An economic evaluation of the school nurse resource in Norwegian primary schools
Mid way results from a randomized controlled trial
Magnus Hodnekvam May 2019
Master thesis in Economics
Department of Economics, University of Oslo
Acknowledgements
This thesis was supervised by Martin Flatø. This thesis was written in collaboration with the Nordic Institute for Studies of Innovation, Research and Education (NIFU). The data analyzed are collected from a randomized controlled trial titled ”Et lag rundt eleven”.
The randomized controlled trial is financed by the Norwegian Directorate of Education.
I would like to thank my supervisor Martin Flatø for good discussions and helpful advice. Thank you Ester Rønsen and Karin Vaagland for help with data administration, discussions, and lunch breaks. I would also like to thank the project leader, Roger Andr´e Federici, for welcoming me at NIFU and letting me be a part of the project.
Abstact
In this thesis, I motivate the school nurse as an input in the education production function.
My research question is whether increasing the school nurse resource can improve the students learning environment. Previous literature show that school leadership, teachers, bullying, and absence are associated with academic performance. I present mid way results from a randomized controlled trial, where treatment schools get funding to increase the school nurse resource by 12.5% of a full time school nurse position, and the resource should collaborate with the school to implement measures to improve the school learning environment, targeting students attending 5th to 7th grade. I test four hypotheses. Can an increased school nurse resource reduce (1) bullying and (2) absence, and increase (3) emotional well being and (4) school belonging? If yes, investing in the school nurse resource could yield a good return on investment as an input in the education production function. The estimated results does not find a significant effect of increasing the school nurse resource on any of the four measures. Also, I do not find any moderation by age or gender for emotional well being, school belonging or bullying, and no moderation by age for absence. Thus, mid way results from the experiment does not support the school nurse as an important input in the education production function. Calculations and estimations were done using Stata.
Contents
1 Introduction 1
2 The school nurse resource as an Input: The Education Production Func-
tion 4
2.1 The Education Production Function. . . 5
2.1.1 Including the SNR in the EPF . . . 6
2.2 Literature review . . . 9
2.2.1 Teacher effects on academic performance . . . 9
2.2.2 School leadership effects on academic performance . . . 10
2.2.3 The effect of absence on academic performance . . . 10
2.2.4 The effect of bullying on academic performance . . . 11
2.2.5 Previous studies of school nurses in primary school . . . 12
2.2.6 School nurse effects on the other inputs in the EPF . . . 13
2.2.7 Grades as predictor . . . 14
2.3 Moderation . . . 15
2.3.1 Moderation by age . . . 15
2.3.2 Moderation by gender . . . 16
3 Data description and Methodology 17 3.1 Sample selection and randomization . . . 17
3.2 Data Description . . . 18
3.2.1 The Pupil Survey . . . 18
3.2.2 Absence . . . 23
3.2.3 Coverage . . . 25
3.3 Methodology . . . 29
3.3.1 Estimation Strategy . . . 29
3.3.2 Randomized Controlled Trials . . . 30
3.3.3 Clustered Standard Errors . . . 32
3.3.4 Instrumental Variables . . . 33
3.3.5 The Local Average Treatment Effect . . . 36
3.3.6 Multiple Hypothesis testing . . . 38
4 Results 39
4.1 Balance test . . . 39
4.2 First Stage Estimates . . . 40
4.3 LATE estimates . . . 41
4.3.1 Emotional well-being . . . 42
4.3.2 School belonging . . . 43
4.3.3 Bullying . . . 44
4.3.4 Absence . . . 46
4.3.5 Multiple Hypothesis Testing . . . 47
5 Discussion and conclusion 48 5.1 Interpreting the results . . . 48
5.1.1 Relating the results to the literature review . . . 49
5.1.2 LATE assumptions . . . 50
5.2 Issues related to RCTs . . . 50
5.3 Conclusion . . . 51
List of Tables
3.1 Construction of emotional well-being . . . 20
3.2 Construction of school belonging . . . 22
3.3 Variable means . . . 28
4.1 Balance tests . . . 39
4.2 2SLS estimates for emotional well-being . . . 41
4.3 Sub group estimates for emotional well-being . . . 43
4.4 2SLS estimates for school belonging . . . 44
4.5 Sub group estimates for school belonging . . . 45
4.6 2SLS estimates for Bullying . . . 45
4.7 Sub group estimates for Bullying . . . 46
4.8 IV Poisson GMM estimates for Absence . . . 47
List of Figures
3.1 Distribution of emotional well-being . . . 19
3.2 Distribution of school belonging . . . 21
3.3 Distribution of absence . . . 24
3.4 Histograms of the student to nurse ratios . . . 26
3.5 Scatter plot of student to nurse ratios . . . 27
Chapter 1 Introduction
Academic performance in primary school is an important predictor for wheter individ- uals complete upper secondary school, as well as how connected individuals will be to the labour force later in life (Falch, 2010). Policy makers are therefore continuously in- terested in evaluations of interventions that can be regarded as inputs in the education production function (EPF). The EPF constitutes all combinations of inputs that produce an output, and the most common output is academic performance, measured with test scores. One such input is the school health service, or more specifically the school nurse.
Maughan (2016) argue that school nurses helps keeping students healthy and ready to learn. Attempting to estimate the value of school nurses, Baisch et al. (2011) call for quality evaluation data to justify that investing in school nurses improve school environ- ments. Thus there is clear motivation to answer questions such as, can an increased school nurse resource (SNR) increase student emotional well-being and school belonging? Can an increased SNR reduce student absence and bullying? These questions are interesting because, as I show in my literature review, bullying, absence, and school environment are all associated with academic performance. In this thesis, I discuss the school nurse as an input in the EPF. I analyze mid-way results from a cluster randomized controlled trial (RCT) investigating the causal effect of an increased SNR on multiple student measures.
The measured analyzed are emotional well-being, school belonging, bullying, and absence.
Motivating the school nurse as an input in the education production function The EPF has been of great economic interest for decades (Brewer et al., 2010). Investment in education can yield high returns, both at the individual level and for society. Research has shown that both income and health are strongly correlated with an individual’s level of education (Johnson, 2010; Griliches and Mason, 1972).
