OULU 2015 D 1292 UNIVERSITY OF OULU P.O. Box 8000 FI-90014...
74
UNIVERSITATIS OULUENSIS MEDICA ACTA D D 1292 ACTA Hannu Vähänikkilä OULU 2015 D 1292 Hannu Vähänikkilä STATISTICAL METHODS IN DENTAL RESEARCH, WITH SPECIAL REFERENCE TO TIME-TO-EVENT METHODS UNIVERSITY OF OULU GRADUATE SCHOOL; UNIVERSITY OF OULU, FACULTY OF MEDICINE
Transcript of OULU 2015 D 1292 UNIVERSITY OF OULU P.O. Box 8000 FI-90014...
UNIVERSITY OF OULU P .O. Box 8000 F I -90014 UNIVERSITY OF OULU FINLAND
A C T A U N I V E R S I T A T I S O U L U E N S I S
Professor Esa Hohtola
University Lecturer Santeri Palviainen
Postdoctoral research fellow Sanna Taskila
Professor Olli Vuolteenaho
University Lecturer Veli-Matti Ulvinen
Director Sinikka Eskelinen
Professor Jari Juga
University Lecturer Anu Soikkeli
Professor Olli Vuolteenaho
Publications Editor Kirsti Nurkkala
ISBN 978-952-62-0792-6 (Paperback)ISBN 978-952-62-0793-3 (PDF)ISSN 0355-3221 (Print)ISSN 1796-2234 (Online)
U N I V E R S I TAT I S O U L U E N S I S
MEDICA
ACTAD
D 1292
ACTA
Hannu Vähänikkilä
OULU 2015
D 1292
Hannu Vähänikkilä
STATISTICAL METHODS IN DENTAL RESEARCH, WITH SPECIAL REFERENCE TO TIME-TO-EVENT METHODS
UNIVERSITY OF OULU GRADUATE SCHOOL;UNIVERSITY OF OULU,FACULTY OF MEDICINE
A C T A U N I V E R S I T A T I S O U L U E N S I SD M e d i c a 1 2 9 2
HANNU VÄHÄNIKKILÄ
STATISTICAL METHODS INDENTAL RESEARCH, WITH SPECIAL REFERENCE TO TIME-TO-EVENT METHODS
Academic dissertation to be presented with the assent of theDoctoral Training Committee of Health and Biosciences ofthe University of Oulu for public defence in Auditorium F202of the Department of Pharmacology and Toxicology (Aapistie5 B), on 29 May 2015, at 12 noon
UNIVERSITY OF OULU, OULU 2015
Copyright © 2015Acta Univ. Oul. D 1292, 2015
Supervised byProfessor Markku LarmasDocent Pentti NieminenProfessor Leo Tjäderhane
Reviewed byDocent Kaisu PienihäkkinenDocent Janne Pitkäniemi
ISBN 978-952-62-0792-6 (Paperback)ISBN 978-952-62-0793-3 (PDF)
ISSN 0355-3221 (Printed)ISSN 1796-2234 (Online)
Cover DesignRaimo Ahonen
JUVENES PRINTTAMPERE 2015
Vähänikkilä, Hannu, Statistical methods in dental research, with special referenceto time-to-event methods. University of Oulu Graduate School; University of Oulu, Faculty of MedicineActa Univ. Oul. D 1292, 2015University of Oulu, P.O. Box 8000, FI-90014 University of Oulu, Finland
Abstract
Statistical methods are an essential part of the published dental research. It is important to evaluatethe use of these methods to improve the quality of dental research. In the first part, the aim of thisinterdisciplinary study is to investigate the development of the use of statistical methods in dentaljournals, quality of statistical reporting and reporting of statistical techniques and results in dentalresearch papers, with special reference to time-to-event methods. In the second part, the focus isspecifically on time-to-event methods, and the aim is to demonstrate the strength of time-to-eventmethods in collecting detailed data about the development of oral health.
The first part of this study is based on an evaluation of dental articles from five dental journals.The second part of the study is based on empirical data from 28 municipal health centres in orderto study variations in the survival of tooth health.
There were different profiles in the statistical content among the journals. The quality ofstatistical reporting was quite low in the journals. The use of time-to-event methods has increasedfrom 1996 to 2007 in the evaluated dental journals. However, the benefits of these methods havenot been fully adopted in dental research.
The current study added new information regarding the status of statistical methods in dentalresearch. Our study also showed that complex time-to-event analysis methods can be utilized evenwith detailed information on each tooth in large groups of study subjects. Authors of dental articlesmight apply the results of this study to improve the study protocol/planning as well as thestatistical section of their research article.
Keywords: article, bibliometrics, research methodology, statistics
Vähänikkilä, Hannu, Tilastolliset tutkimusmenetelmät, erityisesti tapahtumaankuluvan ajan analysointimenetelmät, hammaslääketieteellisessä tutkimuksessa. Oulun yliopiston tutkijakoulu; Oulun yliopisto, Lääketieteellinen tiedekuntaActa Univ. Oul. D 1292, 2015Oulun yliopisto, PL 8000, 90014 Oulun yliopisto
Tiivistelmä
Tilastolliset tutkimusmenetelmät ovat olennainen osa hammaslääketieteellistä tutkimusta. Mene-telmien käyttöä on tärkeä tutkia, jotta hammaslääketieteen tutkimuksen laatua voitaisiin paran-taa. Tämän poikkitieteellisen tutkimuksen ensimmäisessä osassa tavoite on tutkia erilaisten tilas-tomenetelmien ja tutkimusasetelmien käyttöä, raportoinnin laatua ja tapahtumaan kuluvan ajananalysointimenetelmien käyttöä hammaslääketieteellisissä artikkeleissa. Toisessa osassa osoite-taan analysointimenetelmien vahvuus isojen tutkimusjoukkojen analysoinnissa.
Ensimmäisen osan tutkimusaineiston muodostavat viiden hammaslääketieteellisen aikakaus-lehden artikkelit. Toisen osan tutkimusaineiston muodostivat 28 terveyskeskuksessa eri puolellaSuomea hammashoitoa saaneet potilaat.
Lehdet erosivat toisistaan tilastomenetelmien käytön ja tulosten esittämisen osalta. Tilastolli-sen raportoinnin laatu oli lehdissä puutteellinen. Tapahtumaan kuluvan ajan analysointimenetel-mien käyttö on lisääntynyt vuosien 1996–2007 aikana.
Tapahtumaan kuluvan ajan analysointimenetelmät mittaavat seuranta-ajan tietystä aloituspis-teestä määriteltyyn päätepisteeseen. Tämän väitöksen tutkimukset osoittivat, että tapahtumaankuluvan ajan analysointimenetelmät sopivat hyvin isojen tutkimusjoukkojen analysointiin.Menetelmien hyötyä ei ole kuitenkaan vielä saatu täysin esille hammaslääketieteellisissä julkai-suissa.
Tämä tutkimus antoi uutta tietoa tilastollisten tutkimusmenetelmien käytöstä hammaslääke-tieteellisessä tutkimuksessa. Artikkelien kirjoittajat voivat hyödyntää tämän tutkimuksen tulok-sia suunnitellessaan hammaslääketieteellistä tutkimusta.
Asiasanat: artikkeli, bibliometriikka, tilastotiede, tutkimusmetodologia
7
Acknowledgements
This work was carried out at the Department of Restorative Dentistry,
Endodontics and Pedodontics, Institute of Dentistry, University of Oulu. I wish to
express my sincere gratitude to the following co-workers, without whom this
study would not have been completed.
I owe my deepest gratitude to my supervisors, Professor Emeritus Markku
Larmas, DDS, PhD, Docent Pentti Nieminen, PhD, and Professor Leo
Tjäderhane, DDS, PhD. Markku and Leo have guided me to dental research and
how to write scientific articles in English. I want thank Pentti for support and
encouragement and for our conversations during conferences.
I want thank my reviewers Docent Kaisu Pienihäkkinen, DDS, PhD, and
Docent Janne Pitkäniemi, PhD, for sharing their expertise. Your constructive
criticism and comments were valuable for improving this thesis.
I am grateful to my co-authors Docent Vuokko Anttonen, DDS, PhD, and
Professor Jouko Miettunen, PhD, whose guidance, knowledge and understanding
have provided great support during these years. I want to thank my co-authors
Taina Käkilehto, DDS, Joanna Pihlaja, DDS, Jari Päkkilä, MSc, Jorma Suni,
DDS, PhD, and Sinikka Salo, DDS, PhD, for all the scientific guidance and help
throughout the project.
I warmly thank Outi Hiltunen, MA, for revising the language of this thesis
and all the original articles.
I want to express my warmest gratitude to my dear colleagues in the coffee
room in the old building: Jorma Virtanen, Ahti Niinimaa, Mimmi Tolvanen, Paula
Pesonen, Leena Niskanen, Marja-Liisa Laitala, Toni Similä and Ville Vuollo. You
have provided me with the social support which has been essential to complete
this study.
The funding provided by the Finnish Cultural Foundation and Finnish Dental
Society Apollonia is greatly acknowledged. The permission of the publishers to
reprint the original articles is acknowledged.
Finally, my deepest thanks go to my friends and to my mother, sisters and
their families, without their support this thesis would not have been completed.
Oulu, March 2015 Hannu Vähänikkilä
8
9
Abbreviations
AOS Acta Odontologica Scandinavica
ANOVA Analysis of variance
ART Alternative restorative treatment
CDOE Community Dentistry and Oral Epidemiology
CI Confidence Interval
CONSORT Consolidated standard of reporting trials
CR Caries Research
DMF Decayed, missing due to caries and filled teeth
GIC Glass ionomer cement
HR Hazard ratio
ID Identity
IF Impact factor
JD Journal of Dentistry
JDR Journal of Dental Research
NS Non-significant
PRISMA Preferred reporting items for systematic reviews and meta-analyses
SD Standard deviation
STROBE Strengthening the reporting of observational studies in epidemiology
TTE Time-to-event
10
11
List of original articles
This thesis is based on the following articles, which are referred to in the text by
their Roman numerals.
I Vähänikkilä H, Nieminen P, Miettunen J & Larmas M (2009) Use of statistical methods in dental research: comparison of four dental journals during a 10-year period. Acta Odontol Scand 67: 206–211.
II Vähänikkilä H, Tjäderhane L & Nieminen P (2015) The statistical reporting quality of articles published in 2010 in five dental journals. Acta Odontol Scand 73: 76-80.
III Vähänikkilä H, Miettunen J, Tjäderhane L, Larmas M & Nieminen P (2012) The use of time-to-event methods in dental research: a comparison based on five dental journals over a 11-year period. Community Dent Oral Epidemiol 40 suppl 1: 36–42.
IV Vähänikkilä H, Käkilehto T, Pihlaja J, Päkkilä J, Tjäderhane L, Suni J, Salo S & Anttonen V (2014) A data-based study on survival of permanent molar restorations in adolescents. Acta Odontol Scand 72: 380–385.
V Suni J, Vähänikkilä H, Päkkilä J, Tjäderhane L & Larmas M (2013) Review of 36,537 patient records for tooth health and longevity of dental restorations. Caries Res 47: 309–317.
12
13
Table of contents
Abstract
Tiivistelmä
Acknowledgements 7
Abbreviations 9
List of original articles 11
Table of contents 13
1 Introduction 15
2 Review of literature 17
2.1 Statistical methods in dental research ..................................................... 17
2.1.1 Use of statistical methods ............................................................. 17
2.1.2 Quality of statistical reporting ...................................................... 17
2.1.3 Time-to-event methods in dental research .................................... 20
3 Aims of the study 27
4 Material and methods 29
4.1 Study samples ......................................................................................... 29
4.1.1 Bibliometric samples of dental journals (I, II, III) ........................ 29
4.1.2 Survival of restorations based on individual records (IV) ............ 30
4.1.3 Survival of restorations by past caries history (V) ....................... 30
4.2 Variables used in the original studies ...................................................... 31
4.3 Statistical methods .................................................................................. 34
4.3.1 Bibliometric studies (I, II, III) ...................................................... 34
4.3.2 A data-based survival study (IV) .................................................. 34
4.3.3 Survival study (V) ........................................................................ 35
4.4 Ethical considerations and personal involvement ................................... 35
5 Results 37
5.1 Use of statistical methods in dental research articles (I, III) ................... 37
5.2 The quality of statistical reporting in five dental journals (II) ................ 40
5.3 Use of time-to-event-methods in dental research (III) ............................ 41
5.4 A data-based study on survival of permanent molar restorations
in adolescents (IV) .................................................................................. 43
5.5 Survival of teeth sound and longevity of dental restorations in
Finland (V) .............................................................................................. 45
6 Discussion 49
6.1 Main findings .......................................................................................... 49
6.2 Discussion of the results ......................................................................... 49
14
6.2.1 Use of statistical methods and time-to-event methods in
dental research (I, III) ................................................................... 49
6.2.2 The quality of statistical reporting in dental journals and
guidelines for the presentation of TTE methods (II, III) .............. 51
6.2.3 Survival studies (IV, V) ................................................................ 57
6.3 Strengths and limitations of the study ..................................................... 59
6.3.1 Bibliometric studies (I, II, III) ...................................................... 59
6.3.2 Survival studies ............................................................................ 59
6.4 Main conclusions .................................................................................... 60
6.5 Recommendations how to improve statistical analysis in dental
research ................................................................................................... 60
6.6 Future research ........................................................................................ 61
References 63
Original articles 69
15
1 Introduction
Statistics is a science of collecting, analysing and presenting data. Statistical
methods play an important role in empirical and quantitative dental research.
