- Email: [email protected]
Writing a Scientific Paper Is Not Rocket Science!
J Oral Maxillofac Surg 73:S160-S169, 2015
Writing a Scientific Paper Is Not Rocket Science! Thomas B. Dodson, DMD, MPH I elected to update an article on how to write a scientific paper as my contribution to this supplement because Dr Leonard Kaban introduced me to the discipline of writing. Our first collaboration resulted in a paper accepted at its initial submission.1 The present submission is based, with permission, on an article published 8 years ago.2 Ours is a mentor-driven specialty. Like those before me, I learned by observing others. Continuing the tradition, I would like to share strategies and detail tactics I use to write a scientific paper. The best written patientoriented research articles do not make me, the reader, work hard to understand their meanings. It follows that writing a patient-oriented research paper should be a straightforward exercise that translates data into a clear, practical lesson for the clinician. It should not be a burden to write or read. This is not a definitive article on the topic of preparing a scientific paper. Rather, it is an overview of my personal process for writing a paper. This process is dynamic and evolving and has been guided by mentors and associates too numerous to name. Along the way, I have collected 12 aphorisms guiding manuscript preparation: 1. ‘‘There is no such thing as a paper that is too short.’’ Hemingway wrote a story in 6 words: ‘‘For sale: baby shoes. Never worn.’’ 2. ‘‘Write short declarative sentences.’’ 3. ‘‘All studies, no matter how complicated, can be resolved into a 2 2 table.’’ 4. ‘‘Surgeons have the attention span of a flea. You have 30 seconds to get their attention.’’ 5. ‘‘Readers should not have to guess your study purpose.’’
6. ‘‘It is easier to write when you have something to say.’’ 7. ‘‘Avoid the passive voice.’’ 8. ‘ You treat patients and do research with subjects.’’ 9. ‘‘A table should stand on its own.’’ 10. ‘‘Beware the ‘curse of knowledge’’’ The curse of knowledge is the difficulty in imagining what it is like for someone else, the reader, not to know something that you, the writer, know. The curse of knowledge might explain why good people write bad prose. It simply does not occur to writers that readers do not know what they know.3 11. ‘ The reader is king.’’ Another reason why good writing goes bad is that someone is missing when the author is writing. Writing is not a group experience. Writing is an individual experience. We write alone. The audience is missing. In place of the audience is a phantom we have never met, that is, the reader. The reader is king. The writer is his servant. One thing readers want is clear, concise, comprehensible sentences. By outlining a set of organizational principles for drafting a patient-oriented paper, the writer can focus on writing ‘ clear, concise, comprehensible sentences.’’4 12. ‘‘A manuscript submitted for journal publication is not a thesis.’’ A thesis is a long-winded document reviewed by a committee, bound, and filed in the library’s bowels, never to see the light of day. In contrast, a scientific paper is topically focused with clearly delineated observations and recommendations.
Professor and Chair, Department of Oral and Maxillofacial
Health Sciences Building, Room B-241, Box 357134, Seattle, WA
Surgery, Associate Dean for Hospital Affairs, University of
98195-7134; e-mail: [email protected]
Washington School of Dentistry, Seattle, WA.
Received April 10 2015
This work was supported in part by the University of Washington School of Dentistry Department of Oral and Maxillofacial Surgery
Accepted April 10 2015
Research and Training Fund.
0278-2391/15/00499-1
Conflict of Interest Disclosures: The author did not report any
Ó 2015 American Association of Oral and Maxillofacial Surgeons
http://dx.doi.org/10.1016/j.joms.2015.04.039
disclosures. Address correspondence and reprint requests to Dr Dodson: University of Washington School of Dentistry, 1959 NE Pacific Street,
S160
THOMAS B. DODSON
Elements of a Patient-Oriented Manuscript A patient-oriented research paper has 7 elements: title, introduction, materials and methods, results, discussion, conclusion, and abstract. The title grabs the reader. The introduction is short. It focuses the reader on the clinical problem and study goals. The materials and methods section is as long as necessary to cover the elements of the study. Writing the results section is straightforward, especially if you construct and populate the tables first. The discussion can be variable in length, but its structure is formulaic. The conclusion is short. Write the abstract last. Given the space constraints of most abstracts, it might be the most important and challenging aspect of the manuscript to draft. Word count limitations for abstracts preclude wasted words, and the abstract might be the only element of your paper that is read. ELEMENT 1—TITLE
The title’s goal is to engage and entice the reader to read more. The title should be short. Try to limit the study title to 10 to 15 words. To shorten a title, avoid including the study design in the title, eg, ‘‘placebocontrolled, double-blinded, randomized clinical trial’’ (5 words) or ‘‘systematic review and meta-analysis’’ (4 words). Authors can describe the study design in the abstract and body of the paper. Consider using a title composed of a declarative sentence or a question, rather than a technically correct, but verbose title. For example, instead of ‘‘Use of the pyriform ligament for the alar cinch to enhance nasal appearance after Le Fort I osteotomy’’ (16 words), consider a declarative sentence, ‘‘Proper management of the pyriform ligament improves nasal appearance after Le Fort I osteotomy’’ (12 words). An example of constructing the title as a question is ‘‘Does the pyriform ligament exist and why surgeons should care?’’ (10 words). ELEMENT II—THE INTRODUCTION
The purpose of the introduction is to provide enough information to ‘‘hook’’ the reader. The introduction is focused, short, and should be shorter than 2 pages, double-spaced, and with an 11- to 12-point font. The introduction should answer the following questions: 1) Why is this clinical problem of interest? 2) What are the deficiencies or controversies in the current literature? 3) What is the purpose of this paper? An introduction composed of as few as 2 to 3 paragraphs can answer these 3 questions well. Try to avoid including information in the introduction that is better reserved for the discussion section of the paper (eg, a comprehensive literature review).
S161 A key element of the introduction is its final paragraph summarizing the study purpose. The good writer provides readers a study purpose with sufficient specificity to avoid confusion or ambiguity. This is the author’s first, best, and possibly only opportunity to share with the reader exactly what the paper is about and not leave it to the reader to guess. It is dangerous, sometimes fatal to the paper, to let the reader guess the study purpose. The reader might conclude that a good paper is garbage because it failed to address the reader’s purpose or, more specifically, the reader’s perception of the purpose. A well-composed ‘‘purpose’’ paragraph has 3 elements: 1) a purpose or aim statement, 2) a hypothesis statement, and 3) a specific-aim statement. I use 2 different techniques to state the study purpose. The conventional technique is to state the purpose literally (eg, ‘‘The purpose of this study is to measure the efficacy of prophylactic antibiotics in preventing postoperative complications in patients undergoing third molar removal’’). However, this conventional purpose statement could leave ambiguity in the reader’s mind (‘‘What is efficacy?’’ ‘‘How will it be measured?’’ ‘‘What complications will be studied?’’). An alternative approach is to state the research purpose in the form of a clinical question using the PICO acronym. The PICO acronym is composed of 4 elements: patient sample (P), intervention (I), control or comparison (C), and outcome (O).5 Using the earlier example, the purpose can be reformulated as a clinical question: ‘‘The study’s purpose is to answer the following clinical question: ‘Among patients with impacted third molars (patient sample), does the use of prophylactic antibiotics (intervention of interest), when compared with a placebo control (control group or treatment), decrease the frequency of postoperative infections (outcome of interest)?’’’ Restating the purpose in the form of a specific clinical question minimizes ambiguity and focuses the reader on the research goal. Almost all clinical studies (and literature reviews) can have the purpose stated as a clinical question. Even case series can easily have the purpose stated in the form of a clinical question, without the ‘‘C’’ or control or comparison component. Compelling patient-oriented research is hypothesis driven. As such, the second element in the purpose paragraph is a hypothesis statement. Many times, based on the purpose statement, the reader can intuit the hypothesis. However, the reader should not work that hard. Do not leave the reader’s impression of the hypothesis to chance. As such, articulate the hypothesis statement explicitly. Hypotheses statements can be formal or informal. Examples of formal hypothesis statements include:
S162 1. ‘ The frequency of postoperative infections in the treatment group equals the frequency of infections in the control group.’’ This statement posed as the null hypothesis implies a 2-tailed test of hypothesis. 2. ‘‘The frequency of postoperative infections in the treatment group is lower than the frequency of infections in the control group.’’ This statement implies a 1-tailed test of hypothesis. Examples of less formal hypothesis statements include: 1. ‘‘Antibiotics decrease the frequency of postoperative complications.’’ 2. ‘‘There exists a set of at least 1 factor associated with an increased risk for surgical site infections that can be manipulated by the clinician to improve outcomes after third molar surgery.’’ Of note, a case series, by definition, will not have a hypothesis statement. The final element of the purpose paragraph is the specific aim(s) statement. Unambiguously, tell the reader what you did in the study. Specific aims use active verbs such as measure, design, identify, implement, estimate, compare, or identify. Examples of specific aims include: 1. To estimate and compare the frequencies of surgical site infection after third molar removal in subjects who did or did not receive prophylactic antibiotics. 2. To measure the length of hospitalization in a patient sample undergoing orthognathic surgery and identify factors associated with decreased length of stay. 3. To estimate the 1-year survival rates of implants loaded immediately and implants loaded in a delayed manner. ELEMENT III—MATERIALS AND METHODS