In addition to learning course material, the school is an important arena for gaining social skills and interacting with others. Psycho-social environment and mental health are important factors for the learning environment. If a student does not feel welcome or
included at school, it will be more difficult to interact with others, participate in group work and learn in collaboration with peers. If a student does not feel well emotionally, or a student experience bullying, it may affect performance and test scores (Oliveira et al., 2018; Strøm et al., 2013). It may as well have negative effects on the student’s mental health, causing issues such as depression later in life, and affecting performance in the labour market (Drydakis, 2014). This could be very costly for society, and thus it is of great interest to reduce occurrences of bullying in the education system. Students with high level of absence are likely to have lower test scores than students with no absence (Smith and Sherrod, 2013). We want a good learning environment in Norwegian schools, and students that perform well academically. If the SNR can reduce bullying and absence, and improve the schools learning environment, investing in the SNR could prove to be a low cost, effective input in the EPF with a broad target group.
Context of the randomized controlled trial
The project is a clustered RCT where four schools in each of the 12 participating munici- palities were randomly selected to receive a 12.5 % of a full time position increase in their SNR from January 2018 to December 2019. For my analysis, data for both the spring and fall semester of 2018 has been collected. The school nurses are required to work in structured and systemic collaboration with the school and the resource was targeted towards students in 5th to 7th grade. The main hypothesis is that an increased nurse-to- student (NTS) ratio will improve students self-reported measures related to psycho-social aspects of the learning environment and reduce student absence. Norwegian municipali- ties are obliged to provide health station services and school health services to individuals aged 0-20 years. The main purpose of these services is to promote a good environment for children’s physical and psycho-social development through easy access. The school health service is intended to work with schools and other health services to identify and solve health challenges among students. The school nurses are the main provider of these services. To become a full time school nurse, you must be a certified nurse with a spe- cialization in preventive and health promoting work. The activities of the school nurse mainly consist of preventive work or health promotion work, such as providing vaccina- tion, providing information on health issues, offering counseling to individuals or groups, as well as referring children to other health services when necessary. The Norwegian school health services are organized at the municipality level. School nurses typically work at the school’s premises. School nurses may work at several schools and may combine her/his position between activities at the health station and at schools. There is a national norm of minimum one full time nurse per 300 students in primary schools. This norm is not binding, and only 2.5 % of Norwegian primary schools adhered to the norm in 2015. This thesis is a part of the research project ”Et lag rundt eleven”. The RCT is conducted by the Nordic Institute for Studies Innovation, Research and Education (NIFU). The RCT
is funded by the Norwegian Directorate for Education (UDIR). In my analysis I follow a procedure described in the projects protocol (Federici et al., 2017). I describe the method in detail and provide a discussion of the results presented.
Structure of the following chapters
The rest of the thesis is structured as follows. In chapter 2, I present the theoretical concept of the EPF, with the school nurse as an input. Then, I conduct a literature review to substantiate some of the assumptions made in my model. In chapter 3, I describe the data and the econometric methodology used to analyze the RCT. In chapter 4, I present the results from my analysis. In chapter5, I discuss my results in light of the theory presented in chapter 2, and conclude the thesis.
Chapter 2
The school nurse resource as an Input: The Education Production Function
The primary outcomes measured in the RCT are the subjective student measures emo- tional well-being (EWB) , school belonging (SB) and bullying, and the objective measure days of absence. It is reasonable to assume that these outcomes are indicators of how the students will proceed later in life, as well as how students perform on test scores. For example, high absence indicates that you are more likely to gain low grades (Smith and Sherrod, 2013), and thus the probability of failing courses and spending more time in the educational system increases. Bullying can cause psychological issues affecting a person’s ability to contribute to the work force later in life (Samdal (2009), Falch (2010)). Stu- dents that are bullied perform below average on test scores (Oliveira et al., 2018). EWB is also important for how students perform (Samdal, 2009). As with bullying, stress and anxiety can reduce test scores as well as the student’s takeaway from the school setting.
If a student does not have a good sense of SB, meaning that the student does not feel included or welcome at school, or find it hard to make friends, the student may suffer depression, loneliness or similar affecting academic performance, social skills and other traits that are important later in life. If an increased SNR reduces bullying and absence, and/or increases EWB and SB, it might be a good return on investment as an input in the EPF. The EPF constitutes all combinations of inputs that produce a given set of outputs (Harris, 2010)). Test scores and graduation rates typically measure the outputs.
Achievement is a very important measure when evaluating the EPF. I will not present estimates with test scores in this thesis, but in the final evaluation of the RCT a measure of academic achievement constructed by test scores on national tests will be used to es- timate the relation between the SNR and Academic performance. Still, in this chapter I discuss academic achievement as the output. I present a mathematical framework for the
EPF, and consider the SNR and how it affects absence, teacher quality, school leadership and psycho-social environment as inputs in the EPF. Then, I present previous literature that substantiates the assumptions made about the mechanisms at work in the EPF.
2.1 The Education Production Function
I present a theoretical framework for the EPF. This framework includes a mathematical formulation of the EPF along with the assumptions needed for the model to hold. The output of the EPF is increased academic performance. This can be measured through test scores.
Mathematical formulation and mode assumptions
The first assumption that must be in place, is that all available inputs in the EPF are used efficiently. Assuming that the efficiency assumption holds, the EPF can be described as a function f. According to Harris (2010) the EPF can then be written as
Qit =f(Sit, Sit−1, ..., Fit, Fit−1, ..., Ii, it) (2.1) Qit represents individuals i measured academic performance at time t. This outcome depends on several inputs. The first assumption Harris (2010) makes about the EPF is that these inputs can be split into two groups, school inputs (Sit) and family inputs (Fit). Examples of family inputs are parent’s education, parent’s income and number of siblings. Examples of school inputs are teacher education, teacher experience, student- to-teacher ratio, resources (books, computers, quality of building etc). In this case, the input of interest (the SNR) belongs in the school input category. The school nurse works as part of the school health service to better EWB and SB, as well as reducing bullying and absence. The EPF also take inputs from previous time periods. This implies that an intervention implemented in the past still can have an effect on academic performance today. Ii represent a fixed student contribution, often called innate ability. This input is very difficult to measure, as one would like to measure it at the time of birth, before other environmental conditions affect the individual. The lagged variables reflect the cumulative nature of the EPF. Age independence implies that the effect of the inputs in the EPF does not vary with age (Harris, 2010). Fixed family effects implies that all family inputs are constant over time. Additive separability implies that the effects of the inputs in the EPF does not interact with each other. If we, in addition to the efficiency assumption, assume age independence, additive separability and fixed family effects, it is possible to specify the EPF as a value-added model.
