Improved Methods for Predicting Property Prices in Hazard Prone Dynamic Markets
 First Online:
 Accepted:
DOI: 10.1007/s1064001600765
 Cite this article as:
 de Koning, K., Filatova, T. & Bin, O. Environ Resource Econ (2016). doi:10.1007/s1064001600765
 335 Downloads
Abstract
Property prices are affected by changing market conditions, incomes and preferences of people. Price trends in natural hazard zones may shift significantly and abruptly after a disaster signalling structural systemic changes in property markets. It challenges accurate market assessments of property prices and capital at risk after major disasters. A rigorous prediction of property prices in this case should ideally be done based only on the most recent sales, which are likely to form a rather small dataset. Hedonic analysis has been long used to understand how various factors contribute to the housing price formation. Yet, the robustness of its assessment is undermined when the analysis needs to be performed on relatively small samples. The purpose of this study is to suggest a model that can be widely applicable and quickly calibrated in a changing environment. We systematically study four statistical models: starting from a typical standard hedonic function and gradually changing its functional specification by reducing the hedonic analysis to some basic property characteristics and applying kriging to control for neighbourhood effects. Across different sample sizes we find that the latter performs consistently better in the outofsample predictions than other traditional price prediction methods. We present the specific improvements to the traditional spatial hedonic model that enhance the model’s prediction accuracy. The improved model can be used to monitor price changes in riskprone areas, accounting for changes in flood risk and at the same time controlling for autonomous market responses to flood risk.
Keywords
Hedonic analysis Price prediction Flood risk Natural hazards Kriging Climate change Small sample Outofsample prediction1 Introduction
Housing contributes largely to the welfare of individuals. Consequently, housing prices can strongly influence households’ financial decisions (Bostic et al. 2005) making households richer or poorer as prices fluctuate. Changes in property prices are driven by changes in macroeconomic conditions, changes in consumer preferences and incomes, and exogenous shocks (Filatova 2014). At the times of natural disasters price tends to shift significantly and abruptly (Bin and Landry 2013; Atreya 2013) implying that there are systemic changes in property markets. In other words, transactions in the past may not be representative anymore when making current price assessments or projections for the future. Therefore it becomes important to utilize most recent sales in conducting reasonable market price assessments or predictions. Various comprehensive methods have been developed for these purposes in the past decades (Basu and Thibodeau 1998; Case et al. 2004; Dubin 1999; Pagourtzi 2003). Real estate appraisers, local taxation offices, mortgage lenders and insurance companies are eager to know the current value of properties in line with changing market conditions. Models that predict housing values should, thus, be calibrated with most recent sales that represent current developments in the market. There is much demand for models that can detect and predict trends in the market at an early stage, and they require robust predictions while being calibrated with only few observations (Kuntz and Helbich 2014). This is generally problematic in hedonic analysis that may require thousands of transactions to deliver reliable statistically significant estimates for various structural and spatial attributes that influence housing prices in a particular market.
Hedonic analysis is commonly used to asses and predict property prices and to estimate the flood risk premiums. In hedonic analysis a list of housing attributes is combined into a multiple regression with sales price as dependent variable. It can be used to predict future sales prices, yet the main purpose of these models is to calculate the marginal implicit price of specific housing attributes such as neighbourhood amenities, environmental quality or safety against floods (Atreya et al. 2012; Bin and Polasky 2004; Bin and Landry 2013; Hallstrom and Smith 2005). Hedonic studies often employ a large scale crosssectional data measured within a long time frame. The question remains whether these models can effectively predict prices when calibrated with only few recent sales. One of the problems with assessing and predicting future sales prices using traditional hedonic models, is the chosen functional relation between spatial factors and sales prices. The fact that housing location has a strong effect on sales price is widely acknowledged, but the complexity of space as a factor is not captured well enough in the hedonic literature (Dubin 1992). There are several ways to construct regression models that account for spatial and neighbourhood characteristics, including for example spatial error model (Anselin 2001). An extensive analysis of outofsample prediction performance of various spatial (econometric) models has been performed by Voltz and Webster (1990), Bourassa et al. (2007) and Basu and Thibodeau (1998). While usually hedonic analysis (including spatial error models) performs well on large multiyear datasets, there is a need for an improved approach for robust assessment of property prices in highly dynamic markets. As discussed above, dramatic prices changes in property market suffering from a shock, such as flooding for example, require a price prediction model that can work on small samples such as a few months of transaction data.
