The final section in this chapter summarizes the main steps and findings of the study and strives to derive meaningful implications of the multilevel analysis for real estate practice. Moreover, the drawbacks and limitations of the study and its inferences are critically dis- cussed.
Chapter Fehler! Verweisquelle konnte nicht gefunden werden. focused on the problem set and the three main objectives of the study: (1) to investigate the relationship between properties’ brand status and their economic performance in terms of their market value while accounting for relevant factors such as the year of observation, rent, building age, usable area, and the size of the respective city that are commonly considered in valuation practice; (2) to examine potential variances in this relationship between cities; and (3) to identify possible interaction effects between properties’ brand status and its covariates. In Chapter Fehler! Verweisquelle konnte nicht gefunden werden., the IPD Investment Property Databank was introduced as the data source for this study while the definition and scaling of the variables were presented. Afterwards, the hierarchical structure of the dataset was emphasized and the preparatory steps towards the final dataset were also described.
The next chapter introduced multilevel modeling as an appropriate methodology in ac- counting for the nested data structure and discussed model assumptions as well as the sample requirements, model estimation methods, and as relevant approaches for signifi- cance testing and variance explanation that were applied in the analysis. Finally, a step- wise analysis strategy was outlined for exploratory model development
The results of an initial independent group t-test between branded and non-branded ob- servations and a multiple linear regression were briefly summarized in Chapter Fehler!
Verweisquelle konnte nicht gefunden werden. before the successive development of
the final multilevel model was described. On the basis of the pooled dataset, the prelimi- nary t-test indicated a significantly (p < 0.001) larger geometric mean property value for the group of branded observations (3,115.05 EUR/m²) than for non-branded observations (2,230.54 EUR/m²). The regression results supported this finding, suggesting significant relationships for all independent variables, and the interaction effects between ‘Brand Sta- tus’ and ‘Rent’ as well as between ‘Usable Area’ and ‘City Size’. The conditional geometric mean of the untransformed independent variable ‘Value’ was expected to be app. 14.3% (p < 0.001) higher for observations from branded properties than for observations from non-branded properties when all covariates are fixed and interacting variables were kept at zero.
Finally, in the last step of the multilevel analysis, a four-level random intercept random co- efficient model with interaction effects was suggested. In this model, all independent vari- ables showed significant relationships with the outcome variable. Across cities, negative relationships were detected for ‘Year’ and ‘Building Age’, while ‘Rent’, ‘Usable Area’ and ‘City Size’ showed positive regression coefficients. However, significant between-city vari- ances in the regression slopes of ‘Year’, ‘Rent’, and ‘Usable Area’ were found, implying
that the relationship between these predictors and the outcome variable is not stable across the different macro locations.
Regarding the first objective of the study, a significant (p < 0.001) regression coefficient of 0.1659 (95% conf. int.: 0.1125 to 0.2192) was estimated for the explanatory variable ‘Brand Status’, corresponding to an expected 18.1% change (95% conf. int.: 11.9% to 24.5%) in the conditional geometric mean of ‘Value’. With respect to the second study ob- jective, no significant city-level slope variance was detected for the variable, suggesting a stable relationship with the outcome variable across cities. Looking at the third objective, a significant (p < 0.01) interaction effect (-0.0003, 95% conf. int.: -0.0006 to -0.0001) was identified between ‘Brand Status’ and ‘Rent’, implying that, for branded properties, the positive relationship between ‘Rent’ and the dependent variable is weakened in compari- son to non-branded properties. Accounting for all fixed and random effects of the model, a comparison of the fitted values for branded and non-branded observations on city level in- dicated that a property branding is associated with a higher predicted property value in all cities of the sample apart from Leipzig.
The model fit was assessed in comparison to a simple model from the first modeling step that only contained the dependent variable. The total unexplained variance of ‘Logvalue’ was 0.1269 (Model 1: 0.2104), and Maddala ML-R² suggested an improvement of the var- iance explanation by 27.0%. The Bryk/Raudenbush-R² indicated an explained variance of app. 64.1% on city level, 8.2% on postcode area level, 41.8% on property level, and 24.7% on measurement occasion level.
