Choosing day-care centers by location-speciﬁc characteristics: a Conditional Logit model approach

(1)

M

ASTER

T

HESIS

Choosing day-care centers by

location-specific characteristics: a

Conditional Logit model approach

Author: Jildert de Boer University supervisor: Prof. S. Sóvágó Company supervisor: G. Heeringa

A thesis submitted in fulfillment of the requirements for the degree of Master of Science

in the

Faculty of Economics and Business

(2)

i

Master’s Thesis Econometrics, Operations Research and Actuarial Studies Supervisor: prof. S. Sóvágó

(3)

ii

UNIVERSITY OF GRONINGEN

Abstract

Faculty of Economics and Business Master of Science

Choosing day-care centers by location-specific characteristics: a Conditional Logit model approach

by Jildert de Boer

Little is known how geographic characteristics of day-care centers affect the choice for a day-care center. Research has shown that main factors affecting the choice are price, quality and convenience, and that individual characteristics of parents also play a role. This thesis studies how geographic characteristics of day-care centers affect the choice for a day-care center. Specifically, it is analyzed to what extent dis-tance to the center plays a role in choice behavior, and whether day-care centers in primary schools and/or day-care centers with an after-school care center on the same address affect the choice made by parents. Individual characteristics of parents or children are interacted with these geographic features, to explain possible hetero-geneity in choices. To see how these geographic features can be used for practical purposes, counterfactual predictions are made for a fictive day-care center.

(4)

iii

Acknowledgements

Many thanks to my supervisor, prof. S. Sóvágó, for his enthusiasm on this topic, our regular meetings, and the useful feedback you gave me.

(5)

iv

List of Figures

1.1 Decision process for choosing the type of childcare . . . 4 2.1 Mean number of children attending day-care centers over time,

cate-gorized by age and sex. . . 7 2.2 Distances to day-care centers, categorized by municipality. Distances

are categorized in levels. . . 8 2.3 Map of day-care centers provided by SKSG . . . 8 2.4 Relationship of the number of children and the childcare spots . . . 9 2.5 Available childcare spots for day-care centers in Groningen and

West-erkwartier, categorized by whether an after-school care center is present and whether it is located in a primary school. . . 10 2.6 Number of children and utilization rates in day-care centers located

in municipality Groningen. . . 11 2.7 Number of children and utilization rates in day-care centers located

(8)

vii

List of Tables

A.1 Variables and their description . . . 31 A.2 Descriptive statistics of features of day-care centers in municipality

Groningen, aggregrated over years 2016-2019. . . 32 A.3 Descriptive statistics of features of day-care centers in municipality

Westerkwartier, aggregrated over years 2016-2019. . . 32 A.4 Descriptive statistics of children in the data set. . . 32 A.5 The number of children and % of children living in municipality

Gronin-gen, Westerkwartier or another one choosing a day-care center in re-spectively Groningen and Westerkwartier. % of children are calcu-lated over the municipality the child lives in. . . 32 A.6 Estimation results of a CL model on the whole sample with

alternative-specific variables xij and intercept γjfor each day-care center j. . . 33 A.7 Predicted odds ratios in % change (95% CI in parentheses) . . . 34 A.8 Predicted willingness to travel in km (95% CI in parentheses). . . 34 A.9 Estimation results of a CL model with alternative-specific variables xij

and intercept γj for each day-care center j. Subsample: municipality Groningen. . . 35 A.10 Predicted odds ratios in % change, calculated over municipality

Gronin-gen (95% CI in parentheses). . . 36 A.11 Predicted willingness to travel in km over municipality Groningen

(95% CI in parentheses). . . 36 A.12 Estimation results of a CL model with alternative-specific variables

xij,intercept γj and interaction terms for each day-care center j. Sub-sample: municipality Groningen. . . 37 A.13 Estimation results of a CL model with alternative-specific variables xij

and intercept γj for each day-care center j. Subsample: municipality Westerkwartier. . . 38 A.14 Predicted odds ratios for municipality Westerkwartier in % change

(95% CI in parentheses) . . . 39 A.15 Predicted willingness to travel in km over municipality Westerkwartier

(95% CI in parentheses). . . 39 A.16 Estimation results of a CL model with alternative-specific variables

xij,intercept γj and interaction terms for each day-care center j. Sub-sample: municipality Westerkwartier. . . 40 A.17 Summary statistics of the Absolute Error, for each year in which choices

are made. . . 41 A.18 Summary statistics of the Absolute Error with the new day-care center

in the choice set, for each year in which choices are made. . . 41 A.19 Counterfactual predictions for the new day-care center (in bold),

(9)

1

1 Introduction

The choice of daycare for children has been, and still is, a widely researched topic among academics. In recent years, many researchers have tried to explain the choice behavior of parents regarding childcare and have found that the main determinants of childcare choice are price and quality (Blau,1991; Hofferth and Wissoker, 1992; Duncan, Paull, and Taylor,2001; Lehrer,1983). Besides price and quality, research has shown that other important factors for choice of care are convenience - such as hours of care, reliability of care and travel time and characteristics of parents -such as income and education level (Johansen, Leibowitz, and Waite,1996; Kim and Fram, 2009; Huston, Chang, and Gennetian, 2002). Furthermore, a lot of research on this particular topic focuses on explaining the choice where the modes of choice generally exist of three or more parts (Yesil-Dagli,2011; Johansen, Leibowitz, and Waite,1996; Hofferth and Wissoker,1992; Brandon and Hofferth,2003; Peyton et al., 2001). In its most general form, Johansen and his fellow researchers distinguish three parts: care at the child’s home, care in someone else’s home, and center-based care.

1.1 Research question, contribution and hypothesis

Parents interested in center-based care for their preschool child have the option to choose day-care centers. This thesis will focus on one particular mode of choice: care at a day-care center. Little is known about how characteristics of such day-care centers affect the choice for a specific day-care center. Using data from the Dutch childcare organization Stichting Kinderopvang Stad Groningen, shortened SKSG, this paper aims to explain how (geographical) characteristics of day-care centers af-fect the choice caregivers make. In particular, it aims to find evidence as to what ex-tent distance to a day-care center affects the choice of a day-care center. Furthermore, it looks at whether day-care centers located in primary schools and day-care centers with an after-school care center on the same address are more preferred compared to day-care centers that do not have one of these features. Therefore, the goal of this thesis is to determine how these geographical characteristics affect the choices caregivers make regarding day-care centers. This leads to the following research question:

How do geographical characteristics of day-care centers affect the choice for a day-care center? An answer to this research question will not only give insights into how geographi-cal characteristics affect the choice for a day-care center. This thesis will also provide a practical analysis of how these characteristics can be applied to the opening of a new day-care center, by making counterfactual predictions. The research question will be answered by using the features mentioned above, where the following sub questions are listed:

(10)

Chapter 1. Introduction 2 2. How do day-care centers located in primary schools affect the choice for a day-care

center?

