Chapter 30: Data Management

An important part of the research planning process is the development of a data management plan that specifies how data will be recorded, organized, reduced and analyzed. This plan begins with the research proposal, specifying the research question, hypotheses and design. Before any data are collected, the researcher must be able to identify what variables will be measured, using what instruments and units of measurement. Those who will collect data may need to be trained and reliability assessments done. Undoubtedly, some of these plans will change once the project has begun, but nothing should begin without a firm plan in place. This planning requires knowledge of data coding and format requirements, statistics and computers. The purpose of this chapter is to describe procedures for setting up data to be entered into a computer and analyzed with statistical programs.

The research proposal will include a plan for handling data, including maintaining confidentiality of participant information. All subjects should be assigned a unique ID number that is not related to their name, medical unit number, Social Security number or other personal identifier. Documents for data collection should include the subject ID only. A list of subject names, addresses or phone numbers and corresponding ID codes can be kept separate and secured from other files in case participants need to be contacted.

As part of informed consent, subjects should be assured that their personal information, data from medical records and data collected as part of the project will only be accessed as necessary for research. The institutional review board (IRB) that approves the project will want to know the type of data to be collected, the purposes for which the data will be used, who will have access to records, and what safeguards have been put in place for security and confidentiality (see Chapter 3). Many countries have regulations in place that define these standards. In the United States, these are part of the Privacy Rule of the Health Insurance Portability and Accountability Act (HIPAA).¹ In Canada, they are incorporated into the Tri-Council Policy Statement: Ethical Conduct for Research Involving Humans.²

Throughout the project, researchers should have procedures in place to keep accurate and complete records of subject involvement. Records should indicate how many subjects were recruited and why some were not eligible, how many agreed to participate, and how many eventually did participate. Attrition should be monitored, and reasons noted if possible. Changes to the research protocol must be described. Initial group assignments and deviations from these assignments should be documented. This information is relevant to the validity of the project and will be important if the researcher wants to complete an intention to treat analysis (see Chapter 9).

The number of statistical packages for use on a microcomputer has grown dramatically, many at reasonable prices. The two most commonly used programs are SPSS (Statistical Package for the Social Sciences)^∗ and SAS (Statistical Analysis System).^† SPSS has traditionally been used more for the social and behavioral sciences, although its use has increased in health care research. SAS is most useful in medicine and epidemiology as a biostatistics program. These packages, once available only on main frames, have been adapted for use on personal computers. Many other programs are also on the market, and it would be useless to name them here as we are sure more will be published by the time you read this. Even though these packages are all slightly different, they adhere to certain standards that are important for data management. Most programs provide a format for data entry similar to a spreadsheet. Data may also be imported into a statistical program from a spreadsheet such as Microsoft Excel^®.

A data recording system must be carefully developed. Typically, data are collected from each subject and recorded on a separate sheet or directly into a computer program. The subject's identification code is listed, as well other relevant information such as the date, the individual collecting the data (if there is more than one investigator), the subject's group assignment and demographic information such as age, gender and diagnosis. If possible, all data should be listed in the order they will be included in the data file, to facilitate data entry. Figure 30.1 illustrates a data collection form for a study of two diet regimens in patients with diabetes.

The researcher must make decisions about how data will be recorded for each subject. Is there a level of precision in measurements that should be used, such as measuring to the nearest millimeter or half inch? The format for recording open-ended responses or qualitative data should be specified. If data are missing, the reason should be included. The importance of a well organized data collection scheme becomes most evident when the researcher begins to enter data into a computer. If data are not clearly recorded and in a consistent format, data entry will be a difficult and potentially error-ridden process.

An essential part of the data collection plan is the development of a scheme for recording data. Some measurements produce quantitative data, such as range of motion and blood pressure. Variables such as gender, group and race produce categorical data. Surveys and qualitative studies may produce open-ended responses that must be coded.

Types of Variables

Data can be entered as numerals or characters. Quantitative data are numeric, having values of single or multiple digits, sometimes including decimal points, and composed of only numbers. Numeric values can be preceded by a plus or minus sign, although plus signs are assumed and not entered. Character variables, called alphanumeric or string variables, are composed of letters or characters and may include digits. String variables may be letters or words that represent variable values, such as male/female or the names of states or cities. Money values can be coded for different monetary units, with or without decimals places. Variables can also be entered as dates, using one of many acceptable forms, such as MM-DD-YYYY. Date fields can be added or subtracted to determine length of time in days, weeks, months or years.

Codes for Categorical Variables

Data for categorical variables are entered as labels. For instance, if gender is a variable, we enter either male or female as the data value. Although we can enter the full label as the data, it is much easier to code these values. Using character codes, for instance, gender could be coded F for female and M for male. It is generally recommended, however, that codes be entered as numeric data to facilitate statistical analysis, such as coding 1 for female and 0 for male. For dichotomous variables it is conventional to use 1 and 0 as codes, usually signifying the absence of a trait as zero. As a pure label it does not matter whether we code gender as 1 and 0, as 1 and 2, or any other number; however, many statistical procedures will only manipulate categorical data with 1 and 0 as the category codes (see discussion of dummy variables in Chapter 24). When the research design includes group comparisons, each subject's group assignment must be identified by a code for the grouping variable. Decisions about coding categorical variables should be made before data are collected. Codes should be used on data collection forms to expedite transfer of data to the computer.