Descriptive statistics such as frequencies, tables, graphs and other figures are
widely used in dental literature. Descriptive statistics have been complemented
with more formal statistical inference procedures such as test statistics and P-
values. Due to the developments in computer technology, computationally more
demanding and novel statistical methods have also been applied more often.
Bibliometrics is a method of studying written communication by
systematically measuring and analysing research publications (Nieminen 1996). It
covers all aspects of scientific communication; for example, Geminiani et al.
(2013) studied authorship trends in periodontal literature and noticed that the
number of authors per article increased from 1995 to 2010. Bibliometrics can be
used to evaluate methodological aspects of published studies. Statistics and
bibliometrics are both independent sciences but they are also frequently utilized
in other fields of science. The question of how research information is best
communicated to other researches is an important aspect of dissemination of
knowledge. With the help of bibliometrics, statisticians and other researchers can
improve the visibility of their research. Because quantitative studies have become
more common, the role of statisticians in teamwork has increased (Nieminen et
al. 2006b).
As a consequence of the increased use of sophisticated statistical methods
easily available to those that are not experts in statistical analysis, the risk of poor
reporting and the low quality of statistical reporting have increased (Nieminen et
al. 2006b, Baccaglini et al. 2010, Kim et al. 2011). Studies with poor statistical
quality may lead to incorrect conclusions, artificial or false research results and a
waste of valuable resources. Furthermore, the misuse of statistical methods in
medical research can have serious clinical consequences because the evidence
supporting or opposing the research hypothesis is not correctly reported
(Gardenier & Resnik 2002). Poor reporting and errors in the statistical analysis
are very difficult to define because there are no standards on how to assess poor
reporting or misuse of statistical methods.
Some recent review articles have discussed the quality of statistical reporting
and provided related statistical guidelines. This demonstrates an increasing
awareness of the issue in dental research. Krithikadatta and Valarmathi (2012)
attempted to explain the scope of statistics in dental research and provided
16
information on statistical misuse in the Journal of Conservative Dentistry
between the years 2008 and 2011. Hannigan and Lynch (2013) described the
pitfalls of some commonly used statistical techniques in dental research and gave
some recommendations on how to avoid them. Souza (2014) established
guidelines for reporting of data and its statistical analysis.
Time-to-event (TTE) methods measure the follow-up time from a defined
starting point to the occurrence of a given event. The most commonly applied
TTE methods in the estimation of survival experience are the Cox proportional
hazard model and the Kaplan–Meier method. TTE methods have become
increasingly common in medical research over the last few decades (Horton &
Switzer 2005, Strasak et al. 2007). In dental research, TTE methods can be used,
for example, when comparing the survival of amalgam and composite fillings in
the same population (Käkilehto et al. 2009), assessing the effectiveness of fissure
sealants (Leskinen et al. 2008), and examining the survival of ART restorations
with and without cavity disinfection (Farag et al. 2009).
In dental research, decayed (D), missing due to caries (M) and filled (F) teeth
form the very commonly used DMF index values (Klein et al. 1938). The DMF
index has some deficiencies; for example, it is impossible to define the survival of
restorations with the DMF index. For these types of research problems, time-to-
event methods will add tools to investigate dental research problems.
17
2 Review of literature
2.1 Statistical methods in dental research
2.1.1 Use of statistical methods
Several authors have carried out comprehensive studies of medical journals to see
which statistical methods are most frequently used (Emerson & Colditz 1983,
Altman 1991b, Miettunen et al. 2002, Horton & Switzer 2005, Strasak et al.
2007, Yergens et al. 2014). These studies have shown that there has been a great
increase in the use of statistical methods in medical research. The availability of
statistical software packages and the improved computer technology have resulted
in the use of more sophisticated statistical techniques. Horton and Switzer (2005)
reported that in more than half of the articles in the New England Journal of
Medicine in 2004 and 2005 statistical methods such as survival analysis or
multiple regression analysis were used. The use of statistical methods in medical
research has been a topic of considerable debate in recent years, and there has
been a wide consensus that quality is generally low (Strasak et al. 2007).
In dental research, there have been only few studies which have examined the
use of statistical methods. Yang et al. (2001) estimated the availability of
statistical information in paediatric dentistry, and Lesaffre et al. (2007) concluded
that the split-mouth study design and analysis would benefit from improvement in
the use of statistical methods. Kim et al. (2011) reviewed 418 papers published
between 1995 and 2009 in ten well-established dental journals. They reported that
descriptive statistics and simple methods such as chi-square tests or t-tests are still
widely used tools in dental research. Choi et al. (2014) studied the statistical
methods used in articles published by the Journal of Periodontal and Implant
Science and concluded that an increasing trend towards the application of
statistical and non-parametric methods in recent periodontal studies was seen.
2.1.2 Quality of statistical reporting
The importance of statistical quality in medical research has increased in recent
years due to the greater complexity of statistics in medicine (Fernandes-Taylor et
al. 2011). The increase in the use of statistical methods can also be seen in dental
research (Kim et al. 2011). As a consequence of the wider use of statistical
18
methods, the risk of poor reporting and methodological errors may have increased
(Nieminen et al. 2006b, Baccaglini et al. 2010, Kim et al. 2011) despite that
editors and statistical reviewers have long been aware of the existence of these
problems (Altman 2002). The reported proportion of erroneous articles has been
about 30–50% (Kim et al. 2011).
Study design has an important role when planning a research. Study design
can be categorized into experimental studies (clinical trials and animal studies)
and observational studies (cross-sectional studies, cohort studies and case-control
studies). The problem in using observation studies is that they can often ‘happen’
instead of having been designed and data sets which have been collected for
another purpose are used (Altman 1991a). This can result in missing data and also
variation in the measurement methods and the number and quality of the
measurements by different assessors (Hannigan & Lynch 2013). There are
guidelines for researchers regarding their intended research design, such as the
CONSORT guideline (Schulz et al. 2010) for reporting randomized, controlled
trials, the STROBE guideline (von Elm et al. 2008) for observational studies, and
the PRISMA guideline (Liberati et al. 2009) for systematic reviews. Following
guidelines helps researchers to ensure a common minimum standard for reporting
and a critical level of transparency in medical research.
The most common statistical reporting problems are as follows: inadequate
description of statistical procedures, statistical software not reported, lack of
justification for the sample size reported, problems in tables or figures, and
overuse of P-values or confidence intervals with multiple comparisons (Lang &
Secic 1997, Altman 2001, Nieminen et al. 2006a). A sufficient description of the
methods is essential for a research article to be read and cited frequently
(Nieminen et al. 2007). In addition, the reader of the article should be able to
understand how the statistical analysis has been done. Authors are recommended
to report the statistical software they have used in the analyses because different
statistical software packages use different calculation algorithms (Okunade et al.
1993). Sufficient reporting helps critical readers to judge the methodological
aspects of an article and to repeat the statistical analyses when the author
indicates the software used to run the statistical computations.
The lack of sample size calculations is common in medical research (Altman
1998, Lucena et al. 2011). Sample size calculations are important because P-
values evaluating statistical significance are highly dependent on the sample size
and the validity of statistical inference can be affected by a small number of
events per variable. Textbooks of medical statistics require that the sample size
19
should be large enough and some justifications for the sample size should be
given (Armitage et al. 2002). Several journals require the authors to adhere to the
CONSORT guidelines in randomized controlled trials, and those guidelines
require reporting how the sample size was determined. In dental research, the
Journal of Dental Research, Journal of Dentistry, Caries Research, Community
Dentistry and Oral Epidemiology, European Journal of Oral Sciences,
International Endodontic Journal and Journal of Endodontics are some examples
of such journals. However, it is not clear if the journals actually check the
adherence to the guidelines in the reviewing process. This could partly explain the
low use of sample size calculations in dental journals.
Tables and figures are the most efficient way to convey information
especially from a large data set. Tables are used to present background
information related to the methods used and to present the data used. The most
typical problems in tables are that numbers do not add up and that there are no
percentages to help the readers to compare the proportions between the groups.
Figures are useful for presenting data from individual subjects, for example, in a
scatter plot or trends for time (Hannigan & Lynch 2013). Simple rather than
visually too complex effects should be preferred. The main purpose of various
summary statistics is to show essential features of the data, not to steer the
attention of viewers to irrelevant elements of the figure (Tufte 1983). The most
typical problems in figures are insufficient legends and scale problems. Iverson et
al. (1997) give valuable advice in their textbook on how to present tables and
figures in a convenient way in journal articles (e.g. complete titles and complete
row and column headings).
Overuse of P-values increases the risk of making a type I error, which is an
incorrect rejection of a true null hypothesis. It is also called the false-positive
result. In health sciences, it may cause unnecessary worry or treatment. This
multiple testing problem may occur when testing several baseline characteristics
of differences between groups, performing secondary analyses, or performing
subgroup analyses which are not in the original study plan. It may also occur
when performing multiple pairwise comparisons or comparing groups at multiple
time points. If there are hundreds of P-values, the risk of some P-values to be
small just by chance increases (Motulsky 1995). This is a problem especially
when omics and large registry data is available in health research.
20
2.1.3 Time-to-event methods in dental research
Basic concepts
Time-to-event (TTE) methods measure the follow-up time from a defined starting
point to the occurrence of a given event, such as the time from the eruption of a
tooth to the detection of a caries lesion or a filling, or censoring, i.e. no event
during the follow-up (Ollila & Larmas 2007). TTE methods include methods such
as the life-table analysis, Kaplan–Meier survival curves and Cox proportional
hazards models. The use of TTE methods has become increasingly common in
medical research (Horton & Switzer 2005, Strasak et al. 2007). However,
Otwombe et al. (2014) reported that further developments of Cox regression (e.g.
time-dependent models and Cox frailty models) are scarce and concluded that this
motivates more training in the use of advanced time-to-event methods. The
increase in the use of TTE methods has also been seen in dental research
(Hannigan & Lynch 2013).
What makes survival data different from other types of data is the concept of
censoring. The survival time of tooth is said to be censored when, for example,
the event has not been observed for that tooth during the observation period. Each
subject is followed for a certain period of time, and the event must be censored if
it did not occur at the completion of the follow-up or had happened before the
follow-up (Kleinbaum & Klein 2005).
An example of censoring can be taken form the survival of teeth before being
restored. If a tooth has already erupted or been restored at the first examination,
no exact date can be recorded for these events before the time. Such survival data
are called in statistical analyses left-censored. If the observation time is
terminated at a certain time point, the failure time of those teeth are right-
censored after that, i.e. they may happen but are not included in the analysis. If
there is a long time between the examinations, interval-censored observations can
be recorded as having occurred at the first examination following the event, or
alternatively at the midpoint of the two last examinations (Aalen et al. 1995).
Hannigan & Lynch (2013) defined competing risk as an event that is of equal
or more significant clinical importance than the event of interest and alters the
probability of this outcome. An example of a competing risk in dentistry is the
exfoliation of a primary molar in a child, which means that the event of caries
cannot be recorded for that molar. Exfoliation is described as a competing risk,
i.e. caries and exfoliation are competing against each other to occur first for the
21
molar. In this kind of a situation, competing risk survival analysis is
recommended. It means that the exfoliated molars and the time to exfoliation are
accounted for, by appropriate weighting, in a multivariate analysis (Hannigan &
Lynch 2013).