Qit =β1Sit+β2Sit−1+...+γi+it (2.2)
Now, due to the fixed family effects assumption, γi captures both the family inputs and the individual student contribution (innate ability, intelligence). With this specification it is possible to evaluate the marginal effects (reflected by the coefficientsβi, i= 1,2,3, ...) of different inputs. The marginal effects can be described as the change in the outcome variable (academic performance) associated with a marginal change of an input variable, keeping all other inputs constant. The purpose of this framework is to extend it to include the SNR as an input. In the following section I split the school inputs into several categories, evaluating each categories effect on the outcome, as well as the effect of varying the SNR on these categories.
2.1.1 Including the SNR in the EPF
The above model is a simple, additive input-output model. A major limitation in the model is that a lot of the variables affecting the outcome are unobservable. Thus, as is the case for this thesis, the data is not detailed enough to estimate the EPF precisely.
Therefore I focus on the SNR as an input, as well as some inputs that I believed to be affected by the SNR. Figure 1 from the protocol (Federici et al., 2017) displays a “flow- chart” in which the increased SNR affect the school environment, and ultimately, increase the students academic performance. The SNR does so at several levels. At the student level, the teacher level and school level. Keeping this in mind, I consider the following version of the EPF to mathematically represent these ideas.
Qit =β1Nit+β2Mit+β3Tit+β4Eit+β5Ait+γi+it (2.3) The previous time periods are left out for simplicity, but could (and should) be included.
Now, the fixed family effects and the individual student contribution are still included in γi. Qit still denotes academic performance, measured through test scores on National tests. The difference from equation2.2is that the school input has been divided in to a set of several inputs. Nit denotes the SNR available for individualiat timet. If this resource increases, we expect that the student’s contact with the school nurse will increase, and that this contact has a positive effect on academic performance. In light of equation 2.1 this can be interpreted mathematically as
∂Qit
∂Nit
>0 (2.4)
An important note is that this marginal effect does not equalβ1, as I assume and define all other inputs in equation 2.1 to be functions of the SNR. This makes sence since the SNR is the input that are varying (increasing) in the RCTs treatment group. The rest of the inputs denotes other factors in the EPF. In the following I assume that these inputs are functions of the SNR. Mit denotes the quality of the school leadership. The idea is that
systemic and structured collaboration between the school nurse and the school leadership will include the school nurse in planning and decision-making. This will in turn lead to planning and decision making being characterized by health promoting measures. The result of this should be that the school leadership gains increased resources for promoting academic performance. To express this, I assume that the quality of the school leadership is a function of the SNR.Sit =fS(Nit, ...). This is a relaxation of the additive separability assumption, as the effect of the school nurse now interacts with the school leadership input. As increasing the school nurse resource increases the school leaderships resources of promoting academic performance, I assume that
∂Mit
∂Nit >0 (2.5)
In addition, I assume that increased quality of school leadership has a positive effect on students academic performance, i.e.
∂Qit
∂Mit =β2 >0 (2.6)
Tit captures the teacher effect on academic performance. Again, I assume this input to be a function that takes the school nurse resource as input, in addition to measures such as years of experience and teachers education. This assumptions is based upon the systemic nature of the RCT in which the increased resource is intended to increase the teachers support in tasks related to student health and the learning environment. So Tit =fT(Nit, , ...). The increased SNR should free up time for the teacher, so that more time is devoted to teaching. This should increase the quality of teaching, and thus increase academic performance. Mathematically, this implies
∂Tit
∂Nit >0 (2.7)
Or in words, increasing the SNR should increase teacher quality. In addition, we assume that increased teacher quality will have a positive marginal effect on academic perfor- mance,
∂Qit
∂Tit =β3 >0 (2.8)
Eitdenotes the psycho-social environment. Psycho-social means looking at psycholog- ical factors combined with the surrounding social environment when evaluating physical and mental wellness of an individual, as well as an individuals ability to function. EWB and SB, as well as bullying are different measures of the psycho-social environment. I assume that an individuals psycho-social environment is a function that takes the school nurse resource as input, Eit = fE(Nit, ...). I assume that increasing the school nurse resource would increase the psycho-social environment at the school. This because the
school nurse should implement health promoting measures, preventive measures against bullying and so on. Expressed mathematically,
∂Eit
∂Nit >0 (2.9)
The model also assumes that a better psycho-social environment should increase academic performance, as students feeling well and able to function should perform better on test scores,
∂Qit
∂Eit =β4 >0 (2.10)
Aitdenotes the student’s absence. Similarly, I assume that a students absence is a function taking the school nurse resource as input, Ait =fA(Nit, ...). The school nurse is supposed to reduce absence with health promoting work and other preventive measures. Thus increasing the SNR should reduce absence,
∂Ait
∂Nit <0 (2.11)
Increased absence implies that the student is not at school and thus not learning. So we expect higher absence to reduce academic performance,
∂Qit
∂Ait =β4 <0 (2.12)
The way the school nurse works is affected by the leadership style and relationship with teachers working at the school. As we are interested in the causal effect of increasing the SNR, and based on the design of the trial (comparable treatment and control groups) I assume that the variation in the SNR is affecting the other inputs as described, and not the other way around. Assuming this is the case, it is possible to rewrite equation 2.1as
Qit =β1Nit+β2fM(Nit, ...) +β3fT(Nit, ...) +β4fE(Nit, ...) +β5fA(Nit, ...) +γi+it (2.13) And the marginal effect of increasing the SNR is then
∂Qit
∂Nit
=β1+β2∂Mit
∂Nit
+β3∂Tit
∂Nit
+β4∂Eit
∂Nit
+β5∂Ait
∂Nit
>0 (2.14) Which accumulates to a positive effect on academic performance when adding all marginal effects. If we evaluate each part of the sum individually, we see that each part is positive based on the assumptions I have made. Thus summing them yields a positive number. I also highlight the point that
∂Qit
∂Nit > β1 (2.15)
This is indicating that the total marginal effect of increasing the SNR is not only reflected through the nurses direct contact with the students, but also has positive effects through improving the psycho-social environment (at the school level), improve the efficiency of teachers and school leadership, and through preventive measures that reduce bullying and absence.