3. How do day-care centers with an after-school care center on the same address affect the choice for a day-care center?

4. How do these factors differ across characteristics of parents or children?

It is plausible to say that the closer a day-care center is to the parents’ address, the more likely it is to be chosen. However, in regions with heavy competition in child placement for day-care centers among different childcare organizations, the extent to which distance plays a role matters. Parents may consider day-care centers located in primary schools as more preferred compared to day-care centers not located in primary schools. The reason for this may be because parents find it important - or convenient - that the child accustoms to the primary school he/she will attend when he/she reaches the age of four. Furthermore, Davis and Connelly find that choosing center care increases with the age of the child and that this is possibly due to parents preferring school-like environments when the child approaches school age (Davis and Connelly,2005). Therefore, the hypothesis is that parents prefer day-care cen-ters in primary schools over day-care cencen-ters that are not located in primary schools, especially for children approaching school age. In a similar fashion, day-care centers with an after-school care center on the same address may be preferred over day-care centers without an after-school care center on the same address. This is because it allows easier transitioning from the child to the after-school care center.

(11)

Chapter 1. Introduction 3

1.2 Daycare choice in the Netherlands

To introduce the topic properly, a brief description of daycare in the Netherlands is discussed first. As already mentioned, there generally exist three modes of childcare: center-based care, care at the child’s home and care at someone else’s home. All three types are legitimate types of care in the Netherlands. Care at own home or someone else’s home is usually done by other family members, child minders or nannies. There exist different types of facilities for center-based care. For children up to four years old, child care is provided at day-care centers. For children aged two or three years old, the child may attend kindergartens. Technically speaking, a kindergarten simply is a day-care center in the Netherlands.1 For children between four and thirteen years old, there are after-school care centers. As its name suggests, these centers provide care after (or before) school.

In the Netherlands, the market for daycare is competitive and prices vary be-tween different childcare organizations and other providers of daycare. For chil-dren under thirteen, the government subsidies childcare by means of childcare lowances. A child care allowance is a financial contribution (usually a monthly al-lowance) for working parents with a written agreement to a child care organization registered in the National Childcare Register (Landelijk Register Kinderopvang).2 The process of how preschool children are assigned to day-care centers is shown in Figure 1.1. Parents or caregivers interested in using childcare first decide on the type of childcare.3 If parents decide to bring their child to a day-care center, they have to make a decision regarding which day-care center to choose. Typically, the parents contact the childcare organization or immediately contact the day-care cen-ter they are incen-terested in. Subsequently, the parents have the option to undertake a guided tour. Based on the first contact with the childcare organization and the guided tour, they make a decision regarding the day-care center of interest. Only when the parents of the child choose a day-care center belonging to SKSG it will be visible in the data set. Therefore, the children in the data set belong to a selected sample: conditional on choosing care at a care center, the parents choose a day-care center provided by SKSG. This has several implications for the choice set, which will be discussed later on.

The main results in this thesis show that distance to day-care center highly affects the choice. Considering the whole sample, the relative odds of choosing a day-care center 1 km farther away decreases by 41 to 49%. The degree to which this choice is affected depends on densities of day-care centers: the higher the density, the more distance plays a role. The degree to which distance matters differs across parents or children with different characteristics. Most importantly, parents with high income status are less sensitive for the distance to a day-care center, compared to parents with low income status. Day-care centers in primary schools or day-care centers with an after-school care center on the same address are preferred among children approaching school age. However, parents with high income status choose day-care centers with an after-school care center more often compared to parents with low income status.

1_{As of January 1, 2018, a law called "Wet harmonisatie kinderopvang peuterspeelzaalwerk"}

en-forced kindergartens to meet quality requirements of a day-care center.

2_{Childcare is also subsidized when parents choose care at a child minder, provided that the child}

minder is registered in the National Childcare Register.

(12)

Chapter 1. Introduction 4

FIGURE 1.1: Decision process of choosing the type of childcare for preschool children.

(13)

5

2 Data

2.1 Data collection

2.1.1 Child-specific data

Data about parents’ choices is collected. Each observation in the data set reflects a child attending a location provided by SKSG.4 Individual-specific characteristics are provided such as age, gender, home address and application time (i.e. the time between the first contact with SKSG and the actual starting date of the child on the center).

In addition, income data on zip-code level is collected from the Centraal Bureau voor de Statistiek, shortened CBS. In December, 2008 the CBS conducted research and calculated the average fiscal monthly income per household, aggregated on the zip-code level. Over time, this aggregated fiscal monthly income may have changed substantially. Therefore, an imperfect proxy for income per zip-code, either high or low, is created, making income more persistent over time. It is high (i.e. it equals 1) if average income is above or equal to the third quartile of the income distribution.5 It is low (i.e. it equals 0) if average income is below the third quartile of the income distribution. This data is added to the child-specific data, where the income data is matched with the child-specific data based on their zip-codes.

Based on the home address of the child, a variable is added which gives the municipality the child lives in. Again, this could be done by using a data set from the CBS, which provides information about the municipality for a given zip-code. This data set is merged on the main data set using the zip-codes from both data sets. The municipality the child lives in is created in such a way that there are three levels: municipality Groningen, Westerkwartier and other.

2.1.2 Location-specific data

Characteristics specific to the day-care center have been attached to the child-level data. For every day-care center, geographic information was collected, such as the address, zip code and the municipality of the center. Furthermore, characteristics specific to the day-care center were added, by manually scraping them from the of-ficial SKSG website.6 The first variable added gives the daily maximum number of allowed childcare spots per day-care center. Dutch law and regulations concern-ing childcare state that the maximum number of children in a location per day de-pends on the size of that location: the more space a location has, the more children are allowed in that location (Besluit kwaliteit kinderopvang). However, the number of children does not only depend on the size of a location, but also on the number of

4_{Some children attend multiple locations. This issue will be addressed later on.}

5_{The income distribution is based on zip-codes, which means that all parents in the same zip-code}

will either have a high or low income status.

6_For _each _location, _information _about _specific _locations _is _found _on

(14)

Chapter 2. Data 6 employees available.7 _{In addition, two dummy variables are collected: a variable} equalling 1 if the day-care center has an after-school care center on the same address and a variable equalling 1 if the day-care center is located in a primary school.

Distances to day-care centers have been collected by calculating the distance from each child’s home address to all possible day-care centers provided by SKSG.8 The final data set contains choices made by parents, individual-specific characteris-tics of parents and children, and several location-specific characterischaracteris-tics. A descrip-tion of all variables used in this thesis is given in Table A.1.

2.1.3 Sample restrictions

Several restrictions are imposed regarding the sample of children. Firstly, since this thesis focuses on day-care centers, all observations in which children attend after-school care centers have been deleted. This means that the data set focuses on day-care centers that provide childday-care for whole days, usually for children under the age of four.9 Therefore, all children in the data set older than four years old are also deleted, since it is very uncommon that children older than four years attend day-care centers. For some children, child-specific characteristics such as gender, application times and income status are missing. All application times exceeding 365 days are deleted, since it is unrealistic to have application times longer than a year. Furthermore, application times are not calculated for children born before January 1, 2016. The reason for this is that the earliest starting date visible in the data set is January 1, 2016. However, a child may have attended the day-care center already before 2016, which in turn would result in an upward bias of application times. This resulted in a lot of missing values for application times of children. For only 39.8% of the observations, the application time is known. For 13.4% of children in the data set, the income status is unknown. Additionally, children where parents have chosen the following day-care centers are deleted: SKSG Wikke, SKSG Oranjeboog, SKSG Peuterpleintje, SKSG Lutje Stekje and SKSG Lancelot. As of the moment of writing, these day-care centers are all closed.