Time-to-event methods, Cox regression model
An event can be defined as a transition from one state to another. In time-to-event
analysis, it is commonly assumed that the exact timing of occurred events is
known, allowing the use of a broad scope of statistical methods of survival
analysis (Kalbleisch & Prentice 1980). To obtain a first insight into the data, the
association between the patient covariates and response variable measuring the
event under study should be analysed by cross-tabulation. This cross-tabulation
exposes the bivariate relationship between each factor and the response variable,
but does not take into account the possible simultaneous effect of two or more
variables on the response variable. It is useful to tabulate the events by follow-up
time.
At the second stage, the Kaplan–Meier survival analysis procedure (Kaplan
& Meier 1958) can be used to obtain the estimated survival curves of patients.
Several non-parametric statistical tests are available for evaluating the null
hypothesis that in the population, two or more time-to-event curves are equal. The
tests are based on computing the weighted differences between the observed and
expected number of events at each time point. The statistical significance of the
difference between the groups is usually determined by the log-rank statistic.
The Cox proportional hazard model (Kalbleisch & Prentice 1980) is
commonly used in the multivariate analysis of the data to describe the relation
between covariates and time to event, e.g. time to re-restoration. As an alternative
to logistic regression, the Cox regression model can utilize the time to event and
make adjustments for various lengths of follow-up (i.e. not all patients are
followed for the same period of time). The risk set is the number of subjects who
have not yet experienced the response at the beginning of the ith interval. The
general form of the fitted model can be written in terms of the hazard function as:
h(t) = [h (t)] exp(b x +…+b x ),
where b1, …, bp are the estimated parameters and x1, …, xp are the explanatory
factors. For example, if a tooth with the explaining factors x1, …, xp has not been
re-restored by time t, then h(t)dt approximates the risk to be re-restored by
22
time t+dt. Thus, the above hazard function h(t), or re-restoration rate at time t,
tells us how likely a tooth is to experience a re-restoration in the next short time
interval, given that the tooth has remained sound until that time. The baseline
hazard, h0(t), is assumed to be the same for all teeth and to depend only on time. It
is an unspecified function and this is the property that makes the Cox model a
semiparametric model.
Time-to-event methods in dentistry
The first study to use time-to-event methods in dentistry is the classic study by
Carlos and Gittelsohn (1965). They applied life-table estimates of logarithmic
failure intensities separately to different teeth. Afterwards, some results of their
study have been questioned. The follow-up time was divided into four-month
computational intervals, in which interval-censored occurrence times were
denoted to have taken place at the mid-points of the dental examination interval.
Left-censoring so that all permanent teeth that had erupted before the first clinical
examination was conducted leaving only 2,104 children out of 7,400 to be
analysed in the data set. Such exclusion of data due to censoring and not
accounting it properly in the analysis resulted in systematic biases (Härkänen et
al. 2002) because several caries-prone teeth were left-censored.
Time-to-event methods have been applied to re-analysing longitudinal caries
data in many other clinical trials (Hannigan et al. 2001, Baelum et al. 2003).
Recently, multivariable multilevel parametric survival models have also been
applied at the tooth surface level to the analysis of sound-carious and sound-
exfoliation transitions to which first and second primary molar surfaces were
subject (Stephenson et al. 2010a).
More recent studies concerning with TTE methods are by Tonetti and Palmer
(2012), who made specific recommendations to improve the quality of reporting
of clinical research in implant dentistry. In addition, Layton and Clarke (2014a)
assessed the Medical Subject Headings indexing of articles that employed TTE
analyses to report the outcome of dental treatment in patients and concluded that
the Medical Subject Headings allocation within MEDLINE to TTE dental articles
was inaccurate and inconsistent. Layton and Clarke (2014b) have also studied the
quality of reporting of dental survival analyses.
When more sophisticated survival analyses are used, for example, in clinical
trials in which different variables and co-factors are determined in the evidence-
based dentistry, multivariable TTE methods like the Cox regression model are
23
conducted. In these cases, the variables and their impact and timing need to be
carefully determined.
Multivariable TTE methods have been used for analysing the relationship
between caries in primary molars and cavity formation of first permanent molars
in a follow-up study (Leroy et al. 2005a, Leroy et al. 2005b). Stephenson et al.
(2010b) reported that the socioeconomic class, fluoridation status and surface
type were found to be the strongest predictors of caries in primary molars. They
applied the multilevel competing risks method to identify factors associated with
caries occurrence in the presence of a concurrent risk of exfoliation, but not the
effect of different emergence times between the first and second primary molars.
Multivariate multilevel parametric survival models were applied at the surface
level to the analysis of the sound-carious and sound-exfoliation transitions to
which primary tooth surfaces are subject. In earlier studies, the multiple teeth had
been used as indicators or predictors, but Stephenson et al. (2010b) found that
clustering of data had little effect on inferences of parameter significance.
Time-to-event methods in practice-based cariology research
Today, the use of digital dental records, in which the observations are made at the
tooth surface level and the interval of examinations is predetermined
automatically, creates large datasets (big data). This provides excellent
opportunities for large-scale epidemiological studies. In these datasets, the
problem is the registration of caries because it does not meet the requirement of
calibrated observations, although observations and recordings of missing teeth are
very accurate (Larmas 2010).
The tooth-specific system has been used in longitudinal analyses of caries
onset as seen in normal dental records in practice-based dentistry. In these studies,
one tooth is treated as an indicator of dental health of that oral cavity, and these
teeth are compared to their counterparts in other study subjects (Larmas et al.
1995, Suni et al. 1998, Komarek et al. 2005, Ollila & Larmas 2007). All these
studies used the non-parametric Kaplan–Meier method or Bayesian analysis for
comparing tooth-specific survival data.
Non-parametric methods have been used for the determination of the survival
time of each tooth remaining caries-free either from the tooth emergence
(Virtanen & Larmas 1995, Suni et al. 1998) or from the birth of the subject to
20 years of age (Leskinen et al. 2008).
24
The variation in intensities between patients can be analysed by means of a
frailty type model also in non-parametric models (Clayton 1991). In frailty
models with a univariate survival time as an endpoint, a univariate lifetime is used
to describe the influence of unobserved covariates in a proportional hazards
model. When multivariate survival times are used as endpoints, the aim is to
account for the dependence in clustered event times, e.g. in the lifetimes of
patients in study centres in a multi-centre clinical trial, caused by centre-specific
conditions (Andersen et al. 1999). The intensity of frailty of a particular
individual can be determined by a frailty or mixing variable that is a random
variable for the population of individuals. Hence, any given individual has a
specific frailty. Härkänen et al. (2000) applied the Bayesian analysis and frailty
model so that the baseline intensity functions were modelled non-parametrically
and the frailty component was introduced by using cross-validation techniques by
assuming a hierarchical gamma especially prior to finding the ages at which the
intensity model can provide better predictions than the simple models using
DMFS-index values. Chuang et al. (2005) used the frailty approach to account for
the association between the survival times of multiple implants within the same
subject. The problem with applying methods which assume independence
between implants from the same subject is that they may result in invalid
inferences: variances are poorly estimated, P-values are too low and some results
reported as statistically significant are non-significant.
Advantages of using time-to-event methods
As Hannigan (2004) stated in her review of TTE methods applied to caries
outcomes, the additional statistical system provides an approach that can be
understood easily. When the methodology proposed here is either subject- or
tooth-based, it uses all the data collected during the clinical treatment of the
patient during his/her lifetime in practice-based dentistry, or calibrated data in
clinical trials or prevalence studies in evidence-based dentistry. It includes data
collected at clinical examinations and data from subjects in intermediate
examinations as long as the subjects are visiting that dental office or attending the
trial, but censoring can be accounted, too. The results of the analysis are easily
interpreted with the use of survival curves and median survival times.
Survival analysis is time-consuming, and the frailty model approach demands
a lot of computer capacity and is best suited to the scientific analysis of
longitudinal clinical trials in evidence-based dentistry. A simplified approach to
25
the determination of lifetime of sound teeth can be developed for practice-based
dentistry, which could be automatically performed when digital records are
available.
26
27
3 Aims of the study
The bibliometric part of this study (I, II, III) utilized a data set of dental articles
to:
1. Investigate the development of the use of statistical methods in dental
journals.
2. Assess the quality of statistical reporting of high-impact dental journals.
3. Evaluate the reporting of statistical techniques and results in dental research
papers with special reference to time-to-event methods and to create
guidelines for the appropriate reporting of these methods.
The part of this study discussing time-to-event methods (IV, V) utilized empirical
data and aimed to:
4. Demonstrate the strength of time-to-event methods in collecting detailed data
of the development of oral health.
28
29
4 Material and methods
4.1 Study samples
4.1.1 Bibliometric samples of dental journals (I, II, III)
To evaluate the use of statistical methods in dental journals, articles from five
dental journals were reviewed. Editorials, letters, case reports and review articles
were excluded. The author reviewed all the articles. If the interpretation of the
specific article was unclear, the author consulted an experienced biostatistician
(Docent Pentti Nieminen).
The evaluation was limited to articles published in the following dental
journals: Journal of Dental Research (JDR), Journal of Dentistry (JD), Caries
Research (CR), Community Dentistry and Oral Epidemiology (CDOE), and Acta
Odontologica Scandinavica (AOS). JDR is a general dental and oral health and
disease research journal with relatively heavy emphasis in basic biological
research. JDR is also the most visible and cited dental journal. Like JDR, JD is a
general dental and oral health research journal, but has a more specific aim to
influence the practice of dentistry at different levels. CR has the highest impact
factor of cariological journals and CDOE represents an epidemiological approach
to dentistry. AOS is a European dental journal, mainly directed to and subscribed
in the Nordic countries, which are the leading countries in using preventive
methods in dentistry. In 2010, the impact factors of the five journals were: JDR
3.773, JD 2.115, CR 2.926, CDOE 2.328 and AOS 1.130.
In Study I, 1,020 articles from four dental journals (JDR, CR, CDOE, AOS)
were reviewed. The journals were scrutinized for original research articles
published in 1996, 2001 and 2006. The total number of reviewed articles, original
studies reporting based on the statistical analysis were 928, representing about
91% of all articles in the four journals.
In Study II, presuming 40% error rate in dental papers (based on previous
literature review of erroneous articles in Chapter 2.1.2), the sample size of
100 articles was calculated to be the minimum number for the present purpose,
allowing a maximum difference of 10 percentage points between the sample rate
and true population rate at a 95% significance level. However, we anticipated that
20 articles per journal is too low a number of papers to make comparisons
between the journals. Forty subsequent articles published in 2010 from each
30
journal, with the starting articles chosen randomly, were analysed. The total
number of articles reviewed was 200, and each paper underwent careful scrutiny
for the use of statistical methods and reporting.
In Study III, to evaluate the current use of TTE methods and other statistical
methods in dental literature, five dental journals were scrutinized for original
research papers published in the years 1996, 2001 and 2006, following the setup
described in Study I. To obtain a more comprehensive view of the most recent use
of TTE methods, eligible publications from the years 2005 and 2007 were also
evaluated. The total number of original papers reviewed was 1,985, representing
about 94% of papers contained in the five journals (N=2,111).
4.1.2 Survival of restorations based on individual records (IV)
The data were collected from the electronic patient files of oral health documents
of the City of Vantaa, Finland covering the period 1992–2005. The use of an
electronic patient file system started in Vantaa in 1989 enabled the access to data.
Structural data made assessing time to events easier. To study the survival of
restorations on the first permanent molars, we chose two birth cohorts: those born
in 1990 (N=2,975) and in 1995 (N=3,147). This allowed monitoring the
restorations of the first permanent molars starting from their placement for
maximum 10 years in the older cohort and 5 years in the younger cohort.
4.1.3 Survival of restorations by past caries history (V)
All patients born in 1985 (N=11,399), 1990 (N=12,423), or 1995 (N=12,715) who
received dental care in 28 municipal health centres in different parts of Finland
with numerous dental offices and hundreds of operating dentists formed the study
cohort (N=36,537). Subgroups of these patients were formed according to their
past caries experience in any of the first molars before the age of 8 (caries-prone)
or after 10 years (caries-resistant), the rest forming an inter-medial group (Fig. 1).
31
Fig. 1. Distribution of the subjects in each age cohort into caries-prone, inter-medial
and caries-resistant subjects.
4.2 Variables used in the original studies
Study I
The scientific articles were classified into the following groups: non-experimental
studies (including cross-sectional surveys, cohort studies and case-control
studies), experimental and other studies (including reliability, methodology and
basic science studies). A study is non-experimental when the investigator does not
participate in the study arrangements. An experimental study group consists of
clinical trials and animal studies. A study is experimental when one or more
variables is controlled by the investigator in order to monitor its effect on the
process or the outcome. Same classification, with slight modification, was used
by Nieminen et al. (2007) when analysing psychiatric literature.