Comments on the Functional form of the EPF
2.2 Literature review
I present previous research that substantiates – or questions – the mechanisms of the EPF specified above. There are many aspects of the mechanisms evaluated in the RCT that are strongly related to the EPF theory presented above. Studies of how the SNR affects these EPF inputs are scarce. In addition, studies of the SNR in other countries may not be generalizable to the Norwegian primary schools, depending on the structure of the schooling system and how the role of the school nurse is defined. Previous literature is more extensive regarding how teachers, school leadership, absence, and bullying affect academic performance. First, I evaluate estimations of each input’s (Ei, Ai, Ti, Mi) effect on academic performance. Then, I continue to discuss how the school nurse affects both academic performance directly, and how the resource might affect the other inputs in equation 2.3. At last, I discuss moderation by age and gender. Age is interesting as there are ongoing debates about what age is optimal for implementing interventions in the education system, and there is data on the age/grade of the individuals in the RCT.
Gender is interesting as almost every Norwegian school nurse is a female, and thus they may have a greater impact on female student’s, as they are easier to relate to. The psycho- social environment is also very different for boys and girls in this age group (NOU2019:3, 2019).
2.2.1 Teacher effects on academic performance
Studies demonstrate that teacher quality is the single most important factor among all school related inputs in the EPF (Pelayo and Brewer, 2010). One study show that teachers have large effects on reading and mathematics achievement(Rivkin et al., 2005). Results from the same study suggest that class reductions of 10 students are less effective than moving one standard deviation up on the teacher quality scale. This estimate used data with observations from Texas public schools with test scores from 3rd to 7th graders.
Strøm et al. (2013) highlight previous research showing that safe and healthy school environments contribute to improved academic performance. Battistich et al. (1995) finds a significant association between individuals sense of school community and all measured student outcomes. Haynes et al. (1997) argue that school climate is an important variable
for student development and academic achievement. Students that feel supported by their teachers experience an increased sense of belonging, more engagement, have fewer behavioral problems and perform better academically in school. Thus the teacher is an important input in EPF, and research support the claim made in expression 2.8, or in words, that increasing teacher quality should have a positive marginal effect on students academic achievement. A remaining question is what affects teacher quality. Rivkin et al.
(2005) find no evidence that a masters degree improves teacher skills, while the first year of experience has a big impact on teacher quality. The effect of increased experience tends to disappear after a teacher has been working for a few years. One might look for other inputs than experience and level of education to teacher quality. Could collaboration with, and support from the school nurse improve the quality?
2.2.2 School leadership effects on academic performance
School leadership is also a relevant input in the EPF. Maybe most importantly, the leadership is responsible for recruiting, hiring and retaining teachers and staff at the school. Estimates of the school leadership effect on academic performance have yielded mixed results. Grissom et al. (2015) use student test scores to estimate the performance of the schools principal. They test different value added model specifications, some quite similar to the EPF specification I use in this thesis. The authors highlight that models that over-attribute the effect of the principal on student test scores are more aligned with non-test measures (such as school climate, principal ratings) than models that separate the principal effect from other school inputs more convincingly. The effect of the school principal on test scores were not correlated across model specifications, and effect sizes in the different models varied a lot in magnitude. Thus model specification is very important.
In chapter 3 i describe the methodology used in my analysis, which should convincingly isolate the effect of increasing the SNR at a school through the use of randomization and instrumental variables. Jacob et al. (2015) analyze an RCT where principals were randomly assigned to a leadership program. Although they find an effect of the leaderhip program on teacher turnover, they find no significant effect of the program on academic perfomance. The claim2.6, that better school leadership improves academic performance among students is uncertain. There is no evidence showing a negative effect of improved school leadership on academic performance, so I can at least assume that the effect is zero or positive.
2.2.3 The effect of absence on academic performance
It is not possible for a student to utilize the available resources at school if the student does not attend class. These resources are intended to prepare the students for tests, and thus absence should have a negative effect on academic achievement. If students do not attend
school, they naturally do not learn effectively. Lamdin (1996) finds a significant positive effect of student attendance on standardized achievement test performance, establishing absence as an input in the EPF. The findings use data on the average level of attendance at the school level. The author suggest that policy makers should devote resources to reduce absence. Some theory suggest that absence is only an issue for disadvantaged students or disadvantaged schools, implying that socio-economic background interacts with absence.
Hancock et al. (2018) in contrast find that for students attending more advantaged schools the effect of absence on academic performance is larger. Thus the author concludes that reducing absence is important for all schools independent of socio-economic background.
Lamdin (1996) questions what kind of interventions that can have a significant effect on student absence. The author suggest that further research should investigate the ability to influence absence and the cost of different interventions. Results from the RCT can help inform about the policy of increasing the SNR. If the school nurse has an effect on absence, increasing the resource can contribute to better academic performance among students through reducing school levels of absence. To conclude, previous research support the claim 2.12, that increasing absence will decrease academic perfomance, or oposite, decreasing absence will increase academic perfomance, which is the effect we are looking for.
2.2.4 The effect of bullying on academic performance
As with absence, on might expect an efficiency loss if a student attends school but is bul- lied. Being bullied causes low self esteem and problems with paying attention, and hinders a students ability to learn at full potential. This implies that bullying can be very costly for society. For example, In the USA 160,000 children (15 % of all absent children during a school day) miss school everyday due to fear of being bullied (Oliveira et al., 2018).
A study measuring the effect of bullying on students test scores have found a significant negative relationship between bullying and performance in mathematics (Oliveira et al., 2018). The sample consists of 6th graders attending public schools in Brazil. Bullying is measured through a subjective measure collected through a survey. Students that reported that they suffered bullying where compared with a control group consisting of students that did not report suffering bullying using propensity score matching. The authors sug- gest that the results indicate that anti-bullying programs should take into account social emotional skills. The results substantiate the assumption made in expression 2.9, that reduced bullying implies that the psycho-social environment is better. Thus students will perform better academically. In the paper, students who are bullied performs worse than those who where not bullied, implying that less bullying leads to better academic perfor- mance. Oliveira et al. (2018) emphasize that not only those being bullied suffer. Also the ones who cause bullying suffer negative consequences later in life. As mentioned, Oliveira
et al. (2018) studies student at the same age as the students in the RCT, but in Brazillian schools. It would be interesting to see similar results in Norwegian schools. Strøm et al.
(2013) estimates the relationship between bullying and academic achievement conducted in Norway on 15-year-old students. The results show a significant association between bullying and low grades. A question is if the association is a causal one, or if it is selection.
