University of Groningen MSc Economics
Faculty of Economics and Business
Exploring the possibilities of forecasting Hospital turnover
Abstract
This paper aims at developing a forecast model for hospitals which estimates future turnover. The focus in this paper lies on the Dutch healthcare market. The combination of the
complexity of the claiming system and the fact that this system has been changed several times in the last decade make that no decent forecast model is present for hospitals as of yet. In order to make better contract agreements with healthcare insurers, hospitals want to know
what their turnover in the upcoming year will be. Moreover, once the contract has been agreed upon by both the insurer and the hospital, the hospital wants to make sure they will
not exceed the cap on healthcare to be provided for the given insurer, because the excess amount provided will not be paid for by the insurer. In this paper an autoregressive
distributed lag model has successfully been constructed based on historical data on treatment type and specialization. The model circumvents using the claiming system in place
and is able to provide an estimate of future turnover.
Key words: Hospital, Turnover, Forecast, Healthcare, Production JEL code(s): D24, I10, I11,
Author: T.A. Bleeker
Student number: 1874217
Foreword
During my internship at Topicus I wrote this Master’s thesis on forecasting a hospital’s turnover. The field of healthcare economics interests me greatly and this
internship matched this interest perfectly. I would like to thank everyone at Topicus for this useful and interesting internship.
Moreover, I would like to thank both my supervisors Prof. Dr. R.J.M. Alessie, my supervisor at the University of Groningen, and Ir. J. Laagland, my supervisor at Topicus. Both have provided me with useful comments and suggestions on earlier versions of this Master’s Thesis.
1
1. Introduction
1.1. BackgroundIn 2013, Dutch newspapers reported that the Radboud medical center in Nijmegen did not accept any more patients that were insured at Achmea unless it was an emergency from the 25th of November until the 31st of December of that year (Zantingh, 2013). This was a
consequence of the fact that Achmea already reached its cap on the maximum amount of healthcare to be provided by the Radboudumc to Achmea’s clients. In the Netherlands, healthcare insurance companies purchase a certain amount of healthcare prior to the beginning of the new calendar year (1st of November is the deadline for the negotiations
(NVZ, 2014)). Insurers negotiate on the total amount of healthcare to be provided by a given hospital the upcoming year on the basis of last years’ results, a certain growth rate, and various other variables. What is important here is that the amount is fixed. This means that when the cap is reached, a hospital will not provide any more services to clients of a specific healthcare insurer for which the cap has been reached (only emergency treatments will still be performed). When it does it will not be able to claim its remuneration for these treatments from the insurance company. This is exactly what happened in 2013 in Nijmegen. Patients were asked to visit different hospitals or they could postpone their hospital visit to the next calendar year if their illness did not need immediate treatment. The opposite situation could also happen; it might be that an insurance company has purchased an amount that exceeds the actual demand for healthcare in a given year. Unless there was a unique event that caused this temporary low(er) demand, the insurance company will make a downward adjustment in the budget for the upcoming year(s). Hospitals, just like any firm, base the number of employees on demand for their services/products. Not reaching the
2
1.2. Diagnose Behandeling Combinatie (DBC)
In order to improve the comprehension of the upcoming sections in this paper , some background information is presented here. As of 2005 hospitals made use of the so-called DBC systematic. DBC is a Dutch abbreviation for Diagnose Behandeling Combinatie which translates to Diagnose Treatment Combination. Because this is such a specific term in this field of expertise, the Dutch abbreviation DBC will be used throughout the paper. DBC’s are used by hospitals to claim their compensation (for salaries and other costs) from the
healthcare insurance companies. A DBC basically contains all the information regarding a patients treatment from diagnose to treatment to finally claiming the compensation for delivered services. The DBC system has been implemented in the Dutch healthcare sector in order to improve competition in the market for healthcare. Furthermore, the DBC system should increase consumer awareness when choosing in healthcare service/products (NZa, 2014). The DBC system also had its downsides, to mention an important one, it was everything but transparent. Because hospitals were able to document their actions
themselves there were a lot of discrepancies between different hospitals. It might occur that a patient who came into hospital 1 with a given illness received the same treatment as he would have received when he would have gone to hospital 2. However, because both hospitals register their actions differently it might be that the final DBC is different, which is likely to lead to a different price. Because of the inefficiency here various parties (Ministry of Health Welfare and Sports, Healthcare insurers and Healthcare providers, DBC Onderhoud and several others) developed an improved system, to increase transparency and make an important step towards cost homogeneity amongst hospitals (DBC Onderhoud, 2012).
1.3. DOT system
The improved system is called DOT (“DBC op weg naar Transparantie” which means “DBC on its way to Transparency”). As of the first of January 2012 this improved system is used by the Dutch hospitals and various other medical institutions. One important
DBC-3 zorgproduct group (which is specified in accordance with the ICD10-diagnose classification1)
has a unique decision tree in which healthcare products/activities are included. When every action during the patients’ treatment is registered properly the Grouper will deduct a given DBC-zorgproduct from the combination of actions and attach a specific code to it. It will add a hashed identification to the DBC-zorgproduct (so the healthcare insurer has a guarantee the claim is legitimate) and sends it back to the hospital. The hospital can now send the claim to the insurance company in order to receive its compensation for the delivered services. The Grouper detects when a step in the treatment is out of the ordinary and accordingly it sends a notification to the hospital so that it can be corrected. This improves transparency as claims are now the same whether the patient goes to hospital 1 or hospital 2 because the Grouper treats them equally. Because under the DOT systematic the same action boils down to the same DBC-zorgproduct and accordingly the same price, the number of DBC-zorgproducten could drastically be reduced from 45.000 to approximately 4.400 (DBC Onderhoud, 2012 & NZa, 2014). Important to note here is that accurate registration is still important. When this is not done accurately enough the Grouper might derive a different DBC-zorgproduct than the one actually delivered. When this happens the accompanying price is likely to be
incorrect as well. So even with the Grouper which is able to detect discrepancies by means of comparing a decision tree to the actions actually performed by the specialist (leading to a specific DBC-zorgproduct) wrong claims can still be made.
1.4. Aim of paper
The aim of this paper is to develop an econometric model which makes it possible to estimate a hospitals’ turnover prior to the end of the year. This is important to hospitals as they do not want to end up in a similar situation the Radboudumc found itself in, as previously described. So hospitals would like to be able to estimate what their turnover is going to be at the end of the year say in August. So they can act in time if the model suggests the cap will be reached before the year expires. Hospitals can obviously see what their turnover has been in the previous year, however, due to the properties of
1 ICD10 is an internationally used list of diseases which is kept up to date by the World Health
4 zorgproducten it is not as easy to see what it will be in the future. More on this will be
discussed later in the paper.