The types and frequencies of statistical methods were systematically recorded
and classified into three categories. T-tests, simple contingency analysis, non-
32
parametric methods, one-way ANOVA and simple correlation techniques were
considered as basic methods. Multivariate methods included regression models
and other methods such as factor analysis, cluster analysis and discriminant
analysis. Reliability analysis, survival analysis and other statistics were
considered as specific methods.
Papers were categorized as using P-values if the results of statistical
significance testing were reported. The use of confidence intervals and multiple
comparisons was also recorded if these were reported. The usage of statistical
tables and figures was also assessed. The studies were divided into four groups on
the basis of their sample sizes. In the first group, the sample size was under 30, in
the second group 30–99, in the third group 100–300 and in the fourth group
over 300.
In addition, the following information was also recorded: extended
description of statistical procedures, reference to statistical literature and reporting
of the statistical software used.
Study II
The following aspects were recorded (binary variables with codes 1=found, 2=not
found) as poor reporting, as defined by Lang and Secic (1997), Altman (2000)
and Nieminen et al. (2006a):
1. Incomplete description of procedures (e.g. failure to specify all methods used,
wrong names for statistical methods, misuse of technical terms such as
quartile);
2. Statistical software not reported;
3. Sample size not reported;
4. No justification for the sample size reported (power calculations, sample size
calculations, budget restrictions);
5. Inexact P-values reported (P=ns, P>0.05, P=0.0000, etc.);
6. Problems in tables or figures (e.g. insufficient title or legends, numbers not
adding up, no percentages to help comparison, scale problems, using figure
types not informative about the variable scale);
7. Data dredging with multiple comparisons (number of P-values over 50 or
number of P-values greater than the number of subjects in the study).
33
A paper with at least one poor reporting item was classified as ‘problems with
reporting statistics’ and a paper without any poor reporting item was classified as
‘acceptable’.
Study III
The papers were classified as described in Study I, but with slight modifications,
into the following three groups: non-experimental studies (including cross-
sectional surveys, cohort studies and case-control studies), experimental research,
and others (including reliability, methodology and basic scientific studies).
Study IV
The collected data included information/results on both the dental examination
and the treatment of the cohort members on each visit. The relevant time variables
were recorded in accuracy of days. All information on surfaces of the first
permanent molars concerning dental caries needing restorative treatment was
registered (ID, gender, date of birth, all event dates, placement of filling,
replacement of filling, operator, tooth, tooth surface, material, and extraction due
to dental caries). The recorded restoration materials were as follows: composite,
glass ionomer (combining conventional glass ionomer, compomer and resin-
modified glass ionomer), and amalgam.
The findings regarding the restorations by 120 operators on different surfaces
of the first molars were prepared for data for analyses. The date of any restoration
was the starting point for follow-up and re-restoration was the event of interest.
Study V
The inclusion criterion in this study was that each study subject had to have
visited the health centre for dental examination and a dental chart coding the tooth
surface-specific observations had been filled. Patients with only sporadic or
emergency visits were not included. An intermediate file containing the following
information was compiled from the codes in the electronic dental record: date of
birth and gender from the social security number, the presence of a carious lesion
extending into dentin on each tooth surface from the dental chart, and operations
of restorations from the codes in the progress notes. The identification portion of
34
the social security number was excluded from this intermediate file before
analysis and thus the subjects and patients could not be identified from these files.
4.3 Statistical methods
4.3.1 Bibliometric studies (I, II, III)
The reliability of the evaluation was checked by comparing the ratings of two
independent reviewers. The frequencies of statistical procedures used in the
journals were reported. The chi-square test was used to assess differences in the
use of statistical procedures between the journals and the Cochran–Armitage test
to investigate the trend between the publishing years.
4.3.2 A data-based survival study (IV)
The mean age (SD) at which the first restoration of the first molar was placed was
calculated for both cohorts. The proportions of individuals with intact teeth in
both cohorts at the age of 7 and 10 years were calculated. The number and
proportion of restored teeth at baseline and their survival rates at 1 and 3 years
were calculated for both materials and both cohorts.
The Kaplan–Meier survival curves were drawn separately for both cohorts,
for different materials, and for restorations on two or more surfaces. The log-rank
test was used to compare the restoration survival curves. For analysing the
survival of the restorations, right-censoring was used. We considered a
person/teeth to be right-censored when the replacement time was not known,
when the restoration remained in place at the end of the monitoring period, or
when the patient was no longer available for follow-up during the regular check-
up visits (Käkilehto et al. 2009).
The Cox proportional hazards model was used to obtain hazard ratios (HR)
and their 95% confidence intervals to investigate the association of the hazard
between the restorations and the predictor variables (cohort, gender, restorative
material). The assumptions of the Cox regression model were checked graphically
by drawing Kaplan–Meier curves.
35
4.3.3 Survival study (V)
In this study, the probability of avoiding caries was estimated by the Kaplan–
Meier method which was used as a part of the data analyses for each permanent
tooth from the birth of the patient to the diagnosis of a carious lesion extending
into dentin (event of interest) or censoring. This was performed first for each
permanent tooth in the maxilla and mandible, then for all tooth surfaces in each
health centre, after which all surfaces were pooled. The health centres were
analysed both combined and separately for each age cohort.
Another event of interest was the re-restoration of a teeth and therefore the
date of the primary restoration was recorded when a caries lesion recorded in the
dental chart was replaced by a restoration in the progress notes. The re-restoration
was recorded when any of the recorded primary restorations were re-restored for
any reasons. Restorations recorded only in the dental chart were not included in
the survival analyses, because the exact onset time was not available. This
comment concerns those patients that have changed health centres or municipality
during the era of digital records.
Survival of re-restorations for all teeth combined and for each tooth
separately was evaluated for each health centre and each age cohort as well as
separately for caries-prone and caries-resistant subjects.
The age of the subjects and the dates of examination and restorative treatment
were recorded with an accuracy of one day. The dates of caries onset were
recorded to have occurred at the first examination following the event. In practice,
the first observations in permanent teeth were made after 6 years of age
After the data mining procedure, the data from the intermediate file were
transferred to the SAS survival procedure (v. 9.2 SAS Institute, Cary, NC) and
Kaplan–Meier curves were drawn from the intermediate file.
4.4 Ethical considerations and personal involvement
The Committee of Ethical Affairs of the Oulu University Hospital and the
corresponding committees governing the 28 health centres granted their
permission for this study. The children were examined during regular oral health
controls, and the data was confidential, meaning that it was not possible to match
a certain health status to an individual child.
The author of this thesis has evaluated all the articles (N=2,185) used in the
bibliometric part of this thesis and entered the data into a computer file. The
36
author has participated in designing and reporting of all the bibliometric articles
(I, II, III). The author has participated in designing and reporting of both survival
studies (IV, V).
All the statistical analyses, presentations of methods and presentations of the
results in the original articles and this summary were done by the author.
37
5 Results
5.1 Use of statistical methods in dental research articles (I, III)
Of the evaluated articles, only 27% were experimental studies (clinical trials or
animal studies) and a clear majority observational studies. The journals differed
from each other: JDR had an equal amount of experimental and observational
studies, CR had 25% experimental studies and 52% observational studies, while
AOS and CDOE had only about 20% experimental studies and about 70%
observational studies. There were more observational studies in 2006 compared to
1996 and 2001; however, the differences between the years were not statistically
significant by a trend test (Fig. 2).
Fig. 2. Study design by journal and publishing year.
The sample sizes in the studies significantly varied between the journals. In JDR,
almost half of the articles had a small (<30) sample size. In CR and in AOS, 30%
and 18%, respectively, of the articles had a small sample size. In CDOE, 69% of
the articles had a sample size of over 300. The distribution of the sample size was
38
quite similar between the years 1996 and 2001, while the year 2006 articles had
more studies with smaller sample sizes (P<0.001) (Fig 3).
Fig. 3. Study sample size by journal and publishing year.
A total of 343 articles (37%) reported only basic statistical methods. The
exception was CDOE, in which only 18% of the articles reported basic statistical
methods, which is about 50% fewer compared to the other journals. Multivariate
or specific methods were used in 39% of the articles, but in CDOE multivariate or
specific methods were published in 72% of the articles. JDR and CR published
studies with multiple comparisons more than AOS and CDOE (Table 1).
Statistical significance testing was still a generally used method in all four
journals: P-values were reported in 81% of all articles. Confidence intervals were
less frequently reported in JDR, while they were more frequently used in CDOE.
Results were also often presented in graphical form, although there were large
differences between the journals (Table 1).
The journals had differences in the documentation of statistical methods and
software used. An extended description of data analysis procedures was more
frequently presented in CDOE, CR and AOS than in JDR. The software used was
reported in 36% of the evaluated articles. A reference to statistical literature was
39
provided in 50% of the evaluated articles published in CDOE. In CR and AOS
statistical literature was referred to in 30% of the articles, while statistical
references were included only in 15% of the articles in JDR.
Table 1. Statistical methods used in the journals (Study I, published by permission of
Informa Healthcare).
Variable JDR n=404 CR n=193 AOS n=159 CDOE n=172 Total n=928 Sig.1
n % n % n % n % n %
Basic methods only 183 45.3 67 34.7 62 39.0 31 18.0 343 37.0 <0.001
Multivariate methods 105 26.0 74 38.3 62 39.0 124 72.1 365 39.3 <0.001
Multiple comparisons 119 29.5 55 28.5 25 15.7 15 8.7 214 23.1 <0.001
P-value 319 79.0 162 83.9 125 78.6 141 82.0 751 80.9 =0.447
Confidence interval 54 13.4 46 23.8 39 24.5 56 32.6 195 21.0 <0.001
Statistical figure 299 74.0 120 62.2 82 51.6 67 39.0 568 61.2 <0.001
Statistical tables 267 66.1 157 81.3 146 91.8 164 95.3 734 79.1 <0.001
Reference to
statistical literature
60 14.9 59 31.4 50 31.4 86 50.0 255 27.5 <0.001
Software reported 110 27.2 89 46.1 59 37.1 76 44.2 334 36.0 <0.001
Extended description
of procedures
229 56.7 142 73.6 110 69.2 129 75.0 610 65.7 <0.001
1 Significance from Chi-square test
The developments in the use of statistical methods from 1996 to 2006 are
described in Table 2. In 1996, more articles used only basic statistical methods
compared to the years 2001 and 2006. There were no statistically significant
differences between the years 1996–2006 with respect to the use of multivariate
or specific methods. The use of multiple comparisons, P-values or significance
tests and confidence intervals increased significantly from the year 1996 to 2006.
Reporting results in tables was a widespread method, although it decreased from
1996 to 2006. From 1996 to 2006, the use of figures did not change statistically
significantly. Both providing an extended description of procedures and reporting
software increased from 1996 to 2006. There were no differences between the
years 1996, 2001 and 2006 in terms of references to statistical literature.
40
Table 2. Statistical methods used in different years.
Variable 1996 n=310 2001 n=301 2006 n=317 Total n=928 Sig.1
n % n % n % n %
Basic methods only 137 44.2 103 34.2 103 32.5 343 37.0 =0.002
Multivariate methods 123 39.7 121 40.2 121 38.2 365 39.3 =0.698
Multiple comparisons 50 16.1 72 23.9 92 29.0 214 23.1 <0.001
P-value 234 75.5 243 80.7 274 86.4 751 80.9 <0.001
Confidence interval 50 16.1 67 22.3 78 24.6 195 21.0 =0.009
Statistical figure 196 63.2 168 55.8 204 64.4 568 61.2 =0.761
Statistical tables 271 87.4 251 83.4 212 66.9 734 79.1 <0.001
Reference to
statistical literature
81 26.1 86 28.6 88 27.8 255 27.5 =0.650
Software reported 88 28.4 102 33.9 144 45.4 334 36.0 <0.001
Extended description
of procedures
172 55.5 197 65.4 241 76.0 610 65.7 <0.001
1 Significance from Cochran–Armitage test for trend.
5.2 The quality of statistical reporting in five dental journals (II)
Of the 200 papers published in 2010, 182 (91.0%) had problems in statistical
reporting. The most common problem was the lack of justification for the number
of cases (80.0%). About one third of the papers had incomplete description of
statistical procedures or the statistical software was not reported. Inexact P-values
were reported in 29% of the articles. In addition, the following problems were
seldom encountered: sample size was not reported, there was data dredging with
multiple comparisons, and there were problems with the analyses (Fig. 4).