The materials and methods section of a patientoriented paper has a logical predictable structure anticipating readers’ questions. This section of the paper could be the most valuable to the reader and other researchers. The reader needs to understand in detail what was done. Researchers need to know what was done to replicate the study or when abstracting data for a systematic review or meta-analysis. The methods section should be easy to write. The methods have been written before enrolling the first subject or abstracting data from the first chart. Virtually all studies require protocols approved by an institutional review board. As such, with little modification, one can ‘‘cut and paste’’ the protocol
WRITING A SCIENTIFIC PAPER
from the institutional review board application to the manuscript. Failing to draft a written materials and methods section before data collection should give investigators pause, begging the question, ‘‘Is this submission a high-quality, thoughtful research endeavor?’’ The material and methods section has 4 parts: 1) study design, 2) sample identification and selection, 3) variables, and 4) data analyses. Each part should be clearly identified using subheadings for each section. Do not include details of the sample in the materials and methods (eg, sample size, age, and gender) unless the study was designed a priori to enroll a sample that had characteristics of the final study sample—a rare event. The details of the sample should be reported in the results because they are a natural consequence of the methods. Study Design The first sentence of the materials and methods unambiguously states the study design (eg, ‘‘The investigators designed and implemented a double-blind, placebo-controlled, randomized clinical trial’’). Study design has implications for interpreting the results and assessing the study’s level of evidence and associated value in clinical decision making. Do not leave the study design statement to the reader’s imagination. Study design choices are finite. Common study designs, in descending levels of validity, are systematic and meta-analyses of several randomized clinical trials, randomized clinical trials (multi- or single-institutional, double- or single-blinded, placebo-controlled), cohort (prospective or retrospective), case-and-control studies, cross-sectional studies, case series (prospective or retrospective), and case reports. Study Sample The second sentence should summarize clearly the parameters used to select the study sample. These parameters include the population from which the sample is derived and inclusion and exclusion criteria. The reader needs this information to determine whether the patient sample is relevant and of interest. In almost all studies, the study sample (ie, actual subjects) is derived from an underlying patient population (ie, eligible subjects). Owing to inclusion criteria, availability of charts or records, and loss to follow-up, the study population and sample are rarely the same. It is important for the reader to know from which population the sample was drawn. As an example, consider the following: ‘‘The study sample was derived from the population of patients who presented to the department of oral and maxillofacial surgery at a fine hospital for evaluation and management of third molars from January 1, 1998 through December 31, 2000.’’ Based on this information, the reader can assess
THOMAS B. DODSON
the sample’s relevance (eg, academic practice, not a private practice) or applicability (eg, late 20th century might not reflect current practice). In most studies, the investigators do not or cannot enroll every subject. Enrolled subjects meet prespecified inclusion and exclusion criteria. Manuscript authors should tell the reader exactly what these criteria were. Critical readers will use this information to determine whether this study’s sample is similar to other samples of interest (eg, their patients) and choose to continue reading the article. Before reviewing the first record or enrolling the first subject, the inclusion and exclusion criteria should be written down. Quite simply, this information needs to be incorporated into the paper. Inclusion and exclusion criteria are presented in the form of a list. For example, ‘‘Subjects eligible for study inclusion had at least 1 impacted wisdom tooth, were at least 18 years old, underwent operative treatment of the wisdom teeth, agreed to be enrolled in the study, and returned for at least 1 followup visit. Subjects were excluded from study enrollment if they had no impacted teeth, were younger than 18 years, failed to return for follow-up, or refused study enrollment.’’ Study Variables Studies have numerous variables that can be characterized as predictor, outcome, and ‘‘other.’’ In most patient-oriented research, the goal is to establish an association or relation between the predictor (independent) and outcome (dependent) variables, which is adjusted as needed for ‘‘other’’ variables. The investigator should articulate and define the primary predictor variable. In general, predictor variables include exposures, risk or prognostic factors, or treatments of interest (eg, treatment [active vs placebo], implant loading [immediate vs delayed], or dose of radiation therapy). Some studies can have multiple, heterogeneous predictor variables. These predictor variables should be grouped together into logical sets (eg, demographic, anatomic, radiographic, perioperative, etc). There might be numerous outcomes of interest, but the study should be designed to measure 1 primary outcome of interest (eg, postoperative inflammatory complication [yes or no], duration of implant survival [measured in months], or development of osteoradionecrosis [yes or no]). It should be clear to the reader which variable is primary and which variable(s) is a secondary outcome variable. How does the investigator determine the primary outcome variable? Hint: The variable you used to estimate sample size is your primary outcome variable. The ‘‘other’’ variable category includes a listing of all the other variables for which data were collected that help to describe the sample or could affect the relation between the predictor and outcome variables (eg, age,
S163 gender, comorbidities, implant size or coating, exposure to bisphosphonates, duration of operating time, time from injury to fracture repair, education, or socioeconomic status). In some studies, there might be numerous ‘ other’’ variables and it can be tedious to the reader to enumerate them. To make the information more accessible, I group these ‘ other’’ variables into sets. For example, if there are 30 variables, it is tedious to read a list. However, if the variables are grouped into several discrete, logical sets, the list seems less imposing. Commonly used variable categories include demographic (age, gender), socioeconomic (education, annual salary, ZIP codes), medical or dental (comorbidities, medications, status of occlusion, diseases that affect wound healing, steroid use), anatomic (maxilla or mandible or both, level of impaction, simple or compound fracture), time (duration of operation or time from injury to hospital admission), perioperative (open or closed reduction, grafting techniques), and materials (implant type, dimensions, coating, graft materials). The specifics of the investigator’s database will dictate the type and number of variable sets. Be careful not to include too many variable sets or it defeats the value of grouping the variables. All study variables need to be defined unambiguously. It should be clear how the variable is coded (eg, binary, categorical, or continuous). Coding has implications regarding data presentation and analyses. A binary variable only has 2 choices and order is generally unimportant (eg, yes or no, male or female, smoker or nonsmoker). A categorical variable has more than 2 choices (eg, the operation to extract a third molar was categorized as simple extraction, surgical extraction, soft tissue impaction, etc). For categorical variables, order might not be important. A continuous variable is any real number between positive and negative infinity. In reality, most continuous variables are bounded by ceiling or floor values; for example, age is a continuous variable that generally includes positive, nonzero numbers and is generally bounded by the values of 0 and 100. For data collection, collect continuous variables as continuous data, not categorical or binary variables. One can always convert a continuous variable to a categorical or binary variable. The converse is not true. Data Collection, Management, and Analyses Under data collection, I include information regarding the details of randomization, how the standard and experimental treatments were rendered, how subjects were managed, sample size estimates, and details of how data were abstracted from charts. This section would include information regarding methods used to assess intra- and inter-examiner variabilities or standardization of examiners or processes for data collection or abstraction. Data management
S164 would include information regarding who and how data were inputted and stored, what software was used for data storage and analyses, and methods used to assure accurate input (eg, double-entry techniques, software checks for erroneous inputs, and how missing data were managed). Data analyses include a brief description of analytic methods used and specification of level of a error (ie, P value). For example, ‘‘Descriptive statistics (mean, frequency, range, standard deviations) were computed for each study variable.’’ Bivariate analyses (eg, c2 test, t test) are computed to measure the association between any 2 variables of interest (eg, treatment type and age or treatment type and gender). When regression analyses are used, the methods for variable selection should be specified in this section. For example, consider including variables that are statistically or near statistically significant (P < .15) and biologically relevant variables (age and gender) in the regression model. A more sophisticated data analysis section could include a description of sample size estimation. All randomized clinical trials demand an a priori sample size computation before the first subject is enrolled. Details of sample size determination would include explicit statements of the expected treatment effect (50% decrease in complications in treatment group), tails (1- or 2-tailed tests of hypotheses), and a and b errors (P < .05 and b = 0.2 or power of 0.8). If statisticians are involved in the analyses, then, depending on the level of involvement, be sure they are included as authors or acknowledged in the paper. The reader might derive confidence from knowing a statistician was involved in the analyses. ELEMENT IV—RESULTS
The easiest way to write the results section is to complete all the tables first. The text flows naturally from the tables. A case series might have a single table that includes a summary of descriptive statistics of the sample and the outcome variable. Most analytic studies will be composed of 3 to 4 tables: 3 to 4 tables for studies limited to bivariate analyses and 4 to 5 tables for those that include regression analyses. When designing and formatting a table, consider the following assertion: ‘‘A table should stand on its own.’’ The table should not force the reader to flip back and forth in the paper to interpret the table. Make liberal use of footnotes to facilitate interpreting the table. All studies should include at least 1 table called ‘‘Table 1.’’ Using descriptive statistics, ‘‘Table 1’’ summarizes for the reader the characteristics of the study sample (Table 1A). Even case series should have a ‘‘Table 1’’ unless the sample is too small to warrant a table and the sample’s characteristics can be summarized in