1.5 Topicus
During the process of writing this thesis the author will be an intern at Topicus located in Deventer (The Netherlands). Topicus is an innovative company specialized in developing IT-solutions for institutions in Healthcare, Education, Finance and Governmental sectors. The Topicus Fincare division is specialized in designing software for healthcare providers and healthcare insurers. So basically Topicus is an intermediate between both parties and makes sure the claiming of compensations runs smoothly. Because of the fact that Topicus is involved in this market, they have access to an extensive dataset with information on treatments executed in hospitals. This data proves necessary in order to make a prognosis of turnover obviously. Topicus is interested in exploring the possibilities of developing a turnover prognosis model, so they can offer their clients the possibility of monitoring their production and, if necessary, steer production. This is important nowadays as healthcare providers are bound to a predetermined budget.
1.6 Research Questions
In the following subsection the research question and sub-questions will be
formulated along with a brief elaboration on why these questions are relevant. Furthermore, a number of difficulties in answering the research question will be briefly presented. These will be further investigated in the methodology section later on in the paper.
1.6.1. Main research question
The main goal of this paper is to develop a model which is able to estimate future turnover for hospitals based on historical data. This leads to the following research question:
5
1.6.2. Complications
To some extent one can compare a hospital to a regular firm, in the sense that both face a demand for a product/service and they supply this product/service. So in order to generate a model for turnover quantity x price should be sufficient one might say. However, the situation is complicated here. One issue here is the price; hospitals and healthcare
insurers negotiate about prices, leading to differences in prices for equal services for different healthcare insurers (Koninklijke Vereniging voor de Staathuishoudkunde, 2012). So this complicates the situation in the sense that one cannot simply multiply prices and quantity as prices vary. A possible solution could be to calculate an average price for a given DBC-zorgproduct (Rijksoverheid, 2014). In the methodology section this problem will be elaborated on.
Another complication in this research applies to the way DBC-zorgproducten are claimed. In the current system, a DBC-zorgproduct is closed after 42, 90 or 365 days after final treatment of a patient. After it has been closed the payment follows. So it could be the case that someone is being treated in 2012 and the DBC-zorgproduct is closed in 2013. However, though the payment takes place in 2013 it is included in the 2012 turnover
calculation. So in calculating turnover, one has to consider all this so-called Onderhanden werk (Translation: Work-in-progress). A more elaborate explanation of the concept of
Onderhanden werk follows in section 2.1.1..
1.6.3. Sub-questions
Related to the main research question are the following sub-questions:
1. What is the size of the entities to which turnovers can be split up e.g. the entire hospital, healthcare insurers, perhaps specializations within the hospital or per type of treatment? 2. What is the appropriate time period for which turnover can be estimated (monthly, quarterly,
yearly, year of claiming?)
3. Is it possible to generalize the model so that it can be applied in several types of hospitals?
6 regarding the contracting procedure between the hospitals and healthcare insurers. So when clear forecasts are available to the NZa they can target specific bottlenecks in the contracting process more efficiently.
Sub-question 2 addresses the question of what is the time period for which turnover can be estimated. Is it possible to determine turnover for November 2013 or can turnover be forecasted for the entire consecutive year? By being able to asses turnover per month/quarter it might be possible to detect seasonal peaks and troughs in demand for certain types of healthcare. Preventative healthcare can be targeted better if it is known when certain diseases are spreading, to give an example of why it might be interesting to determine monthly turnovers. Moreover, hospitals can steer their production if it is likely they will exceed their production cap. Or they can (if there is sufficient demand obviously) increase production so they meet the cap and avoid receiving a smaller budget for the next year.
The final sub-question applies to the degree of generality. Obviously hospitals differ to a large extent. Not only in size, they also differ in the types of treatments they offer. University Medical Centers (UMC’s) are very different from ‘regular’ hospitals as one can imagine. So this third sub-questions is meant to asses if it is possible to apply the model to different types of hospitals.
This paper continues with the institutional background, which describes the situation in the Dutch healthcare market. In the third section the data used in this research will be discussed. Next, the methodology section follows in which the developed model will be elaborated on. The fifth section will present the results of the regressions. The final section will conclude the paper and will summarize the research. At the end of the final section some recommendations and shortcomings of this research will also be presented.
2. Institutional background
In the following section the institutional background will be described. Due to the unique and complicated financial system in the Dutch healthcare sector providing
7 Dutch healthcare financing system. Furthermore, it will shed light on a number of issues that are considered to be important for this research or are considered to be interesting possible complements to it. As a consequence of the specific situation in the Dutch healthcare market literature regarding this topic is scarce. To begin with the claiming system in place will be described in section 2.1. and what the complexity of this system is. In subsection 2.1.1. the concept of Onderhanden werk will be described. After this section 2.1.2. will briefly describe the concept of asymmetric information and how it applies to the healthcare market. 2.1.3. describes the lack of transparency that comes along with the claiming system in the Dutch healthcare market. After this, section 2.2. will describe the change in demand for healthcare the past years and its main purpose is to show that the market for healthcare is changing partly as a consequence of the claiming system (and the changes made to it) in the Netherlands. Section 2.3. describes the importance of the emergency room for the total turnover generated in hospitals. Though its effect cannot be tested in this research, it would be an interesting addition for future research. Seasonal demand for healthcare is discussed in section 2.4. as it will be included in the model constructed in this paper later on. The
importance of turnover prognoses will be discussed in section 2.5. and finally the main results of a brainstorm session are described in section 2.6. The purpose of section 2.6. is to describe information and insights gathered from professionals who work with financial managers in hospitals on a daily basis with respect to turnover prognoses.
2.1 The implementation of the DOT system
To start off this analysis providing a more extensive explanation on the DOT system and why it was implemented seems helpful. As was briefly mentioned in the introduction, the system preceding the DOT system, the DBC system, was implemented in 2005 in order to increase competition on the market for healthcare and at the same time increase the quality of the healthcare provided (DBC Onderhoud, 2011). Another important reason to implement the DBC system was to provide better insight into the amount of healthcare actually
8 necessary, as the budget for healthcare is limited and prices kept increasing in the sector. Increased competition is a tool used to stop the increasing prices in the healthcare sector (DBC Onderhoud, 2011). In practice however, this system was not as successful as planned, as described in the introduction. Then came the DOT system (DBC on its way to
transparency). Now as explained earlier, in the basis this system is the same. Still a patient comes into the hospital and is being diagnosed with a certain illness. The suitable treatment is started until the patient is cured. All this information, along with information about the age and gender of the patient is registered as accurately as possible. The information is sent to the Grouper which derives a certain DBC-zorgproduct from the registered actions by means of a decision tree. A price is being attached to this DBC-zorgproduct and then the hospital claims its compensation from the healthcare insurance company (DBC Onderhoud, 2011). Now this seems to work better, as DBC-zorgproducten are now derived automatically and will be equal for every hospital. However, the importance of accurate registration is still not completely eliminated. Though the Grouper warns the hospital if it seems to be the case that a mistake has been made in the registration process, it is still possible to claim different DBC-zorgproducten than actually have been performed. For example when several actions are accidentally registered wrongly, a different DBC-zorgproduct may be derived by the Grouper. So the system is not completely waterproof and it might lead to asymmetric information (see subsection 2.1.2.) problems between hospitals and healthcare insurers.