When comparing the journals, problems with poor reporting varied between
82.5–97.5% of the papers (Table 3). The sample size was quite well reported in
all journals (80–100% of the articles) and there were only few problems in tables
and figures (0–15% of the articles). A notable variation between the journals was
an incomplete description of statistical procedures (17.5–45.0%) and data
dredging with multiple comparisons (0–32.5% of the articles).
41
Fig. 4. Percentage of problems in five dental journals by reporting items.
Table 3. The misuse rate of reporting problems in five dental journals.
Journal Misuse rate
N %
Acta Odontologica Scandinavica 33 82.5
Community Dentistry and Oral Epidemiology 36 90.0
Caries Research 36 90.0
Journal of Dentistry 38 95.0
Journal of Dental Research 39 97.5
All journals 182 91.0
5.3 Use of time-to-event-methods in dental research (III)
Fifty-six articles (2.8% of the total 1,985) used TTE methods, with a significant
increase in their frequency from 1996 (2.2%) to 2007 (4.2%) (P=0.034). TTE
methods were used more in experimental studies (clinical trials or animal
experiments) than in the other types of research, but the observational articles
formed a clear majority in both groups. There were significant differences in the
42
sample sizes between the majority of articles in the article set using TTE-methods
and in the article set not applying TTE-methods. Of the articles using TTE
methods, 52.7% reported having a sample size of over 300 and only 1.8% a
sample size under 30. The sample size distribution of the other articles was much
more even.
The percentage distributions of the reporting characteristics between the TTE
articles and the others are shown in Table 4. A justification for the number of
cases examined was given in 11% of the TTE papers and 8% of the other papers.
An extended description of the data analysis procedures was also given more
frequently in the TTE articles. The statistical software used was reported in 64%
of the TTE articles but only in 38% of the other papers. Traditional standard
software packages (SPSS, SAS, BMDP and Stata) were mainly used (in 85.4% of
the papers that reported the software used), and similarly a reference to statistical
literature was made in 68% of the TTE articles but only in 22% of the other
papers. Reporting the results in the form of tables was widespread in both groups.
Almost all the TTE articles used tables, and a higher percentage of the TTE
articles also used figures.
Table 4. Statistical methods and reporting in time-to-event articles and other articles.
Variable TTE-articles (N=56) Other articles (N=1,929) Sig1
N % N %
Statistical tables 55 98.2 1518 78.7 <0.001
Statistical figure 44 78.6 1172 60.8 0.008
Software reported 36 64.3 733 38.0 <0.001
Reference to statistical literature 38 67.9 433 21.9 <0.001
Description of procedures 53 94.6 1225 63.4 <0.001
Justification for number of cases 5 8.9 147 7.6 0.547
1Significance from Fisher’s exact two-sided test
The most commonly used statistical technique was the log rank test to compare
Kaplan–Meier curves, which was used in 28 articles (50% of the TTE articles),
while the Cox proportional hazards method was employed for multivariate
modelling in 15 papers (26.8%) and the Weibull modulus method in 9 papers
(16.1%). Details about the reporting in the TTE articles are given in Fig. 5.
43
Fig. 5. Reporting of methodological details in articles applying time-to-event methods.
5.4 A data-based study on survival of permanent molar
restorations in adolescents (IV)
The sizes of both birth cohorts and numbers of male and female participants were
almost equal. The proportion of participants with sound first molars at the age of
7 years was 96.6% in the 1990 birth cohort and 94.6% in the 1995 birth cohort. At
the age of 10, the proportion of participants with sound first molars for the 1990
cohort was 49.0%. The mean age of placing the first restoration was 8.4 years
(SD 0.9) in the 1990 birth cohort and 8.2 years (SD 0.9) in the 1995 birth cohort.
Boys had their first restorations placed on average at the age of 8.2 (0.9) years,
and the girls at the age of 8.3 (0.9) years. The proportions of different restorative
materials were similar in both birth cohorts. Almost 4 out of 5 restorations were
composites.
The survival of the restorations in the 1995 birth cohort was lower compared
to the 1990 birth cohort for both the glass ionomer and the composite materials.
(Fig. 6). In the 1995 birth cohort, the risk of the replacement of restorations was
higher compared to the 1990 birth cohort (HR 3.02 (1.97–4.64)). The risk of
44
replacement of glass ionomer was two-fold (HR 2.27 (1.62–3.16)) and of
amalgam seven-fold compared (HR 6.79 (3.03–15.18)) to composite restorations.
Fig. 6. Survival of composite and glass ionomer restorations of the participants in the
1990 and 1995 birth cohorts (Study IV, published by permission of Informa
Healthcare).
During the follow-up period, the survival of tooth-coloured restorations on the
occlusal surface of molars was lower in the 1995 cohort compared to the 1990
cohort. In terms of the restorations on the occlusal surface with the longest
monitoring period (the 1990 cohort), the composite restorations achieved the
highest survival rate. The median survival time for the amalgam restorations was
1.2 years, for the glass ionomer restorations 4.5 years, and for the composite
restorations 6.9 years, respectively. The survival time of the glass ionomer
restorations on the interproximal surfaces of permanent molars was shorter than
that of the composites. In the 1995 cohort, data for the tooth-coloured restorations
on the interproximal surfaces of permanent molars were almost non-existent. In
both cohorts, the survival of the tooth-coloured restorations on the occlusal
surfaces was significantly lower than on the interproximal surfaces of molars
(P=0.023 for glass ionomer and P=0.034 for composite, log-rank test overall)
(Table 5).
45
Table 5. Survival rates (%) and 95% confidence intervals of the composite and glass
ionomer (GIC) restorations on the occlusal and interproximal surfaces of molars
combining the 1990 and 1995 cohorts.
Surface and material Survival rates (95% confidence intervals)
1 year 3 year 5 year
Occlusal GIC 83.7 (83.5–83.9) 60.8 (60.6–61.0) 44.0 (43.6–44.4)
Occlusal composite 91.2 (91.1–91.3) 71.0 (70.8–71.2) 66.2 (66.0–66.4)
Interproximal GIC 99.8 (99.7–99.9) 94.6 (94.3–94.9) 82.0 (81.7–82.3)
Interproximal composite 99.9 (99.8–99.9) 93.8 (93.6–94.0) 88.8 (88.6–89.0)
5.5 Survival of teeth sound and longevity of dental restorations in
Finland (V)
Survival of each permanent tooth remaining caries-free indicated a shortening of
the survival of dental health in Finland. The age cohort born in 1985 had the
longest healthy survival in each tooth, the shortest being in the first mandibular
molars (Fig. 7A), then in the maxillary first molars (Fig. 7B) and rapidly
progressing caries in the second molars. Second molars reached the same level of
morbidity as first molars in a much shorter time (Fig. 7A, 7B).
Caries prevalence in all first molars was 10% at about nine years in the 1990
and 1995 cohorts, whereas 10% prevalence was reached later, at 12 years in the
1985 cohort. Maxillary incisors and all premolars had the same order and level of
survival regardless of the tooth type, as shown for the maxillary first premolars in
the Fig. 7C. Tooth surface-specific analysis revealed that proximal caries was the
type of caries in maxillary incisors and all premolars (Fig. 7C), but a similar curve
was seen in the proximal surfaces of first molars as well. No mandibular incisors
or canines became carious during the follow-up.
46
Fig. 7. Survival of permanent caries-free teeth from the birth of the subject in the
different age cohorts. A = Mandibular first and second molars. B = Maxillary first and
second molars. C = Maxillary central incisors and maxillary first premolars. (Study V,
published by permission of Karger).
47
Because the 1990 cohort had the longest reliable follow-up time, this cohort
was selected for a more detailed analysis. When retrospectively grouped into
three categories according to caries activity around the time of tooth eruption, the
caries-prone subjects indeed developed caries the most rapidly in the first and
second molars (Fig. 8A). In the caries-resistant group, the second molars followed
the curve of the first molars (Fig. 8A). The curves for the second molars fell faster
than that of the first molars (Fig. 8A) and were close to each other in the caries-
prone and inter-medial groups. The curves for premolars and maxillary incisors
were similar and revealed that the caries-resistant subjects had the longest
survival whereas the caries-prone and inter-medial subjects had similar survival,
as shown for the maxillary first premolars (Fig. 8B) and central incisors (Fig. 8C)
as an example.
The median longevity of the restorations of all permanent teeth combined
was 11.7 years. In the health centre with the longest restoration survival
(maximum), about 95% of the restorations remained in the oral cavity over
4 years but in the health centre with the shortest (minimum) restoration survival
less than 2 years.
When the survival curves of restorations were drawn separately for each
permanent tooth, a great variation was seen. Restoration survival was the shortest
in all molars in both jaws: around 70% remained in the oral cavity for 6 years.
The restorations in canines, incisors and premolars survived the longest: 70%
more than ten years.
When the survival of restorations was used as an indicator of the quality of
the restorative work in the health centre, it was noted that tooth-specific dental
health followed the same pattern: in health centres with the longest survival of
restorations, the caries-free survival was also the longest both in the maxillary and
mandibular molars, whereas no major differences were seen in the other teeth (not
shown).
48
Fig. 8. Caries-free survival of A. maxillary first and second molars, B. second
premolars and C. maxillary central incisors in caries-prone, inter-medial and caries-
resistant subjects in the 1990 cohort. (Study V, published by permission of Karger).
49
6 Discussion
6.1 Main findings
The main study findings related to the presented aims are:
1. The use of multivariable or specific methods did not increase from 1996 to
2006 (I).
2. There are problems in reporting statistical methods and findings in dental
journals (II).
3. The proportion of dental research papers using time-to-event methods is
lower than in many other fields of medical research (III).
4. Time-to-event methods allow the use of big data in a reliable way (IV, V)
6.2 Discussion of the results
6.2.1 Use of statistical methods and time-to-event methods in dental
research (I, III)
Because of the increasing availability of statistical computer packages, there has
been a trend to use new sophisticated and more complex statistical methods in
many fields of medical science (Wang & Zhang 1998, Miettunen et al. 2002,
Horton & Switzer 2005, Strasak et al. 2007). However, this study indicates that
these methods have not been applied considerably in dental research. The reasons
for this are not evident, but it is possible that sophisticated statistical methods may
not have been applied more frequently because of following reasons: it appears
that only a few dental journals have published their statistical guidelines for
authors; there are very few journals which have statistical reviewers; and the
awareness of the value of the use of appropriate statistical methods is inadequate
among editors and reviewers of journals and researchers.
Another reason may be that sample sizes have decreased from 1996 to 2006.
Advanced statistical methods require larger sample sizes than basic methods. In
addition, smaller sample sizes may require use of sophisticated methods and there
is even more need for these methods. Furthermore, animal studies have become
more common, and the increasing costs of animal studies and the ethical
requirements that demand using as few animals as possible are examples of
reasons for decreasing the sample size. It is also noteworthy that there were only a
50
few articles which had sample size calculations or other justification for their
sample sizes. The reasons for that could be that editors and referees do not require
sample size calculations, or other justifications (e.g. budget restrictions). The lack
of available large epidemiological data may also decrease the number of study
subjects.
This study showed that dental journals had different profiles in their statistical
analyses. For example, confidence intervals were less frequently reported in JDR,
while they were more frequently used in CDOE. In general, the low use of
confidence intervals is a very alarming phenomenon. In an experimental study
design, it is very important to report confidence intervals. Furthermore, in JDR,
multivariate or specific methods were used in 26% of the articles, but in CR and
AOS in 39% of the articles and in CDOE as much as in 72% of the articles. One
explanation may be that use of statistical techniques differs between basic and
clinical research, which can be seen when comparing statistical methods used in
the New England Journal of Medicine with methods used in the Nature Medicine
(Strasak et al. 2007). Animal studies use experimental designs which include less
intra-individual variation due to the usage of genetically identical species and the
other confounding factors being easily controlled. Consequently, there is no need
for the application of multivariate analysis to adjust for possible confounding
factors, which is typical in clinical and epidemiological settings. For the same
reason, other sophisticated methods which were frequently used in clinical
research studies probably are less likely to fit for basic research studies. Animal
studies have also smaller sample sizes, probably due to well-planned study
design.
Use of TTE methods was found to be low (3%) in dental research papers.
Although the use seems to be increasing, it is still behind the trend in many other
fields of medical research (Nieminen et al. 1995, Goldin et al. 1996, Rigby et al.