2.2 Data description

This section will present some statistics about the children in the data set and the day-care centers provided by SKSG. Furthermore, by means of box plots it shows how the number of children and utilization rates of day-care centers with different characteristics are distributed.

2.2.1 Child-specific characteristics

Table A.4 shows descriptive statistics of children in the data set. There are almost as many males as females in the data set: the percentage of children being male is 51.4%. Not surprisingly, the ages of children attending day-care centers vary be-tween zero and four years with the average child attending a day-care center being 1.4 years. On average, the first contact with SKSG is 142 days before the date the

7_{In the Netherlands, the ratio number of children per pedagogical employee may not exceed some}

threshold c, depending on the age of the child (Besluit kwaliteit kinderopvang).

8_{Bing Maps enables users to create an API key for free, which allows calculating distances from}

address A to address B. Source: https://www.bingmapsportal.com.

9_{In the Netherlands, education starts at the age of four. Therefore, children aged four and up usually}

(15)

Chapter 2. Data 7

(A) Age (B) Sex

FIGURE 2.1: Mean number of children attending day-care centers over time, categorized by age and sex.

child attends a day-care center, or 5 months. The average child travels 3.5 km to a day-care center. The income variable, denoting either zero or one, can be interpreted as approximately 30% of parents having a high-income status.

To elaborate on the demographics of children attending day-care centers at SKSG, the number of children is plotted over time, categorized by age and sex. Figure 2.1 shows the mean number of children for each age- and sex category over the past four years, aggregated on the monthly level. Both figures show heterogeneity in the mean number of children across age and sex and an increasing number of chil-dren in day-care centers. Regarding the mean number of chilchil-dren categorized by age, Figure 2.1a shows that children aged two and three years old are more likely to attend day-care centers compared to other age groups, which is not surprising.10 Furthermore, the increase in number of children in day-care centers is mainly due to children that are two or three years old: the trend line appears to be steeper for those groups. Regarding the differences in sex of children, Figure 2.1b shows that there have always been more males than females attending day-care centers in the past four years, a result that is rather remarkable. The most obvious reason for this is simply more births in males than in females.

Next, we visualize distances of children to day-care centers, categorized by the municipality they live in. Figure 2.2 shows the percentage of children travelling a certain distance, categorized by distance interval. On average, parents of children in Westerkwartier travel a higher distance compared to parents of children in Gronin-gen. Where the average distance is 2.2 km for parents in municipality Groningen, it is 2.8 km for parents in Westerkwartier. The distribution of distances does not vary a lot between Groningen and Westerkwartier. Approximately 45% of children living in a municipality other than Groningen and Westerkwartier travel a distance longer than 10 km. More than 40% of children living in Groningen and Westerkwartier travel a distance no longer than 1 km. Almost 70% of these children do not travel a distance longer than 2 km.

10_{Frost and Schneider,}₁₉₇₁_{show that pre school children represent the largest group of care, either}

(16)

Chapter 2. Data 8

FIGURE2.2: Distances to day-care centers, categorized by municipal-ity. Distances are categorized in levels.

2.2.2 Day-care centers

Figure 2.3 shows us how the day-care centers are divided over different regions within the province of Groningen. The day-care centers are located in municipalities Groningen and Westerkwartier. There are 47 different day-care centers, of which 35 are located in Groningen and 12 in Westerkwartier. The size of the dots is propor-tional to the maximum number of children allowed in that location.

FIGURE 2.3: An overview of day-care centers in municipalities Groningen and Westerkwartier.

(17)

Chapter 2. Data 9 Groningen. In Westerkwartier, this is slightly lower with 37 children on average. More than half of the day-care centers in Groningen have an after-school care center on the same address. In Westerkwartier this is the case for exactly half of the day-care centers. The fraction of day-day-care centers located in primary schools is roughly 37% in Groningen, and 17% in Westerkwartier.

Conditional on deciding that the type of childcare is care at a day-care center, the parents’ first choice for a day-care center may not be chosen due to capacity con-straints (Carlsson and Thomsen,2015). These capacity constraints depend on the daily number of employees available and the maximum number of allowed child-care spots in a day-child-care center. The daily number of employees and the maximum number of allowed childcare spots are likely to depend on each other: the higher the established maximum number of allowed children, the more employees are avail-able on a daily basis, which in turn affects the daily number of children attending a day-care center. Figure 2.4 shows this presumption is correct: the higher the mean number of children attending a day-care center, the "bigger" the day-care center is in terms of childcare spots. This also confirms that more employees are working on "bigger" day-care centers. Therefore, the maximum number of allowed child-care spots can be used as an imperfect proxy for capacity: it controls to some extent whether parents choose a day-care center based on availability of that center.

Fig-FIGURE2.4: Scatter plot of the mean number of children and the max-imum number of allowed childcare spots, including a fitted

regres-sion line by OLS.

(18)

Chapter 2. Data 10

(A) After-school care centers. (B) Primary schools.

FIGURE2.5: Available childcare spots for day-care centers in Gronin-gen and Westerkwartier, categorized by whether an after-school care

center is present and whether it is located in a primary school.

that offers space for 31 children. Day-care centers located in primary school allow fewer children on a daily basis compared to other day-care centers in the municipal-ity Groningen. This is not the case for day-care centers in Westerkwartier.

Based on Figure 2.5, one would, therefore, expect that day-care centers with an after-school care center on the same address and those located in primary schools are chosen more often in the municipality Groningen, simply because of more capacity. For Westerkwartier, exactly the opposite is expected.

To obtain insight into how day-care centers with different geographical characteris-tics are utilized over the past four years, several box plots are plotted. Two measures for the utilization of a day-care center are given: the daily number of children and the daily utilization rate of a day-care center. The utilization rate for day-care center j at time t is defined as

Utilization Ratejt =

Number of childrenjt

(19)

Chapter 2. Data 11

(A) Number of children, grouped by BSO. (B) Utilization rates, grouped by BSO.

(C) Number of children, grouped by PRIMS. (D) Utilization rates, grouped by PRIMS.

FIGURE2.6: Number of children and utilization rates in day-care cen-ters located in municipality Groningen.

Regarding day-care centers located in primary schools and those that are not, the differences are more pronounced as we can see in Figures 2.6c and 2.6d. The daily number of children attending day-care centers located in primary schools is significantly less compared to the number of children attending day-care centers not located in primary schools. This can partly be explained by capacity constraints. It becomes clear from Figure 2.6d that utilization rates of day-care centers in primary schools are no less different from day-care centers not located in primary schools. However, this figure also shows that utilization rates of these day-care centers have increased over time.

(20)

Chapter 2. Data 12

(A) Number of children, grouped by BSO. (B) Utilization rates, grouped by BSO.

(C) Number of children, grouped by PRIMS. (D) Utilization rates, grouped by PRIMS.

