Descriptive research or statistical research provides data about the population or universe being studied. But it can only describe the "who, what, when, where and how" of a situation, not what caused it. Therefore, descriptive research is used when the objective is to provide a systematic description that is as factual and accurate as possible. It provides the number of times something occurs, or frequency, lends itself to statistical calculations such as determining the average number of occurences or central tendencies.

One of its major limitations is that it cannot help determine what causes a specific behaviour, motivation or occurrence. In other words, it cannot establish a causal research relationship between variables.

The two most commonly types of descriptive research designs are

1. Observation and
2. Surveys

If the objective is to determine which variable might be causing a certain behaviour, i.e. whether there is a cause and effect relationship between variables, causal research must be undertaken. In order to determine causality, it is important to hold the variable that is assumed to cause the change in the other variable(s) constant and then measure the changes in the other variable(s). This type of research is very complex and the researcher can never be completely certain that there are not other factors influencing the causal relationship, especially when dealing with people’s attitudes and motivations. There are often much deeper psychological considerations, that even the respondent may not be aware of.

There are two research methods for exploring the cause and effect relationship between variables:

1. Experimentation, and
2. Simulation

Observation is a primary method of collecting data by human, mechanical, electrical or electronic means. The researcher may or may not have direct contact or communication with the people whose behaviour is being recorded. Observation techniques can be part of qualitative research as well as quantitative research techniques. There are six different ways of classifying observation methods:

1. participant and nonparticipant observation, depending on whether the researcher chooses to be part of the situation s/he is studying (e.g. studying social interaction of tour groups by being a tour participant would be participant observation)
2. obtrusive and unobtrusive (or physical trace) observation, depending on whether the subjects being studied can detect the observation (e.g. hidden microphones or cameras observing behaviour and doing garbage audits to determine consumption are examples of unobtrusive observation)
3. observation in natural or contrived settings, whereby the behaviour is observed (usually unobtrusively) when and where it is occurring, while in the contrived setting the situation is recreated to speed up the behaviour
4. disguised and non-disguised observation, depending on whether the subjects being observed are aware that they are being studied or not. In disguised observation, the researcher may pretend to be someone else, e.g. "just" another tourist participating in the tour group, as opposed to the other tour group members being aware that s/he is a researcher.
5. Structured and unstructured observation, which refers to guidelines or a checklist being used for the aspects of the behaviour that are to be recorded; for instance, noting who starts the introductory conversation between two tour group members and what specific words are used by way of introduction.
6. Direct and indirect observation, depending on whether the behaviour is being observed as it occurs or after the fact, as in the case of TV viewing, for instance, where choice of program and channel flicking can all be recorded for later analysis.

The data being collected can concern an event or other occurrence rather than people. Although usually thought of as the observation of nonverbal behaviour, this is not necessarily true since comments and/or the exchange between people can also be recorded and would be considered part of this technique, as long as the investigator does not control or in some way manipulate what is being said. For instance, staging a typical sales encounter and recording the responses and reactions by the salesperson would qualify as observation technique.

One distinct advantage of the observation technique is that it records actual behaviour, not what people say they said/did or believe they will say/do. Indeed, sometimes their actual recorded behaviour can be compared to their statements, to check for the validity of their responses. Especially when dealing with behaviour that might be subject to certain social pressure (for example, people deem themselves to be tolerant when their actual behaviour may be much less so) or conditioned responses (for example, people say they value nutrition, but will pick foods they know to be fatty or sweet), the observation technique can provide greater insights than an actual survey technique.

On the other hand, the observation technique does not provide us with any insights into what the person may be thinking or what might motivate a given behaviour/comment. This type of information can only be obtained by asking people directly or indirectly.

When people are being observed, whether they are aware of it or not, ethical issues arise that must be considered by the researcher. Particularly with advances in technology, cameras and microphones have made it possible to gather a significant amount of information about verbal and non-verbal behaviour of customers as well as employees that might easily be considered to be an invasion of privacy or abusive, particularly if the subject is unaware of being observed, yet the information is used to make decisions that impact him/her.