Finally, the first-level residuals and higher level random effects were inspected on the ba- sis of an examination of quantile-quantile plots, histograms, and the corresponding values for skewness and kurtosis. Additionally, a scatter plot of first-level residuals against the fit- ted values was considered. Altogether, no substantial violations of common boundaries were detected, leading to the conclusion that assumptions of normality and homoscedas- ticity were sufficiently fulfilled.
From a methodological perspective, the multilevel analysis proved appropriate in investi- gating the relationship between the brand status of a property and its economic perfor- mance in terms of its market value, while accounting for relevant covariates and the hier- archical structure of the dataset. Considering the obviously nested structure of real estate market data, it seems surprising that this method is not more commonly applied in real es- tate research. In fact, some studies relying on pooled datasets might suffer from violations of the independency assumption, leading to biases in the estimation results and to incor- rect calculations of confidence intervals and significance levels. A comparison of the re- sults from the single-level multiple regression and the multilevel analysis in this study em- phasizes the potential fallacies (see Table ). While the overall direction of the relationship between the explanatory variables and the outcome is identical, there are substantial dif- ferences in the size of the regression coefficients. Ignoring the hierarchical data structure, the single-level regression estimated a higher intercept and expected weaker relationships for ‘Year’, ‘Building Age’, ‘Usable Area’, and ‘City Size’ and a stronger relationship for ‘Rent’. Furthermore, the significance level of ‘Usable Area’ and ‘Rent’ was lower according to the multilevel approach. For ‘Brand Status’, the multilevel model also estimated a
stronger relationship with ‘Logvalue’, expecting an increase of 18.0% in the conditional geometric mean of ‘Value’ instead of only 14.3%. The single-level regression model indi- cated significant interaction effects of ‘Brand Status’ with ‘Rent’, ‘Usable Area’, and ‘City Size’. Only the interaction with ‘Rent’ proved to be significant in the hierarchical linear model. Looking at the apparent differences in the regression estimates, a more frequent consideration of data hierarchies seems useful for real estate studies dealing with spatial or longitudinal information.
Table 18: Comparison of Single-Level and Multilevel Model Estimates
Source: Own illustration.
From a real estate perspective, the multilevel analysis findings have three main implica- tions with respect to the objectives of the study. For one thing, real estate practitioners’ in- tuitive perceptions and assumptions of differences in the value of branded and non- branded office properties seem to be justified. The significant positive relationship be- tween ‘Brand Status’ and ‘Value’ indicates that a property brand is associated with higher property values in comparison to a non-branded property that is comparable with respect to the year of observation, its age, contract rent, usable area, and the size of its respective macro location. Secondly, this positive relationship seems to be stable across cities, since no significant between-city variance was detected. And finally, the negative significant in- teraction effect between ‘Brand Status’ and ‘Rent’ suggests that, for branded buildings, the relationship between their contract rent and their overall value is reduced in compari- son to non-branded properties.
Beyond the original objectives of this study, the multilevel analysis highlighted that there are significant between-city variances in the relationship between properties’ value and their contract rent, usable area, and the year of observation. Across the city sample, a
Si ngl e-l evel mul tipl e l i nea r regres s i on model
Hi era rchi ca l l i nea r model (model 19) Year -0.0079*** -.0103*** Rent 0.0049*** .0013*** Building Age -0.0020*** -.0045*** Brand Status 0.1339*** .1659*** Usable Area 0.0204*** .0515** City Size 0.0989*** .1128** Brand Status*Year 0.0071 e Brand Status*Rent -0.0010*** -.0003**
Brand Status*Building age -0.0018 e
Brand Status*Usable Area -0.0391*** e
Brand Status*City Size -0.0341* e
Constant 7.7166*** 7.5531***
*** P>|z| = 0.001 ** P>|z| = 0.01
* P>|z| = 0.05
e not retai ned i n the cours e of the model devel opment due to non- s i gni fi ca nce
Estimated coefficient Variable
higher rental level was associated with a higher value. However, the extent of this relation depends on the respective city. To the contrary, properties with larger usable areas might even see lower values in some cities, whereas higher values can be expected in other cit- ies. The regression slope variance for the year of observation indicates that property val- ues saw different trends across time depending on the macro location. While values de- clined in some cities, others realized an upward tendency. Together, these findings draw attention to the macro location as a source of substantial variance in the appraisal of properties while additionally emphasizing the importance of a detailed analysis of the macro location in the course of investment decisions and property valuations.