2.1.1. Onderhanden werk (OHW)
Another important subject in the field of hospital finance is the Onderhanden werk (OHW). As briefly mentioned in the introduction, OHW is the work-in-progress in a
9 other hand, in general a small part of the DBC-zorgproducten will drop out in the end
(before they are claimed) which lowers the final payout/claim (NVZ & NFU, 2013). According to Dutch Hospital Data (2014) the amount of OHW at every point in time contributes between 15% and 20% to the final turnover of that particular year. So one can imagine that it is important that the OHW is determined with great accuracy. The OHW can be used to provide an indication of what is to be expected of turnover for that given period and turnover growth. Hospitals can make use of a special Grouper (the Onderhanden werk-Grouper) which estimates the final DBC-zorgproduct after medical specialists have indicated what they have undertaken so far for the patient. The OHW is related to a concept known as the claiming year. A claiming year is not the same as a turnover year. Insurance companies tend to think in claiming years whereas many firms as well as hospitals think in terms of turnover years. Figure 1 below helps explaining the difference between both concepts:
Figure 1: Difference between turnover year and claiming year
2012 2013 2014
Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2
Treatment (A) opened in 2012, closed in 2013 A1 A2
Treatment (B) started in 2013 and closed in 2013 B1 B2
Treatment (C) started in 2013 and closed in 2014 C1 C2
Turnover 2013 consists of turnover in periods: A2 B1 B2 C1
Claiming 2013 consists of turnover in periods: B1 B2 C1 C2
(Dutch Hospital Data, 2014)
10 year. This discussion becomes more complicated because of the fact that prices differ over time. So even when A and C are equal in terms of quantities, prices differ which in turn leads to differences in turnovers. Because of the long lead times of DBC-zorgproducten it is likely they will open and close in different years. Moreover, it takes a long time before the final data is available to parties such the Ministry of Health Welfare and Sports, healthcare insurers and providers, DBC Onderhoud and several others active in the healthcare market. Interesting to see is that the Ministry of Health Welfare & Sports recognizes the problem of the delayed claims. In order to increase the speed at which information is available to the involved parties, the maximum lead time will be decreased from 365 to 120 days from 2015 and onwards according to the Nederlandse Zorgautoriteit (translation: Dutch Healthcare authority)(NZa)(NZa, 2013). In this paper turnovers in a given year will be used, so not the claiming years used by healthcare insurers. Since there is no information in the dataset on when patients come in for their first treatment or whether their treatment is a follow-up on an earlier one it is not possible to work with claiming years (even though it might have been interesting to see the differences between the both).
2.1.2. Asymmetric information
Information asymmetry exists when an individual is better informed about a given situation/subject than another person (Ruwaard et al, 2014). Information asymmetry also exists between hospitals and healthcare insurers. The larger the gap in knowledge between both parties the more difficult it will be to reach efficient prices during negotiations. When the hospital is better able to assess the situation they will aim for compensations that are higher than actually needed. Healthcare insurers can counter this by closely monitoring the hospitals. This way they increase their knowledge about the hospitals’ situation and
11
2.1.3. Lack of transparency
Due to the fact that the DOT system is being improved/adapted on a frequent basis, making a turnover prognosis based on historical data becomes complicated. The product structures undergo changes regularly so it is not (always) possible to compare a DBC-zorgproduct of 2014 with ‘the same’ one in 2013, as it might consist of different elements in both years. As a consequence, it is likely that the prices are different in both years (NZa, 2012). In the data section will be described how the data is going to be structured in order for it to be used in the regressions in this research. This is required in order to compare data across different years.
2.2 Lower demand for healthcare
According to De Bruijn (2012), the number of patients in hospitals has dropped the past few years. Whereas it used to be the case that demand for healthcare grew at
approximately 5-7% a year, there are now hospitals facing a negative growth rate of 4%. It is unclear what causes this lower demand, though there are several suggestions. The first suggestion is that general practitioners send less patients to a hospital. Another explanation could be that people are deterred from visiting a hospital due to the higher own contribution (De Bruijn, 2012). Over the years the own contribution has increased significantly; where the own contribution was €150,- in 2008, for 2015 the own contribution will be €375,-
12
2.3. Importance of ER
According to the NVZ (2013) estimates show that around 40% of hospitals’ turnovers can be, directly and indirectly, attributed to the emergency room (ER). This can be explained in the following way; an ER stay generates turnover on its own obviously, as the more severely injured patients often need expensive treatments. This is the direct contribution of the ER to turnovers. The indirect part can be explained in the following way: when a patient enters the ER he/she is likely to stay within the same hospital for further treatment. To compensate for lower turnovers, as a consequence of closing of an ER, in other departments of the hospital proves to be challenging (NVZ, 2013). For this paper it would have been interesting to see if the model estimated here shows a similar estimate for the contribution of an ER to a hospital’s turnover, however, data providing information on ER stays is not available here so unfortunately this will not be possible. In order to test the effect of an ER one would need data on individual patients that provides information on whether a person entered a hospital via the ER and remained in the hospital for further treatments. Moreover, patients would have to be monitored over several years to see if they came to the same hospital later in time with a different illness. As mentioned, unfortunately, this kind of data is not available here, so the effect of the ER cannot be tested, though it would have been interesting because when indeed the contribution of the ER to total turnover is that large, hospital managers know they have to make sure that the ER keeps running as efficiently as possible in order to generate a sufficiently high turnover and thereby assuring (to a large extent) the right of existence of the hospital.