WRITING A SCIENTIFIC PAPER
the text (eg, ‘‘The sample was composed of 3 subjects whose ages ranged from 10 to 14 and 2 were female’’). The value of this ‘‘Table 1’’ is to help readers ascertain whether the study sample is relevant and composed of patients much like their own. If the study sample deviates sharply from the readers’ own samples, then they might want to stop investing time in this paper and move to a paper with a more relevant sample. A challenge of interpreting any study is to ascertain issues regarding selection bias. In the materials and methods section, the investigators describe their inclusion and exclusion criteria that outline their ideal study sample. Needless to say, the ideal sample and the study sample are rarely the same. Subjects are lost to follow-up, charts are not found, data collection is incomplete, or data are missing. All these issues result in a final sample that deviates in small (or large) ways from the planned study sample. If the investigators are missing a ‘‘significant’’ amount of data or subjects, an important variant of a ‘‘Table 1’’ is to summarize the bivariate associations between subjects included and excluded from the study (Table 1B). This information permits readers to ascertain important selection biases that could affect the interpretation of the results. Ideally, the investigators would like to see the results of Table 1B show no statistical differences between subjects included and excluded from the final study sample. Differences between included and excluded samples should be addressed in the discussion because key differences could affect interpretation of the study findings. Table 2 summarizes the bivariate associations between the primary predictor variable and all other study variables, except the primary outcome variable, and includes a column of P values. Sometimes to meet space requirements, I will not use a ‘ Table 1.’’ Instead, I will summarize the key descriptive statistics in the text and start with a ‘ Table 2.’’ The value to Table 2 is that it helps identify variables that are not equally distributed between or among the primary predicator variables, creating unequal samples. For example, age might be associated with an adverse outcome. After assembling the study groups, it might become apparent that one group is older than the other group. This difference in age could bias one group toward having a better outcome than the other group. Of note, for randomized studies, there is controversy as to whether P values should be included in a ‘ Table 2.’’6,7 The argument is that randomization should create groups with the known and unknown variables equally distributed between groups. As such, testing the hypothesis of equal groups is unnecessary. From a personal standpoint, I like to test the hypothesis that the treatment groups have their study variables equally distributed. Especially in a small sample, there is a chance that the variables are not
S165
THOMAS B. DODSON
Table 1. EXAMPLES—DESCRIPTIVE SUMMARY OF STUDY SAMPLE
A—Summary of Study Variables for Entire Sample Study Variable (Name and Type)
Descriptive Statistics
Sample size (n) Gender (binary)—male* Anatomic location (categorical) Maxilla Mandible Both Age—(continuous)
n (report total sample size) n (x%)y n1 (x1%) n2 (x2%) n3 (x3%) Mean SD or median quartiles or range
B—Summary of Study Variables Grouped by Study Inclusion or Follow-Up Status Variable Name (and Type)
Sample
Lost to Follow-Up
P Value
Sample size (n) Gender—male (binary) Anatomic location (categorical) Maxilla Mandible Both Age (continuous)
n1 (x1%) n1 (x1%)
n2 (x2%) n2 (x2%)
Not applicable c2z
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
c2z t test or ANOVAx
Abbreviations: ANOVA, analysis of variance; SD, standard deviation. * For binary variables, when the choices are obvious (eg, yes or no, male or female), just 1 variable can be listed to simplify the table. With 1 variable, the numeric details of the other variable can be computed by subtraction. y The n is the number of male subjects and x% is the total percentage of male subjects in the sample. The number of female subjects can be computed by subtracting the number of male subjects from the total sample size. The percentage of female subjects can be computed by subtracting the percentage of male subjects from 100%. z On average, c2 would be the appropriate test to compute the P value. x In most cases, t test or ANOVA is robust enough to be the appropriate test to compute the P value. There are situations in which nonparametric tests might be more appropriate, but that discussion is beyond the scope of this article. Thomas B. Dodson. Writing a Scientific Paper. J Oral Maxillofac Surg 2015.
distributed equally between or among treatment groups. If I know this, then I have the opportunity to address this inequality in analyses that adjust for the relevant variables or in the discussion. Table 3 summarizes the bivariate associations between the primary outcome variable and all other study variables, except the primary predictor variable, and includes a column for P values. The purpose of a ‘ Table 3’’ is to identify any variables, other than the primary predictor variable, associated with the outcome variable. ‘ Table 3’’ should be formatted just like ‘ Table 2,’’ substituting the outcome variable for the predictor variable in the column headings. I use the findings from Tables 2 and 3 to identify a subset of variables common to the 2 tables that have small P values (eg, <.15). This subset can include variables that need to be addressed in the regression analyses (Table 5). Table 4 summarizes the bivariate relation between the primary predictor and the outcome variables. I like to isolate this table from the other tables to highlight the primary analysis of interest from all other analyses. It might be the only table readers need to use.
If the predictor and outcome variables are binary, the relation can be summarized in a 2 2 table with the predictor variable in the rows and the outcome variable in the columns (Table 4A). A 2 2 table is a simple but powerful tool to communicate information. I strive to construct studies designed to use 2 2 tables whenever possible. It permits the computation of a statistic (c2 and its associated P value) measuring the association between the primary predictor and outcome variables. Depending on the study or purpose, it facilitates computation of odds ratios or relative risks, treatment (or harm) effect and numbers needed to treat (or harm), or measurements of a diagnostic test (ie, specificity, sensitivity, and positive and negative predictive values). However, many times the predictor and outcome variables are not binary. If the predictor variable is binary and the outcome variable is continuous, a t test could be computed (Table 4B). If the predictor variable is categorical and the outcome variable is continuous, analysis of variance is a good tool to measure the association between variables (Table 4C).
S166
WRITING A SCIENTIFIC PAPER
Table 2. EXAMPLES—STUDY VARIABLES VERSUS PREDICTOR VARIABLE
A—With Binary Predictor Variable Variable Name (and Type) Sample size (n) Gender—male (binary) Anatomic location (categorical) Maxilla Mandible Both Age (continuous)
Antibiotic
Placebo
P Value
n1 (x1%) n1 (x1%)
n2 (x2%) n2 (x2%)
Not applicable c2*
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
c2* t test or ANOVAy
B—With Categorical Predictor Variable Variable Name (and Type) Sample size (n) Gender—male (binary) Anatomic location (categorical) Maxilla Mandible Both Age (continuous)
Preoperative Antibiotics
Postoperative Antibiotics
Placebo
P Value
n1 (x1%) n1 (x1%)
n2 (x2%) n2 (x2%)
n3 (x3%) n3 (x3%)
Not applicable c2*
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
c2* ANOVAz
C—With Continuous Predictor Variable Variable Name (and Type) Sample size (n) Gender Male Female Anatomic location (categorical) Maxilla only Mandible Both Duration of disability (days) (continuous)
Antibiotic Dose (Units of Penicillin)
P Value
n1 (x1%)
Not applicable
Mean SD Mean SD
t test
Mean SD Mean SD Mean SD Report the correlation or regression coefficients (with confidence intervals)
ANOVA Pearson correlation or regression analyses
Abbreviations: ANOVA, analysis of variance; SD, standard deviation. * On average, c2 would be the appropriate test to compute the P value. y In most cases, t test or ANOVA is robust enough to be the appropriate test to compute the P value. There are situations in which nonparametric tests might be more appropriate, but that discussion is beyond the scope of this article. z In most cases, ANOVA is robust enough to be the appropriate test to compute the P value. There are situations in which nonparametric tests might be more appropriate, but that discussion is beyond the scope of this article. Thomas B. Dodson. Writing a Scientific Paper. J Oral Maxillofac Surg 2015.
If the predictor and outcome variables are continuous, correlation or regression coefficients can be computed and reported (Table 4D). Unless the study is a case series or there are very few variables, it is difficult to summarize the findings in fewer than 3 to 4 tables. More tables are used if the investigators propose using regression techniques to measure the relation between the predictor and outcome variables adjusted for other important variables. The variables age and gender carry a large
amount of biologic information. As such, I frequently will create a preliminary regression model that measures the association between the primary predictor and outcome variables adjusted for age and gender. Studies designed to match for age have already been adjusted by the nature of the study design. Table 5 summarizes the results of regression analyses. ‘‘Table 5’’ includes all the study variables that are included in the final regression model with their associated coefficients, confidence intervals, and
S167
THOMAS B. DODSON
Table 3. EXAMPLES—STUDY VARIABLES VERSUS OUTCOME VARIABLE
With a Binary Outcome Variable Variable Name (and Type) Sample size (n) Gender—male (binary) Anatomic location (categorical) Maxilla Mandible Both Age (continuous)
Infection (Yes)
Infection (No)
P Value
n1 (x1%) n1 (x1%)
n2 (x2%) n2 (x2%)
Not applicable c 2*
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
c 2* t test or ANOVAy
Note: Tables with categorical or continuous outcome variables are formatted and analyzed exactly as the Table 2 examples except the columns are labeled with the primary outcome variable rather than the primary predictor variable. Abbreviations: ANOVA, analysis of variance; SD, standard deviation. * On average, the c2 would be the appropriate test to compute the P value. y In most cases, t test or ANOVA is robust enough to be the appropriate test to compute the P value. There are situations in which nonparametric tests might be more appropriate, but that discussion is beyond the scope of this article. Thomas B. Dodson. Writing a Scientific Paper. J Oral Maxillofac Surg 2015.