FIGURE2.7: Number of children and utilization rates in day-care cen-ters located in municipality Groningen.

(21)

Chapter 2. Data 13

2.3 Restrictions on the choice set

Estimating discrete choice models requires the choice set to fulfill three criteria (Train, 2009, p. 15-18). First, the choice set needs to be mutually exclusive, that is, choos-ing one alternative necessarily implies not chooschoos-ing other alternatives. Second, the choice set must be exhaustive, which means that all possible alternatives should be in there. Third, the choice set must be finite. This criteria is satisfied already.

The first criteria is not necessarily satisfied: the parent may have chosen multiple day-care centers for their child. 7.7% of children in the data set went to multiple day-care centers. As this is a substantial portion of children in the data set, deleting these observations would bias the estimates. Therefore, for those children the "pri-mary" choice of day-care center is picked, which is the day-care center that is most visited by those children over the years 2016-2019.11

The second restriction is not satisfied, since parents may also choose formal day-care centers provided by competitors. These alternatives should also be in the choice set. Unfortunately, there is no data available about parents choosing a day-care center from competitors. To overcome this problem, I impose the following assumption.

Assumption 1 The parent chooses one of the available day-care centers provided by SKSG,

given that the parent decided to choose daycare at SKSG.

Suppose a parent is interested in taking his/her child to a day-care center. Then this assumption implies that the parent first decides that daycare is provided by SKSG. Subsequently, the parent chooses one of the available day-care centers provided by SKSG. Thus, conditional on choosing formal childcare organization SKSG, the par-ent selects one out of the available options. This assumption is somewhat restricted in that a parent may not choose according to this two-step procedure. If it is decided that the type of childcare is care at a day-care center, then parents do not necessarily select on the formal childcare organization, but rather select on the day-care cen-ter immediately. However, the assumption might hold because prices differ among formal day-care services, and parents choose the formal childcare organization that offers low prices (Davis and Connelly,2005). Furthermore, there might be other dif-ferences specific to the formal childcare organization, such as trustworthiness, addi-tional services provided by the organization etc. that are preferred by the parent. Estimating discrete choice models becomes more complicated and computationally more expensive when the number of parameters to be estimated becomes larger. This is the case when the choice set is large, that is, the number of alternatives a par-ent can choose from is large. Allowing heterogeneity in regions is more fair, because we have reason to believe that the effect of distance on choice of day-care center is different in regions where the total available options are small. Therefore, the choice set is restricted for children living in municipality Groningen and Westerkwartier to the day-care centers located in Groningen and Westerkwartier, respectively. Let S = {Groningen, Westerkwartier}. Imposing such a restriction requires the follow-ing assumption:

11_{Johansen, Leibowitz, and Waite,}₁₉₉₆_{faced a similar problem in that 14% of mothers reported}

(22)

Chapter 2. Data 14

Assumption 2 The parent living in municipality r chooses a day-care center in

municipal-ity r, where r∈ S.

(23)

15

3 Methodology

This chapter provides an overview of the mathematical foundation behind discrete choice models. Based on the general framework of Additive Random Utility Models (ARUM), the multinomial logistic models are used for explaining choices of day-care centers. First, the mathematics behind ARUM is covered. Then, the theory behind multinomial logistic models is explained by specifically considering Condi-tional Logit models.

3.1 Theoretical framework

3.1.1 Additive Random Utility Models

Under the framework of ARUM, the utility of parent i choosing between different day-care centers j is decomposed into a deterministic component and a random com-ponent. Let the utility of parent i choosing day-care center j be denoted by Uij. The utility that parent i chooses option j is given by

Uij =Vij+eij, j=1, ..., m, (3.1) where Vijdenotes the deterministic component of utility and eijthe random compo-nent of utility. From the perspective of an economist, the parent selects the option which yields the highest utility (Mas-Colell, Whinston, Green, et al.,1995). The util-ity of parent i is not observed for each option. Rather, the option picked by the parent is observed, which is denoted by yij =1 if parent i chooses option j. The probability pij that parent i chooses day-care center j is given by

pij =P[yij =1] =P[Uij ≥Uik,∀k6= j] =P[˜eikj ≤Vij−Vik,∀k6=j], (3.2) where ˜eikj = eik−eij is the error difference with respect to reference alternative k for parent i choosing option j. In case of m alternatives, the expression for the probability that parent i chooses option j is an m−1-variate integral. Suppose we are interested in the probability that parent i chooses option 1, pi1. The expression for pi1is given by pi1= Z V_i1−V_i2 −∞ Z V_i1−V_i3 −∞ · · · Z V_i1−V_im

−∞ f(˜ei21, ˜ei31, ..., ˜eim1)d˜ei21˜ei31... ˜eim1. (3.3)

3.1.2 Conditional Logit models

The expression for the probability in Equation 3.3 becomes computationally tractable when the errors eij are independent and identically extreme value. Mathematically, the probability distribution of the error terms are assumed to be given by

f(eijk) =e−eije−e

−_eij

(24)

Chapter 3. Methodology 16 for i= 1, ..., N and j = 1, ..., m. This distribution has the convenient feature that the cumulative density of the difference between two error terms is given by the logistic function, i.e.

F(˜eijk) = e˜eijk

1+e˜eijk. (3.5)

It turns out that after some algebraic computations, the expression for the probability that parent i chooses day-care center j simplifies to

p_ij = e

Vij

∑m k=1eVik

. (3.6)

A Conditional Logit model allows alternative-specific regressors in the specification for the deterministic term Vijin Equation 3.1. Equation 3.1 can be written as

Uij =γj+x0ijβ+eij, (3.7) with xij being a K×1 vector of alternative-specific variables and β being a K×1 vector of parameters. γjis the intercept corresponding to option j and can be inter-preted as an average unobserved effect on utility. The alternative-specific variables in xijare variables related to day-care center j. Variables included in xijare distance, the maximum number of allowed childcare spots, indicators for after-school care center on the same address and primary school, and the number of collaborations with primary schools.

In order to capture heterogeneous effects across parents, the specification in Equa-tion 3.7 is modified by adding interacEqua-tion terms. In its most general form, the utility that parent i obtains from choosing day-care center j is given by

Uij =γj+x0ijβ+ (xij×zi)0α+eij, (3.8) with zi denoting an individual-specific characteristic (regarding the parents or the child) and with α being a J×1 vector of parameters.

Estimation of parameters is done by means of Maximum Likelihood. Since no closed-form solution exists, the estimates are calculated numerically using the Newton-Raphson algorithm. More about estimating paramaters of a Conditional Logit model can be found in Appendix B.