2.4 Seasonal demand for healthcare
According to Fullerton and Crawford in their 1999 paper “The winter bed crisis –
quantifying seasonal effects on hospital bed usage”, there is a relationship between seasons and
13 the variability of the demand for hospital beds in hospitals. So the variability of general medicine is buffered to a large extent by the rather stable demand for surgical specialties (Fullerton & Crawford, 1999)
Figure 2: Daily bed occupancy and fitted seasonal curve in a General Medicine and b Surgical specialties
(Fullerton & Crawford, 1999)
2.5 Importance of turnover prognoses
The upcoming section will shed some light on the question of why turnover
14 indicates a significant drop in patients this may lead them to investigate what the cause of this sudden drop is. On the other hand, because the hospital has made agreements with various healthcare insurance companies on the quantity of healthcare to be provided, they would like to know if they are going to exceed this maximum amount. Optimally, they know this before the healthcare is actually being provided so that they can act in time in order to prevent penalties that will follow on exceeding the caps. Another reason according to Hakbijl (2012) is that hospitals need to negotiate with banks in order to get loans. Banks require insight into what the prognosis is for the upcoming period before they grant the loan to whoever applies for a loan. When hospitals cannot provide reliable information to the bank they will not receive a loan or the hospital will be considered to be a high(er) risk and has to pay higher interest rates on the loan. As mentioned before, not exceeding the caps on the amount of healthcare to be provided is vital as hospitals will be punished if they do so. When a single hospital exceeds the cap, and claims more than it was allowed to, it has to pay back the surplus it received to the healthcare insurer. So for the next year the healthcare insurer will have a smaller budget as it has to pay back what it received in excess the year before. This is clear to hospitals and they know beforehand exactly what the size of their budget is. However, there is also something called the Macrobeheersingsinstrument (MBI) which loosely translates to macro controlling instrument (NVZ, 2014). Each year the government sets a macro budget for the healthcare sector in order to control the growth of the total healthcare sector. So the sector is not allowed to exceed a predetermined maximum amount of healthcare to be provided. When the total sector does exceed this cap, the
government uses the MBI to penalize the hospitals. Every hospital will receive a fine proportional to its market share, even if a specific hospital did not exceed the cap by itself. This might lead hospitals to increase their prices in order to compensate for potential losses due to other hospitals exceeding the cap set by the government (moral hazard) and the possibility of being punished for it.
2.6.1. Brainstorm session with Verhoeve BV.
15 firm. Verhoeve BV also employs financial specialists in hospitals in the Netherlands, so they have extensive knowledge of what is important for hospitals when it comes to their finances. Now the first of October a brainstorm session was organized between Topicus and Verhoeve BV to gain some insight in why it is important for hospitals to have an accurate turnover prognosis model. Furthermore, some complications were discussed in designing a prognosis model. In order to guide the session, some questions were posed beforehand. The session will be discussed according to these questions in the following subsections.
2.6.2. What are the consequences if a hospital over or under produces?
The first possible consequence is the fact that hospitals might lay off certain divisions and/or treatments. Because hospitals are bound to a certain budget, they might specialize in certain treatments and let other types of treatments (that occur less frequent) to other
hospitals. This brings forth possible scale advantages and might increase the quality of these treatments as they perform certain activities more often than before. For the patients
however, this means that they have to investigate whether the hospital of their choice offers the treatment they need. If this is not the case they will have to look for another hospital that does offer the treatment and also has a contract with their healthcare insurer. This in turn might incur patients to travel more to get the treatment needed. So it is not a bad thing per se if a hospital lays off certain treatments, however, this also depends on the presence of other hospitals in the area and the treatments they offer. When this specialization is coordinated between hospitals it might be a positive development, however, then the Dutch competition authority (Autoriteit Consument en Markt, ACM) also has to agree with these plans.
16 negative effects of under producing. These goods can be used in the next year and profits will be larger in that year.
The third and final consequence, as was also mentioned in the introduction, hospitals might deter patients of certain healthcare insurers when the cap for that year at that
particular insurer has been reached (so overproducing). This is obviously a bad development for patients because they now have to wait until the next year or have to go to a different hospital, which may not have their preference due to travel distance/quality offered/price differences.
2.6.3. What information do hospitals use now to monitor production?
Hospitals make extensive use of excel in registering their production. They also tend to make adjustments in documents after they have been approved by Topicus for claiming. This makes that the uniformity is compromised and that different hospitals and probably even different managers within the same hospital send differently formatted excel files for final registration and claiming. This all makes it more time consuming to check registrations before they are claimed at the insurance company.
A second important issue is that hospitals do not consider the already planned activities in their prognosis. They base their estimates mainly on historical data. These planned activities give a good approximation of what can be expected the coming period in terms of turnover. So the fact that these activities are not taken into account seems
objectionable.
Another important observation is that independent doctors and financial controllers do not seem to have the same interests all the time. The doctors consider the patients’ interests to be more important than the budgets that are given the highest priority by the financial controllers. Furthermore, it might seem obvious that doctors have another
17
2.6.4. What is necessary information to steer the production of hospitals?
The main conclusion here is that it depends very much on the level at which steering has to take place. A hospital’s management has a completely different view on what is optimal than the staff that actually treats the patients. As such the way in which steering has to take place has to be adapted to a specific level within the hospital. Registration of
healthcare for example has to be done by the person who actually performed a certain activity. However, nurses’ main priority is the wellbeing of the patients and not the administration that comes with it. When another person has to register the healthcare provided, mistakes are easily made, because the person who registers might not know exactly what actions took place.
2.6.5. Have there been any attempts in constructing a turnover prognosis model before?
Due to the complexity of the system in place in the Netherlands, there has no model been developed that can be used regardless of the claiming system in place. So there is definitely a need for such a model. Another important complication is that different types of treatment (clinical/diagnostic/surgery etc.) all have very different prices. So to develop a model with prices that differ and vary over time is challenging.
2.6.6. Main results
From the session it becomes clear that a wish for a decent forecast model definitely exists within hospitals. For now mainly the financial controllers within the hospitals are interested in such a model, as they negotiate with insurance companies about contracts and make sure the caps on healthcare to be provided are not exceeded. By having a decent forecast model, the financial controllers are better able to steer production in order to stay within their budget. Providing the best possible healthcare for the patients and at the same time remaining within the available budget do not always go hand in hand. This has to be
18
3. Data
In the following section a description of the dataset used will be provided. For this paper a panel dataset has been used with monthly data from January 2010 until May 2014. Due to the presence of long lead times in the hospital claiming system the dataset had to be reduced to a set including data ranging from January 2010 until October 2013. The set had to be reduced, as the dataset was not yet complete for the original time period (claims still had to be claimed). Using data over a longer time period would increase the quality of the
regressions, however there is no data available prior to the period under consideration here. By using panel data it becomes possible to use group and time fixed effects in the model. The cross-sectional variable here is specialization (index s) and the time variable is ym (index t, see below for a more elaborate explanation of this variable). In the dataset information is captured from one Dutch hospital and it was provided by Topicus. Data on each variable was split up between the various healthcare insurers that have contracts with the hospital under consideration. The dataset originally contains information on 18 different
specializations, of which some will be dropped as will become clear later in this section. For a complete list of the specializations used here see Table A.1 in the appendices. All data was aggregated per specialization and the data is not split up between the healthcare insurers anymore. This leads to an average price for a given treatment at a given specialization. When the data would be split up between healthcare insurers as well, the number of observations per insurer and specialization would become small which negatively influence the
significance given that the dataset is not that large to begin with. Basically this also implies that average prices will be used for forecasting because there is no separation anymore between healthcare insurers. So aggregating the data per specialization was the first step in making the dataset suitable for regression purposes.