The methodology used is a linear regression with self reported grades in four courses as the dependent variable, and bullying as independent variable. This regression is run on a sample of 7050 students. Thus, the results can only be interpreted as a correlation. If it is bullying that causes lower grades, or if low grades and bullying are correlated because of something else, say bad teachers or a bad school environment, is uncertain. Strøm et al.
(2013) also evaluate the effect of bullying at the school level. Students attending schools with high levels of bullying perform worse academically than schools with lower levels of bullying when controlling for sociodemographic characteristics. This negative relationship points in the same direction as the Brazilian study of 6th graders. The results presented indicate that bullying on the school level is important for individual grades. Students attending schools with high levels of bullying reported lower grades than those in schools with less bullying. This was true both for both bullied and non-bullied students.
2.2.5 Previous studies of school nurses in primary school
Previous studies of how school nurses affect students learning environment and learning outcomes are scarce (Federici et al., 2017). A systematic review of literature and research regarding school nurses in American primary schools (Lineberry and Ickes, 2015) states that the biggest gap in this field of research is a lack of rigorous methodology to evaluate the efficacy of school nurses. The review covers 30 articles, of which 22 are descriptive in nature and 8 are quasi-experimental.
The literature mentioned above is solely based on American primary schools. There has also been some Norwegian literature on the topic, summarized in a research project titled ”A team around the teacher” (Borg, 2016). This research resulted in two reports, the first being a literature review of previous research, and the latter suggesting five models for interdisciplinary interaction between teachers and other professions in the schooling system. One of these models is called the The school nurse model, and describe how the school nurse can work in collaboration with teachers and school leadership to improve the school environment (Borg, 2016). This literature differs from the design of the RCT in many ways. It mostly comments research in very specific aspects of the school nurses role, for example solely focusing on overweight children (Bergan, 2013) or asthma (Trollvik and Severinsson, 2005) . In addition, there is mostly qualitative methods in use, interviewing a small sample of students. In contrast, the RCT evaluates the overall effect of increasing the SNR, as well as using quantitative data to conduct inference. Thus the RCT will be
an important contribution to the literature.
Abrahamsen et al. (2018) investigate how improved access to low threshold health services offered in school affect educational attainment, teenage pregnancies, labor mar- ket attachment, welfare dependency and health care utilization of young adults in the Norwegian school system. The paper exploits a reform gradually implemented between 1999 and 2008 in Norway. The reform had an explicit goal of increasing the availability of school nurses. The results show that increasing access to health professionals may have substantial positive impacts on later-life outcomes. In particular, exposure to a larger expansion in school nurse coverage increases the likelihood of high school graduation by age 20. In addition they observe an increased likelihood of starting early in higher edu- cation, and graduating before age 25. Abrahamsen et al. (2018) investigates quite similar research questions as I do in this thesis, and the results they present are significant.
An important difference is that Abrahamsen et al. (2018) conduct a natural experiment, while in this thesis we compare randomized treatment and control groups. They rely on a differences-in-differences approach utilizing different levels of school nurse coverage in different municipalities in different points in time. As the authors point out themselves, the results presented could be driven by other unobserved municipality characteristics due to selection bias. I return to discuss this, and thus the need for an RCT, in 3. It will be interesting to see if the results from the RCT support the findings of Abrahamsen et al.
(2018), with similar positive estimates for the measures described.
2.2.6 School nurse effects on the other inputs in the EPF
According to Maughan (2016) increasing the SNR has a positive effect on student health, well being, bullying, and absence. Other American studies point in the same direction.
Smith and Sherrod (2013) exploits a policy change regarding school nurse coverage in North Carolina. The setting is quite different from the Norwegian educational system, especially as the student to nurse ratio is much larger. Even so, the study observes that the percentage of students making grades below B (on an scale from A to F, where F is failing), are increasing with absence, i.e. if a student has a lot of absence he or she is more likely not gaining the top grades. Students with 2 or more weeks of absence during a semester are also more likely to have poor health. The findings suggest that if nurse staff levels are increased, student absence will decrease. The groups compared in the study are not equal, as the experiment is not randomized. The author compares schools that met the new required student to nurse ratio set by the policy change, with schools that did not. They differ as the schools that received an increased nurse staff level tended to have a larger student population, as well as a larger fraction of students living in poverty. According to Allen (2003) schools with a full time nurse has less health related checkouts during the day compared to schools without a full time nurse. The study
estimates an effect size of -0.77. The effect size is interpreted as follows. Schools with full time nurses has -0.77 standard deviations less health related check outs when compared to school without a full time nurse. This is indicating a large impact of the presence of a full time school nurse on the percentage of enrolled students who check out for medical reasons during the day. The paper defines an interesting tool, as it is difficult to measure performance of or place a monetary value on services produced by public organizations, as the service is not sold to the public at a given price. The tool is called X-inefficiency and it consists of comparing the actual output produced with the maximum possible output for a given set of inputs. In the school setting this indicates that all enrolled students should be in school at all times for the school to be X-efficient. If a child is sick or away from school for other reasons, the resources the school offers cannot be fully utilized, as the learning for this student is not efficient. If the child experiences such issues during the school day, the school nurse is its main resource.
2.2.7 Grades as predictor
The literature review substantiates the different mechanisms described by the SNR as an input in the EPF. Now, grades are believed to be very good predictors for future level of education, income, and health status. One question is if the school nurse effect on student’s future prospects only comes through increased academic performance, or if the work of the school nurse affects students beyond this measure. For instance if the school nurse is able to avoid a student suffering depression later in life through counseling and guidance at school, this might not be reflected in the student’s test scores, but still has a large value, both for the student and for society. There is evidence that support the relationship between bullying and learning environment and academic achievement.
Then it is also relevant to investigate how this affects the student’s future prospects. A study of all Norwegian students that graduated primary school in 2002 looks for reasons for completing upper secondary school (Falch (2010)). The most important indicator for completing upper secondary school is the student’s average grade at the end of primary school (10th grade). As with (Strøm et al., 2013), the methodology used does not guaran- tee a causal interpretation of this indicator. The authors conclude that the causal nature of the findings is uncertain. They estimate that a 1-grade increase in the student’s average grade increases the probability of completing upper secondary school by 30 percentage points. The study also looks for consequences of not completing upper secondary school.