Missing Data

It is not unusual for some pieces of data to be missing from a subject's record because of errors in recording, unavailability of information, nonresponses on surveys, or problems in data collection. To identify missing values, blanks are used as the default in most computer programs. Others have specific rules for identifying missing values, such as the use of a period in place of a missing datum. It is not advisable to use zeros to represent missing values, as zeroes will be read as a number and there may be true zeroes in the data. It is often useful to assign specific codes for missing values, to identify the reason for the missing information. For instance, separate codes might be used to distinguish a refusal to answer a question, a response of "Don't know," a question that was not asked, investigator error and so on. Such distinctions can be helpful for interpretation of results, especially when there are many missing data points. Missing data should be coded using numeric values that are out of range of any actual data values. For example, the code of −99 is commonly used.

The standard structure for data entry requires that each variable is entered in a separate column, and each row represents an individual subject. Data may be typed directly into a statistical program, or it may be entered in a spreadsheet first and later imported into the statistical program. No matter how information is entered, the wise researcher will save the data often and back up the data file regularly. We have suffered with too many colleagues who have lost hours of work to take this advice lightly!

When data originate as a spreadsheet, the first row in each column should contain the variable name, which will then be read by the statistical program. To facilitate this transfer, the variable names should conform to the restrictions of the statistical program. Other than this first row, all other rows in the spreadsheet should contain only data. No embedded formulas or charts should be included. If formulas are used in specific cells, they should be converted to the actual data values before transferring to a statistical program.

Variable Names

Variable names identify each data point in a file. Every variable in the file must have a unique name. When variable names are long, abbreviations can be used. As much as possible, variable names should be readily identifiable. Certain rules apply to variable names, depending on the statistical package being used.

Many programs require that a variable name be no more than eight characters (numbers or letters), although more recent versions of some packages allow for longer variable names. Variable names typically must begin with a letter and have no spaces. Some programs allow hyphens, underscores, dollar signs or number signs within a variable name. Generally special characters such as !, ? and / cannot be used in variable names. For example, a pretest and posttest value for pain could be coded PAIN1 for the pretest and PAIN2 for the posttest. Researchers should be familiar with the requirements for the statistical package they use.

Variable Fields

Each row of data, representing a single subject's scores, is called a record or case. Each individual score, or variable value, is identified as a field. A case is composed of several fields of data. Fields are described according to their width, that is, the number of digits or spaces needed for the maximum possible value. The field width is described according to the format Fw.d, where w is the total number of spaces (or field width), and d is the number of digits within the field that follow a decimal point (the F is for Format). For example, the value 7.85 takes up four spaces (including the decimal point), for a field width of F4.2. The value 3560 also takes up four spaces with no decimal places, for a field width of F4.0. The value 136.45 takes up six spaces, described as a field width of F6.2. Many programs set a default field width that can be changed by the researcher.

Labels

Because variable names must be kept short and categories are coded, it is sometimes confusing to read a printout of an analysis with many abbreviations. To facilitate reading the output, most programs will allow the researcher to specify labels for variable names and for category value codes. These labels can usually extend to 40–60 characters or longer. They allow the researcher to customize the printout in a way that will be convenient for interpretation. To make this happen, however, the researcher must take the time to type in all the labels. But it is worth the effort when reams of paper are sitting in front of you and you can't remember whether males are coded 1 or 0! Labels are not required, but with large data sets they are extremely useful. Labels may be listed in the data collection form.

Code Books

Code books are used to organize data and to catalog the order of entry of all variables. Variable names are listed with their abbreviations. Codes are listed to identify their values. Figure 30.2 shows a sample page from a code book for the study examining the effectiveness of two diet regimens on fasting blood sugar in patients with diabetes. Data were collected on the subjects' age, gender, and baseline and follow-up blood sugar levels. Two trials were performed for each test. Codes were developed for gender and group assignment. Subjects were also asked how often they exercised and if they were compliant with their medications. The code book is a necessary reference for all those who are involved with the study, most especially those who will analyze the data. SPSS provides this information in the Variable View of the data file.

Once data are entered into the computer, and before analyses are run, the data should be checked against the raw data to be sure there are no discrepancies or coding errors. This process is called data cleaning, and although it may be time consuming and tedious, it is essential to ensure validity of the data analysis. The data file can be printed out or displayed on a computer screen and visually checked for accuracy against the original data.

Running descriptive statistics on the data will allow the researcher to see if there are obvious discrepancies. Frequency counts should be checked for all categorical variables. The output will list all the codes for each variable and the number of times that code appears in the data. It will also indicate how many subjects are counted, and if there are missing data for that variable. This allows the researcher to determine if there are mistakes in codes, or if the variable has too few entries to be useful. For continuous variables, descriptive statistics and graphs, such as histograms or plots, should be run to analyze means, minimums and maximums, to be sure that the range of scores is appropriate. In this way, the researcher can ascertain if values out of the possible range have been entered. For instance, if the maximum blood sugar score is printed as 560, the researcher knows there is an error and can go back and correct that entry. Sometimes it is useful to sort data, reordering the subjects according to the value of a particular variable, to determine if appropriate numbers have been entered.