19 It might be insightful to explain the names of the variables. As the dataset used is from a Dutch hospital and applies to the Dutch hospital market the variables mainly have Dutch names to keep the paper easily readable for Dutch professionals in this field of expertise. To increase comprehensibility for an audience not active in the medical world the various terms used for the variables will be translated and briefly explained below:
POL – Consultation at the outpatient clinic (Dutch: Poli)
DAG – Treatments at the outpatient clinic or a day treatment (Dutch: Dagbehandeling) KLI – Clinical (patients who stay overnight) (Dutch: Klinisch)
DIA – Diagnostic treatment (research in a lab, MRI-scans etc.) (Dutch: Diagnostiek) OPE – Surgery and other high-expense procedures (Dutch: Operatief)
Also see the Terminology list and the end of the paper for various other terms used in this paper
Data on each of the above mentioned variables are retrieved from the dataset provided by Topicus. In the model the variables are referred to as POL, DAG, KLI, DIA and OPE respectively. In the dataset the number of every type of treatment is included, so for example in January 2011 x outpatient clinic treatments were performed by specialization s. Because the actual numbers are included there is no need to make dummy variables for the type of treatment and because of that the effects on turnover of every type of treatment can be estimated by the model (since there is no treatment type being omitted).
Another variable that might need some explanation is the ym variable. This variable is a combination of year and month, hence ym. In the dataset year and month were included separately so a new variable had to be constructed to combine the two variables. The
variable starts at 600 which is January 2010 and ends with 645 which represents October 2013. This variable as the time variable of the panel.
20 more accurate insight in the data actually used here. Now the question is of course, what happened to the 188 observations that have been dropped from the set. First of all is not complete as for specializations 304, 318, 327, 362 and 389 there are various months for which there is no data available. Starting with specialization 304, there were only 4 observations so this data cannot be used. As no data was present for the other 42 months we are already down 46 observations from the 828 the dataset would include if it were balanced. In case of 318 this is only the case for 4 months, for 327 this is 10 months, for 362 data is missing for 28 months and finally for 389 data is missing for 43 months. In total there are 46 months so specializations 327, 362 and 389 have been dropped from this analysis as there are too many observations missing for these variables to provide reliable estimates. So in total there are 4*46 + 4 = 188 observations less than there could have been and thus 640 observations
remain. Table A.2 shows that, on average, there are many visits to the outpatient clinic (POL) compared to the other types of treatments. This could be expected beforehand, as treatments here are relatively cheap and quickly handled. Aside from the missing observations for three of the specializations nothing seems to be out of the ordinary in the dataset by looking at the summary statistics.
When examining figure A.1 in the appendices, it becomes clear that for certain specializations there is a time trend in the data whereas other series are fairly constant over time.
Finally, the data has been lagged with 12 months and when the lagged variants of the original ones are used the variable will have -12 as a subscript. After restructuring the panel is almost balanced, only for specialization 318 there are 4 observations missing
21
4. Methodology
In the following section the derivation of the model itself will be explained. Because of the specific situation in the Dutch healthcare market and the lack of financial literature regarding this topic there is no benchmark model to work with. So the model has to be built from scratch. To some extent the turnover model will show similarities with turnover models used for firms. Basic economic theory learns that the quantity multiplied by the price
determines a firm’s turnover. In hospitals this is basically the same; the medical specialists deliver a certain service to a patient and they receive a compensation for this. Turnover in this case is determined by multiplying the total number of treatments by the price of this type of treatment. In hindsight this is possible obviously because it is known how many patients have been treated during that year. Though the situation will be slightly more complicated because of the long lead times of the DBC-zorgproducten, it is still possible to calculate what the turnover was in the previous year. To make a prognosis for the upcoming year is more complex however. First of all, one has to make an estimation on the number of treatments that will take place next year. For this one can look at what the growth rates were in the past. When the number of treatments grew fairly stable over the past years one can expect a similar growth rate for the next year (assuming no major changes will take place in the sector).
As described extensively above, the Dutch healthcare market was subject to major changes when it comes to the claiming system for hospitals and other healthcare providers. This complicates the matter of making accurate prognoses, as the situation is different before the implementation and after the implementation of the DOT-system in 2012, when it comes to the prices of zorgproducten and what they consist of (they cannot be compared one-to-one). This paper attempts to construct a model that forecasts turnovers for a hospital on the basis of care provided in the time period from January 2010 until October 2013. This is done by avoiding the use of DBC-zorgproducten as indicators of what types and quantities of healthcare have been provided. In order to make the model independent of claiming
22 This paper attempts to address these uncertainty problems by developing a model that provides us with an estimate of what turnover can be expected by estimating the types of treatments to be expected next period for every specialization and, obviously, their impacts on turnover. This all is based on historical data. To start off the analysis the following subsection describes how to deal with the uncertainty.
Uncertainty of future healthcare activities
First of all, to counter the problem of long lead times, the model’s estimations will be based on type of treatments. The advantage of this is that it does not matter when the money is paid, because you know directly for which year the turnover is generated at that moment. Furthermore, the type of treatment is known the day the patient enters the hospital, or at least close to this date (and can be registered as Onderhanden Werk (OHW)). So more patients could be included in the model because of this, which increases the accuracy of the estimates. A downside of using this method is that one does not know which treatments a patient with a certain diagnose will receive in the future. So though it is known what the amount of patients is that is being treated at this moments, one does not know which activities will follow in the near future for these patients as the development of the illness might differ between patients or change unexpectedly for a given patient. It is not
23 the data. More on this will follow later on in this section. When it is indeed the case that a relatively constant distribution is observed in the data it can be assumed that this will continue to exist in the (near) future. For this paper only historical data is available, so healthcare that provided and has been claimed by the hospital at the insurer. So no data is available on the amount of OHW. Later on in section 4.3. it will be discussed how the model can be extended in case this data is available, as it would provide us with more accurate information as it is know that healthcare registered as OHW has already been provided, though not yet claimed.