P values. ‘‘Table 5’’ has the ‘‘take-home’’ information regarding the study. As such, time and care should be spent constructing it. As the reader, time and care should be spent reviewing and interpreting it. Once the tables have been completed, drafting the results section should be straightforward. Because the details of the study results are in the tables for the readers to peruse at their leisure, the text should summarize only the high points of each table and should be presented in the same order as the tables are laid out. The first paragraph of the results summarizes the characteristics of the study sample. For example, ‘‘During the study interval, 250 subjects were screened for eligibility. The final sample was composed of 125 subjects with a mean age of 25 4.2 years and 60% (75) were female.’’ The subsequent 1 or 2 paragraphs should summarize the important findings from Tables 2 and 3. For example, ‘‘There were no statistically significant differences in the distribution of any of the study variables between subjects assigned to the prophylactic antibiotics and those subjects assigned to placebo groups.’’ If there are differences in the distribution of study variables between the 2 groups, the investigators should describe in detail these variables. For example, ‘‘The mean ages of subjects assigned to the antibiotic and place groups were 18.2 3.2 years and 26.8 3.5 years, respectively (P = .002).’’ The sentences or paragraphs that summarize ‘‘Table 3’’ should parallel the summary of ‘‘Table 2,’’ with the emphasis on the primary outcome variable. The next paragraph should summarize the results of ‘‘Table 4.’’ This is an important paragraph to write. It summarizes the key finding of the study, namely the unadjusted relation between the primary predictor
and outcome variables. It should include a simple statement that summarizes the magnitude of the association between the predictor and outcome variables and the degree of statistical significance. Additional paragraphs are written as needed to summarize the findings of regression analyses. ELEMENT V—DISCUSSION
Writing the discussion can be challenging.8 To start, consider preparing the discussion section using a formulaic structure composed of 4 sections. The goal of the first section, usually limited to a single paragraph, is to draw the reader’s attention back to the goal of the study. This paragraph has 3 sentences. The first sentence restates the study purpose, the second summarizes the hypothesis, and the third recapitulates the specific aims. Many times, it resembles the last paragraph of the introduction. Don’t bury the lead! The second section summarizes the key results as they relate to the study purpose or hypothesis. This is an important paragraph to compose because it includes the study’s take-home message. For example, ‘ The results of this study confirm the hypothesis that antibiotics administered after the operation are as effective as no antibiotics in preventing postoperative inflammatory complications. The frequency of complications in the antibiotic and no-antibiotic groups were 6.2 and 5.8%, respectively (P = .8). Based on the results of this study, the postoperative use of oral antibiotics is unwarranted and could be associated with an increased risk of undesirable side effects such as nausea, allergic reactions, or unnecessary patient expense.’’ In lieu of including a literature review in the introduction, the third section should summarize how
S168
WRITING A SCIENTIFIC PAPER
Table 4. EXAMPLE—SUMMARY OF BIVARIATE ASSOCIATION BETWEEN PREDICTOR AND OUTCOME VARIABLES
A—2 2 Table—Predictor and Outcome Variables Are Binary Outcome Variables Predictor Variable Antibiotic Placebo Total
‘‘Infection’’
‘‘No Infection’’
Total
a c a+c
b d b+d
a+b c+d a+b+c+d
B—Binary Predictor and Continuous Outcome Variable Name
Duration of Disability (days)
P Value
Mean SD Mean SD
t test
Treatment Antibiotic Placebo
C—Categorical Predictor and Continuous Outcome Variable Name (and Type)
Duration of Disability (days)
P Value
Antibiotic management Preoperative antibiotic Postoperative antibiotic Placebo
Mean SD Mean SD Mean SD
ANOVA
D—Continuous Predictor and Outcome Variables Outcome—Duration of Disability (days) Predictor—antibiotic dose (penicillin units)
P Value
Report the correlation coefficient or b coefficient of regression analyses with appropriate confidence levels
Abbreviations: ANOVA, analysis of variance; SD, standard deviation. From this 2 2 table, one can compute the following measurements of association (with appropriate confidence intervals): 1. Relative risk (RR) for a cohort study—(a/a + b)/(c/c + d) 2. Odds ratio (OR) for a case-control study—(a d)/(b c) 3. For a diagnostic test: A. Sensitivity—a/a + c B. Specificity—d/b + d C. Positive predictive value (PPV)—a/a + b 4. Negative predictive value (NPV)—d/c + d 5. Treatment effect (TE)—(a/a + b)—(c/d + d) A. number needed to treat (NNT)—1/TE Thomas B. Dodson. Writing a Scientific Paper. J Oral Maxillofac Surg 2015.
Table 5. EXAMPLE—SUMMARY OF REGRESSION MODEL
Study Variable
b Coefficient
95% Confidence Interval
P Value
Thomas B. Dodson. Writing a Scientific Paper. J Oral Maxillofac Surg 2015.
this study’s results compare with those of other studies published on the topic. This section should be as long as necessary, but focused. If the investigator starts on tangents or addresses distantly related topics, this reader quickly loses interest. The mark of a more sophisticated paper is in the final section that summarizes the weaknesses and strengths of the study. I organize this section by addressing a study weakness and how it was offset or neutralized
S169
THOMAS B. DODSON
by a strength, such as study design or analytic method. If the weakness persists, I outline how it could affect the reader’s interpretation of the results.
tion is achieved ‘ . not when there is nothing more to add, but when there is nothing left to take away.’’10 Acknowledgments
ELEMENT VI—CONCLUSION
Depending on the structure of the discussion or journal’s parameters for a manuscript, the conclusion might be superfluous. When writing the conclusion, I will generally summarize the key findings of the study. It will closely resemble an abbreviated second paragraph of the discussion. This section will commonly include a brief outline of future research questions that have been raised by the current study. ELEMENT VII—ABSTRACT
Writing the abstract is the most challenging aspect of a well-composed paper. The abstract must articulate for the reader all key points of the study within prescribed space limitations (eg, <250 words). By the time I have prepared a paper for publication, I have internalized the study, its results, outcomes, and take-home message. The abstract becomes much easier to write. ‘‘The very act of writing for publication imposes a discipline and forces issues to be thought through in a logical manner..’’9 The preparation of an excellent patient-oriented research paper requires structured procedures and organization paralleling the design and implementation of the original research project. ‘ In today’s world of evidence-based medicine, a clinician’s ability to critique research publications . is a crucial skill; one that is greatly enhanced by conducting research and especially by writing it up for publication.’’9 A thoughtfully prepared patientoriented research paper will facilitate clinicians’ review activities by being accessible and permitting them to grasp the study’s relevance and appropriateness to enhance their patients’ care. Remember, perfec-
I thank Dr James Hupp for the motivation to prepare the initial paper.2 I also recognize my research mentors and colleagues who have helped me develop the manuscript style outlined in this article: Dr Leonard B. Kaban, the faculty from the Division of Clinical Epidemiology, University of California–San Francisco, and the faculty teaching EBM Workshops at McMaster University, Hamilton, Ontario, Canada. A very special thank you is extended to authors submitting papers to the Journal of Oral and Maxillofacial Surgery and dozens of residents, junior faculty, and students who I mentored and obligated me to be explicit about my expectations when editing their works.
References 1. Dodson TB, Kaban LB: California mandatory seatbelt law: The effect of recent legislation on motor vehicle accident-related maxillofacial injuries. J Oral Maxillofac Surg 46:875, 1988 2. Dodson TB: A guide for preparing a patient-oriented research manuscript. Oral Surg Oral Med Oral Pathol Oral Radiol Endod 104:307, 2007 3. Pinker S: The curse of knowledge, in The Sense of Style. New York, NY, Viking, 2014, pp 57–76 4. Casagrande J: It Was the Best of Sentences, It Was the Worst of Sentences. Berkeley, CA, Ten Speed Press, 2010, pp 3–5 5. National Collaborating Centre for Methods and Tools. Defining your question: PICO and PS; 2012. Hamilton, ON, Canada, McMaster University, 2012. Available at: http://www.nccmt. ca/registry/view/eng/138.html. Accessed April 8, 2015. 6. Austin PC, Manca A, Zwarenstein M, et al: A substantial and confusing variation exists in handling of baseline covariates in randomized controlled trials: A review of trials published in leading medical journals. J Clin Epidemiol 63:142, 2010 7. Wang R, Lagakos SW, Ware JH, et al: Statistics in medicine. Reporting of subgroup analyses in clinical trials. N Engl J Med 357:2189, 2007 8. Jenicek M: How to read, understand, and write ‘ Discussion’’ sections in medical articles. An exercise in critical thinking. Med Sci Monit 12:SR28, 2006 9. Rosenfeldt FL, Dowling JT, Pepe S, et al: How to write a paper for publication. Heart Lung Circ 9:82, 2000 10. Antoine de Saint-Exupery. Available at: http://en.thinkexist.com/ quotation/a_designer_knows_he_has_achieved_perfection_not/ 221998.html. Accessed April 7, 2015
Writing a Scientific Paper Is Not Rocket Science! Thomas B. Dodson, DMD, MPH I elected to update an article on how to write a scientific paper as my contribution to this supplement because Dr Leonard Kaban introduced me to the discipline of writing. Our first collaboration resulted in a paper accepted at its initial submission.1 The present submission is based, with permission, on an article published 8 years ago.2 Ours is a mentor-driven specialty. Like those before me, I learned by observing others. Continuing the tradition, I would like to share strategies and detail tactics I use to write a scientific paper. The best written patientoriented research articles do not make me, the reader, work hard to understand their meanings. It follows that writing a patient-oriented research paper should be a straightforward exercise that translates data into a clear, practical lesson for the clinician. It should not be a burden to write or read. This is not a definitive article on the topic of preparing a scientific paper. Rather, it is an overview of my personal process for writing a paper. This process is dynamic and evolving and has been guided by mentors and associates too numerous to name. Along the way, I have collected 12 aphorisms guiding manuscript preparation: 1. ‘‘There is no such thing as a paper that is too short.’’ Hemingway wrote a story in 6 words: ‘‘For sale: baby shoes. Never worn.’’ 2. ‘‘Write short declarative sentences.’’ 3. ‘‘All studies, no matter how complicated, can be resolved into a 2 2 table.’’ 4. ‘‘Surgeons have the attention span of a flea. You have 30 seconds to get their attention.’’ 5. ‘‘Readers should not have to guess your study purpose.’’