There are many different ways for the investigator to collect data from subjects by communicating directly with them either in person, through others or through a document, such as a questionnaire. Direct communication is used in both qualitative and quantitative research. Each has a number of different techniques:

1. Qualitative research techniques

• In-depth interview
• Focus group
• Projective methods
• Case study
• Pilot study

2. Quantitative research techniques

• Telephone
• Self-administered
• In-person Interview

In quota sampling, the population is first segmented into mutually exclusive sub-groups, just as in stratified sampling. Then judgement is used to select the subjects or units from each segment based on a specified proportion. It is this second step which makes the technique one of non-probability sampling.

Let us assume you wanted to interview tourists coming to a community to study their activities and spending. Based on national research you know that 60% come for vacation/pleasure, 20% are VFR (visiting friends and relatives), 15% come for business and 5% for conventions and meetings. You also know that 80% come from within the province. 10% from other parts of Canada, and 10% are international. A total of 500 tourists are to be intercepted at major tourist spots (attractions, events, hotels, convention centre, etc.), as you would in a convenience sample. The number of interviews could therefore be determined based on the proportion a given characteristic represents in the population. For instance, once 300 pleasure travellers have been interviewed, this category would no longer be pursued, and only those who state that one of the other purposes was their reason for coming would be interviewed until these quotas were filled.

Obvious advantages of quota sampling are the speed with which information can be collected, the lower cost of doing so and the convenience it represents.

In convenience sampling, the selection of units from the population is based on easy availability and/or accessibility. The trade-off made for ease of sample obtention is the representativeness of the sample. If we want to survey tourists in a given geographic area, we may go to several of the major attractions since tourists are more likely to be found in these places. Obviously, we would include several different types of attractions, and perhaps go at different times of the day and/or week to reduce bias, but essentially the interviews conducted would have been determined by what was expedient, not by ensuring randomness. The likelihood of the sample being unrepresentative of the tourism population of the community would be quite high, since business and convention travellers are likely to be underrepresented, and – if the interview was conducted in English – non-English speaking tourists would have been eliminated.

Therefore, the major disadvantage of this technique is that we have no idea how representative the information collected about the sample is to the population as a whole. But the information could still provide some fairly significant insights, and be a good source of data in exploratory research.

Validity determines whether the research truly measures that which it was intended to measure or how truthful the research results are. In other words, does the research instrument allow you to hit "the bull’s eye" of your research object? Researchers generally determine validity by asking a series of questions, and will often look for the answers in the research of others.

Starting with the research question itself, you need to ask yourself whether you can actually answer the question you have posed with the research instrument selected. For instance, if you want to determine the profile of Canadian ecotourists, but the database that you are using only asked questions about certain activities, you may have a problem with the face or content validity of the database for your purpose.

Similarly, if you have developed a questionnaire, it is a good idea to pre-test your instrument. You might first ask a number of people who know little about the subject matter whether the questions are clearly worded and easily understood (whether they know the answers or not). You may also look to other research and determine what it has found with respect to question wording or which elements need to be included in order to provide an answer to the specific aspect of your research. This is particularly important when measuring more subjective concepts such as attitudes and motivations. Sometimes, you may want to ask the same question in different ways or repeat it at a later stage in the questionnaire to test for consistency in the response. This is done to confirm criterion validity. All of these approaches will increase the validity of your research instrument.

Probing for attitudes usually requires a series of questions that are similar, but not the same. This battery of questions should be answered consistently by the respondent. If it is, the scale items are said to have high internal validity.

What about the sample itself? Is it truly representative of the population chosen? If a certain type of respondent was not captured, even though they may have been contacted, then your research instrument does not have the necessary validity. In a door-to-door interview, for instance, perhaps the working population is severely underrepresented due to the times during which people were contacted. Or perhaps those in upper income categories, more likely to live in condominiums with security could not be reached. This may lead to poor external validity since the study results are likely to be biased and not applicable in a wider sense.

