 513Floor level 5140.004*** (0.000) 5150.005*** (0.000) 516Has parking 517−0.015* (0.009) 518−0.076*** (0.003) 519Has management 520−0.094*** (0.011) 521−0.114*** (0.005) 522Elevator building 5230.130*** (0.013) 5240.149*** (0.006) 525High-price area 5261.241*** (0.013) 5270.845*** (0.007) 528Mid-price area 5290.788*** (0.013) 5300.397*** (0.007) 5312022 5320.092*** (0.005) 5330.095*** (0.003) 5342023 5350.145*** (0.006) 5360.144*** (0.003) 5372024 5380.227*** (0.009) 5390.209*** (0.005) 540Intercept 5412.743*** (0.019) 5423.033*** (0.009) 543R² 5440.591 5450.526 546Adjusted R² 5470.590 5480.526 549N 55027,708 55180,963 552Difference in policy coefficient 553−0.005 (0.008), z = −0.67, p = 0.503 554 555Note 1: The dependent variable is the log of the housing unit price. Note 2: Heteroskedasticity-robust (HC1) 556standard errors in parentheses. Note 3: *** significant at 1%, ** at 5%, * at 10%. Note 4: The difference in the 557policy coefficient is tested as an unpaired comparison of the two sub-sample estimates. 558 5594.5.2 Difference-in-differences Based on Differential Exposure 560Following the design set out in Section 3.4, dwellings of 35 ping or less are treated as the 561high-exposure group (N = 59,428) and dwellings larger than 50 ping as the low-exposure 562comparison group (N = 20,341), with the intermediate band excluded. The estimating 563equation adds a treatment indicator, period fixed effects for the five sub-periods, and the 564interaction of the treatment indicator with the post-policy indicator, which is the coefficient of 565interest. Table 9 reports the results. 566Specification (1) constrains the coefficients on the control variables to be equal across the 567two exposure groups and yields an estimate of 0.015, significant at the 1% level. 568Specification (2) relaxes that constraint by interacting the treatment indicator with the full set 569of controls. The estimate falls to −0.001 with a p value of 0.825. Given the evidence in Table 5708 that the price structures of the two segments differ sharply, specification (2) is the 571appropriate one, and specification (1) is misspecified: it imposes a common linear floor-area 572effect across a range running from below 35 ping to above 50 ping, and a common locational 573premium, neither of which the data support. The null result in specification (2) is not an 574artefact of the chosen threshold. Rows (5) to (7) re-estimate the same specification with three 575alternative definitions of the exposure groups, and the estimate remains between −0.001 and 5760.000 with p values above 0.83 throughout. Rows (3) and (4) show, in addition, that even the 577restricted estimate of 0.015 ceases to be statistically significant once standard errors are 578clustered to allow for serial correlation within group-period cells, which is the concern raised 579by Bertrand et al. (2004). We note that our setting is the canonical two-group, multiple-period 580case rather than a staggered-adoption design, so the aggregation problems documented for 581two-way fixed-effects estimators under heterogeneous timing do not arise here; the binding 582583584===== PAGE 18 (score 29, numeric-paren 8, table-words 9, var-words 1) =====585586Business and Economic Research 587ISSN 2162-4860 5882026, Vol. 16, No. 4 589http://ber.macrothink.org 59091591constraint is instead the credibility of the parallel-trends assumption itself (Baker et al., 5922026). 593Table 9. Difference-in-differences estimates of the policy effect 594Specification 595DID estimate 596p value 597(1) Control coefficients restricted to be equal across groups 5980.015*** (0.005) 5990.002 600(2) Control coefficients allowed to differ across groups 601−0.001 (0.005) 6020.825 603(3) Specification (1), SE clustered by group × period (10) 6040.015 (0.015) 6050.296 606(4) Specification (1), SE clustered by group × period × tier (30) 6070.015 (0.017) 6080.359 609(5) Threshold ≤ 30 / > 55 ping, specification (2) 610−0.001 (0.006) 6110.889 612(6) Threshold ≤ 40 / > 50 ping, specification (2) 613−0.001 (0.005) 6140.832 615(7) Threshold ≤ 35 / > 45 ping, specification (2) 616−0.000 (0.005) 6170.991 618Note 1: The dependent variable is the log of the housing unit price. The reported coefficient is the interaction of 619the high-exposure indicator with the post-policy indicator. Note 2: All specifications include period fixed effects 620for the five sub-periods and the full set of control variables. Note 3: HC1 robust standard errors in parentheses 621unless clustering is stated. Note 4: *** significant at 1%, ** at 5%, * at 10%. 