6. ‘‘It is easier to write when you have something to say.’’ 7. ‘‘Avoid the passive voice.’’ 8. ‘ You treat patients and do research with subjects.’’ 9. ‘‘A table should stand on its own.’’ 10. ‘‘Beware the ‘curse of knowledge’’’ The curse of knowledge is the difficulty in imagining what it is like for someone else, the reader, not to know something that you, the writer, know. The curse of knowledge might explain why good people write bad prose. It simply does not occur to writers that readers do not know what they know.3 11. ‘ The reader is king.’’ Another reason why good writing goes bad is that someone is missing when the author is writing. Writing is not a group experience. Writing is an individual experience. We write alone. The audience is missing. In place of the audience is a phantom we have never met, that is, the reader. The reader is king. The writer is his servant. One thing readers want is clear, concise, comprehensible sentences. By outlining a set of organizational principles for drafting a patient-oriented paper, the writer can focus on writing ‘ clear, concise, comprehensible sentences.’’4 12. ‘‘A manuscript submitted for journal publication is not a thesis.’’ A thesis is a long-winded document reviewed by a committee, bound, and filed in the library’s bowels, never to see the light of day. In contrast, a scientific paper is topically focused with clearly delineated observations and recommendations.
Professor and Chair, Department of Oral and Maxillofacial
Health Sciences Building, Room B-241, Box 357134, Seattle, WA
Surgery, Associate Dean for Hospital Affairs, University of
98195-7134; e-mail: [email protected]
Washington School of Dentistry, Seattle, WA.
Received April 10 2015
This work was supported in part by the University of Washington School of Dentistry Department of Oral and Maxillofacial Surgery
Accepted April 10 2015
Research and Training Fund.
0278-2391/15/00499-1
Conflict of Interest Disclosures: The author did not report any
Ó 2015 American Association of Oral and Maxillofacial Surgeons
http://dx.doi.org/10.1016/j.joms.2015.04.039
disclosures. Address correspondence and reprint requests to Dr Dodson: University of Washington School of Dentistry, 1959 NE Pacific Street,
S160
THOMAS B. DODSON
Elements of a Patient-Oriented Manuscript A patient-oriented research paper has 7 elements: title, introduction, materials and methods, results, discussion, conclusion, and abstract. The title grabs the reader. The introduction is short. It focuses the reader on the clinical problem and study goals. The materials and methods section is as long as necessary to cover the elements of the study. Writing the results section is straightforward, especially if you construct and populate the tables first. The discussion can be variable in length, but its structure is formulaic. The conclusion is short. Write the abstract last. Given the space constraints of most abstracts, it might be the most important and challenging aspect of the manuscript to draft. Word count limitations for abstracts preclude wasted words, and the abstract might be the only element of your paper that is read. ELEMENT 1—TITLE
The title’s goal is to engage and entice the reader to read more. The title should be short. Try to limit the study title to 10 to 15 words. To shorten a title, avoid including the study design in the title, eg, ‘‘placebocontrolled, double-blinded, randomized clinical trial’’ (5 words) or ‘‘systematic review and meta-analysis’’ (4 words). Authors can describe the study design in the abstract and body of the paper. Consider using a title composed of a declarative sentence or a question, rather than a technically correct, but verbose title. For example, instead of ‘‘Use of the pyriform ligament for the alar cinch to enhance nasal appearance after Le Fort I osteotomy’’ (16 words), consider a declarative sentence, ‘‘Proper management of the pyriform ligament improves nasal appearance after Le Fort I osteotomy’’ (12 words). An example of constructing the title as a question is ‘‘Does the pyriform ligament exist and why surgeons should care?’’ (10 words). ELEMENT II—THE INTRODUCTION
The purpose of the introduction is to provide enough information to ‘‘hook’’ the reader. The introduction is focused, short, and should be shorter than 2 pages, double-spaced, and with an 11- to 12-point font. The introduction should answer the following questions: 1) Why is this clinical problem of interest? 2) What are the deficiencies or controversies in the current literature? 3) What is the purpose of this paper? An introduction composed of as few as 2 to 3 paragraphs can answer these 3 questions well. Try to avoid including information in the introduction that is better reserved for the discussion section of the paper (eg, a comprehensive literature review).
S161 A key element of the introduction is its final paragraph summarizing the study purpose. The good writer provides readers a study purpose with sufficient specificity to avoid confusion or ambiguity. This is the author’s first, best, and possibly only opportunity to share with the reader exactly what the paper is about and not leave it to the reader to guess. It is dangerous, sometimes fatal to the paper, to let the reader guess the study purpose. The reader might conclude that a good paper is garbage because it failed to address the reader’s purpose or, more specifically, the reader’s perception of the purpose. A well-composed ‘‘purpose’’ paragraph has 3 elements: 1) a purpose or aim statement, 2) a hypothesis statement, and 3) a specific-aim statement. I use 2 different techniques to state the study purpose. The conventional technique is to state the purpose literally (eg, ‘‘The purpose of this study is to measure the efficacy of prophylactic antibiotics in preventing postoperative complications in patients undergoing third molar removal’’). However, this conventional purpose statement could leave ambiguity in the reader’s mind (‘‘What is efficacy?’’ ‘‘How will it be measured?’’ ‘‘What complications will be studied?’’). An alternative approach is to state the research purpose in the form of a clinical question using the PICO acronym. The PICO acronym is composed of 4 elements: patient sample (P), intervention (I), control or comparison (C), and outcome (O).5 Using the earlier example, the purpose can be reformulated as a clinical question: ‘‘The study’s purpose is to answer the following clinical question: ‘Among patients with impacted third molars (patient sample), does the use of prophylactic antibiotics (intervention of interest), when compared with a placebo control (control group or treatment), decrease the frequency of postoperative infections (outcome of interest)?’’’ Restating the purpose in the form of a specific clinical question minimizes ambiguity and focuses the reader on the research goal. Almost all clinical studies (and literature reviews) can have the purpose stated as a clinical question. Even case series can easily have the purpose stated in the form of a clinical question, without the ‘‘C’’ or control or comparison component. Compelling patient-oriented research is hypothesis driven. As such, the second element in the purpose paragraph is a hypothesis statement. Many times, based on the purpose statement, the reader can intuit the hypothesis. However, the reader should not work that hard. Do not leave the reader’s impression of the hypothesis to chance. As such, articulate the hypothesis statement explicitly. Hypotheses statements can be formal or informal. Examples of formal hypothesis statements include:
S162 1. ‘ The frequency of postoperative infections in the treatment group equals the frequency of infections in the control group.’’ This statement posed as the null hypothesis implies a 2-tailed test of hypothesis. 2. ‘‘The frequency of postoperative infections in the treatment group is lower than the frequency of infections in the control group.’’ This statement implies a 1-tailed test of hypothesis. Examples of less formal hypothesis statements include: 1. ‘‘Antibiotics decrease the frequency of postoperative complications.’’ 2. ‘‘There exists a set of at least 1 factor associated with an increased risk for surgical site infections that can be manipulated by the clinician to improve outcomes after third molar surgery.’’ Of note, a case series, by definition, will not have a hypothesis statement. The final element of the purpose paragraph is the specific aim(s) statement. Unambiguously, tell the reader what you did in the study. Specific aims use active verbs such as measure, design, identify, implement, estimate, compare, or identify. Examples of specific aims include: 1. To estimate and compare the frequencies of surgical site infection after third molar removal in subjects who did or did not receive prophylactic antibiotics. 2. To measure the length of hospitalization in a patient sample undergoing orthognathic surgery and identify factors associated with decreased length of stay. 3. To estimate the 1-year survival rates of implants loaded immediately and implants loaded in a delayed manner. ELEMENT III—MATERIALS AND METHODS
The materials and methods section of a patientoriented paper has a logical predictable structure anticipating readers’ questions. This section of the paper could be the most valuable to the reader and other researchers. The reader needs to understand in detail what was done. Researchers need to know what was done to replicate the study or when abstracting data for a systematic review or meta-analysis. The methods section should be easy to write. The methods have been written before enrolling the first subject or abstracting data from the first chart. Virtually all studies require protocols approved by an institutional review board. As such, with little modification, one can ‘‘cut and paste’’ the protocol
WRITING A SCIENTIFIC PAPER