4.1. Turnover model
Now obviously the one patient is different from another patient, so it is likely that patient x diagnosed with illness y receives a different treatment for this illness than patient z with the same illness y. This is a direct consequence of the condition a patient is in (how far along is the illness), a patient’s age, gender, various personal characteristics and preferences etcetera. In order to be able to estimate a model, a number of assumptions has to be made regarding the classification of different patients. Each treatment has its own price obviously, depending on whether or not a patient needs surgery, expensive medicine or can a patient be treated with relatively cheap medicine and perhaps outpatient care is sufficient for the patient. Since every patient receives different treatments which consists of various activities depending on the patient’s needs, it is not possible to attach one price to a given diagnose. Also, prices differ every year and between claiming systems (DBC vs. Dot). So a fixed price for a given diagnose will not suffice. To estimate the model and make insightful what type of treatment contributes which amount to total turnover a differentiation has to be made
24 data. An autoregressive distributed lag model will be estimated here, in which not only a lagged term of the dependent variable is included, also other lagged independent variables that are expected to have an impact on turnover enter the right-hand side of the equation. Estimating ADL models for forecasting purposes is common when forecasting time series data. In other words data on the number of treatments at a specific specialization in the past years will be used to predict the number of treatments at that specialization one year from now. Also a lagged term of the dependent variable is included, which captures previous values of turnover. The autoregressive distributed lag (ADL) model for turnover takes on the following form:
(1) 𝑇𝑂𝑠𝑡 = 𝛽1𝑇𝑂𝑠𝑡−12+ 𝛽2s𝑃𝑂𝐿𝑠𝑡−12+ 𝛽3𝑠𝐷𝐴𝐺𝑠𝑡−12+ 𝛽4𝑠𝐾𝐿𝐼𝑠𝑡−12+ 𝛽5𝑠𝐷𝐼𝐴𝑠𝑡−12+
𝛽6𝑂𝑃𝐸𝑠𝑡−12+ 𝛽7𝑃𝑂𝑂𝑠𝑡+ 𝛾𝑡+ 𝑐𝑠+ 𝜀st
, where subscript s is an indicator for the various specializations (and can run from 0 to 14, where zero means no specialization) and subscript t is an indicator for time. 𝑐𝑠 represents
individual fixed effects (within specializations) and 𝜀st is the error term.
As can be seen the dependent variable here is turnover (TOst), as we want to estimate future turnovers. The hospital’s turnover depends on the independent variables that are stated on the right-hand side of the equation. All variables in the model with -12 in their subscripts represent the lagged variants of that specific variable. These lagged variables are included in order to be able to make an out-of-sample prediction later on after the regression has been run. This is the final goal, because by doing so a forecast of turnover will be created. So basically by transforming the current observations into their lagged variants it seems as if the observations took place a year ago (12 months, hence subscript -12). The first independent variable is the lagged term of turnover. It is included as it is likely that turnover in this period, depends partially on turnover in the previous period. The next 5 independent variables (POL, DAG, KLI, DIA, OPE) are interaction terms between the specialization and the type of treatment, this in order to control for price differences between specializations and types of treatments. The interpretation of these variables is the effect of an extra patient visiting the outpatient clinic at specialization s increases turnover by 𝛽𝑥 for example. The
25 (not dependent on a specialization). The next independent variable is specialization (POOst)
which is a variable indicating the individual effect of an extra patient coming to that specific specialization, regardless of the type of treatment (so it is an average contribution to turnover for every specialization per additional patient). By including this variable it is as if group fixed effects are used in the regression (since the group effects are controlled for
individually). The final independent variable is dummy variable 𝛾𝑡which is included to
detect seasonal effects. A test with an interaction between specialization and months (𝛾𝑠𝑡) has been done, however the standard errors became very large in this case. Because of this the author chose to use normal month dummies here. As from the paper by Fullerton and Crawford (1999) it became clear that some seasonal effect exists in demand for healthcare it seemed interesting to test whether this effect was found in this data as well. By including this variable one is able to detect in which months of the year more turnover is being generated. Beforehand the expectation on each of the independent variables is that they will have a positive effect on turnover as more production will, normally, lead to higher turnover. The expectation on the 𝛾𝑡 dummy is that in the winter months (January, February and December)
more turnover is generated for certain specializations. The expectation is that most types of illnesses do not depend on the weather, however it could be that demand is higher for certain types of healthcare during the winter because more people fall and break a leg for example or due to the fact that people are likely to be more vulnerable in this period. More importantly, since people have to pay a deductible once they make a claim on their
healthcare insurance, it is likely that once this deductible has been paid in a given year, people plan a hospital visit before that year expires so that they do not have to pay the deductible again. People will often have to come back after a short period to check if the treatment was successful or for further treatment. It is likely these treatments will take place in the new year (hence higher demand in January and February). Finally the error term is captured by
ε
st in the model.The model will be estimated using time fixed effects by including dummies for each month, as mentioned above, except the first month of the year (which is omitted).
26 specialization code 301, which is omitted). Because of a heteroskedasticity problem, White standard errors are used in the regression. For the test results of the heteroskedasticity test see Table A.3 in the appendices. As can be seen from the test results heteroskedasticity is present in the data, as the null hypothesis is rejected. A joint-significance test has to prove that the coefficients are not jointly equal to zero, and hence have no impact on turnover.
As will become clear in section 5, the result of the joint-significance test is that indeed the hypothesis of all coefficients being jointly equal to zero is rejected, hence they do have an impact on turnover. However, as will also become clear, there are various coefficients
insignificant. Since the output table is fairly large and includes many insignificant coefficients the possibility of reducing the model has been explored. By testing various variants for significance and by comparing the outcomes to the original model the following (reduced) model has also been estimated:
(2) 𝑇𝑂𝑠𝑡 = 𝛽2𝑃𝑂𝐿𝑠𝑡−12+ 𝛽3𝑠𝐷𝐴𝐺𝑠𝑡−12+ 𝛽4𝐾𝐿𝐼𝑠𝑡−12+ 𝛽5𝑠𝐷𝐼𝐴𝑠𝑡−12+ 𝛽6𝑂𝑃𝐸𝑠𝑡−12+
𝛽7𝑃𝑂𝑂𝑠𝑡+ 𝛾𝑡+ 𝑐𝑠𝑡+ 𝜀𝑠𝑡
27 turnover. The other interactions with KLI have been dropped from the model and the results did not change massively. Cleaning up the model this way led to an output table that
contained less irrelevant information. Both will be presented in order to show the differences in the estimations and the final turnover predictions.
4.2. Forecast
As mentioned before, in order to be able to make an out-of-sample prediction, the independent variables have been transformed (lagged) in such a way that it seems as if all observations occurred a year prior to when they actually occurred (so 12 months as the time variable is months here). This makes it possible to make a within sample prediction right from the point where the last observation occurred minus one year. In this case the actual data ran to October 2013, which has been transformed to October 2012. Hence, the prediction will start at November 2012 until October 2013. Since the software remembers that the data used to run until October 2013, it is able to calculate the within sample prediction. This is necessary since it is not possible to calculate an out-of-sample in a different way. By using this technique the software uses the data of the original first year (2010) to calculate the lagged variables, though for the actual regression the first year’s observations are not included anymore (that is why the output table will show a smaller number than the 640 observations mentioned in the data section). Now after the prediction was generated, all observations in the dataset (including the predicted ones) are transformed back into their actual time of occurrence. The predicted observations will now look like as if they occurred in November 2013 until October 2014, even though actual data was not available for this period as of yet. So the data is now extended into the future and the fluctuations in the data are based on fluctuations in the actual, historical data. The predicted values of turnover can now be extracted for the period of November 2013 until October 2014.