To mitigate the problem of a careful and robust assessment of the influence of spatial factors, Dubin (1992) suggests to omit all spatial variables in the hedonic analysis and to interpolate the spatial correlation in property prices by using kriging. Kriging is a spatial statistics method used to perform spatial interpolation, and is used for a wide range of applications in environmental sciences also based just on few observation points (Alemi et al. 1988; Delhomme 1978; HernandezStefanoni and PonceHernandez 2006; Webster and Burgess 1983). Yet just a few hedonic studies have adopted this method despite the fact that it can significantly improve the prediction performance compared to the traditional regressionbased hedonic analysis (Case et al. 2004; Kuntz and Helbich 2014). Some studies applied the technique to correct for spatial autocorrelation (Basu and Thibodeau 1998; Bourassa et al. 2007; Militino et al. 2004), and other studies also validated the method through outofsample predictions (Case et al. 2004; Kuntz and Helbich 2014). The model specifications examined in this study are based on hedonic analysis and kriging. While the literature suggests that kriging improves the prediction performance of spatial hedonic models, the performance of these models over a range sample sizes have yet to be tested. Therefore, the main purpose of this paper is to assess the robustness of the prediction performance of spatial hedonic models, either enhanced or not with kriging, under different sample sizes.
Another method used to predict property prices is artificial neural networks (Nguyen and Cipps 2001). While housing attributes in hedonic models are typically fitted with linear, log or squared relationships with price, artificial neural networks are used to fit more complex functional relationships. This works well with large housing transactions samples, but is sensitive to overfitting when calibrated with small samples. Nguyen and Cipps (2001) have compared the performance of multiple regression models with artificial neural network models across sample sizes, and have concluded that the multiple regression models perform better than artificial neural network models at small sample sizes. Moreover, regression models are far less complicated and more widely used than artificial neural networks, thus we do not consider it further in our paper.
Given the number of different methods to assess and predict property prices, the purpose of this study is to understand which model can be widely applicable across a range of sample sizes and can be quickly calibrated in a changing environment. Our main research objective is to determine which specification of a spatial hedonic model is the best in predicting sales prices when calibrated with a small sample of recent sales. We test various hedonic models by systematically changing the size of the insample set of transactions based on which the models are calibrated. The sales prices of outofsample properties are predicted with the calibrated models. We perform this analysis on the dataset with residential property transactions between 1992 and 2002 in a housing market in North Carolina. We also analyse the reliability of the flood risk discount assessed with different statistical models under various insample sizes. Given the challenges of assessing capital at risk and flood risk discount in particular in a changing environment, the outcomes of the current paper may be of interest for policy makers conducting CBA of flood risk management policies, for monitoring developments in insured and uninsured property values and capitalatrisk, and for assessing structural changes in property markets in response to natural hazards. The analysis in this paper can be applied to a wide range of natural hazards and real estate appraisal in general. Our results demonstrate when and why the krigingenhanced hedonic model performs better. Especially the prediction performance with small sample sizes is interesting, because this is where the model can quickly be calibrated and be applied for price predictions under changing market conditions. We present the specific improvements to the traditional spatial hedonic model that enhance the prediction accuracy, especially when it is calibrated with few observations.
The paper proceeds as follows. We start by giving a description of the data and the four different models, which are systematically compared for different samples sizes. Then, we explain how the analysis is done to compare the models. We conclude by discussing the results and their implications.