from the institutional review board application to the manuscript. Failing to draft a written materials and methods section before data collection should give investigators pause, begging the question, ‘‘Is this submission a high-quality, thoughtful research endeavor?’’ The material and methods section has 4 parts: 1) study design, 2) sample identification and selection, 3) variables, and 4) data analyses. Each part should be clearly identified using subheadings for each section. Do not include details of the sample in the materials and methods (eg, sample size, age, and gender) unless the study was designed a priori to enroll a sample that had characteristics of the final study sample—a rare event. The details of the sample should be reported in the results because they are a natural consequence of the methods. Study Design The first sentence of the materials and methods unambiguously states the study design (eg, ‘‘The investigators designed and implemented a double-blind, placebo-controlled, randomized clinical trial’’). Study design has implications for interpreting the results and assessing the study’s level of evidence and associated value in clinical decision making. Do not leave the study design statement to the reader’s imagination. Study design choices are finite. Common study designs, in descending levels of validity, are systematic and meta-analyses of several randomized clinical trials, randomized clinical trials (multi- or single-institutional, double- or single-blinded, placebo-controlled), cohort (prospective or retrospective), case-and-control studies, cross-sectional studies, case series (prospective or retrospective), and case reports. Study Sample The second sentence should summarize clearly the parameters used to select the study sample. These parameters include the population from which the sample is derived and inclusion and exclusion criteria. The reader needs this information to determine whether the patient sample is relevant and of interest. In almost all studies, the study sample (ie, actual subjects) is derived from an underlying patient population (ie, eligible subjects). Owing to inclusion criteria, availability of charts or records, and loss to follow-up, the study population and sample are rarely the same. It is important for the reader to know from which population the sample was drawn. As an example, consider the following: ‘‘The study sample was derived from the population of patients who presented to the department of oral and maxillofacial surgery at a fine hospital for evaluation and management of third molars from January 1, 1998 through December 31, 2000.’’ Based on this information, the reader can assess
THOMAS B. DODSON
the sample’s relevance (eg, academic practice, not a private practice) or applicability (eg, late 20th century might not reflect current practice). In most studies, the investigators do not or cannot enroll every subject. Enrolled subjects meet prespecified inclusion and exclusion criteria. Manuscript authors should tell the reader exactly what these criteria were. Critical readers will use this information to determine whether this study’s sample is similar to other samples of interest (eg, their patients) and choose to continue reading the article. Before reviewing the first record or enrolling the first subject, the inclusion and exclusion criteria should be written down. Quite simply, this information needs to be incorporated into the paper. Inclusion and exclusion criteria are presented in the form of a list. For example, ‘‘Subjects eligible for study inclusion had at least 1 impacted wisdom tooth, were at least 18 years old, underwent operative treatment of the wisdom teeth, agreed to be enrolled in the study, and returned for at least 1 followup visit. Subjects were excluded from study enrollment if they had no impacted teeth, were younger than 18 years, failed to return for follow-up, or refused study enrollment.’’ Study Variables Studies have numerous variables that can be characterized as predictor, outcome, and ‘‘other.’’ In most patient-oriented research, the goal is to establish an association or relation between the predictor (independent) and outcome (dependent) variables, which is adjusted as needed for ‘‘other’’ variables. The investigator should articulate and define the primary predictor variable. In general, predictor variables include exposures, risk or prognostic factors, or treatments of interest (eg, treatment [active vs placebo], implant loading [immediate vs delayed], or dose of radiation therapy). Some studies can have multiple, heterogeneous predictor variables. These predictor variables should be grouped together into logical sets (eg, demographic, anatomic, radiographic, perioperative, etc). There might be numerous outcomes of interest, but the study should be designed to measure 1 primary outcome of interest (eg, postoperative inflammatory complication [yes or no], duration of implant survival [measured in months], or development of osteoradionecrosis [yes or no]). It should be clear to the reader which variable is primary and which variable(s) is a secondary outcome variable. How does the investigator determine the primary outcome variable? Hint: The variable you used to estimate sample size is your primary outcome variable. The ‘‘other’’ variable category includes a listing of all the other variables for which data were collected that help to describe the sample or could affect the relation between the predictor and outcome variables (eg, age,
S163 gender, comorbidities, implant size or coating, exposure to bisphosphonates, duration of operating time, time from injury to fracture repair, education, or socioeconomic status). In some studies, there might be numerous ‘ other’’ variables and it can be tedious to the reader to enumerate them. To make the information more accessible, I group these ‘ other’’ variables into sets. For example, if there are 30 variables, it is tedious to read a list. However, if the variables are grouped into several discrete, logical sets, the list seems less imposing. Commonly used variable categories include demographic (age, gender), socioeconomic (education, annual salary, ZIP codes), medical or dental (comorbidities, medications, status of occlusion, diseases that affect wound healing, steroid use), anatomic (maxilla or mandible or both, level of impaction, simple or compound fracture), time (duration of operation or time from injury to hospital admission), perioperative (open or closed reduction, grafting techniques), and materials (implant type, dimensions, coating, graft materials). The specifics of the investigator’s database will dictate the type and number of variable sets. Be careful not to include too many variable sets or it defeats the value of grouping the variables. All study variables need to be defined unambiguously. It should be clear how the variable is coded (eg, binary, categorical, or continuous). Coding has implications regarding data presentation and analyses. A binary variable only has 2 choices and order is generally unimportant (eg, yes or no, male or female, smoker or nonsmoker). A categorical variable has more than 2 choices (eg, the operation to extract a third molar was categorized as simple extraction, surgical extraction, soft tissue impaction, etc). For categorical variables, order might not be important. A continuous variable is any real number between positive and negative infinity. In reality, most continuous variables are bounded by ceiling or floor values; for example, age is a continuous variable that generally includes positive, nonzero numbers and is generally bounded by the values of 0 and 100. For data collection, collect continuous variables as continuous data, not categorical or binary variables. One can always convert a continuous variable to a categorical or binary variable. The converse is not true. Data Collection, Management, and Analyses Under data collection, I include information regarding the details of randomization, how the standard and experimental treatments were rendered, how subjects were managed, sample size estimates, and details of how data were abstracted from charts. This section would include information regarding methods used to assess intra- and inter-examiner variabilities or standardization of examiners or processes for data collection or abstraction. Data management
S164 would include information regarding who and how data were inputted and stored, what software was used for data storage and analyses, and methods used to assure accurate input (eg, double-entry techniques, software checks for erroneous inputs, and how missing data were managed). Data analyses include a brief description of analytic methods used and specification of level of a error (ie, P value). For example, ‘‘Descriptive statistics (mean, frequency, range, standard deviations) were computed for each study variable.’’ Bivariate analyses (eg, c2 test, t test) are computed to measure the association between any 2 variables of interest (eg, treatment type and age or treatment type and gender). When regression analyses are used, the methods for variable selection should be specified in this section. For example, consider including variables that are statistically or near statistically significant (P < .15) and biologically relevant variables (age and gender) in the regression model. A more sophisticated data analysis section could include a description of sample size estimation. All randomized clinical trials demand an a priori sample size computation before the first subject is enrolled. Details of sample size determination would include explicit statements of the expected treatment effect (50% decrease in complications in treatment group), tails (1- or 2-tailed tests of hypotheses), and a and b errors (P < .05 and b = 0.2 or power of 0.8). If statisticians are involved in the analyses, then, depending on the level of involvement, be sure they are included as authors or acknowledged in the paper. The reader might derive confidence from knowing a statistician was involved in the analyses. ELEMENT IV—RESULTS
The easiest way to write the results section is to complete all the tables first. The text flows naturally from the tables. A case series might have a single table that includes a summary of descriptive statistics of the sample and the outcome variable. Most analytic studies will be composed of 3 to 4 tables: 3 to 4 tables for studies limited to bivariate analyses and 4 to 5 tables for those that include regression analyses. When designing and formatting a table, consider the following assertion: ‘‘A table should stand on its own.’’ The table should not force the reader to flip back and forth in the paper to interpret the table. Make liberal use of footnotes to facilitate interpreting the table. All studies should include at least 1 table called ‘‘Table 1.’’ Using descriptive statistics, ‘‘Table 1’’ summarizes for the reader the characteristics of the study sample (Table 1A). Even case series should have a ‘‘Table 1’’ unless the sample is too small to warrant a table and the sample’s characteristics can be summarized in
WRITING A SCIENTIFIC PAPER