The study underlies different limitations and drawbacks resulting from the data sample and the methodology that should be taken into account. As a matter of the principles of data collection from secondary sources, information quality in this study depends on the quality and soundness of the original data collector. The IPD database in turn relies on the correctness and completeness of the data provided by its cooperating companies. Look- ing at the outliers that were identified, erroneous data entries obviously cannot be ruled out. Thus, it is possible that the dataset includes information that is erroneous, but was not identified as an outlier.
Several multivariate outliers were eliminated by the dataset inspection. Using a dummy variable as an identifier for outlying cases, a multiple linear regression suggested that a low market value, an early year of observation, and a high rental level are associated with a higher probability of being an outlier. Consequently, the study results will have a limited transferability to properties that show extreme parameter values in these variables or, specifically, a combination of these characteristics. From a real estate perspective, this might be the case for properties where the contract rent is clearly above market level or substantial incentives were promised to the tenants, resulting in gaps between contract rent and market value. Similarly, study inferences might be limited for cases in which there are discrepancies between the rental situation of a property at the time of data delivery and the last valuation, or in which higher shares of non-office space lead to differences in the relationship between the rental level and market value.
The complete anonymization of the dataset that was required due to data protection standards was an obvious restriction of the analysis. On the one hand, it was not feasible to review the correctness of the data entries in detail and thus further evaluate the primary independent variable ‘Brand Status’. On the other hand, no direct conclusions could be drawn at the single asset level. Because of this, an identification of individual properties and an in-depth analysis of respective branding activities could not be realized.
Another point that needs to be considered when interpreting the study results, is the un- derlying definition of the ‘Brand Status’. By applying a broad definition approach, a wide range of properties is covered, whose brandings might differ with respect to their estab- lishment, their strength in the market, and the overall effort and resources invested in their development. For this reason, the study only enables a binary comparison of branded and non-branded properties, but it does not allow for a differentiated evaluation of respective variations in the overall kind and strength of branding activities.
Regarding the selection and number of independent variables, the study mainly relied on factors that are commonly included in standardized valuation methods and for which in- formation was available in the IPD dataset. As a result, not all potential covariates of brand status that might be relevant were covered. Future studies in this field might include explanatory variables at the postcode area level while also containing additional predictors at the city level.
On the highest level, the dataset was comprised of only 20 different cities, potentially lead- ing to smaller standard errors at this level. Interpretations of higher-level random effects should therefore be considered carefully as already discussed in Section Fehler! Ver-
weisquelle konnte nicht gefunden werden.. Thus, the significance levels for the be-
tween-city variances that were expected for the variables ‘Year’, ‘Rent’, and ‘Usable Area’ might be overestimated.
Finally, the study underlies a general limitation of regression-based methods: The meth- odology is appropriate to ascertain relationships, but not to determine underlying causali- ties. As a consequence, the multilevel analysis alone does not provide evidence about whether the branding of a property leads to a higher value or whether a higher value causes properties to be branded. Nevertheless, from a real estate perspective, the first causal chain seems to be more comprehensible and logic. As outlined in chapter Fehler!
Verweisquelle konnte nicht gefunden werden., a property’s value is commonly deter-
mined on the basis of standardized valuation methods that build upon all relevant building characteristics. Thus, one can assume that these approaches also implicitly account for the property branding. From this point of view, property brands obviously precede the property value that is determined only after the brand is already established. A reversed causality is also imaginable where a property’s high value drives an owner’s decision to engage in establishing a property brand. However, this causal chain seems less likely, since the decision to establish a brand is commonly part of new developments or refur- bishments in the property lifecycle where properties do not exhibit a high value.526
With this in mind, the study findings provide initial empirical support for the relevance of brands in an office property context and generally support decisions to engage in the de- velopment of property brands, regardless of the building’s economic year of construction, its rental level, its usable area, or the size of its macro location.
526
See ROTTKE/WERNECKE (2008), pp. 211-212; SCHULTE/BRADE (2001), pp. 38-41; ISENHÖFER