622 6234.5.3 Event Study and Placebo Tests 624Table 10 reports an event-study specification in which the treatment indicator is interacted 625with each period, taking 2021 as the base, together with two placebo tests estimated on 626pre-policy data alone. Under the unrestricted specification the interaction for the first half of 6272023 — a period that precedes the policy — is −0.029 and significant at the 1% level, which 628is a direct violation of the parallel-trends assumption: the two exposure groups were already 629diverging before the policy took effect. The interaction for the first post-policy period, the last 630five months of 2023, is −0.038 and also significant, meaning that the relative price of the 631high-exposure segment fell rather than rose in the months immediately following 632implementation. The coefficient for 2024 is 0.007 and not significant. Under the restricted 633specification the pattern is similarly non-monotonic, with a significantly negative coefficient 634in the first post-policy period followed by a positive coefficient a full year later. Neither 635timing pattern is consistent with the immediate capitalization of a subsidy into the prices of 636the dwellings that the subsidy finances. 637The placebo tests point in the same direction. Assigning a false cut-off at the start of 2022 638produces a coefficient of 0.002 that is not significant, which is reassuring. Assigning a false 639cut-off at the start of 2023, however, produces a coefficient of −0.016 that is significant at the 6405% level, even though no policy change occurred at that date. Taken with the event-study 641evidence, this indicates that differential movements between the two exposure groups arise in 642the pre-policy period as well, and that the identifying assumption required for a causal 643reading of the difference-in-differences estimate does not hold in these data. 644 645 646 647 648649650===== PAGE 19 (score 31, numeric-paren 15, table-words 3, var-words 8) =====651652Business and Economic Research 653ISSN 2162-4860 6542026, Vol. 16, No. 4 655http://ber.macrothink.org 65692657Table 10. Event study and placebo tests 658Period (base = 2021) 659Restricted specification 660Unrestricted specification 6612022 (pre-policy) 6620.006 (0.006) 663−0.009 (0.006) 6642023 Jan-Jul (pre-policy) 665−0.006 (0.007) 666−0.029*** (0.007) 6672023 Aug-Dec (post-policy) 668−0.017** (0.008) 669−0.038*** (0.008) 6702024 (post-policy) 6710.035*** (0.006) 6720.007 (0.006) 673Placebo: cut-off set at 2022 6740.002 (0.005) 675— 676Placebo: cut-off set at 2023 Jan 677−0.016** (0.007) 678— 679Note 1: The dependent variable is the log of the housing unit price. Entries are the interaction of the high-exposure 680indicator with the stated period. Note 2: The restricted specification constrains control coefficients to be equal 681across groups; the unrestricted specification allows them to differ. Note 3: Placebo tests are estimated on 682pre-policy data only (2021 to July 2023). Note 4: HC1 robust standard errors in parentheses. *** significant at 1%, 683** at 5%, * at 10%. 684 6854.5.4 Interpretation 686Three findings emerge, and they point the same way. The split-sample estimation finds no 687significant difference in the policy coefficient between the segment the policy targets and the 688segment it does not. The difference-in-differences estimate, once the specification is allowed 689to accommodate the different price structures of the two segments, is indistinguishable from 690zero across every threshold we examined. And the identifying assumption that would license 691a causal reading of that estimate is rejected by both the event study and one of the two 692placebo tests. We therefore do not find robust evidence that the 4.4% differential reported in 693Section 4.3 is attributable to price capitalization of the New Youth Policy. 694This conclusion requires one important qualification, which follows from the third 695assumption discussed in Section 3.4. A difference-in-differences estimator recovers a 696differential effect, not a level effect. If the chain move-up mechanism documented by Acolin 697et al. (2016) operates in the New Taipei City market, price pressure originating among 698first-time buyers of small dwellings would be transmitted to the larger dwellings that 699constitute our comparison group, and a uniform market-wide policy effect would be 700differenced away. Our results therefore establish that the policy did not raise prices in its 701target segment relative to the non-target segment; they do not establish that the policy had no 702effect on the market as a whole. What they do remove is the principal empirical support for 703the capitalization interpretation as it is usually stated, namely that the subsidy is bid into the 704prices of exactly those dwellings that eligible buyers compete for. The more parsimonious 705reading of the evidence is that the post-August-2023 price differential mainly reflects the 706broader upswing of the New Taipei City market during the observation window, which the 707year dummies capture only imperfectly. 