the text (eg, ‘‘The sample was composed of 3 subjects whose ages ranged from 10 to 14 and 2 were female’’). The value of this ‘‘Table 1’’ is to help readers ascertain whether the study sample is relevant and composed of patients much like their own. If the study sample deviates sharply from the readers’ own samples, then they might want to stop investing time in this paper and move to a paper with a more relevant sample. A challenge of interpreting any study is to ascertain issues regarding selection bias. In the materials and methods section, the investigators describe their inclusion and exclusion criteria that outline their ideal study sample. Needless to say, the ideal sample and the study sample are rarely the same. Subjects are lost to follow-up, charts are not found, data collection is incomplete, or data are missing. All these issues result in a final sample that deviates in small (or large) ways from the planned study sample. If the investigators are missing a ‘‘significant’’ amount of data or subjects, an important variant of a ‘‘Table 1’’ is to summarize the bivariate associations between subjects included and excluded from the study (Table 1B). This information permits readers to ascertain important selection biases that could affect the interpretation of the results. Ideally, the investigators would like to see the results of Table 1B show no statistical differences between subjects included and excluded from the final study sample. Differences between included and excluded samples should be addressed in the discussion because key differences could affect interpretation of the study findings. Table 2 summarizes the bivariate associations between the primary predictor variable and all other study variables, except the primary outcome variable, and includes a column of P values. Sometimes to meet space requirements, I will not use a ‘ Table 1.’’ Instead, I will summarize the key descriptive statistics in the text and start with a ‘ Table 2.’’ The value to Table 2 is that it helps identify variables that are not equally distributed between or among the primary predicator variables, creating unequal samples. For example, age might be associated with an adverse outcome. After assembling the study groups, it might become apparent that one group is older than the other group. This difference in age could bias one group toward having a better outcome than the other group. Of note, for randomized studies, there is controversy as to whether P values should be included in a ‘ Table 2.’’6,7 The argument is that randomization should create groups with the known and unknown variables equally distributed between groups. As such, testing the hypothesis of equal groups is unnecessary. From a personal standpoint, I like to test the hypothesis that the treatment groups have their study variables equally distributed. Especially in a small sample, there is a chance that the variables are not
S165
THOMAS B. DODSON
Table 1. EXAMPLES—DESCRIPTIVE SUMMARY OF STUDY SAMPLE
A—Summary of Study Variables for Entire Sample Study Variable (Name and Type)
Descriptive Statistics
Sample size (n) Gender (binary)—male* Anatomic location (categorical) Maxilla Mandible Both Age—(continuous)
n (report total sample size) n (x%)y n1 (x1%) n2 (x2%) n3 (x3%) Mean SD or median quartiles or range
B—Summary of Study Variables Grouped by Study Inclusion or Follow-Up Status Variable Name (and Type)
Sample
Lost to Follow-Up
P Value
Sample size (n) Gender—male (binary) Anatomic location (categorical) Maxilla Mandible Both Age (continuous)
n1 (x1%) n1 (x1%)
n2 (x2%) n2 (x2%)
Not applicable c2z
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
c2z t test or ANOVAx
Abbreviations: ANOVA, analysis of variance; SD, standard deviation. * For binary variables, when the choices are obvious (eg, yes or no, male or female), just 1 variable can be listed to simplify the table. With 1 variable, the numeric details of the other variable can be computed by subtraction. y The n is the number of male subjects and x% is the total percentage of male subjects in the sample. The number of female subjects can be computed by subtracting the number of male subjects from the total sample size. The percentage of female subjects can be computed by subtracting the percentage of male subjects from 100%. z On average, c2 would be the appropriate test to compute the P value. x In most cases, t test or ANOVA is robust enough to be the appropriate test to compute the P value. There are situations in which nonparametric tests might be more appropriate, but that discussion is beyond the scope of this article. Thomas B. Dodson. Writing a Scientific Paper. J Oral Maxillofac Surg 2015.
distributed equally between or among treatment groups. If I know this, then I have the opportunity to address this inequality in analyses that adjust for the relevant variables or in the discussion. Table 3 summarizes the bivariate associations between the primary outcome variable and all other study variables, except the primary predictor variable, and includes a column for P values. The purpose of a ‘ Table 3’’ is to identify any variables, other than the primary predictor variable, associated with the outcome variable. ‘ Table 3’’ should be formatted just like ‘ Table 2,’’ substituting the outcome variable for the predictor variable in the column headings. I use the findings from Tables 2 and 3 to identify a subset of variables common to the 2 tables that have small P values (eg, <.15). This subset can include variables that need to be addressed in the regression analyses (Table 5). Table 4 summarizes the bivariate relation between the primary predictor and the outcome variables. I like to isolate this table from the other tables to highlight the primary analysis of interest from all other analyses. It might be the only table readers need to use.
If the predictor and outcome variables are binary, the relation can be summarized in a 2 2 table with the predictor variable in the rows and the outcome variable in the columns (Table 4A). A 2 2 table is a simple but powerful tool to communicate information. I strive to construct studies designed to use 2 2 tables whenever possible. It permits the computation of a statistic (c2 and its associated P value) measuring the association between the primary predictor and outcome variables. Depending on the study or purpose, it facilitates computation of odds ratios or relative risks, treatment (or harm) effect and numbers needed to treat (or harm), or measurements of a diagnostic test (ie, specificity, sensitivity, and positive and negative predictive values). However, many times the predictor and outcome variables are not binary. If the predictor variable is binary and the outcome variable is continuous, a t test could be computed (Table 4B). If the predictor variable is categorical and the outcome variable is continuous, analysis of variance is a good tool to measure the association between variables (Table 4C).
S166
WRITING A SCIENTIFIC PAPER
Table 2. EXAMPLES—STUDY VARIABLES VERSUS PREDICTOR VARIABLE
A—With Binary Predictor Variable Variable Name (and Type) Sample size (n) Gender—male (binary) Anatomic location (categorical) Maxilla Mandible Both Age (continuous)
Antibiotic
Placebo
P Value
n1 (x1%) n1 (x1%)
n2 (x2%) n2 (x2%)
Not applicable c2*
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
c2* t test or ANOVAy
B—With Categorical Predictor Variable Variable Name (and Type) Sample size (n) Gender—male (binary) Anatomic location (categorical) Maxilla Mandible Both Age (continuous)
Preoperative Antibiotics
Postoperative Antibiotics
Placebo
P Value
n1 (x1%) n1 (x1%)
n2 (x2%) n2 (x2%)
n3 (x3%) n3 (x3%)
Not applicable c2*
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
c2* ANOVAz
C—With Continuous Predictor Variable Variable Name (and Type) Sample size (n) Gender Male Female Anatomic location (categorical) Maxilla only Mandible Both Duration of disability (days) (continuous)
Antibiotic Dose (Units of Penicillin)
P Value
n1 (x1%)
Not applicable
Mean SD Mean SD
t test
Mean SD Mean SD Mean SD Report the correlation or regression coefficients (with confidence intervals)
ANOVA Pearson correlation or regression analyses
Abbreviations: ANOVA, analysis of variance; SD, standard deviation. * On average, c2 would be the appropriate test to compute the P value. y In most cases, t test or ANOVA is robust enough to be the appropriate test to compute the P value. There are situations in which nonparametric tests might be more appropriate, but that discussion is beyond the scope of this article. z In most cases, ANOVA is robust enough to be the appropriate test to compute the P value. There are situations in which nonparametric tests might be more appropriate, but that discussion is beyond the scope of this article. Thomas B. Dodson. Writing a Scientific Paper. J Oral Maxillofac Surg 2015.
If the predictor and outcome variables are continuous, correlation or regression coefficients can be computed and reported (Table 4D). Unless the study is a case series or there are very few variables, it is difficult to summarize the findings in fewer than 3 to 4 tables. More tables are used if the investigators propose using regression techniques to measure the relation between the predictor and outcome variables adjusted for other important variables. The variables age and gender carry a large
amount of biologic information. As such, I frequently will create a preliminary regression model that measures the association between the primary predictor and outcome variables adjusted for age and gender. Studies designed to match for age have already been adjusted by the nature of the study design. Table 5 summarizes the results of regression analyses. ‘‘Table 5’’ includes all the study variables that are included in the final regression model with their associated coefficients, confidence intervals, and
S167
THOMAS B. DODSON
Table 3. EXAMPLES—STUDY VARIABLES VERSUS OUTCOME VARIABLE
With a Binary Outcome Variable Variable Name (and Type) Sample size (n) Gender—male (binary) Anatomic location (categorical) Maxilla Mandible Both Age (continuous)
Infection (Yes)
Infection (No)
P Value
n1 (x1%) n1 (x1%)
n2 (x2%) n2 (x2%)
Not applicable c 2*
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
n1 (x1%) n2 (x2%) n3 (x3%) Mean SD
c 2* t test or ANOVAy
Note: Tables with categorical or continuous outcome variables are formatted and analyzed exactly as the Table 2 examples except the columns are labeled with the primary outcome variable rather than the primary predictor variable. Abbreviations: ANOVA, analysis of variance; SD, standard deviation. * On average, the c2 would be the appropriate test to compute the P value. y In most cases, t test or ANOVA is robust enough to be the appropriate test to compute the P value. There are situations in which nonparametric tests might be more appropriate, but that discussion is beyond the scope of this article. Thomas B. Dodson. Writing a Scientific Paper. J Oral Maxillofac Surg 2015.