The fall 2007, at a time where the 2002-cohort should have graduated if they finish up- per secondary within 5 years 2.6 % had status as looking for work, 4.7 % was receiving public aid and 1 % was or had been in jail. The individuals that did not complete upper secondary school was over-represented in all these groups. At the same time, over half of the cohort was still in the educational system. The students that graduated upper
secondary school within 5 years are over represented in this group, indicating that they have moved on to higher education. This implies that student’s with low grades are more likely to be a cost for society instead of contributing in the work force or higher education.
The grades achieved in junior high school are of crucial importance for acceptance into selective schools, and can subsequently impact long-term career prospects. The results of the paper emphasize the need for preventive efforts that focus not only on vulnerable groups, but also on all students and the school context. This is exactly what the increased SNR aim to do.
2.3 Moderation
2.3.1 Moderation by age
In equation 2.2, the previous time periods in the value-added formulation is skipped for simplification. Even so, it is likely to believe that inputs from previous time periods (t−1, t−2, ...) also matter for the outcome in the current time period (t). There is an ongoing debate on whether intervention are moderated by age, i.e. that an intervention has higher marginal return on the investment in human capital the earlier in a child’s life the intervention is implemented. If this is the case, one could argue that the RCT should focus on a younger target group, for example 1st to 3rd grade instead of 5th to 7th grade. Heckman (2006) discusses early life interventions targeting disadvantaged chil- dren. According to the findings, the rate of return to investment in human capital is a decreasing function of age. Interventions taking place after pre-school or very early in school may have a rate of return that is lower than the opportunity cost of funding the intervention. Thus the need for investing in disadvantaged young children is stressed, as it has much higher returns than later interventions such as reduced student-teacher ratio, public job training, convict rehabilitation programs, tuition subsidies or expenditure on police. Gardner et al. (2019) analyze the quantitative support of the statement “the ear- lier the better”. The findings justify the age group in focus in this RCT. They analyze a parenting intervention across Europe, conducting meta-analysis making use of implemen- tations of the same intervention at different ages. Across the studies analyzed they find a significant effect on child conduct problems (the interventions dependent variable) of the parenting intervention. But when looking for moderation by age effect, they discover no differences across ages. The trials under study include ages from 2 to 11 years old.
They found no evidence that child age moderated the benefit of the intervention. It is not certain that these findings apply to other interventions, but for the intervention analyzed, the findings findings suggest no significant moderation by age. If the findings do apply to other interventions, they imply that the child’s age at intervention does not matter as much as previous human capital theory suggests. Thus focusing the intervention on 5th
to 7th graders instead of younger children is justifiable also in the context of investing in human capital.
2.3.2 Moderation by gender
It is also interesting to look for moderate by gender. I present two reasons for this moderation. First, females are highly over represented among school nurses, and 99.7% of all school nurses in Norway are female (Federici et al., 2017). If almost all school nurses are female, it is likely that it is easier for the school nurse to interact with girls than with boys. Many children may find it easier to talk to an adult of the same sex when discussing personal issues, and when the school nurse work with preventive measures at the school level, it may be easier to catch the attention of girls than of boys. The second reason is that the psycho-social environment is quite different for boys and girls. The way the social groups work are quite different. NOU2019:3 (2019) describes these differences. On the basis of these differences one might expect different effects of an intervention. Therefore it is interesting to look for a moderation by gender and conduct sub-group analysis for each gender.
Chapter 3
Data description and Methodology
In this chapter, I present the data used for analysis and the methodology used to calculate estimates. I begin with describing how the sample was selected and randomized. Then, I describe the data and present the econometric framework and estimation strategy used to analyze the RCT.
3.1 Sample selection and randomization
The sample in the mid-way analysis consists of 107 schools from 12 municipalities. To be invited, the municipality had to have at least eight primary schools with at least 20 students in grade 5 to 7. There was approxemately 9000 students attending these schools during the 2018/2019 school year. Municipalities engaged in other large NIFU projects or similar projects conducted by other institutes were excluded to avoid contamination.
Out of 31 invited, 15 municipalities agreed to participate. Before implementation, 5 municipalities withdrew from the study, while NIFU recruited 2 new municipalities. I have data from 3 school semesters. The fall 2017 is the baseline, before the intervention was implemented. The intervention started in January 2018 and will last until December 2019.
In addition to the fall 2017, my analysis consists of available data from the spring 2018 and fall 2018. Random assignment was conducted after the participating municipalities signed an agreement of collaboration. In each municipality, two strata were constructed based on measures of well being and school environment, so that the lowest scoring schools are in one stratum, and the highest scoring schools in the second. Then, two schools from each stratum were randomly assigned to treatment.
Each participating municipality receive a 50 % full time school nurse position increase.
This additional resource is to be used to increase the presence of the school health service at the schools in the treatment group, which consist of 4 randomly selected schools in each municipality. These schools each receive an additional 12.5 % of a full time position increase in the SNR, targeting students in the 5th, 6th, and 7th grade.
The school health service should work in systemic collaboration with the staff at
school. Systemic means working with psycho-social aspects of the learning environment in a universal and preventive manner. The intervention is structured by a set of criteria, described by Federici et al. (2017). Treatment schools must receive at least 3.25 additional hours of an actively present school nurse per week. The school nurse should fulfil the tasks and themes described in the guidelines for the school health service. The increased SNR should identify specific health promoting and preventive measures that promote the students’ psycho-social environment at each treatment school. The municipalities should follow guidelines for an initial meeting and further meeting series between the school nurse and staff at the treatment schools. The school nurse and the staff at the treatment schools have extensive autonomy to decide upon how to use the additional resource, as long as the measure are systemic, the collaboration is structured, and the resource is used in line with the principles and guidelines of the project.
3.2 Data Description
I describe the data used in the analysis. The outcomes of the effect study are comprised of both subjective and objective measures. The data used in the analysis are collected from three sources. The subjective measures for EWB, SB and bullying are collected through the Pupil Survey. The objective measure of absence is collected from each school. The measure of school nurse coverage are collected from UDIR (2019).
The measures are designed to estimate the SNR effect on psychological and social out- comes. The outcomes are constructed from previously internationally validated measures (Federici et al., 2017).