Odds ratios

A natural representation of interpreting estimated coefficients ˆβ is in terms of odds ratios. The odds of event A are defined as the probability of that event happening divided by the probability of the event not happening:

odds= P(A)

1−P(A). (3.9)

(25)

Chapter 3. Methodology 17 center j given that x stays the same is then given by

odds(x) = P[yij =1|x not incremented]

1−P[yij =1|x not incremented]

. (3.10)

Suppose x is incremented by 1. The odds are subsequently given by odds(x+1). The odds ratio is the odds if the corresponding variable x is incremented by 1 di-vided by the odds if variable x is not incremented:

OR= odds(x+1)

odds(x) . (3.11)

Plugging in the expression for the odds as defined in Equation 3.10 and simplifying, we obtain OR= P[yij=1|x+1] 1−P[yij=1|x+1] P[yij=j|x] 1−P[yij=1|x] =eβx_, _(3.12)

where βxis the parameter corresponding to explanatory variable x. The odds ratio can be interpreted as the effect of a 1 unit change in x on the predicted odds ratio, keeping other factors constant. Suppose x=DISTij. The odds ratio is then the odds of choosing day-care center j when the distance to day-care center j remains the same divided by the odds when the distance to day-care center j is incremented by 1. An estimated parameter of ˆβDIST_ij = 0.3 implies that the relative odds of choosing day-care center j subsequently decrease by(1−e0.3_{) ×}_100%₌_{35% when the distance to} day-care center j increases by 1 km. The expression in 3.12 can be written in a more general way as OR= P[yij=1|x+∆x] 1−P[yij=1|x+∆x] P[yij=j|x] 1−P[yij=1|x] =eβx×∆x_, _(3.13)

where∆x is the increment in x. In case x = DISTij again, this expression allows us to calculate odds ratios for different increments in distance to day-care center j. Confidence intervals for odds ratios are constructed based on a symmetric interval. They are calculated using

exp(ˆβ_k±zα/2× q

Var(ˆβ_k)), (3.14) with zα/2 being the (1−a/2)th percentile of the standard normal distribution and Var(ˆβ_k)being the variance of ˆβk(Dietz et al.,2002).

Willingness to travel

Another useful interpretation of estimated coefficients is in terms of willingness-to-travel. Suppose utility is specified as Uij = γj+β1DISTij+β2BSOj+eij. The ratio β2/β1represents the willingness to travel for a day-care center with an after-school care center on the same address. If β1and β2are estimated as -0.4 and -0.1 respec-tively, then parents are willing to travel an additional−₋0.1_0.4 = −0.25 km for a day-care center without an after-school care center on the same address.

Let a point estimate for the willingness to travel be denoted by dWTT= ˆβk

(26)

Chapter 3. Methodology 18 denoting the estimate for variable k and ˆβDISTthe estimated parameter for distance. A confidence interval for dWTT is given by

d

WTT±zα/2× q

Var(WTTd ), (3.15)

where Var(WTTd )is a function of ˆβDIST, ˆβ_k, the estimated variances of ˆβDIST and ˆβ_k and estimated covariances of ˆβDIST and ˆβ_k(Gatta, Marcucci, and Scaccia,2014).12

Adding interaction terms

In case interaction effects are incorporated in the deterministic component of util-ity, the interpretation of odds ratios and willingness to travel is slightly different. Suppose utility is specified as Uij = γj+β1xij+β2(xij×zi) +eij. The odds ratio, defined as the odds that xijis incremented by 1 divided by the odds that xijstays the same, is then given by

OR= eβxij+α_xijzizi

, (3.16)

where αxijzi is the parameter corresponding to interaction xij×zi.

Suppose Uij = γj+β1DISTij +β2BSOj+α(BSOj ×AGEi) +eij. The willingness to travel does not only depend on the estimated coefficients β1and β2, but also on the AGEi variable. Now the ratio−(β2+αAGEi)/β1 represents the willingness to travel if a day-care center has an after-school care center on the same address.

Predictions

Conditional Logit models can be used for predicting the choice for a day-care cen-ter. Suppose Uij = x0ijβ+ (xij ×zi)0α+eij. The predicted probability that parent i chooses day-care center j is then given by

ˆpij = ex0ijβˆ+(xij×zi)0ˆα ∑m k=1ex 0 ikβˆ+(xik×zi)0ˆα . (3.17)

A popular way to predict the number of children choosing day-care center j is by means of sample enumeration (Train,2009). Denote the actual number of children attending day-care center j by Sj. An estimate of the total number of parents choos-ing day-care center j is then given by

ˆ Sj =

∑

i

wiˆpij, (3.18)

where we set wi = 1 for all i. The predictive accuracy is evaluated based on the Mean Absolute Error (MAE). Mathematically, the MAE is defined as

MAE= AE

m , (3.19)

where AE=_∑m_j₌₁|Sj−Sˆj|.

(27)

19

4 Empirical Analysis

This chapter focuses on estimating the Conditional Logit model and the analyses of findings by looking at estimated coefficients, odds ratios and the willingness to travel. Furthermore, interaction effects are included to incorporate heterogeneous effects among children or parents. This chapter consists of five parts. First, the whole sample is estimated. This will be followed by the estimation of the models for children in Groningen and Westerkwartier in the second and third part, respectively. Fourth, a comparison is made in the effect of distance across municipalities. Lastly, the estimated models will be put into practice by making counterfactual predictions for a new day-care center.

4.1 Full sample

As mentioned above, the full sample will be estimated first. Estimates for the full sample are found in Table A.6, based on the specification as in Equation 3.7. Different models have been specified for each year in which parents choose a day-care center. In the last column, estimates are provided for all years. The estimated coefficients in Table A.6 are not directly interpretable in terms of a marginal effect. However, negative signs indicate a negative effect on the probability of choosing a day-care center; a positive sign indicates a positive effect. Furthermore, the higher the esti-mated coefficient in absolute value terms, the stronger the effect on the probability. In columns (1)-(5) we find a negative significant effect of distance on the probability of choosing a day-care center. The longer the distance to a day-care center, the lower the probability of it being chosen, which is not a surprising result. Furthermore, we can see that the effect that distance has on the probability of choosing a day-care center becomes stronger over the years. Where the coefficient for distance is esti-mated -0.528 in 2016, it has decreased to -0.607 in 2019. A possible reason for this is that more day-care centers have opened over time, meaning distances to the clos-est day-care centers have decreased. The positive clos-estimated coefficient for childcare spots implies that a day-care center with a higher maximum number of childcare spots has a higher probability of being chosen. This is simply because day-care cen-ters with more childcare spots allow more children. When comparing estimates of day-care centers located in primary school and day-care centers with an after-school care center on the same address in columns (1)-(5), we only find significant effects in 2016 and the "full" model. The effect that day-care centers in primary schools have on the probability of choosing such day-care centers is negative for parents that chose a day-care center in 2016 or before. Furthermore, day-care centers with an after-school care center on the same address have a positive effect on the probabil-ity of choosing such centers. This might be because it allows for an easier transition to an after-school care center when the child reaches the age of four.

(28)

Chapter 4. Empirical Analysis 20 The odds ratios are calculated based on Equation 3.12. The estimated odds ratio for distance can be interpreted as follows: if distance to day-care center A increases by 1 km, the relative odds of choosing day-care center A decrease by 41% up to 49%. Roughly speaking, this means that if the probability of choosing day-care center A would be 10%, then the probability decreases to almost 5% if day-care center A would be located 1 km farther away.