2 Methods
2.1 Data
Summary statistics of the property attributes
Variables  Summary (\(\mathrm{N} = 4779\))  

Mean  SD  
Sales price  USD 156,612  87,354 
Age of the house  22.4 years  19 
Number of bedrooms  3.2  0.59 
Total structure square feet  2391  993 
Lot size in acres  0.64  2.4 
Gas heating (\(=1\))  0.35  0.48 
Fireplace (\(=1\))  0.77  0.42 
Face brick (\(=1\))  0.48  0.50 
Hard wood flood (\(=1\))  0.25  0.43 
Good quality (\(=1\))  0.031  0.17 
Vacant home (\(=1\))  0.0048  0.069 
Distance to creek  854 feet  596 
Distance to airport  33,966 feet  17,859 
Distance to major road  135 feet  99 
Distance to business centre  4632 feet  2452 
Distance to railroad  5498 feet  6378 
Distance to Tar River  20,999 feet  17,587 
Distance to park  7490 feet  7051 
Sold between Fran and Floyd (\(=1\))  0.34  0.48 
Sold after Floyd (\(=1\))  0.37  0.47 
Floodplain (\(=1\))  0.064  0.24 
2.2 Model Specifications
A spatial hedonic model from Bin and Landry (2013) (M1)
An adjusted version of M1 with different functional forms (M2)
M2 with a reduced number of input variables (M3)
M3 whereby spatial variability in property prices is predicted with kriging (M4)
Input variables and their functional forms for the hedonic models
Variables  M1  M2  M3  M4 

Age of the house  \(\hbox {X} + \hbox {X}^{2}\)  \(\mathrm{X} + \sqrt{\hbox {X}}\)  X  X 
Number of bedrooms  \(\hbox {X} + \hbox {X}^{2}\)  \(\mathrm{X} + \sqrt{\hbox {X}}\)  X  X 
Lot size in acres  \(\hbox {X} + \hbox {X}^{2}\)  \(\mathrm{X} + \sqrt{\hbox {X}}\)  \(\sqrt{\hbox {X}}\)  \(\sqrt{\hbox {X}}\) 
Total structure square feet  \(\hbox {X} + \hbox {X}^{2}\)  \(\mathrm{X}\, + \) ln(X)  ln(X)  ln(X) 
Gas heating (\(=\)1)  X  X  
Fireplace (\(=\)1)  X  X  
Face brick (\(=\)1)  X  X  
Hard wood flood (\(=\)1)  X  X  
Good quality (\(=\)1)  X  X  
Vacant home (\(=\)1)  X  X  
Log of distance to creek  X  X  Kriging  
Log of distance to airport  X  X  Kriging  
Log of distance to major road  X  X  Kriging  
Log of distance to business centre  X  X  Kriging  
Log of distance to railroad  X  X  Kriging  
Log of distance to Tar River  X  X  Kriging  
Log of distance to park  X  X  Kriging  
Sold between Fran and Floyd (\(=\)1)  X  X  
Sold after Floyd (\(=\)1)  X  X  
Floodplain (\(=\)1)  X  X  X  X 
Floodplain \(\times \) sold btw Fran and Floyd  X  X  
Floodplain \(\times \) sold after Floyd  X  X 
The hedonic analysis is used to estimate the coefficients of the input variables \(\beta _k \) (Eq. 2). In M4 hedonic analysis is used to understand the influence of the core spatial variable of interest (flood risk) on property prices while the rest of the spatial variability in prices is captured by interpolating the residuals (\(E^{i}\), Eq. 2) using kriging. A list of all the input variables can be found in Table 2. M1 is same model with the same specifications as used in Bin and Landry (2013). However, the model in Bin and Landry (2013) has more dummy variables than M1 in this paper, as they distinguish the 100 and 500year flood zones. The reason this separation is not made in M1 is that only 2 % properties in the 500year floodplain were sold, so that the 500year floodplain properties are many times absent in this subset. Therefore, we merge the 100 and 500year flood zone properties in one floodplain variable.
M2 was built with the same input variables as M1, while changing some of the functional forms in order to better describe the saturation behaviour of the variable’s influence on price. The variables \(bedrooms^{2}\), \(age^{2}\), \(square~footage^{2}\) and \(acres^{2}\) were substituted by \(\sqrt{bedrooms}\), \(\sqrt{age}\), \(\ln \left( {square~footage} \right) \) and \(\sqrt{acres}\) respectively. These variables often enter the hedonic analysis function in the quadratic specification (Case et al. 2004). Yet, in our dataset price dependence on them does not necessary follow the parabolic form (Do and Grudnitski 1993; Goodman and Thibodeau 1995) (Fig. 5, Appendix 1). In M3 and M4 we consider a reduced regression that contains only the main characteristics of the properties—sq. footage, bedrooms, acres and age—that have a clear, straightforward and always statistically significant effect on price.^{2}
2.3 Analysis
We split the entire data set (4779 observations) into insample and outofsample parts. The coefficients of the hedonic regressions are calculated using insample data, which is constructed as a subset of the sales data ranging from 0.1 to 20 %. We systematically vary this share to understand how small the insample subset can be to be able to deliver an acceptable predictive power for each of the models. The remaining transactions form the outofsample dataset, which we use to compare the prices predicted by the four models against the actual sales price. The coefficients of the insample hedonic regressions are used to form the predicted prices of the outofsample properties in M1, M2 and M3. In M4 we sum the regression estimates and the kriged residuals.