3.2.1 The Pupil Survey
Data from the Pupil Survey is used to construct the subjective measures of the learning environment. This survey is administrated by the Norwegian Directorate for Education and Training and administered twice each school year (autumn and spring). The survey consists of questions regarding well-being, motivation, teacher support, cooperation be- tween family and school, evaluation, environment etc. It is compulsory in the autumn for 7th, 10th, and 11th grade. However, all schools with students from 5th to 13th grade are encouraged to include all grades in the survey and the decision whether to participate or not is usually taken at the municipal level or school level. All municipalities participating in the RCT are required to conduct the Norwegian pupil survey twice a year, so that it is possible to conduct analysis for all schools each semester. The project has also added specific questions to the Pupil Survey to gain internationally validated measures.
Figure 3.1: Distribution of emotional well-being
The histograms display the distribution of the standardized measure of emotional well-being using 5 bins, for each semester
English Norwegian
Recall how you‘ve felt last week in class. Tenk p˚a hvordan du har hatt det i klassen How often have you felt the following? den siste uken. Hvor ofte har du?
Been happy Har du vært glad
Been sad Har du vært trist
Been stressed Har du vært stresset
Been bored Har du kjedet deg
Had fun Har du hatt det gøy
Table 3.1: Construction of emotional well-being
The table list the Items in the pupil survey used to construct the measure of emotional well-Being
Emotional well-being
This outcome focuses on affective states and emotional responses in class during the last week. The measure consists of five items comprising of both positive and negative affect.
Responses are given on a five-point Likert scale ranging from ‘never’ (1) to ‘always’ (5).
The scale is used as a composite measure indicative of students’ EWB. The questions are asked in Norwegian. Table 3.1 lists all items in Norwegian along with the English translation. The variable used in the analysis is constructed in the following way. First, for each semester, each item is standardized using the baseline mean and standard deviation (The baseline means items in the Pupil survey the fall 2017, prior to the intervention). For items containing negative questions, the scale is reversed before the items are standardized.
The standardization means that the responses are transformed from the Likert scale into a variable with a mean of 0 and a standard deviation of 1. After this is done, the items are summarized. Then the sum is again standardized using the baseline mean and standard deviation of the sum of standardized items. This is the variable that is constructed for each semester. Figure 3.1contain histograms of the standardized variable. The histograms use 5 bins. In each semester, we see that the answers are left-skewed, indicating that most students have a good sense of EWB. It does not look normally distributed. There are 7,809 observations of EWB the spring 2018, and 8,254 observations from the fall 2018, totalling 7,809+8,254 =16,063 observations of this variable.
School belonging
The protocol defines SB as a feeling of acceptance and being liked by the rest of the group, feeling connected to others and feeling like a member of a community. The measure consists of six items from the Norwegian pupil survey. Responses are given on a five-point Likert scale ranging from ‘strongly disagree’ (1) to ‘strongly agree’ (5). Table 3.2 lists all the items used to construct the measure of SB in Norwegian, along with the English translation.
The variable used in the analysis is constructed in a similar manner as for the EWB variable. First, the scale of the items are converted. Then, each item is standardized separately using the baseline mean and standard deviation. Then, they are summarized
Figure 3.2: Distribution of school belonging
The histograms display the distribution of the standardized measure of school belonging using 5 bins, for each semester.
English Norwegian
Other students seem to like me Det virker som de andre elevene liker meg I make friends easily at school Jeg f˚ar lett venner p˚a skolen
I feel like I belong at school Jeg føler at jeg hører til p˚a skolen 1 Do you sometimes feel lonely at school Hender det at du føler deg ensom p˚a skolen
I feel different than others Jeg føler meg annerledes and out of place in my school og at jeg ikke passer inn p˚a skolen
I feel like an outsider Jeg føler at jeg blir (or excluded out of things) at school holdt utenfor p˚a skolen
Table 3.2: Construction of school belonging
The table list the Items in the pupil survey used to construct the measure of school belonging
and standardized again. Figure 3.2 shows histograms for the standardized value of SB.
We see that most respondents each semester have a strong sense of SB. Furtermore, as with EWB, the variable is skewed. There are 7,341 observations of SB the spring 2018 and 8,098 observations from the fall of 2018, totalling 7,341+8,098=15,439.
Bullying
Bullying is measured by means of one item and the question is compulsory in the Nor- wegian Pupil Survey. The students are asked whether they have been bullied by other students at school during the past few months. The response categories are ”not at all”,
”rarely”, ”2 or 3 times a month”, ”about once a week”, and ”several times a week”. A student is defined as bullied if he or she experiences bullying 2 or 3 times a month or more. This definition is provided by Olweus (1994), and is an internationally regonized definition. Respondents who state that they experience bullying from other pupils, from teachers and cyberbullying several times a week, and in addition state that they bully others on the same questions, are excluded from the analysis. In the Pupil Survey 2018 this amounted to 0.1 %, or 623 of the respondents (Wendelborg, 2017). In our sample, 1 respondent the fall 2017, 6 respondents the spring of 2018, and 3 respondents the fall of 2018 were excluded from the sample due to this.
The variable used in the analysis is a dummy variable, equal to 1 if the student has experienced bullying 2 to 3 times a month or more, and equal to 0 otherwise. Summary statistics for the variable is presented in table X above. Looking at the semesters sepa- rately, there are 7,512 observations with a mean of 0.05 during the spring semester, and 8,286 observations with a mean of 0.06 during the fall semester. This means that in the sample, approximately 376 pupils experienced bullying during the spring semester, and 571 pupils experienced bullying during the fall semester. In the baseline, there is 8,465 observations with a mean of 0.07. Thus 577 experienced bullying in the baseline sample.
Since the observations are not linked over time, it is not possible to say if it is the same pupils that experienced bullying in each semester. In total, combining the semesters giver 7,512 + 8,286 = 15,798 observations of the bullied variable.
3.2.2 Absence
The only outcome that is not collected from the Pupil Survey is absence. It differs from the other primary outcomes as it is not subjective. Data on absence was collected in two different ways. A subgroup of the participating municipalities uses the services of a company called Visma. In this subgroup, absence is extracted from Visma’s database.
The data collected from Visma consist of recorded full days of absence as well as recorded hours of absence. The data is on the individual level, with id numbers identifying each student, and the students grade. For the remaining municipalities, absence is reported manually by teachers at the school using a excel template sheet distributed by NIFU to record absence. Each student has a unique identification number, as was the case for data collected from Visma. The variable constructed for the analysis is a count of total days of absence for each student during the semester, divided by total days of school during the semester.