P values. ‘‘Table 5’’ has the ‘‘take-home’’ information regarding the study. As such, time and care should be spent constructing it. As the reader, time and care should be spent reviewing and interpreting it. Once the tables have been completed, drafting the results section should be straightforward. Because the details of the study results are in the tables for the readers to peruse at their leisure, the text should summarize only the high points of each table and should be presented in the same order as the tables are laid out. The first paragraph of the results summarizes the characteristics of the study sample. For example, ‘‘During the study interval, 250 subjects were screened for eligibility. The final sample was composed of 125 subjects with a mean age of 25 4.2 years and 60% (75) were female.’’ The subsequent 1 or 2 paragraphs should summarize the important findings from Tables 2 and 3. For example, ‘‘There were no statistically significant differences in the distribution of any of the study variables between subjects assigned to the prophylactic antibiotics and those subjects assigned to placebo groups.’’ If there are differences in the distribution of study variables between the 2 groups, the investigators should describe in detail these variables. For example, ‘‘The mean ages of subjects assigned to the antibiotic and place groups were 18.2 3.2 years and 26.8 3.5 years, respectively (P = .002).’’ The sentences or paragraphs that summarize ‘‘Table 3’’ should parallel the summary of ‘‘Table 2,’’ with the emphasis on the primary outcome variable. The next paragraph should summarize the results of ‘‘Table 4.’’ This is an important paragraph to write. It summarizes the key finding of the study, namely the unadjusted relation between the primary predictor
and outcome variables. It should include a simple statement that summarizes the magnitude of the association between the predictor and outcome variables and the degree of statistical significance. Additional paragraphs are written as needed to summarize the findings of regression analyses. ELEMENT V—DISCUSSION
Writing the discussion can be challenging.8 To start, consider preparing the discussion section using a formulaic structure composed of 4 sections. The goal of the first section, usually limited to a single paragraph, is to draw the reader’s attention back to the goal of the study. This paragraph has 3 sentences. The first sentence restates the study purpose, the second summarizes the hypothesis, and the third recapitulates the specific aims. Many times, it resembles the last paragraph of the introduction. Don’t bury the lead! The second section summarizes the key results as they relate to the study purpose or hypothesis. This is an important paragraph to compose because it includes the study’s take-home message. For example, ‘ The results of this study confirm the hypothesis that antibiotics administered after the operation are as effective as no antibiotics in preventing postoperative inflammatory complications. The frequency of complications in the antibiotic and no-antibiotic groups were 6.2 and 5.8%, respectively (P = .8). Based on the results of this study, the postoperative use of oral antibiotics is unwarranted and could be associated with an increased risk of undesirable side effects such as nausea, allergic reactions, or unnecessary patient expense.’’ In lieu of including a literature review in the introduction, the third section should summarize how
S168
WRITING A SCIENTIFIC PAPER
Table 4. EXAMPLE—SUMMARY OF BIVARIATE ASSOCIATION BETWEEN PREDICTOR AND OUTCOME VARIABLES
A—2 2 Table—Predictor and Outcome Variables Are Binary Outcome Variables Predictor Variable Antibiotic Placebo Total
‘‘Infection’’
‘‘No Infection’’
Total
a c a+c
b d b+d
a+b c+d a+b+c+d
B—Binary Predictor and Continuous Outcome Variable Name
Duration of Disability (days)
P Value
Mean SD Mean SD
t test
Treatment Antibiotic Placebo
C—Categorical Predictor and Continuous Outcome Variable Name (and Type)
Duration of Disability (days)
P Value
Antibiotic management Preoperative antibiotic Postoperative antibiotic Placebo
Mean SD Mean SD Mean SD
ANOVA
D—Continuous Predictor and Outcome Variables Outcome—Duration of Disability (days) Predictor—antibiotic dose (penicillin units)
P Value
Report the correlation coefficient or b coefficient of regression analyses with appropriate confidence levels
Abbreviations: ANOVA, analysis of variance; SD, standard deviation. From this 2 2 table, one can compute the following measurements of association (with appropriate confidence intervals): 1. Relative risk (RR) for a cohort study—(a/a + b)/(c/c + d) 2. Odds ratio (OR) for a case-control study—(a d)/(b c) 3. For a diagnostic test: A. Sensitivity—a/a + c B. Specificity—d/b + d C. Positive predictive value (PPV)—a/a + b 4. Negative predictive value (NPV)—d/c + d 5. Treatment effect (TE)—(a/a + b)—(c/d + d) A. number needed to treat (NNT)—1/TE Thomas B. Dodson. Writing a Scientific Paper. J Oral Maxillofac Surg 2015.
Table 5. EXAMPLE—SUMMARY OF REGRESSION MODEL
Study Variable
b Coefficient
95% Confidence Interval
P Value
Thomas B. Dodson. Writing a Scientific Paper. J Oral Maxillofac Surg 2015.
this study’s results compare with those of other studies published on the topic. This section should be as long as necessary, but focused. If the investigator starts on tangents or addresses distantly related topics, this reader quickly loses interest. The mark of a more sophisticated paper is in the final section that summarizes the weaknesses and strengths of the study. I organize this section by addressing a study weakness and how it was offset or neutralized
S169
THOMAS B. DODSON
by a strength, such as study design or analytic method. If the weakness persists, I outline how it could affect the reader’s interpretation of the results.
tion is achieved ‘ . not when there is nothing more to add, but when there is nothing left to take away.’’10 Acknowledgments
ELEMENT VI—CONCLUSION
Depending on the structure of the discussion or journal’s parameters for a manuscript, the conclusion might be superfluous. When writing the conclusion, I will generally summarize the key findings of the study. It will closely resemble an abbreviated second paragraph of the discussion. This section will commonly include a brief outline of future research questions that have been raised by the current study. ELEMENT VII—ABSTRACT
Writing the abstract is the most challenging aspect of a well-composed paper. The abstract must articulate for the reader all key points of the study within prescribed space limitations (eg, <250 words). By the time I have prepared a paper for publication, I have internalized the study, its results, outcomes, and take-home message. The abstract becomes much easier to write. ‘‘The very act of writing for publication imposes a discipline and forces issues to be thought through in a logical manner..’’9 The preparation of an excellent patient-oriented research paper requires structured procedures and organization paralleling the design and implementation of the original research project. ‘ In today’s world of evidence-based medicine, a clinician’s ability to critique research publications . is a crucial skill; one that is greatly enhanced by conducting research and especially by writing it up for publication.’’9 A thoughtfully prepared patientoriented research paper will facilitate clinicians’ review activities by being accessible and permitting them to grasp the study’s relevance and appropriateness to enhance their patients’ care. Remember, perfec-
I thank Dr James Hupp for the motivation to prepare the initial paper.2 I also recognize my research mentors and colleagues who have helped me develop the manuscript style outlined in this article: Dr Leonard B. Kaban, the faculty from the Division of Clinical Epidemiology, University of California–San Francisco, and the faculty teaching EBM Workshops at McMaster University, Hamilton, Ontario, Canada. A very special thank you is extended to authors submitting papers to the Journal of Oral and Maxillofacial Surgery and dozens of residents, junior faculty, and students who I mentored and obligated me to be explicit about my expectations when editing their works.
References 1. Dodson TB, Kaban LB: California mandatory seatbelt law: The effect of recent legislation on motor vehicle accident-related maxillofacial injuries. J Oral Maxillofac Surg 46:875, 1988 2. Dodson TB: A guide for preparing a patient-oriented research manuscript. Oral Surg Oral Med Oral Pathol Oral Radiol Endod 104:307, 2007 3. Pinker S: The curse of knowledge, in The Sense of Style. New York, NY, Viking, 2014, pp 57–76 4. Casagrande J: It Was the Best of Sentences, It Was the Worst of Sentences. Berkeley, CA, Ten Speed Press, 2010, pp 3–5 5. National Collaborating Centre for Methods and Tools. Defining your question: PICO and PS; 2012. Hamilton, ON, Canada, McMaster University, 2012. Available at: http://www.nccmt. ca/registry/view/eng/138.html. Accessed April 8, 2015. 6. Austin PC, Manca A, Zwarenstein M, et al: A substantial and confusing variation exists in handling of baseline covariates in randomized controlled trials: A review of trials published in leading medical journals. J Clin Epidemiol 63:142, 2010 7. Wang R, Lagakos SW, Ware JH, et al: Statistics in medicine. Reporting of subgroup analyses in clinical trials. N Engl J Med 357:2189, 2007 8. Jenicek M: How to read, understand, and write ‘ Discussion’’ sections in medical articles. An exercise in critical thinking. Med Sci Monit 12:SR28, 2006 9. Rosenfeldt FL, Dowling JT, Pepe S, et al: How to write a paper for publication. Heart Lung Circ 9:82, 2000 10. Antoine de Saint-Exupery. Available at: http://en.thinkexist.com/ quotation/a_designer_knows_he_has_achieved_perfection_not/ 221998.html. Accessed April 7, 2015