All statistical programs include processes for data modification or transformation to create new variables or to assign new codes to existing variables. For example, we might want to compute the mean of several trials to use for data analysis. Or we might have scores for several items on a scale and want to get the sum. Perhaps a continuous variable will be converted to categories. When these types of transformations are performed, a new variable is created, and must be given a new and unique variable name.

Computing New Variables

Computing a new variable requires that some arithmetic operation be performed on the existing data. All programs use the same symbols to represent logical operations. These symbols, known as operators, are used to create expressions that are instructions to the computer. The following symbols are used for arithmetic operations:

These expressions are considered simple expressions because they contain one operator. When more than one operator is used, a compound expression is created, for instance,

is a compound expression. This expression is equal to

When compound expressions are used, specific rules apply to the order in which operations take place. First, all expressions within parentheses are carried out. Second, adjacent operations are carried out in the following order: (1) exponentiation, (2) division and multiplication and (3) addition and subtraction. Within each of these levels, operations proceed from left to right. Therefore, in the preceding expression, the first operation will be to complete the addition (C + 1.0) within the parentheses. Next, the value of A will be squared. This value will then be multiplied by B. Lastly, this product will be divided by the sum (C + 1.0). If the parentheses had been left out, the expression would be read differently. Using

the expression would read

To illustrate the application of these arithmetic operators, we might want to compute a mean baseline and follow-up score to use for analysis for the data in Figure 30.3. To do this, we tell the computer we want to create two new variables called BASEMEAN and FOLLMEAN using the following expressions:

Note the importance of the parentheses, so that the sum of the two items is divided by 2, and not just the value for BASE2 or FOLLOW2. When these computations are done, the values for the new variables will appear as new columns in the data file, as shown in Figure 30.3. These new variables can now be used in statistical procedures. We could, for instance, get a difference score between BASEMEAN and FOLLMEAN, and subject these values to a t-test. This type of data modification can also be done within spreadsheet programs.

Recoding Variables

We can also use comparison operators to recode variables by specifying relationships between them. Comparison operators may be specified as symbols or letter combinations:

Comparison operators are usually used with an IF statement, which specifies a specific operation to be carried out if a given relationship exists. For instance, we have a variable called AGE in our data set (see Figure 30.2). We can create two age groups for a comparison analysis, as follows:

These statements illustrate how we specify values for a new variable called AGEGRP (shown in the last column in Figure 30.3). The actual method for setting recode values will depend on the statistical program.

When assigning values to a new variable, the researcher must be careful not to overlap any categories, or the computer will not be able to perform the desired functions. In addition, groupings should reflect the full range of values that is present in the data.

Statistical Procedures

Many statistical procedures also provide a mechanism for creating and saving new variables. For example, when running a factor analysis, factor scores are created for each subject on each factor. These values can be saved and used as variables in future analyses. When regression procedures are run, residual scores can be calculated and saved. Most programs require specific instructions for these options.

Data collection is complete, all the data are entered and saved (and backed up!), and you are set to begin data analysis. If the research proposal was done well, you are ready to approach this phase of the research process in an organized way. It is a good idea to start by becoming familiar with the data by looking at descriptive statistics—frequencies for categorical variables and means for continuous variables. Histograms, line plots, stem-and-leaf plots or box plots are helpful to visually assess the shape of a distribution, and to identify gaps or outliers. For correlational data, scatterplots should be created to get a sense of the linearity and degree of relationship in the data. These initial steps are necessary to understand the scope of the data, and may suggest alternative statistical approaches. For example, transformations may be needed for nonlinear variables (see Appendix D).

The next step is the culmination of all the research efforts—to apply statistical procedures to answer the research question. This is the fun part. Some helpful hints:

To make this process efficient, prepare a list of specific hypotheses, variables and appropriate statistical procedures to guide your time at the computer. Be specific. For instance, if you intend to compare two groups, specify the t-test, paired or unpaired, and which variables will be used. If you run several regressions, list which are the independent and dependent variables for each one. Then you won't have to sit at the computer, faced with columns and columns of data, and wonder where to start.

Look at the output as you generate it. Examine your findings. Often, additional questions emerge and you may choose to run further tests. For instance, you may find relationships among some variables that you did not anticipate. Groups may end up having different characteristics than planned. It may be of interest to perform certain analyses on subgroups within the data. Statistical programs provide different filtering options to select subjects according to a specified criterion. You might specify that an analysis be done only on those coded 1 for gender, or only those coded for group 1.

Finally, most statistical programs include choices for creating tables or charts directly from the data. Many of these programs provide fairly sophisticated options, with a variety of fonts and colors to customize your presentation. These charts and tables can be imported into word processing or presentation programs. Many different types of charts are usually available, and it is often helpful to try out different formats to see which presents the data best.

Be sure you save your data and output so you can play with options and prepare your project for the final phase of the process—dissemination as a journal article or presentation as a platform or poster.