2004, Horton & Switzer 2005, Strasak et al. 2007). One possible reason for the
limited use of TTE methods is the high proportion of cross-sectional studies in
dental research. Since retrospective collection of data manually from patient
records is costly and extremely time-consuming, it may have limited the use of
the longitudinal approach. A longitudinal design in many cases can assess the
dynamic nature (time related process) of oral diseases better than a cross-sectional
design. Data mining programs based on electronically stored patient data are
powerful tools that allow extensive longitudinal datasets to be compiled
effectively and economically (Käkilehto et al. 2009). The use of TTE methods
with such data provides a much more comprehensive picture of the development
51
of the disease or the outcome of its treatment than does a cross-sectional
approach.
6.2.2 The quality of statistical reporting in dental journals and
guidelines for the presentation of TTE methods (II, III)
Eight out of ten articles published in dental journals contained at least one
problem in statistical reporting. The finding of poor quality of statistical reporting
is not directly comparable to the previous study in dental literature (Kim et al.
2011) because the present study concentrated on the quality of reporting statistics
whereas the goal in the previous study was to assess the level of statistical
misuses and abuses in dental literature. The reason that the present work did not
take statistical errors into account was that it is difficult to define if the correct
statistical methods or appropriate statistical test were used if the original data is
not available. Instead, we investigated how the methods and results were reported,
and that is not dependent of the availability of data.
The statistical procedures of the evaluated articles were mainly adequately
described. A sufficient description of the statistical methods used is essential for a
scientific research article (Nieminen et al. 2007) and for it to be included in
systematic reviews or meta-analyses. When a published study does not clearly
report its methods, a reader attempting to evaluate its scientific validity may rely
largely on the authors’ reputation and style of writing, or simply on the journal’s
reputation. This should not take place in good science. Statistical tests and
methods should therefore be identified in the methods section (Lang & Secic
1997). It is also useful to report the statistical software used in multivariate
analyses because different software packages use different calculation algorithms
(Okunade et al. 1993). The reporting of software also helps critical readers to
evaluate and understand details that are specific to certain statistical software
packages.
The sample size was usually reported in the evaluated papers. However, the
sample size calculations were seldom reported in the evaluated articles. This is a
common problem in medical research. Altman (1998) reported a review of
100 consecutive papers published in the British Medical Journal in 1991–1993
(excluding controlled trials) and noted that there were no sample size calculations
in any of them. This is an important issue as the frequently used P-values
evaluating statistical significance are highly dependent on the sample size, and the
validity of often-used statistical inferences based on approximations is affected by
52
a small number of events per variable. To evaluate the findings, it is beneficial for
the reader to know the justifications for the sample size selected. Nowadays, there
is easy-to-use software to calculate sample size commonly available. These
should be taught for all researches as part of the scientific training in dental
schools. Mathematical justifications for the intended sample size using statistical
power calculations are valuable (Friedman et al. 1996). Textbooks of medical
statistics simply require that the sample size should be large enough (or as large
as possible) (Nieminen et al. 2006b).
Inexact P-values (e.g. P<0.01, P>0.05, P=ns) were often represented. This is
not a statistical error, but hides evidence from the reader. To provide more
accurate information, it is advisable to provide the exact P-values, e.g. P=0.469.
When the P-value is presented as ‘P>0.05’, it implies that the P-value is anywhere
between 0.05 and 1.0. It is important to state the level of significance for the test
as this determines whether or not to reject or accept the null hypothesis at the
given level (Kim et al. 2011). There is no reason that the P-value should be
degraded into this less-informative dichotomy (Rothman 2012). This convention
is inherited from the days before computers and should not be encouraged
anymore.
There were only few problems with figures and tables in this study. Figures
are visual means of conveying information and have a strong visual impact. The
most typical problems in figures were that there were insufficient legends, there
were no measurement units and there were scale problems. Tables are commonly
used for presenting background information related to the used methods, and a
well-structured table is perhaps the most efficient way to convey a large amount
of data in a scientific paper (Iverson et al. 1997). Problems in tables and figures
do not help readers to evaluate the findings. The most common problems in tables
were that the numbers did not add up and the percentages were not presented. The
presented percentages should be relevant to the studied hypothesis. When the
results are presented in figures or tables, statistically significant differences
should also be indicated. The mode of presentation should be selected according
to the data: large data sets with multiple statistical comparisons are often
confusing when presented in one figure, and a table should then be preferred.
If studies do multiple testing on multiple hypotheses and report several P-
values, it increases the risk of making a type I error, such as saying that a
treatment is effective when chance is a more likely explanation for the results.
This is often referred to as the multiple testing problem (Motulsky 1995). These
53
“data dredging” analyses involve computing many P-values to find something
that is statistically significant (Motulsky 1995).
If studies generate hundreds of P-values, interpreting multiple P-values is
difficult. If authors make many comparisons, reader may expect some P-values to
be small just by chance, and it is difficult to know how to deal with reports that do
not concentrate on the planned main outcomes. To make sense of this kind of
study, readers need to look at the overall pattern of results and not interpret any
individual P-value too closely. Naively applying multiple correction methods such
as Bonferroni could lead to a loss of significant effects. However, modern
statistical tools allow researchers to use more sophisticated methods to avoid
multiple testing problem (Bayes methods, simulation based assessment of P-
values such as permutation).
Proper use of a powerful and sophisticated modelling technique such as the
TTE approach requires considerable care both in the decisions made during data
analysis and in the reporting of the final results, and it is essential to have a full
understanding of the strengths and limitations of the various methods available.
Given the two sets of papers described here (Study II, Study III), those employing
TTE more often stated the essential information, but substantial shortcomings
were still frequently found. Even though there is an excellent body of literature
available on the subject (e.g. “Critical Thinking – Understanding and Evaluating
Dental Research” (Brunette 1996)), detailed instructions on the use of TTE
methods are not readily available. Recommendations for the presentation of TTE
methods and data are given in Fig. 9 and below.
54
Fig. 9. Checklist for reporting statistical methods and results with special focus on
time-to-event methods.
55
For a hypothesis-testing paper, the reader needs to know what the question or
hypothesis of the study was and what it was based on. An advantage of stating the
question as a hypothesis is that the question is precise. In these cases, the results
can be interpreted in light of a priori hypotheses. Furthermore, unless the research
question is clearly stated, the appropriateness of the study design, data collection
methods and statistical procedures cannot be judged.
Justification of the number of cases, i.e. sample size calculations, inclusion of
all the patients treated within the specific time limit, or budget restrictions, was
seldom reported in this study. This is an important issue because P-values
evaluating statistical significance are highly dependent on the sample size, and the
validity of statistical inferences can also be affected by a small number of events
per variable. To evaluate the findings, it is beneficial for the reader to know the
justifications for the sample size selected. In multivariable modelling (such as
Cox regression) it is also important that the events per variable ratio are near or
above 10 (Peduzzi et al. 1995).
The description of statistical procedures is described earlier in this chapter. It
is vital to report the statistical software used in multivariate analysis because
different software use different calculation algorithms (Okunade et al. 1993). The
reporting of software also helps to evaluate and understand the details specific to
some statistical software.
The interpretation for an explanatory variable depends on how that variable is
coded (i.e. how each possible value is presented numerically). Continuous
variables are often converted into categorical factors by grouping into two or
more categories. This is mostly done to simplify the analysis and interpretation of
results. From a methodological point of view, the disadvantages of grouping a
predictor also have to be considered (Royston & Sauerbrei 2009). Grouping
introduces an extreme form of rounding, with an inevitable loss of information
and power. The selection of cut-off points should be based on clinically and/or
methodologically relevant reasoning and should be explained in the text. Both
Chang and Pocock (2000) and Altman and Royston (2006) recommend the use of
splines for complex functional dependencies.
Even when only a few values are missing for each variable, the number of
subjects with at least one missing value can be large. Analysis based on the
complete cases only leads to biased estimates when the exclusion rates are
different for subgroups (Blettner et al. 1999). Additionally, restricting the analysis
to a subset of the data may lead to the loss of valuable information. The
frequently used approach of defining an additional category of missing response
56
to a variable is not sensible and leads to biased estimates. Other techniques
include probability imputations and replacing missing data with values generated
randomly from the distribution of the available data (Donders et al. 2006).
Multiple imputation methods are aimed to obtain valid inference. “Missingness”
needs a careful evaluation when incorporated in the statistical analysis. In any
case, the process leading to missing information needs to be described as
carefully as possible.
Usually explanatory variables are chosen based on earlier research.
Sometimes they are selected by data-driven method, i.e. by the statistical
significance of bivariate analysis between potential explanatory variables and
outcome variable. In the articles evaluated in our study, the selecting of
explanatory variables was mainly based on earlier studies. The variables included
in the reported Cox model may be determined either by an automatic procedure
(usually one of forward inclusion or backward elimination, or best subset) or be
specified a priori, either collectively or in hierarchically grouped subsets (Bagley
et al. 2001). Regardless of the procedure, it should be explicitly stated, preferably
with some description for the appropriateness of that choice. Sensitivity of the
results with respect of the variable selection method or procedure is good to
evaluate.
The presentation of results in figures or tables has been discussed earlier in
this chapter. Additionally, the Kaplan–Meier estimate of the survival is an
essential with the TTE approach, and the reporting of basic data is recommended
in addition to coefficients and hazard ratios.
The proportionality assumption is relevant to the proportional hazards
analysis. The Cox regression model assumes that for any two patients A and B,
the ratio of the hazard functions across time will be a constant. This means that if
patient A has twice the risk of response at any given time than patient B does,
then patient A will be twice as likely to experience the event at all time points.
This proportionality assumption of the Cox regression model can be assessed
graphically, the commonly used methods being Kaplan–Meier curves and a log-
minus-log survival plot (Kleinbaum & Klein 2005). Such proportionality
assumptions were poorly reported in the articles reviewed in the present study.
There are methods available in most statistical packages that should be utilized.
A common goal of the TTE regression modelling is to investigate the
association between an explanatory variable and an outcome while controlling the
possible influence of additional variables. Modelling can be used to adjust for the
effect of many variables simultaneously in order to determine the independent
57
effect of the main explanatory factor. If authors report both the unadjusted and
adjusted effects, the readers can evaluate whether the association has been
adequately adjusted for confounding effects and consider the complexity of these
effects.
6.2.3 Survival studies (IV, V)
In dental research, several indexes have been developed to describe the status of
oral health. Examples of these are plaque and gingival indexes (Loe 1967, Barnes
et al. 2008), tooth wear indexes (Lopez-Frias et al. 2012, Vered et al. 2014) and
different indexes of periodontal diseases (Beltran-Aguilar et al. 2012). Statistical
approach of these indexes has mainly been to report only summary statistics like
mean or median values. Time-to-event methods would offer a more informative
approach for these indexes, allowing the evaluation of the process of the disease
status in the mouth.
As an epidemiological measure of caries prevalence used in dental research,
the DMF (decayed (D), missing (M) or filled (F)) index has been used since 1938
(Klein et al. 1938). When age cohorts have been compared to each other, the
mean values of the DMF index have been used to describe dental caries even
though the index describes carious, restored or extracted teeth, not patients with
caries. It would be better to measure the caries intensity as a function of time,
relating disease process to a statistic (caries hazard) that describes the event of
primary interest.
Caries data are usually collected employing the tooth or tooth surface as a
unit, and the statistical methods applied aggregate data at the subject level to
obtain a measure of the amount of disease at certain age. The aggregated subject-
based outcome measure is typically derived by summarizing the original tooth- or
surface-based observations, for example, by means of calculating a DMF score
per subject. As a result, important information may be lost since different tooth
surfaces or teeth in the same mouth have different exposure times due to the great
variation in their eruption times and different risks.
The benefits of using TTE methods compared with using cross-sectional
studies are that while cross-sectional studies are reports on existence of certain
time point, TTE methods can reveal if there is some critical point where
progression of the disease will change or results of a given treatment will
decrease.
58
In Study IV the survival of restorations was studied. The life span of a
restoration begins when it is placed on a tooth (Lucarotti et al. 2005), but the
point in time when the life of a certain filling comes to its end is not that clear. In
survival analyses, which have been used only recently for studying the longevity
of dental restorations, right-censoring and left-censoring are used to reduce
uncertainty at both ends of the lifespan of restorations. One could argue that a
patient often only seeks for professional help when a filling is badly fractured,
fallen out, looks unaesthetic, or is making the patient feel discomfort or pain. This
means that a patient can go on for years with a badly fractured or missing filling,
distorting the statistics as to when the filling actually failed. With such data, one
should consider interval-censored survival methods. In Finland, however,
adolescents are called for check-ups at regular intervals, which make the
registering of failures of restorations more reliable and add value to the present
study. It should be kept in mind that in retrospective studies based on patient
record evaluations, little or no information is available on how the restorations
have failed and why they need to be replaced (Jokstad et al. 2001).