7085. Conclusions and Recommendations 7095.1 Conclusions 710This study investigates the impact of the New Youth Policy on housing prices in New Taipei 711City, using transaction-level actual price registration data as the sample, building a multiple 712regression model through OLS with the log of the housing unit price as the dependent 713variable, and incorporating building age, floor area, floor level, building type, parking, 714715716===== PAGE 20 (score 12, numeric-paren 0, table-words 0, var-words 12) =====717718Business and Economic Research 719ISSN 2162-4860 7202026, Vol. 16, No. 4 721http://ber.macrothink.org 72293723management organization, district, and year as control variables to reduce the interference of 724housing heterogeneity. 725According to the pooled empirical results, the New Youth Policy variable is significantly and 726positively associated with housing prices, indicating that after controlling for building 727characteristics, locational conditions, and year effects, housing prices after implementation 728remain significantly higher than before. This suggests that the policy may stimulate housing 729demand reflected in rising prices by lowering the purchase threshold, raising loan ceilings, 730and easing the financial pressure of home purchase. The further tests reported in Section 4.5, 731however, do not sustain that reading. Splitting the sample by housing unit size yields policy 732coefficients of 0.040 and 0.045 for small and large dwellings respectively, a difference that is 733statistically indistinguishable from zero. A difference-in-differences estimator that exploits 734the NT$10 million loan ceiling to separate high-exposure from low-exposure dwellings 735returns an estimate of −0.001 once the control coefficients are allowed to differ between the 736two segments, and this null is stable across four alternative thresholds. An event study shows 737a significant divergence between the two groups in the first half of 2023, before the policy 738took effect, and a placebo cut-off placed at the start of 2023 is significant at the 5% level, so 739the parallel-trends assumption underpinning the design is not supported. Taken together, these 740results indicate that housing prices in New Taipei City rose markedly over the observation 741window, but that the data do not provide robust evidence that the increase is attributable to 742price capitalization of the New Youth Policy in the segment the policy targets. The 4.4% 743pooled differential should therefore be read as a conditional association that most probably 744overstates the policy contribution, and the more parsimonious explanation is that it largely 745reflects the broader market upswing that the year dummies capture only imperfectly. We note 746one qualification: because a difference-in-differences estimator recovers a differential rather 747than a level effect, and because the chain move-up mechanism may transmit price pressure 748from the target segment to the comparison segment, our results do not rule out a uniform 749market-wide policy effect. What they remove is the principal empirical support for the 750capitalization interpretation as it is conventionally stated. This finding is consistent with the 751view in the literature that the price response to a demand-side subsidy is conditional on local 752supply elasticity and on the design of the scheme rather than automatic, and it shows that 753when housing supply is difficult to increase quickly in the short run, a purchase subsidy 754policy may be partly converted into a housing price increase. 755For building characteristics, building age has a significant negative effect on housing prices, 756reflecting depreciation and market preference; floor level has a significant positive effect, 757indicating that higher floors are more favored owing to better lighting, ventilation, and views; 758and floor area has a significant negative relationship, confirming the “small unit, high unit 759price” phenomenon, as smaller products with lower total-price thresholds are favored by 760first-time buyers and investment demand. In terms of product characteristics, the prices of 761elevator buildings are significantly higher than those of apartments, indicating that product 762type has a clear influence on price; although the parking and management variables are 763significant, their direction differs from expectation, reflecting the effects of product structure, 764pricing method, and overlapping explanatory power under the per-ping unit price model, and 765