The effect of the increase of childcare spots of a day-care center by 1 is repre-sented by an increase in relative odds of choosing that center by approximately 1.5%. The "bigger" the day-care center, the more likely it is to be chosen, but this does not necessarily mean that bigger day-care centers are more preferred by parents. The last two rows measure the change in odds ratios when the day-care center has an after-school care center on the same address, or is located in a primary school. In 2016 (or before), the relative odds of day-care centers with an after-school care cen-ter increase by 18.3%, compared to day-care cencen-ters without an afcen-ter-school care center. In a similar way, we find that the odds of choosing day-care centers located in primary schools decrease by 17.1% compared to day-care centers not located in primary schools. No significant effect is found in the years 2017-2019.

Estimated willingness to travel and 95% confidence intervals are tabulated in Ta-ble A.8. For a day-care center where the maximum number of childcare spots is increased by 1, a parent is willing to travel an additional 30 m. For all years, par-ents are willing to travel an additional 90 to 320 m for a day-care center with an after-school care center on the same address. For day-care centers located in pri-mary schools chosen in 2016 or before, a parent is more likely to choose such centers if the distance decreases by 350 m or more, on average. This effect is not persistent over time: in the years 2017 and 2018, an average customer is willing to travel an additional 80 to 90 m for a day-care center located in a primary school. However, these effects are small.

4.2 Subsample Groningen

The model estimated in A.6 is based on the full sample: no distinction is made in different municipalities the parents live in. In this subsection, models for children living in the municipality Groningen are estimated.

Estimates are found in Table A.9. Compared to estimates based on the whole sample in Table A.6, we see no surprising differences in terms of signs in front of the esti-mated coefficients. Compared to the whole sample, several differences in estiesti-mated coefficients exist. First, the estimated coefficient for distance in Table A.9 reflects a stronger effect of distance on choice for a day-care center. The reason for this is that, compared to day-care centers in Westerkwartier, there are more of them in a smaller area and people live closer to the nearest day-care center. Furthermore, we can see that the effect of distance to a day-care center on the probability of choosing it be-comes stronger over the years. The effects of day-care centers located in primary schools and/or with an after-school care center on the same address also increased for the choices made in 2016, and are also significant in 2018 and 2019.

(29)

Chapter 4. Empirical Analysis 21 center A would be chosen with high probability (80%), and if then that day-care cen-ter moves to another location 1 km farther away, this probability decreases to less than half. Note that the relative odds of choosing a day-care center with an after-school care center on the same address are positive for all years and that the relative odds of choosing a day-care center located in a primary school are negative.

Estimated willingness to travel is tabulated in Table A.11. Increasing the number of childcare spots by 1 implies that parents are willing to travel an additional 20 m. For day-care centers with an after-school care center on the same address, parents are now willing to travel an additional 120 to 250 m. The effect of willingness to travel on day-care centers not located in primary schools is stronger: parents are willing to travel an additional 140 to 460 m for such day-care centers.

4.2.1 Individual-specific heterogeneous effects

(30)

Chapter 4. Empirical Analysis 22

4.3 Subsample Westerkwartier

Continuing with the subsample of Westerkwartier, the estimation results are shown in Table A.13. Because of singularity issues (i.e. multicollinearity) in calculating the Hessian matrix, the alternative-specific coefficients γj have not been incorporated in these models. Compared to parents in the municipality Groningen, the nega-tive effect that distance has on choice of day-care center is weaker. This is likely due to the differences in the number of day-care centers that exist in the municipal-ity Groningen compared to Westerkwartier. SKSG provides less day-care centers in Westerkwartier compared to Groningen. This means that it is very plausible that the closest distance to a day-care center for an average parent in Westerkwartier is greater compared to the closest distance to a day-care center for a parent in Gronin-gen.13 Interestingly, the effect of distance on the choice becomes stronger over the years the day-care center is chosen. This means that the effect that distance to a day-care center has on the choice becomes weaker. This could possibly suggest the establishment of more day-care centers. In the chosen years 2016 and 2019, day-care centers with an after-school care center on the same address have a very high prob-ability of being chosen. This is simply because the total capacity that these day-care centers offer is much higher compared to day-care centers without an after-school care center (see Figure 2.5a).

Estimated odds ratios are found in Table A.14. If distance to day-care center A in-creases by 1 km, then the relative odds of choosing that day-care center decrease by 40% to 45%. Predicted odds ratios for day-care centers with an after-school care cen-ter on the same address exceed 100% in 2016 and 2019. This indicates the probability of choosing a day-care center more than doubles when an after-school care center is established at that day-care center.

In terms of willingness to travel, we find that the average parent in Westerkwartier is willing to (or simply has to) travel an additional 0.51 to 1.96 km for a day-care center with an after-school care center on the same address, as Table A.15 shows.

4.3.1 Individual-specific heterogeneous effects

Estimation results with heterogeneous effects for parents in Westerkwartier are found in Table A.16. Similar to the findings reported for the municipality Groningen, we find that day-care centers in primary schools are preferred for parents with children aged two to three years. On the contrary, day-care centers with an after-school care center result in a higher probability of being chosen among young children (zero to one years old). There is no explanation yet for this result. Again, we can see that the effect of distance on day-care center is weaker for boys. We do not see significant interactions of the geographic characteristics with application times in column (3). For the full model, however, we find evidence that parents signing up early tend to prefer day-care centers without an after-school care center on the same address. Probably this has nothing to do with the fact that day-care centers have an after-school care center or not. Instead, the total daily capacity of these day-care centers is very low compared to day-care centers with an after-school care center on the same address, meaning that there is a higher probability that these day-care centers are

13_{Of course, the number of day-care centers provided by competitors also contribute to the closest}

(31)

Chapter 4. Empirical Analysis 23 occupied. This in turn would lead to parents signing up early for those day-care centers, which could explain the fact that these parents have a slight tendency to choose day-care centers without an after-school care center on the same address. Last but not least, we find that the effect of distance on parents with high income status is weaker compared to parents with low income status, implying that it is less important for such parents to have the day-care center close to their home address. Similar to the reasoning for children in municipality Groningen, these parents find it more important that day-care centers offer good quality, and are less concerned about distance.

4.4 The effect of distance across municipalities

To elaborate on the predicted odds ratios for distance in different municipalities, it is interesting to determine how the relative odds change when the change in distance to day-care centers is varied. The relative odds are calculated based on Equation 3.13, which allows for different changes in distance to a day-care center. Figure 4.1 presents the results. Not surprisingly, we see that the relative odds increase as a function of∆x but at a decreasing rate. Note the differences in relative odds across municipalities. When the distance to a day-care center is incremented by 2.5 km relative to its current location, the relative odds of choosing that day-care center decrease by 75% in Westerkwartier and 88% in Groningen. When∆x > 5, i.e. the day-care center is located 5 km farther away relative to its current location, then the relative odds are close to 100%, meaning the probability that such a day-care center is chosen is approaching 0%. These results suggest that the location of day-care centers is very important. If a day-care center is located 5 to 10 km farther away relative to its current location and there are no potential customers in a radius of 5 to 10 km of that day-care center then it is very likely that little to no children will attend this day-care center.