Further, we apply Monte Carlo method by taking 50,000 random subsets ranging from 0.1 % (\(\mathrm{N}\,=\,48\)) to 20 % (\(\mathrm{N}\,=\,956\)) of the dataset, with constant density.^{4} Each model is calibrated with the same subsets, so that we can make pairwise comparisons. There is no limit to the number of subsets we can take, so we decide to take enough to cover the full range of model performances. The performance of a model may strongly depend on the subset it is calibrated on. Therefore we choose to look not only at average performance, but also at the 95 % confidence interval of the performance range across sample sizes.
First, M1 and M2 are compared to assess how changing functional forms affect the prediction performance across sample sizes. Second, M2 and M3 are compared to assess the effect of reducing the number of input variables in the hedonic analysis, which should be more suitable for predicting prices based on small samples. Third, M3 and M4 are compared to see the effect of kriging for different sample sizes. And finally, the performances of all models are compared to see which model performs best across various sample sizes.
3 Results
3.1 Comparing Functional Specifications of the Full Hedonic Model (M1 and M2)
The difference in performance between M1 and M2 can be explained by the functional forms of the input variables. The functional forms of some input variables of M1 are described by a squared relation with price, which is meant to represent the saturation behaviour of the characteristic’s effect on price. However, a squared or second degree polynomial function is not a saturating function. Rather, it has a peak (minimum value or maximum value depending on the sign of the coefficient) and can only approximate saturation behaviour on a local scale, but it can neither describe nor predict actual saturation behaviour overall. This results in a low model fit, which is measured by the Adjusted Rsquared metric (Fig. 1). When the properties in the subset only contain a limited range of the characteristics compared to their entire range within the population, it can lead to large errors in the predicted price of properties with characteristics on the extreme ends of the range. Changing the squared functions to square root and log functions as in M2 could therefore considerably reduce the extreme prediction errors.
3.2 Reducing the Number of Explanatory Variables (M2 and M3)
At sample fractions higher than 0.05 (or \(\hbox {N}>240\)) M2 generally has a lower mean absolute error than M3, suggesting that M2 outperforms M3 when they are calibrated with samples of \(\hbox {N}>240\). However, this conclusion does not hold when looking at the other performance metrics. The upper part of the 95 % confidence interval of RMSE and SDE shows that the predictions of M2 can still be quite volatile compared to M3. Thus even though M2 performs better than M3 on average with samples of \(\hbox {N}>240\), the precision of M2’s predictions is still lower than that of M3, where M2 has a higher probability of strongly inaccurate predictions.
3.3 Explaining Spatial Variability in Prices Through Kriging (M3 and M4)
Kriging offers an alternative way to account for the influence of spatial complexity in price assessments. Namely, it captures any systematic variation in prices through the analysis of residuals. In our dataset we find a clear spatial correlation in the semivariogram of the insample residuals of M3 (Fig. 3). The semivariance increases with distance between properties, which shows that property prices are spatially correlated. This indicates that the prediction performance of M3 can be improved with regression kriging (Basu and Thibodeau 1998). Regression kriging is done with the same variables as the hedonic model M3, after which the residuals are interpolated, so that only the remaining variation is addressed.
Comparing M3 (reduced hedonic model without kriging) and M4 (with kriging), we see that the model’s performance consistently improves when kriging captures the spatial autocorrelation in residuals (Fig. 4). SDE, MEA and RMSE are consistently lower when kriging is added to M3 (Wilcoxon SignedRank Test, \(\hbox {P}<0.001\)), and Adjusted Rsquared is consistently higher with kriging (Wilcoxon SignedRank Test, \(\hbox {P}<0.001\)). Most importantly, this result is consistent across all sample fractions.