In Study V, a sub-grouping of the subjects into three categories according to
caries activity was also made. It showed that the caries-prone (9.5% of the cohort)
and inter-medial (25.7%) patients had almost identical survival curves, whereas
the caries-resistant (64.8%) subjects had markedly longer survival. In the caries
activity determinations, the word “caries-prone” subject instead of “caries-active”
was preferred because activity would suggest active measures of prevention while
the caries-resistant subjects may be passive in that respect. On the other hand, the
“inter-medial” group can be combined to the “caries-prone” group that forms the
caries-active population. Therefore, the declining caries-resistant proportion with
more recent age cohorts (Fig. 8) indicates well the deterioration of dental health.
This conclusion can also be more reliably drawn and the findings demonstrated
with the use of TTE methods and Kaplan–Meier survival curves than through
cross-sectional studies. At present, the health centres in Finland use almost
exclusively composite materials in all teeth (Antalainen et al. 2013), and therefore
the present survival curves of restorations can be compared to those with the same
material type. The present values are about the same as reported earlier in Finland
(Käkilehto et al. 2009).
59
6.3 Strengths and limitations of the study
6.3.1 Bibliometric studies (I, II, III)
Strengths of the studies
This sample of dental articles was more extensive than most of the previous
studies. The selected journals included a large variation of dental journals, so the
selection was not restricted to a subfield of dental research, as has been the case in
most of the previous studies. We have not just scanned the abstract of the article
but we have read and evaluated all the articles. The relatively long follow-up of
eleven years enables making a good estimation about the development of
statistical techniques. As far as the author knows, this was also the first study
where the quality of statistical reporting in dental research was assessed.
Limitations of the studies
We have not compared the results with other subfields in medicine. In Study II, it
was difficult to evaluate the statistical quality because there is no generally
accepted quality index in common use. In Study III, we evaluated only whether or
not TTE methods were used and did not consider how many articles there were in
which the possibility of a TTE method had been ignored when it could have been
an appropriate approach.
Only one rater (the author of this thesis) evaluated the articles. This may have
resulted in more incorrect ratings than would have been the case with the several
raters, but on the other hand this guaranteed that the articles were rated similarly.
In addition, if the interpretation of a specific article was unclear, its classification
was subsequently discussed with an experienced biostatistician.
6.3.2 Survival studies
Strengths of the studies
The sample sizes in both of the studies (IV, V) were very large and suited well to
applying TTE methods. Especially the data in Study V covers well the population
of Finnish adolescents. The participant rate is very high because in Finland the
60
municipality is responsible for arranging oral health care free of charge to people
under the age of 18 years.
Limitations of the studies
In Study IV, the data was collected from only one health centre. Thus, the study
was local and one could argue that the data does not cover well the entire Finnish
population. In addition, the differences between operators would have been
interesting to study. However, that was impossible because the only information
about operators was initials and there was no reason for comparing all
120 operators against each other.
6.4 Main conclusions
The current study provided new information regarding the past and current status
of applying/use of the statistical methods in dental research. The last 20 years
have been marked by a rapid expansion in computing capability, and therefore
available computer-intensive statistical methods can be expected to make
significant contributions towards the use of statistics in dentistry. Authors as well
as referees/editorial boards of dental journals could utilize the results of this study
to improve the statistical section of their research articles and to present the
results in such a way that it is in line with the policy and presentation of the
leading dental journals.
The availability of statistical software packages, even free of charge, has
enabled using time-to-event methods for the analysis and interpretation of large
amounts of data provided by longitudinal designs in modern dental projects. In
the late 1980s, the use of electronic dental records started to become more
common, and by the late 1990s all municipalities in Finland used them. This
resulted in huge longitudinal data sets on individual oral health examinations. Our
study showed that data mining together with time-to-event methods are suitable
for investigating patient records for the benefit of public dental health.
6.5 Recommendations how to improve statistical analysis in dental
research
Most of the shortcomings in the reporting of statistical information in the dental
articles were related to topics included even in most introductory medical
61
statistics books. Apparently these problems in one of the key issues in high-
quality research are so common and wide-spread that the scientific dental
community should take action to improve statistical reporting. For example, the
importance of the use of appropriate statistical methods and the accuracy of their
description in the manuscript should be highlighted in the section “Instructions to
Authors”. Journals should also inform the authors of the guidelines developed for
different kinds of study designs by either describing the guidelines or providing a
reference to the most recent version of their guidelines in the “Instructions to
Authors” section. The reviewers could be separately asked to evaluate the quality
and reporting of the issues dealing with statistical analysis. The editors and
editorial boards should also consider consulting a statistician either for every
manuscript with statistical analysis, or at least when the reviewer raises questions
or doubts about the quality and/or reporting of the statistical analysis. The
manuscript should also include a statement that clearly states who is responsible
for the statistical analysis and the quality of statistical reporting.
6.6 Future research
It would be interesting to follow up the use of statistical methods by selecting a
new sample, for example, from the original articles published in the same journals
in 2014. Frequencies of different statistical methods in general dental journals
may change rapidly. The methodological quality of articles could also be studied
in general and, for example, in terms of studies concerning time-to-event
methods. A standard form to help the readers and reviewers of the journals to
investigate the use and quality of statistical methods could also be created. In the
future studies, the reliability and validity of the form should be examined.
62
63
References
Aalen OO, Bjertness E & Soonju T (1995) Analysis of dependent survival data applied to lifetimes of amalgam fillings. Stat Med 14(16): 1819–1829.
Altman DG (1991a) Practical Statistics for Medical Research. London, Chapman and Hall. Altman DG (1991b) Statistics in medical journals: developments in the 1980s. Stat Med
10(12): 1897–1913. Altman DG (1998) Statistical reviewing for medical journals. Stat Med 17(23): 2661–2674. Altman DG (2001) Statistics in medical journals: some recent trends. Stat Med 19(23):
3275–3289. Altman DG (2002) Poor-quality medical research: what can journals do? JAMA 287(21):
2765–2767. Altman DG & Royston P (2006) The cost of dichotomising continuous variables. BMJ
332(7549): 1080. Andersen PK, Klein JP & Zhang MJ (1999) Testing for centre effects in multi-centre
survival studies: a Monte Carlo comparison of fixed and random effects tests. Stat Med 18(12): 1489-1500.
Antalainen AK, Helminen M, Forss H, Sandor GK & Wolff J (2013) Assessment of removed dental implants in Finland from 1994 to 2012. Int J Oral Maxillofac Implants 28(6): 1612–1618.
Armitage P, Berry G & Matthews J (2002) Statistical Methods in Medical Research. Oxford, Blackwell Science.
Baccaglini L, Shuster JJ, Cheng J, Theriaque DW, Schoenbach VJ, Tomar SL & Poole C (2010) Design and statistical analysis of oral medicine studies: common pitfalls. Oral Dis 16(3): 233–241.
Baelum V, Machiulskiene V, Nyvad B, Richards A & Vaeth M (2003) Application of survival analysis to carious lesion transitions in intervention trials. Community Dent Oral Epidemiol 31(4): 252–260.
Bagley SC, White H & Golomb BA (2001) Logistic regression in the medical literature: standards for use and reporting, with particular attention to one medical domain. J Clin Epidemiol 54(10): 979–985.
Barnes VM, Vandeven M, Richter R & Xu T (2008) The Modified Gingival Margin Plaque Index: a predictable clinical method. J Clin Dent 19(4): 131–133.
Beltran-Aguilar ED, Eke PI, Thornton-Evans G & Petersen PE (2012) Recording and surveillance systems for periodontal diseases. Periodontol 2000 60(1): 40–53.
Blettner M, Sauerbrei W, Schlehofer B, Scheuchenpflug T & Friedenreich C (1999) Traditional reviews, meta-analyses and pooled analyses in epidemiology. Int J Epidemiol 28(1): 1–9.
Brunette DM (1996) Critical Thinking: Understanding and Evaluating Dental Research. Quintessence Publishing Co., Inc.
Carlos JP & Gittelsohn AM (1965) Longitudinal studies of the natural history of caries. II. A life-table study of caries incidence in the permanent teeth. Arch Oral Biol 10(5): 739–751.
64
Choi E, Lyu J, Park J & Kim HY (2014) Statistical methods used in articles published by the Journal of Periodontal and Implant Science. J Periodontal Implant Sci 44(6): 288-292.
Chang BH & Pocock S (2000) Analyzing data with clumping at zero. An example demonstration. J Clin Epidemiol 53(10): 1036–1043.
Chuang SK, Cai T, Douglass CW, Wei LJ & Dodson TB (2005) Frailty approach for the analysis of clustered failure time observations in dental research. J Dent Res 84(1): 54–58.
Clayton DG (1991) A Monte Carlo method for Bayesian inference in frailty models. Biometrics 47(2): 467–485.
Donders AR, van der Heijden GJ, Stijnen T & Moons KG (2006) Review: a gentle introduction to imputation of missing values. J Clin Epidemiol 59(10): 1087–1091.
Emerson JD & Colditz GA (1983) Use of statistical analysis in the New England Journal of Medicine. N Engl J Med 309(12): 709–713.
Farag A, van der Sanden WJ, Abdelwahab H, Mulder J & Frencken JE (2009) 5-Year survival of ART restorations with and without cavity disinfection. J Dent 37(6): 468–474.
Fernandes-Taylor S, Hyun JK, Reeder RN & Harris AH (2011) Common statistical and research design problems in manuscripts submitted to high-impact medical journals. BMC Res Notes 4: 304-0500-4-304.
Friedman L, Furberg C & DeMets D (1996) Fundamentals of Clinical Trials. St. Louis, Mosby.
Gardenier JS & Resnik DB (2002) The misuse of statistics: concepts, tools, and a research agenda. Account Res 9(2): 65–74.
Geminiani A, Ercoli C, Feng C & Caton JG (2014) Bibliometrics study on authorship trends in periodontal literature from 1995 to 2010. J Periodontol 85(5): e136-43.
Goldin J, Zhu W & Sayre JW (1996) A review of the statistical analysis used in papers published in Clinical Radiology and British Journal of Radiology. Clin Radiol 51(1): 47–50.
Hannigan A (2004) Using survival methodologies in demonstrating caries efficacy. J Dent Res 83 Spec No C: C99-102.
Hannigan A & Lynch CD (2013) Statistical methodology in oral and dental research: pitfalls and recommendations. J Dent 41(5): 385–392.
Hannigan A, O’Mullane DM, Barry D, Schafer F & Roberts AJ (2001) A re-analysis of a caries clinical trial by survival analysis. J Dent Res 80(2): 427–431.
Härkänen T, Virtanen JI & Arjas E (2000) Caries on permanent teeth: a non-parametric Bayesian analysis. Scand J Stat 27: 577–588.
Härkänen T, Larmas MA, Virtanen JI & Arjas E (2002) Applying modern survival analysis methods to longitudinal dental caries studies. J Dent Res 81(2): 144–148.
Horton NJ & Switzer SS (2005) Statistical methods in the journal. N Engl J Med 353(18): 1977–1979.
Iverson C, Christiansen S & Flanagin A (1997) AMA Manual of Style: A Guide for Authors and Editors. New York, Oxford Press.
65
Jokstad A, Bayne S, Blunck U, Tyas M & Wilson N (2001) Quality of dental restorations. FDI Commission Project 2-95. Int Dent J 51(3): 117–158.
Käkilehto T, Salo S & Larmas M (2009) Data mining of clinical oral health documents for analysis of the longevity of different restorative materials in Finland. Int J Med Inform 78(12): e68–74.
Kalbleisch JD & Prentice RL (1980) The statistical analysis of failure time data. New York, John Wiley & Sons.
Kaplan EL & Meier P (1958) Nonparametric estimation from incomplete observations. J Am Statist Assoc 53: 457–481.
Kim JS, Kim DK & Hong SJ (2011) Assessment of errors and misused statistics in dental research. Int Dent J 61(3): 163–167.
Klein H, Palmer C E & Knutson JW. (1938) Dental status and dental needs of elementary school children. In Anonymous, Public Health Reports: 751–55.
Kleinbaum D & Klein M (2005) Survival analysis: a self-learning text. New York, Springer.
Komarek A, Lesaffre E, Härkänen T, Declerck D & Virtanen JI (2005) A Bayesian analysis of multivariate doubly-interval-censored dental data. Biostatistics 6(1): 145–155.
Krithikadatta J & Valarmathi S (2012) Research methodology in dentistry: Part II – The relevance of statistics in research. J Conserv Dent 15(3): 206–213.