FIGURE4.1: Estimated relative odds of choosing a day-care center as a function of∆x, the increment in distance to a day-care center. Based

(32)

Chapter 4. Empirical Analysis 24

4.5 Policy evaluation: opening a new day-care center

It is of interest to evaluate what happens if SKSG opens a new day-care center K at a predetermined location.14 In this section, we make counterfactual predictions for a fictive day-care center for the years 2016-2019 as if it already existed in 2016, using the whole sample of children. This because it provides insight into the potential of that center.In order to do so, we first have to decide on the specification for the util-ity that parent i obtains from choosing day-care center j, Uij.15 Second, the model is estimated and used for predictions under the current situation. Predictions into a given year are made based on the set of children that went to a day-care center in that year. Third, the model is estimated under the new situation where the new day-care center is considered a viable option among already existing day-day-care centers. First, we have to decide on the specification for the utility that parent i obtains from choosing day-care center j, Uij. This specification, based on Equation 3.8, allows for interactions between location-specific characteristics and a single individual-specific characteristic. It is given by

Uij =xij0β+ (xij×zi)0α+e_ij, (4.1) where xijand ziare chosen as

xij =     DIST_ij ACSj BSO_j PRIMS_j     , zi =AGEi.

The main reason for using this specification is because 1) estimated parameters (col-umn (1) of Tables A.9 and A.13) are almost all significant and 2) we have reason to believe that after-school care center and primary school interact with the age of the child. Furthermore, note that the alternative-specific constant γj is not included. This is because, when estimating the probabilities of choosing the new day-care cen-ter for each child, we need an estimate of γK. An estimate of γK is not possible, because we do not have any observed outcomes for this new day-care center. Based on the MAE, the predictive accuracy of the model in each year is assessed. Table A.17 gives summarized statistics of the AE for every year. We see that the MAE and the variability of the error is high in 2016, compared to other years. In years 2017-2019, the MAE is approximately five to six. This means that on average, the observed number of children on a day-care center differs by about five or six from the predicted number of children. The main reason for this difference is that 7.7% of children choose multiple locations, and thus the observed numbers Sjare bi-ased downwards. Overall, the model does well in predicting the number of children on each location.

Next, we include the new day-care center in the choice set. We set the location-specific characteristics equal to ACSK = 32, BSOK = 0 and PRIMSK = 0. Predictive performance of the model with the new day-care center included are found in Table

14_{The coordinates of this new day-care center are not reported.}

15_{A better option is to estimate demand using different model specifications, evaluate performance}

(33)

Chapter 4. Empirical Analysis 25 A.18. Compared to predictions made in the current situation, the model performs equally or even better. The counterfactual predictions for the new day-care center can be found in Table A.19. In order to assess how these predictions compare with the observed number of children of existing day-care centers, several statistics are shown as well. In 2016, we see that the prediction is lower than the median, mean-ing more than half of the existmean-ing day-care centers have more children. In years 2017-2018, the number of children predicted for the new day-care center is higher than the median. The predicted number of children in those years is higher than the observed number of children, for more than half of the existing day-care centers. In 2019, the prediction is slightly lower than the median. This means that for more than half of the day-care centers the number of children is higher than the predicted number of children for the new center in that year.

(34)

26

5 Discussion

This thesis has researched how geographical characteristics of day-care centers affect the choice for a particular day-care center. In addition, a practical method for estab-lishing new day-care centers is carried out by making counterfactual predictions for a fictive day-care center. The research question is answered by means of four sub questions.

First, it aims to provide an answer to what extent distance to day-care center affects the choice for a day-care center. Using common sense, it is plausible to say that a day-care center located farther away is less likely to be chosen. The extent to which distance plays a role in choosing day-care centers is interpreted in terms of relative odds: if the distance to a day-care center A is increased by 1 km, the relative odds of choosing that day-care center decrease by 41 to 49%. This means that there is a higher probability of selecting other day-care centers than A. Therefore, this result is only applicable for choices between day-care centers: nothing can be said about how this affects the probability of selecting other modes of care, such as care at home or care at someone else’s home.

Since the data set only contains children choosing day-care centers provided by SKSG, it is questionable whether the extent to which distance plays a role in choosing day-care center is applicable for all day-care centers. Camasso and Roche (1991) argue that parents reported convenience to be the main factor in explaining choice for a day-care center. Using this information, it is argued that the extent to which distance plays a role in choosing day-care centers provided by SKSG can be partly generalized to choices between all day-care centers. This can only be generalized to some extent, because besides convenience, it is also argued that other factors are important. For example, price differences between childcare organizations play a role in choosing day-care centers (Davis and Connelly, 2005). Due to the fact that only data is acquired from one particular childcare organization, price differences could not be taken into account.

The extent to which distance plays a role varies with individual-specific charac-teristics of the child, or the parent. Surprisingly, these variations do not give consis-tent results across municipalities. In Groningen, it is found that the older the child, the less it is affected by distance to day-care center. In Westerkwartier, the older the child, the more it is affected by distance to day-care center. There is no clear-cut rea-son for these differences across municipalities. One would expect that older children (two to three years old) are less affected by distance to day-care center, because they prefer a day-care center located in a primary school. Because there are not many of such day-care centers yet, these children are likely to cover a bigger distance to attend such locations.

(35)

Chapter 5. Discussion 27 The differences in preferences for children of different ages can be explained as fol-lows. Parents with children aged zero to one years old might not yet think about the primary school the child will attend when he/she reaches school age.

Third, it is investigated how day-care centers with an after-school care center on the same address affect the choice for a day-care center. It is hypothesized that day-care centers with an after-school care center on the same address are preferred, because it allows easier transitioning from the child to the after-school care center. From the results it is clear that day-care centers with an after-school care center on the same address are chosen more often. This is likely due to higher total daily ca-pacity of these locations. However, parents with older children are more likely to choose day-care centers with an after-school care center on the same address. It can be argued that this is because of additional convenience. When the child reaches school age, he/she can subsequently move on to the after-school care center.

There are several limitations regarding this study. First, the sample contains only children selecting day-care centers provided by a specific childcare organization. Therefore, assumption 1 might only be partly true. It might not be entirely correct, because parents select day-care centers based on their distance to day-care centers and thus select out of all available day-care centers within a certain boundary (Ca-masso and Roche,1991). On the other hand, the process of choosing a day-care cen-ter is also affected by price, which is a good reason why the assumption might hold (Davis and Connelly,2005). For future research, it would therefore be recommended to have a data set containing children for multiple childcare organizations, so that other factors, such as prices, can be taken into account as well. A second limitation is that it was a mistake to exclude day-care centers from the choice set that did not exist anymore in 2020. These day-care centers were valid options to choose from when they were not closed yet. The reason why these options could not be added any-more, is because of time constraints. However, exclusion of these day-care centers is likely not to produce a lot of biases in the estimates. This is because the children having chosen those day-care centers only represent 1.5% of the sample. A third limitation is that only three geographic characteristics of locations are included in the analysis. For future research, it is therefore recommended to identify additional location-specific characteristics explaining the choice for a day-care center.