When comparing the four models, we find that M4 with kriging performs consistently better across various sample sizes. Thus, it delivers a more robust model to be used in hedonic analysis without a need for a researcher to worry about meeting a particular threshold of an insample size: it simply performs well for a large variety of sample sizes. Looking at the performance metrics, M3 with kriging is the best model in 96.2 % of the cases with RMSE and SDE, 97.2 % of the cases with MAE and 98.9 % of the cases with Adjusted Rsquared. It also implies that kriging explains the spatial variability in property prices better than the spatial variables that are included in M1 and M2 and releases a researcher from worrying about functional specifications of the spatial variables of secondary importance. At the same time, using kriging in combination with traditional hedonic analysis allows disentangling spatial attributes of a particular interests for economic analysis—e.g. a location within a flood zone in our case—to be studied with precision.
3.4 Implications for Policy Making: Example of the Flood Risk Discount
Stability of a regression coefficient of the floodplain variable over various insample sizes
Sample fraction  M1  M2  M3 & M4  

Mean  SD  Mean  SD  Mean  SD  
\({<}\)0.05  0.150  17.034  −0.055  0.447  −0.050  0.109 
\({<}\)0.10  −0.063  0.103  −0.061  0.056  −0.052  0.055 
\({<}\)0.15  −0.061  0.040  −0.061  0.041  −0.053  0.041 
\({>}\)0.15  −0.060  0.033  −0.060  0.033  −0.053  0.033 
Total population  −0.056  –  −0.056  –  −0.054  – 
The full regression model with the traditional functional specification M1 provides a rather unstable estimation of the flood risk coefficient since it varies greatly with the sample size. In fact for small insample sizes the flood dummy coefficient is positive but its standard deviation across random Monte Carlo sampling sets is huge undermining its statistical insignificance. The same can be said about M2 for small insample sizes, although the standard deviation is already much lower than M1. M3 and M4 have the lowest standard deviation for small insample fractions \(({<}0.05)\). With larger sample fractions \(({>}0.10)\) it does not matter anymore which model is used to predict the flood coefficient.
Results in Table 3 suggest that the sales price differential between inside and outside the floodplains ranges from 5.0 to 6.3 % with an exception of M1 with less than 0.05 sample fraction. Several previous studies have documented the price reduction from location in a floodplain (MacDonald et al. 1987; Bin and Polasky 2004; Hallstrom and Smith 2005; Bin et al. 2008; Daniel et al. 2009). A common finding in these studies is that location within a floodplain lowers property value anywhere from 4 to 12 %. As shown in Table 3, our approach to limit the number of variables that enter the hedonic regression can be quite useful in determining the risk premiums associated with flooding especially with small samples. Our results may help insurance practitioners and policy makers make informed decisions on the flood risk management especially when the available data set is very limited.
4 Discussion
Across all sample sizes we see that M4 with kriging performs best in the outofsample predictions regardless of the insample size. This model differs from M1 and M2 in the way the spatial variables enter the price estimation. In M3 the spatial variables were omitted completely, whereas kriging was used in M4 to predict spatial variability and spatial autocorrelation in property prices by analysing the residuals. Krigingbased M4 is thus more powerful in predicting the spatial patterns in property prices. This may not be surprising, as kriging is used for spatial interpolation by explicitly accounting for spatial autocorrelation, which is often present the property market (Basu and Thibodeau 1998; Bourassa et al. 2007; Dubin 1992; Militino et al. 2004). Case et al. (2004) have already shown that kriging can, for this reason, enhance the outofsample prediction performance of spatial hedonic models. Yet, their models were calibrated with very large samples \((\hbox {N}\approx 50,000)\) and the question remained whether these conclusions hold for small samples. We have specified a model that is consistent in assessing housing prices and predicting future sales prices, even when calibrated with a limited number of recent sales. We can zoom into the mechanisms of why this model performs best by comparing M1, M2, M3 and M4 with kriging pairwise across different sample sizes.
Comparing M1 and M2 reveals that the squared terms in the hedonic model can cause large errors in outofsample predictions of the property prices, with estimated prices that sometimes deviate even by several orders of magnitude from the actual price. This problem especially occurred at small sample sizes for which M1 was calibrated. The poor outofsample prediction performances were expressed by a high variability in prediction errors and a low model fit. The latter indicates that the chosen functional forms of the variables may not represent their actual effect on price. Squared functional forms, and sometimes even cubed functional forms, are currently widely used in hedonic literature (Bin and Landry 2013; Case et al. 2004). These models are based on datasets that are usually large enough to sanitise the effects. Yet, we found that even with samples of \(\mathrm{N}\,=\,900\) the predictions can be inaccurate in the outofsample predictions as a result of a low model fit.