Lang T & Secic M (1997) How to report statistics in medicine. Philadelphia, American College of Physicians.
Larmas M (2010) Has dental caries prevalence some connection with caries index values in adults? Caries Res 44(1): 81–84.
Larmas MA, Virtanen JI & Bloigu RS (1995) Timing of first restorations in permanent teeth: a new system for oral health determination. J Dent 23(6): 347–352.
Layton DM & Clarke M (2014a) Accuracy of medical subject heading indexing of dental survival analyses. Int J Prosthodont 27(3): 236-244.
Layton DM & Clarke M (2014b) Quality of reporting of dental survival analyses. J Oral Rehabil 41(12): 928-940.
Leroy R, Bogaerts K, Lesaffre E & Declerck D (2005a) Effect of caries experience in primary molars on cavity formation in the adjacent permanent first molar. Caries Res 39(5): 342–349.
Leroy R, Bogaerts K, Lesaffre E & Declerck D (2005b) Multivariate survival analysis for the identification of factors associated with cavity formation in permanent first molars. Eur J Oral Sci 113(2): 145–152.
Lesaffre E, Garcia Zattera MJ, Redmond C, Huber H, Needleman I & ISCB Subcommittee on D (2007) Reported methodological quality of split-mouth studies. J Clin Periodontol 34(9): 756–761.
Leskinen K, Ekman A, Oulis C, Forsberg H, Vadiakas G & Larmas M (2008) Comparison of the effectiveness of fissure sealants in Finland, Sweden, and Greece. Acta Odontol Scand 66(2): 65–72.
66
Liberati A, Altman DG, Tetzlaff J, Mulrow C, Gotzsche PC, Ioannidis JP, Clarke M, Devereaux PJ, Kleijnen J & Moher D (2009) The PRISMA statement for reporting systematic reviews and meta-analyses of studies that evaluate healthcare interventions: explanation and elaboration. BMJ 339: b2700.
Loe H (1967) The Gingival Index, the Plaque Index and the Retention Index Systems. J Periodontol 38(6): Suppl:610–6.
Lopez-Frias FJ, Castellanos-Cosano L, Martin-Gonzalez J, Llamas-Carreras JM & Segura-Egea JJ (2012) Clinical measurement of tooth wear: Tooth wear indices. J Clin Exp Dent 4(1): e48–e53.
Lucarotti PS, Holder RL & Burke FJ (2005) Analysis of an administrative database of half a million restorations over 11 years. J Dent 33(10): 791–803.
Lucena C, Lopez JM, Abalos C, Robles V & Pulgar R (2011) Statistical errors in microleakage studies in operative dentistry. A survey of the literature 2001–2009. Eur J Oral Sci 119(6): 504–510.
Miettunen J, Nieminen P & Isohanni M (2002) Statistical methodology in general psychiatric journals. Nordic Journal of Psychiatry 56(3): 223–228.
Motulsky H (1995) Intuitive Biostatistics. In Anonymous Oxford, Oxford University Press: 117–126.
Nieminen P (1996) Therapeutic community research and statistical data analysis. Oulu: University of Oulu, Acta Universitatis Ouluensis D 360.
Nieminen P, Bloigu A, Kukkonen J & Isohanni M (1995) Role of bibliometrics in the evaluation of medical research. Duodecim 111(2): 134–143.
Nieminen P, Carpenter J, Rucker G & Schumacher M (2006a) The relationship between quality of research and citation frequency. BMC Medical Research Methodology 6: 42.
Nieminen P, Miettunen J, Koponen H & Isohanni M (2006b) Statistical methodologies in psychopharmacology: a review. Hum Psychopharmacol 21(3): 195–203.
Nieminen P, Rucker G, Miettunen J, Carpenter J & Schumacher M (2007) Statistically significant papers in psychiatry were cited more often than others. J Clin Epidemiol 60(9): 939–946.
Okunade AA, Chang C & Evans R (1993) Comparative Analysis of Regression Output Summary Statistics in Common Statistical Packages. The American Statistician 47: 298–303.
Ollila P & Larmas M (2007) A seven-year survival analysis of caries onset in primary second molars and permanent first molars in different caries risk groups determined at age two years. Acta Odontol Scand 65(1): 29–35.
Otwombe KN, Petzold M, Martinson N & Chirwa T (2014) A review of the study designs and statistical methods used in the determination of predictors of all-cause mortality in HIV-infected cohorts: 2002–2011. PLoS One 9(2): e87356.
Peduzzi P, Concato J, Feinstein AR & Holford TR (1995) Importance of events per independent variable in proportional hazards regression analysis. II. Accuracy and precision of regression estimates. J Clin Epidemiol 48(12): 1503–1510.
67
Rigby AS, Armstrong GK, Campbell MJ & Summerton N (2004) A survey of statistics in three UK general practice journal. BMC Med Res Methodol 4(1): 28.
Rothman K (2012) Epidemiology. An Introduction. Oxford, Oxford University Press. Royston P & Sauerbrei W (2009) Multivariable model-building: a pragmatic approach to
regression analysis based on fractional polynomials for continuous variables. Chichester, UK, John Wiley & Sons Ltd.
Schulz KF, Altman DG, Moher D & CONSORT Group (2010) CONSORT 2010 statement: updated guidelines for reporting parallel group randomised trials. BMJ 340: c332.
Souza E (2014) Research that matters: setting guidelines for the use and reporting of statistics. Int Endod J 47(2): 115–119.
Stephenson J, Chadwick BL, Playle RA & Treasure ET (2010a) A competing risk survival analysis model to assess the efficacy of filling carious primary teeth. Caries Res 44(3): 285–293.
Stephenson J, Chadwick BL, Playle RA & Treasure ET (2010b) Modelling childhood caries using parametric competing risks survival analysis methods for clustered data. Caries Res 44(1): 69–80.
Strasak AM, Zaman Q, Marinell G, Pfeiffer KP & Ulmer H (2007) [MEDICINE] The Use of Statistics in Medical Research: A Comparison of The New England Journal of Medicine and Nature Medicine. The American Statistician 61(1): 47–55.
Suni J, Helenius H & Alanen P (1998) Tooth and tooth surface survival rates in birth cohorts from 1965, 1970, 1975, and 1980 in Lahti, Finland. Community Dent Oral Epidemiol 26(2): 101–106.
Tonetti M, Palmer R & Working Group 2 of the VIII European Workshop on Periodontology (2012) Clinical research in implant dentistry: study design, reporting and outcome measurements: consensus report of Working Group 2 of the VIII European Workshop on Periodontology. J Clin Periodontol 39 Suppl 12: 73-80.
Tufte E (1983) The visual display of quantitative information. Cheshire, Graphics Press. Vered Y, Lussi A, Zini A, Gleitman J & Sgan-Cohen HD (2014) Dental erosive wear
assessment among adolescents and adults utilizing the basic erosive wear examination (BEWE) scoring system. Clin Oral Investig (in press: Epub ahead of print).
Virtanen JI & Larmas MA (1995) Timing of first fillings on different permanent tooth surfaces in Finnish schoolchildren. Acta Odontol Scand 53(5): 287–292.
von Elm E, Altman DG, Egger M, Pocock SJ, Gotzsche PC, Vandenbroucke JP & STROBE Initiative (2008) The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. J Clin Epidemiol 61(4): 344–349.
Wang Q & Zhang B (1998) Research design and statistical methods in Chinese medical journals. JAMA 280(3): 283–285.
Yang S, Needleman H & Niederman R (2001) A bibliometric analysis of the pediatric dental literature in MEDLINE. Pediatr Dent 23(5): 415–418.
Yergens DW, Dutton DJ & Patten SB (2014) An overview of the statistical methods reported by studies using the Canadian community health survey. BMC Med Res Methodol 14: 15-2288-14-15.
68
69
Original publications
I Vähänikkilä H, Nieminen P, Miettunen J & Larmas M (2009) Use of statistical methods in dental research: comparison of four dental journals during a 10-year period. Acta Odontol Scand 67: 206–211.
II Vähänikkilä H, Tjäderhane L & Nieminen P (2015) The statistical reporting quality of articles published in 2010 in five dental journals. Acta Odontol Scand 73: 76-80.
III Vähänikkilä H, Miettunen J, Tjäderhane L, Larmas M & Nieminen P (2012) The use of time-to-event methods in dental research: a comparison based on five dental journals over a 11-year period. Community Dent Oral Epidemiol 40 suppl 1:36–42.
IV Vähänikkilä H, Käkilehto T, Pihlaja J, Päkkilä J, Tjäderhane L, Suni J, Salo S & Anttonen V (2014) A data-based study on survival of permanent molar restorations in adolescents. Acta Odontol Scand 72: 380–385.
V Suni J, Vähänikkilä H, Päkkilä J, Tjäderhane L & Larmas M (2013) Review of 36,537 patient records for tooth health and longevity of dental restorations. Caries Res 47:309–317.
Reprinted with permissions from Informa Healthcare (I,II,IV), John Wiley &
Sons Ltd (III) and Karger (V).
Original publications are not included in the electronic version of the dissertation.
70
A C T A U N I V E R S I T A T I S O U L U E N S I S
Book orders:Granum: Virtual book storehttp://granum.uta.fi/granum/
S E R I E S D M E D I C A
1276. Savela, Salla (2014) Physical activity in midlife and health-related quality of life,frailty, telomere length and mortality in old age
1277. Laitala, Anu (2014) Hypoxia-inducible factor prolyl 4-hydroxylases regulatingerythropoiesis, and hypoxia-inducible lysyl oxidase regulating skeletal muscledevelopment during embryogenesis
1278. Persson, Maria (2014) 3D woven scaffolds for bone tissue engineering
1279. Kallankari, Hanna (2014) Perinatal factors as predictors of brain damage andneurodevelopmental outcome : study of children born very preterm
1280. Pietilä, Ilkka (2015) The role of Dkk1 and Wnt5a in mammalian kidneydevelopment and disease
1281. Komulainen, Tuomas (2015) Disturbances in mitochondrial DNA maintenance inneuromuscular disorders and valproate-induced liver toxicity
1282. Nguyen, Van Dat (2015) Mechanisms and applications of disulfide bond formation
1283. Orajärvi, Marko (2015) Effect of estrogen and dietary loading on rat condylarcartilage
1284. Bujtár, Péter (2015) Biomechanical investigation of the mandible, a related donorsite and reconstructions for optimal load-bearing
1285. Ylimäki, Eeva-Leena (2015) Ohjausintervention vaikuttavuus elintapoihin jaelintapamuutokseen sitoutumiseen
1286. Rautiainen, Jari (2015) Novel magnetic resonance imaging techniques for articularcartilage and subchondral bone : studies on MRI Relaxometry and short echotime imaging
1287. Juola, Pauliina (2015) Outcomes and their predictors in schizophrenia in theNorthern Finland Birth Cohort 1966
1288. Karsikas, Sara (2015) Hypoxia-inducible factor prolyl 4-hydroxylase-2 in cardiacand skeletal muscle ischemia and metabolism
1289. Takatalo, Jani (2015) Degenerative findings on MRI of the lumbar spine :Prevalence, environmental determinants and association with low back symptoms
1290. Pesonen, Hanna-Mari (2015) Managing life with a memory disorder : the mutualprocesses of those with memory disorders and their family caregivers following adiagnosis
1291. Pruikkonen, Hannele (2015) Viral infection induced respiratory distress inchildhood
UNIVERSITY OF OULU P .O. Box 8000 F I -90014 UNIVERSITY OF OULU FINLAND
A C T A U N I V E R S I T A T I S O U L U E N S I S
Professor Esa Hohtola
University Lecturer Santeri Palviainen
Postdoctoral research fellow Sanna Taskila
Professor Olli Vuolteenaho
University Lecturer Veli-Matti Ulvinen
Director Sinikka Eskelinen
Professor Jari Juga
University Lecturer Anu Soikkeli
Professor Olli Vuolteenaho
Publications Editor Kirsti Nurkkala
ISBN 978-952-62-0792-6 (Paperback)ISBN 978-952-62-0793-3 (PDF)ISSN 0355-3221 (Print)ISSN 1796-2234 (Online)
U N I V E R S I TAT I S O U L U E N S I S
MEDICA
ACTAD
D 1292
ACTA
Hannu Vähänikkilä
OULU 2015
D 1292
Hannu Vähänikkilä
STATISTICAL METHODS IN DENTAL RESEARCH, WITH SPECIAL REFERENCE TO TIME-TO-EVENT METHODS
UNIVERSITY OF OULU GRADUATE SCHOOL;UNIVERSITY OF OULU,FACULTY OF MEDICINE