(36)

Chapter 5. Discussion 28 What this thesis contributes with its findings is that it can be used for establish-ing new day-care centers, for SKSG and other childcare organizations. Preferences in choice of day-care centers differ across individuals and this difference manifests itself in various ways. It is therefore hard to establish day-care centers that meets the wishes of all customers. Nonetheless, the results in this thesis clearly show that, when positioning a new day-care center, it is important to ensure enough children fall within a certain radius of that day-care center. This is, of course, because dis-tance to a day-care center highly affects the choices made by parents. Furthermore, the more day-care centers in a certain region, the more distance to day-care center plays a role, and the lower the potential customers.

(37)

29

6 Conclusion

This thesis examined how location-specific characteristics of day-care centers con-tribute to the choices customers make between day-care centers. The following fac-tors were used: distance to a day-care center, whether a day-care center is located in a primary school or not, and whether a day-care center has an after-school care cen-ter on the same address. Furthermore, this thesis examined hecen-terogeneity in choice behavior across parents or children with different individual-specific characteristics. Differences between day-care centers having a location-specific feature or not were compared. Furthermore, Conditional Logit models have been estimated. A distinc-tion has been made between the regions Groningen and Westerkwartier to incorpo-rate regional differences in the amount of day-care centers. Lastly, counterfactual predictions are made for a fictive day-care center.

In general, the box plots showed that differences in location-specific features of day-care centers do not necessarily imply differences in choice behavior regard-ing day-care centers with specific features. However, day-care centers in primary schools showed to be more occupied in region Westerkwartier, suggesting a possible preference towards day-care centers in primary schools. There is a huge difference in average distances to day-care center for municipalities Groningen and Westerk-wartier, which is partly explained by the number of available day-care centers to choose from in each region.

Even though location-specific characteristics might not be the main factor for choosing a day-care center, this thesis did show that distance is of great impact. To answer the first sub question, distance affects the choice of day-care center in the following ways. First, considering the whole sample, the relative odds of choosing a day-care center that is located 1 km farther away decreases by 41 to 49%. There are, however, significant differences in the magnitude of the effect of distance on choice of day-care center across the two municipalities. The more day-care centers in a re-gion, the stronger the effect of distance to day-care center on the choice becomes. Furthermore, the negative effect of distance on the choice of a day-care center has increased over time, suggesting the establishment of more day-care centers over the years. Considering heterogeneous characteristics of parents or children, the effect of distance on the choice of day-care center is different. Parents for which the child is a boy find distance to day-care center less important compared to parents where the child is a girl. And parents with high income status find distance less impor-tant compared to parents with low income status, suggesting that these parents find quality and/or development of the child more important.

(38)

Chapter 6. Conclusion 30 schools more than parents registering relatively late.

(39)

31

A Tables

TABLEA.1: Variables and their description Variable Abbreviation Definition

Location-specific variables

Childcare spots ACS The maximum number of children allowed to stay in a day-care center on a daily ba-sis.

After-school care center BSO Dummy variable indicating if the day-care center also has an after-school care center on the same address of the day-care center.

Primary school PRIMS Dummy variable indicating if the day-care center is located in a primary school.

Child-specific variables

Gender SEX Variable indicating gender of the child.

Age AGE Variable indicating age of the child.

Application time APPLT Variable indicating the time (in days) between the first contact with SKSG and the actual starting date of the child on a day-care center. Income INCOME Dummy variable equalling 1

if the child lives within a high-income household, and 0 if low-income.

(40)

Appendix A. Tables 32

TABLE A.2: Descriptive statistics of features of day-care centers in municipality Groningen, aggregrated over years 2016-2019.

Statistic N Mean St. Dev. Min Pctl(25) Pctl(75) Max Childcare spots 35 44.286 24.718 8 23 55.5 100 After-school care center 35 0.571 0.502 0 0 1 1

Primary school 35 0.371 0.490 0 0 1 1

TABLE A.3: Descriptive statistics of features of day-care centers in municipality Westerkwartier, aggregrated over years 2016-2019.

Statistic N Mean St. Dev. Min Pctl(25) Pctl(75) Max Childcare spots 12 36.917 26.763 16 16 45.5 100 After-school care center 12 0.500 0.522 0 0 1 1

Primary school 12 0.167 0.389 0 0 0 1

TABLEA.4: Descriptive statistics of children in the data set.

Statistic N Mean St. Dev. Min Pctl(25) Pctl(75) Max Male 7,163 0.514 0.500 0.000 0.000 1.000 1.000

Age 7,167 1.411 1.122 0 0 2 4

Application time 2,851 142.086 99.466 1.000 45.000 231.000 365.000 Distance 7,616 3.539 11.361 0.001 0.731 2.867 330.263 Income 6,206 0.303 0.460 0.000 0.000 1.000 1.000

TABLEA.5: The number of children and % of children living in mu-nicipality Groningen, Westerkwartier or another one choosing a day-care center in respectively Groningen and Westerkwartier. % of

chil-dren are calculated over the municipality the child lives in. Groningen Westerkwartier Children % of Children Children % of Children

Groningen 4,930 99.7% 17 0.3%

Westerkwartier 63 3.5% 1,712 96.5%

(41)

(42)

Appendix A. Tables 34

TABLEA.7: Predicted odds ratios in % change (95% CI in parentheses)

Year 2016 2017 2018 2019

Distance -41.0% -45.8% -48.6% -45.5%

(-42.2%,-39.8%) (-47.6%,-43.8%) (-50.6%,-46.6%) (-47.4%,-43.6%)

Childcare spots 1.7% 1.7% 1.5% 1.5%

(1.5%,1.8%) (1.4%,1.9%) (1.2%,1.8%) (1.2%,1.8%) After-school care center 18.3% 5.7% 14.2% 6.9%

(7.3%,30.3%) (-10.2%,24.5%) (-3.3%,34.9%) (-9.0%,25.6%)

Primary school -17.1% 5.2% 6.0% -3.5%

(-26.2%,-6.7%) (-13.3%,27.6%) (-12.8%,28.9%) (-20.1%,16.6%)

TABLEA.8: Predicted willingness to travel in km (95% CI in paren-theses).

Year 2016 2017 2018 2019

Childcare spots 0.03 0.03 0.02 0.02

(-0.01,0.07) (-0.03,0.08) (-0.04,0.08) (-0.03,0.08) After-school care center 0.32 0.09 0.20 0.11

(0.25,0.39) (0.02,0.15) (0.12,0.28) (0.05,0.17)

Primary school -0.35 0.08 0.09 -0.06

(43)

Choosing day-care centers by location-speciﬁc characteristics: a Conditional Logit model approach

M

T

Choosing day-care centers by

location-specific characteristics: a

Conditional Logit model approach

Abstract

Acknowledgements

Contents

List of Figures

List of Tables

1 Introduction

1.1

Research question, contribution and hypothesis

1.2

Daycare choice in the Netherlands

2 Data

2.1

Data collection

2.2

Data description

2.3

Restrictions on the choice set

3 Methodology

3.1

Theoretical framework

∑

4 Empirical Analysis

4.1

Full sample

4.2

Subsample Groningen

4.3

Subsample Westerkwartier

4.4

The effect of distance across municipalities

4.5

Policy evaluation: opening a new day-care center

5 Discussion

6 Conclusion

A Tables