4.3. Extension of the model
28 could be extended with the OHW data in order to get a more accurate estimate of turnover, because more recent data is available. Unfortunately, for this research there was only data on claimed healthcare activities, so the final amount of care provided, which is known at the moment at which the patient has been released from the hospital and hasn’t come back for 365 days for the same illness. At that moment the DBC-zorgproduct is derived by the grouper and the claim is made. The OHW says something about the care already provided, though not yet claimed because the patient is still being treated (or the 365 days have not passed yet). So it tells us something about the period between now and one year ago.
Though this data is not yet complete, one does know the amount of healthcare that has been provided with certainty. In this paper however, as mentioned earlier, OHW data is not available so it cannot be included in the research. Although the OHW data is not actually being used here, below a description will follow on how the data could have been used in case it was available.
The OHW data can be used, by performing a truncated estimation. This means that one can re-estimate the model afterwards conditional on the fact that the estimates of the original model are at least as large as the OHW data says they actually were. So observations below the threshold are excluded from the estimation, increasing the accuracy of the
estimates as the values that have not been observed in reality are now excluded from the predicted data used in the original regression.
The expected value of TOj|xj, conditional on TOj|xj being present on the interval (lj, uj), is
defined as follows:
E(l
j,u
j) = E(x
jb + e
j|lj
< x
jb + e
j< u
j) = ŷ
j- s
Ø (𝑢𝑗− ŷ𝑗)𝑠 − Ø (𝑙𝑗− ŷ𝑗)𝑠 Φ (𝑢𝑗− ŷ𝑗)𝑠 − Φ (𝑙𝑗− ŷ𝑗)𝑠
, where ŷ𝑗 is the predicted value of turnover (= xjb) and s is the regression error. Ø is the
normal density and Φ is the cumulative normal. lj is the lower bound of the prediction, which
is OHW turnover here. uj is the upper bound of the prediction of turnover and can run to
infinity in principle.
29 in the original model. So the expectation is e(OHW, uj), which means as much as turnover is
at least equal to OHW turnover and can run towards infinity (in theory).
So by using a truncated estimation one could use the OHW data, given that this data is available, in order to make the estimate more accurate.
5. Results
The regression results of model (1) can be found in the appendices in Table A.4. Because of the many dummy and interactions variables the output has become quite extensive so not every coefficient will be discussed individually. Moreover, since the coefficients are mainly calculated in order to be able to make the turnover forecast and are not the final outcomes of this paper they are briefly discussed before we move on to the actual forecast results. Before starting off the discussion on the results, recall from previous section that a T-test has been performed on the joint hypothesis that all coefficients are equal to 0. The T-test rejected the null hypothesis of all coefficients jointly being equal to 0 with a p-value of 0.000. Hence we can conclude that indeed the variables in the model have an impact on the hospital’s turnover. The first result to be discussed is the outcome for the lagged term of turnover, 𝑇𝑂𝑡−12. As can be seen in the table the coefficient is equal to -0.027, however, it is
not significant with a p-value of 0.820. Because of this, we cannot draw any conclusions from this coefficient. For the following variables discussed below, keep in mind that every result is relative to specialization 301, since 1 dummy will always be omitted and the omitted dummy is 301 here (so 301 is the reference group). So a negative sign does not imply a negative impact on turnover, it means that for example an extra outpatient visit at specialization 302 contributes 𝛽1 less to turnover than an outpatient visit at specialization 301. So continuing
30 different specializations and KLI do provide many significant results. Only for specialization 310 the result is insignificant with a p-value of 0.786. Every other interaction is significant at the 1% significance level and every specialization contributes less to turnover than
specialization 301 does when it comes to clinical treatments. The diagnostic treatments for every specialization show only 1 (only just) significant result (at the 5% significance level) namely for specialization 316 with a p-value of 0.050. Once again, specialization 301
contributes most to turnover relative to 316. The other coefficients are insignificant. For the interaction between surgery (OPE) and the various specializations there are two omitted results for which no surgeries were registered. Here, only the coefficient for 320 is significant at the 1% level with a p-value of 0.005. Continuing with the results for the different
specializations independently (variable POO), it can be seen that 6 coefficients are not significant at the 5% level and the rest is. It turns out that in general specialization 301 does not contribute the most to turnover since 307 contributes most every month followed by 305 which is significant only at the 10% level with a p-value of 0.090. The lagged variables for every type of treatment show that only KLIt-12 is significant with a p-value of 0.000. The idea of seasonal demand in healthcare does seem to be reflected in the data as most coefficients are negative, which imply that they contribute less than January (which is the omitted variable and hence results are relative to January). However, most results are not significant so we cannot draw solid conclusions from these results. Only the result for July is negative and significant at the 1% level. So overall numerous results were significant, though many were also insignificant. As mentioned above a T-test on the joint hypothesis of all coefficients being equal to 0 was performed and the test showed that the coefficients are not all equal to 0 at the same time. Now that we have these intermediate results, we can move on to the actual 1 year ahead forecast results. Based on the previous results this forecast is predicted and results can be found in table A.5 in the appendices. So as can be seen from the table, a forecast has been calculated and the estimate is that in the next twelve months (starting at November 2013 until October 2014) the hospital will realize a turnover of roughly €
31 many observations were dropped and that might have an impact. However, the turnovers realized by these specializations were not extremely large according to the original dataset. In 3 years’ time 327 generated € 316.500,- in turnover. So around a € 100.000,- each year. 362 generated even less namely € 21.550,- in 18 months for which data was available. So probably the results are not influenced to a very large extent by leaving out these specializations completely. When comparing the result found it resembles turnovers realized by this hospital in previous years for which data was available as the turnovers ranged between € 57.000.000,- and slightly over € 59.000.000,-. So according to the forecast, turnover will slightly grow over time. Bear in mind that the calculated turnover is not the turnover for the entire year of 2014, instead it is turnover from November 2013 until October 2014. So from this prediction we cannot say if the hospital will remain within the budget set for that year, however if a few extra (more recent) months are added to the dataset the prediction can be made for January 2014 until December 2014 and then the results can be compared to a budget cap and the maximum growth rate of 1.5% set by the government for 2014. Since more recent data was not available at the time of writing this paper the result found is the best estimate for the upcoming period. The results have also been plotted in figure A.2 in the appendices, in which first the actual observed data is plotted and the second part of the graph shows the predicted values for the extended time period. When examining figure A.2 no major irregularities stand out except perhaps for specialization 320 where one can observe a large drop. As mentioned earlier the quality of this research could be improved by
increasing the size of the dataset. The standard errors in table A.6 also suggest using a larger dataset, as they are large. For specialization 335 in particular, the standard errors are (too) large. This is likely to be a consequence of the fact that there were not many patients treated at this specialization, compared to the others. However, as of yet more data is not available to the author. When extra data is added perhaps the major decrease in the graph would also be somewhat smaller as the prediction is based on more observations and is less sensitive because of this.