A comparison between M2 and M3 reveals that overfitting is the main cause of the poor outofsample prediction performance of M2 at very small sample sizes. A similar mechanism called overtraining is also causing poor prediction performances of artificial neural network models (Nguyen and Cipps 2001). Reducing the number of input variables in M3 results in a consistent prediction performance across sample sizes. Moreover, it leads to a decrease of volatility in the model’s predictions. However, we also observe that M2 scores better on the metric mean absolute error when a sample size increases. This implies that some of the model’s parameters can only be estimated when the number of sales is high. When trying to calibrate a model based on only a few recent sales, it is better to focus only on a few explanatory variables.
M4 with kriging outperforms M3 across the range of sample sizes, but performs only slightly better than M3 with sample sizes of \(\hbox {N}\approx 50\). In fact, the improvement of kriging with respect to the prediction performance of M4 increases with sample sizes. The strong prediction performance of M3 and M4 at small sample sizes is mainly due to the reduction of variables in the hedonic model, whereas the role of kriging in improving the prediction performance of M4 becomes more important with increasing sample sizes. The latter is caused by an increase in density of the sample, so that the 15 nearest properties that are selected for spatial interpolation are on average closer to the predicted location, i.e.: their actual transaction prices are more correlated with the predicted price. To summarize: we observe that changes in functional forms of the input variables, a reduction of input variables and kriging improve the prediction performance at different parts of the range of sample sizes. Together, these specifications complement each other to form a model that is consistently better in prediction performance across sample sizes.
Despite that M4 consistently produces the best price predictions across a range of sample sizes, we do not suggest that it can replace the hedonic analysis when it is used for other purposes, namely to assess marginal implicit prices of specific housing attributes (Janssen et al. 2001). In this case the spatial attributes of interest should be kept in the hedonic function part of the analysis while the impact of other spatial neighbourhood attributes on price may be captured by kriging. The limitation of kriging is that all involved spatial attributes go into the black box of spatial interpolation. Thus, one cannot trace back which spatial factors exactly affect property prices and to what extent.
For the purpose of predicting property values, for example in real estate appraisal (Pagourtzi 2003), it can be useful to work with models that are consistent in their performance even when calibrated with few current sales. Our model, which performs consistently well across a range of sample sizes, is particularly useful for this application. For policy makers that deal with management of natural hazards it is important that a good assessment of the capital at risk is made. Housing markets are affected by macroeconomic changes as well as changes in consumer preferences, incomes and WTP for various property attributes. Housing markets in hazard areas experience structural changes in price trends after disastrous events, and these changes are expected to accelerate with climate change. Thus, when conducting a CBA for flood risk management policies it is essential that the assessment of capital at risk or a flood risk premium is based on the most recent sales to better reflect current market conditions. For example when dealing with flood risk, it is important to account for changing risk perceptions, which is driven by flood events and changing flood probabilities, influencing how risk is capitalized into property prices (Pryce et al. 2011; Atreya et al. 2012; Bin and Landry 2013). Our algorithm allows for a rapid updating of property price assessments and predictions. Therefore, it can quickly capture a market response to potential changes in location preferences, market conditions and flood risk perceptions. This approach can be used to monitor price changes in riskprone areas, accounting for changes in flood risk and at the same time controlling for autonomous market responses to flood risk.
Note, that the original study of Bin and Landry (2013) employed a spatial error model. Kriging and spatial error models differ in the way in which the spatial weight matrix is constructed. The spatial weight matrices in spatial error models are constructed based on the assumptions of the user, whereas in kriging they are based on the spatial structure of the error, which is defined in the construction of the semivariogram. The spatial error model is particularly relevant for proper estimation of the coefficients in the hedonic price estimation, whereas kriging focuses on the prediction of the dependent variable.
We did a thorough analysis with various combinations of log and square root functional forms to identify the functional forms that best fitted the transaction data.
We did a sensitivity analysis and concluded that 10–20 nearby properties is a good number to use since it does not change the model’s performance. More than 20 does not change the performance but enhances computation time, while less than 10 properties reduces the prediction performance. The results of our sensitivity analysis are available upon request.
Funding information
Funder Name  Grant Number  Funding Note 

NWO VENI 

Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.