There has been delays in reporting on absence from the schools in municipalities not using Visma, at this point I only have access to baseline data (November-December 2017) and the first semester of intervention, spring 2018. In the baseline, I have reported absence from 93 out of 107 schools. Out of the 14 schools missing, 10 are in the control group and 4 are in the treatment group. In the spring semester 2018, I have reported absence from 74 schools, thus missing 33 schools. Out of the 33 missing, 23 schools are in the control group and 10 schools are in the treatment group. Thus, both in the baseline and for the spring 2018, control schools are over-represented among the schools that did not report absence. This may lead to some selection bias in the analysis, as the schools that have not yet reported their absence may have a certain set of school characteristics correlated with not being able to report absence.
Some schools are missing observations for students with zero days of absence. For schools where this is the case, I have imputed zero-observations in the dataset, so that the number of observations in each cohort at each school matches the total number of students reported by the Norwegian Directorate for Education (UDIR, 2019). The data does not contain information about gender. Thus I am not able to control for gender when analysing this variable. The variable used in the analysis is constructed in the following way. A count variable of days absent for each student is divided by the total number of school days during a semester, which I assume to be approximately 95 days for Norwegian primary schools according to the Norwegian Directorate of Education.
Figure3.3displays histograms of the sample distribution of days absent. Observe that there is more zero days absent observations during Nov-Dec 2017. This is due to absence being recorded for two months, or approximately 38 days, while for the spring semester absence was recorded for the whole semester, approximately 95 days. The longer period observed, the less likely it is that a student has no days of absence.
Figure 3.3: Distribution of absence
The upper histogram show the distribution of absence during November and December 2017. The lower histogram show the distribution of absence during the spring semester 2018. The number of observations decrease fast when days of absence increase, but there are long tails in both distributions
3.2.3 Coverage
Data on treatment status is given implicit by the randomization in the project. How the actual resource was distributed is reported by the municipalities, collected from UDIR (UDIR, 2019). In addition, time registration data are collected from school nurses both at treatment and control schools through surveys. This is done so the intended coverage increase and actual time spent at the treatment schools can be compared. In my analysis, I use the coverage data collected from the municipalities to estimate the nurse to student Ratio (NTS ratio). In each municipality, there is a head school nurse responsible for distributing the available resource among the schools in the municipality. The variable for coverage is constructed in the following way. The student to nurse ratio at each school is calculated for pupils in 5th to 7th grade. The baseline mean from November 2017 is 638 students per full time school nurse. The mean student population at the schools in the sample is 110 students. The average school nurse coverage in the sample population would be 17 % (110/638) of a full time school nurse position, prior to the treatment.
During time of treatment, I evaluate the treatment and control group separately.
In the control group, the mean from November 2017 was 563 students per full time school nurse. In April 2018 it was 547, and in November 2018 it was 538. It is keeping almost constant, with small reductions each semester. For the treatment group, the number of students was much higher in the baseline of November 2017, with 714 student per full time school nurse. Looking into why this mean is so much higher than for the control group, there is some large outliers in the treatment group with observations of 1500-2000 students per full time school nurse. Such outliers are not in the control group.
In April 2018 the mean is reduced drastically to 444 student per full time school nurse, and in November 2018 the mean is reduced to 311. From November 2017 to April 2018 the increased coverage in the treatment group would be 110444 − 1110714 = 24%−15% = 9%.
In the control group, the increase would be 110547 − 110564 = 0.5%. Taking the difference, 9%−0.5% = 8.5% is quite close to the intended increase of 12.5 %, but a few percentages lower. From November 2017 to November 2018, the increased coverage would be 110311 −
110
714 = 34%−15% = 19% in the treatment group. In the control group, it was 110538−110564 = 1%
Taking the difference 19% −1% = 18%, we see that the average increase is actually above the intended treatment by 5.5%. Why the increase is bigger than intended may be explained by some of the outliers in the treatment group. It may also indicate threats to internal validity. I return do discuss this in chapter 5. As can bee seen in figure 3.4, there are some observations of students with a student to school nurse ratio of approximately 2000 in the baseline. Figure 3.4 show histograms of the number of students per full time nurse in each observation. In November 2017, the treatment and control group look quite similar, except for the outliers in the treatment group. It is possible to observe a shift to the left in the treatment group, indicating an overall reduction in the student to school
Figure 3.4: Histograms of the student to nurse ratios
The two upper histograms display the distribution of student to nurse ratios in the treatment and control group during November 2017. The two lower histograms show the same distribution for November 2018
Figure 3.5: Scatter plot of student to nurse ratios
The x axis denotes the number of students per full time school nurse in November 2017. The y axis denotes the number of students per full time school nurse in November 2018. The size of the circles reflect the number of students attending that school. The lines are fitted values.
nurse ratio. The differences between November 2017 and November 2018 is much smaller for the control group. Figure 3.5 show a scatter plot of the student to school nurse ratio in November 2017 and November 2018, as well as fitted value lines. The red colour reflect the control group and the blue colour the treatment group. The size of he circles reflect the number of students at the school.The predicted relationship is quite different for the groups. The control group line is steeper than the treatment group line. This means that if the school is a treatment school it is more likely to have a larger reduction in the student to school nurse ratio. Also observe that the majority of the control group schools has had a constant student to school nurse ratio.
In the actual analysis, I use the NTS ratio (the inverse of the student to nurse ratio).
This is because the NTS-ratio increases when the coverage increase, opposed to the student to nurse ratio, that decreases. It is easier to infer and evaluate an increasing variable, instead of a decreasing one.
To summarize, I provide a table with means of the different variables, for treatment and control schools, in November 2017 and November 2018. Table 3.3 shows that most
Variable November 2017 November 2018
Treatment Control Treatment Control
Emotional well-being 3.89 3.86 3.88 3.84
School belonging 4.33 4.31 4.29 4.26
Bullying 0.07 0.07 0.06 0.06
Student to nurse ratio 714 561 311 538
November and december 2017 Spring 2018
Absence 1.33 1.46 4.0 4.37
Student to nurse ratio 714 561 547 444
Table 3.3: Variable means
The table display the mean of the measures described. The means of school belonging and emotional well-being are not standardized, but reported on the Likert scale in this table.
of the measures are approximately equal across control and treatment groups. EWB, SB, and bullied seem to be constant over time. The student to nurse ratio is as seen in the figure 3.4 and 3.5. In table 3.3, EWB and SB are not standardized, but measured on the Likert scale.