32 to the exclusion of numerous insignificant variables. Once again, the regression results are intermediate results needed for making the prediction of turnover. All interaction terms between POL, KLI and OPE with the relevant specializations together are jointly significant with a p-value of 0.000. Testing the interaction terms with the specializations individually showed that they are significant as well (p-values of 0.000, 0.001 respectively. Note that OPE*320 is the only interaction term included for this treatment type, so no individual joint significance test is performed here, as there are no other interactions with other
specializations included. Now as can be seen from the regression output most coefficients differ compared to the original ones. This is due to the different specification of the model. Evaluating the coefficients for the lagged variants of treatment type, it can be seen that more coefficients are significant than in the original specification, where only KLI turned out to be significant. Now DAG is also significant at the 5% significance level. OPE and DIA are only significant at the 10% level, so we have to be careful drawing conclusions from these coefficients. So apparently for DAG and KLI there is some persistence in their effects on turnover; these variables indicate the effect on this year’s turnover (holding everything else constant) of a one-unit increase in last year’s number of treatment type DAG or KLI
respectively. So when these variables turn out to be significant it suggests that turnover in this year can be predicted by the number of observed treatment types in the previous period (amongst other variables of course). Once again the results seem to suggest that January turnover is high(er) compared to the other months of the year. As only February, July and September are significant (at the 5% level) we cannot draw solid conclusions for the other months, however all coefficients are negative which point in the direction of a high turnover in January.
Turning to the actual turnover predictions in table A.8 we see that total turnover is now approximately € 59.450.000,- which is almost € 150.000,- less than the original
33 some differences to be observed. For a number of months the forecast is lower and for other the forecast is higher than in the original model. So due to the different (reduced)
specification of the model the monthly results have changed however, the final result is acceptable. Some of the variables that were dropped were only significant at the 1% and therefore considered not to be very important.
The results per specialization have been graphed in figure A.3. When comparing figure A.3 to figure A.2 (in which the original predictions were graphed), what can be seen is that the predicted turnovers in the reduced model are more constant than the ones found in the original model. Figure A.2 seems to reflect the actual fluctuations better then figure A.3. It seems that figure A.3 shows more averaged results. Important to note here is that the scale of both graphs is the same, so that does not cause the differences between the graphs. So apparently by leaving out the dropped variables the results are ‘smoother’ than before. Concluding, when interested in monthly results it is preferred to use the original model as it seems to explain the true situation better than the reduced model.
6. Conclusion
This paper examined the possibilities of forecasting a hospital’s turnover and in doing so avoiding the use of specific claiming systems that are in place. This way the forecasting model does not have to be adapted every time a new system comes into place and
34 into place. In this paper proxies for prices were used in the form of the type of treatment (as an indicator of the price level; outpatient visits are generally cheaper than surgery) and the various specializations at which the treatments were performed. Since it is not possible to make out-of-sample predictions an autoregressive distributed lag (ADL) model was estimated in order to be able to make turnover predictions for the future. The data was lagged 12 months, which makes it possible to make predictions about the upcoming 12 months. Because hospitals are interested in their turnover at the end of the upcoming year and also want to know if they will stay within their budget for that year, a prediction period of 12 months was chosen. If hospitals have an idea of how turnovers are likely to develop over the upcoming year they can steer production more efficiently and, if necessary, decrease production in time so that they will not exceed their budget. Two models were estimated in this paper, one of which is a reduced variant of the original model that does not include the insignificant variables of the original model and drops the variables that were only
significant at the 10% significance level (as it is less strict than the 5% significance level, so results that are significant at the 10% level are less reliable than the results that are significant at the 5% significance level). The original model predicted turnover for the upcoming 12 months to be € 59.600.000,- and the reduced variant predicted a turnover of € 59.450.000,-. So a difference of around € 150.000,-. The final yearly predicted turnovers found by both models are quite similar, so the dropped variables indeed do not influence the results (to a large extent) as could be expected since most of them were insignificant or only significant at the 10% significance level. Additionally the model tested if seasonal effects were present though the effects were mostly insignificant for this variable.
35 original model. So monthly or yearly forecasts for the entire hospital seem possible, smaller entities are not preferred.
The second sub-question related to what the time frame was for which turnover could be forecasted. This paper predicted turnover for the next 12 months and given the available data it is not recommended to forecast for more than 1 year. As discussed above it is not preferable to estimate monthly data per specialization individually, however, for the entire hospital the monthly outcomes were more similar between both model specifications. So monthly or yearly total hospital turnover can be used, though for the individual
specializations the estimates seem to be too sensitive for changes in the model specification.
Finally the model can most likely be applied to different hospitals within the Netherlands as well as it includes variables that can be observed in any normal hospital. Obviously, the data for these hospitals has to be structured individually so that estimations can be made. When using the model for a different hospital, one should use the original model, as it might turn out that for another hospital some of the insignificant variables in this research do provide significant estimates for another hospital. This depends on the data from the new hospital obviously. So when using the reduced model, one might miss out one some important variables (which are important for explaining the new hospital’s turnover).
Concluding this paper the main result is that it is possible to find an estimate of future turnover and to make the model independent of the claiming system in place. Whether claims are made on the basis of DOT’s or DBC’s, or perhaps a new system which might make its way to the Dutch healthcare market in the future, it is possible to make a forecast of turnover. In order to estimate the model one has to be in the possession of historical data of the hospital under consideration. Preferably the dataset is longer than the one used in this paper in order to make the prediction less sensitive and more accurate.
Recommendations
36 discussed briefly here. The main recommendation would be to extend the model with extra data or perhaps also test the model using data from other hospitals. Perhaps there are more extensive datasets available for these other hospitals. So it would be interesting to increase the scale of this paper by testing the model for more hospitals and also to include more data for the hospital under consideration here, once it becomes available. When more data is available it would perhaps be possible to provide turnover forecasts per healthcare insurer. This is useful for negotiations about future contracts between hospitals and healthcare insurers. Regarding the data, the importance of accurate registration cannot be stressed enough. In order to provide accurate estimates it is of upmost importance that registration takes place accurately.
Another suggestion follows from subsection 2.3 in which the importance of the emergency room for a hospital’s turnover is discussed. For further research it might be interesting to investigate this relationship between turnover and ER treatments.