Most field research has relatively poor external validity since the researcher can rarely be sure that there were no extraneous factors at play that influenced the study’s outcomes. Only in experimental settings can variables be isolated sufficiently to test their impact on a single dependent variable.

Although an instrument’s validity presupposes that it has reliability, the reverse is not always true. Indeed, you can have a research instrument that is extremely consistent in the answers it provides, but the answers are wrong for the objective the study sets out to attain.

The extent to which results are consistent over time and an accurate representation of the total population under study is referred to as reliability. In other words, if the results of a study can be reproduced under a similar methodology, then the research instrument is considered to be reliable.

Should you have a question that can be misunderstood, and therefore is answered differently by respondents, you are dealing with low reliability. The consistency with which questionnaire items are answered can be determined through the test-retest method, whereby a respondent would be asked to answer the same question(s) at two different times. This attribute of the instrument is actually referred to as stability. If we are dealing with a stable measure, then the results should be similar. A high degree of stability indicates a high degree of reliability, since the results are repeatable. The problem with the test-retest method is that it may not only sensitize the respondent to the subject matter, and hence influence the responses given, but that we cannot be sure that there were no changes in extraneous influences such as an attitude change that has occurred that could lead to a difference in the responses provided.

Probing for attitudes usually requires a series of questions that are similar, but not the same. This battery of questions should be answered consistently by the respondent. If it is, the instrument shows high consistency. This can be tested using the split-half method, whereby the researcher takes the results obtained from one-half of the scale items and checks them against the results of the other half.

Although the researcher may be able to prove the research instrument’s repeatability and internal consistency, and therefore reliability, the instrument itself may not be valid. In other words, it may consistently provide similar results, but it does not measure what the research proposed to determine.

The questionnaire is a formal approach to measuring characteristics, attitudes, motivations, opinions as well as past, current and possible future behaviours. The information produced from a questionnaire can be used to describe, compare or predict these facts. Depending on the objectives, the survey design must vary. For instance, in order to compare information, you must survey respondents at least twice. If you are comparing travel intentions and travel experience, you would survey respondents before they leave on vacation and after they return to see in which ways their perceptions, opinions and behaviours might have differed from what they thought prior to experiencing the destination.

Everything about a questionnaire – its appearance, the order the questions are in, the kind of information requested and the actual words used – influences the accuracy of survey results. Common sense and good grammar are not enough to design a good questionnaire! Indeed, even the most experienced researchers must pre-test their surveys in order to eliminate irrelevant or poorly worded questions. But before dealing with the question wording and design and layout of a questionnaire, we must understand the process of measurement.

Some researchers actually question whether a nominal scale should be considered a "true" scale since it only assigns numbers for the purpose of catogorizing events, attributes or characteristics. The nominal scale does not express any values or relationships between variables. Labelling men as "1" and women as "2" (which is one of the most common ways of labelling gender for data entry purposes) does not mean women are "twice something or other" compared to men. Nor does it suggest that 1 is somehow "better" than 2 (as might be the case in competitive placement).

Consequently, the only mathematical or statistical operation that can be performed on nominal scales is a frequency run or count. We cannot determine an average, except for the mode – that number which holds the most responses - nor can we add and subtract numbers.

Much of the demographic information collected is in the form of nominal scales, for example:

In nominal scale questions, it is important that the response categories must include all possible responses. In order to be exhaustive in the response categories, you might have to include a category such as "other", "uncertain" or "don’t know/can’t remember" so that respondents will not distort their information by trying to forcefit the response into the categories provided. But be sure that the categories provided are mutually exclusive, that is to say do not overlap or duplicate in any way. In the following example, you will notice that "sweets" is much more general than all the others, and therefore overlaps some of the other response categories:

Which of the following do you like: (check all that apply):

When items are classified according to whether they have more or less of a characteristic, the scale used is referred to as an ordinal scale (definition of ordinal scale). The main characteristic of the ordinal scale is that the categories have a logical or ordered relationship to each other.These types of scale permit the measurement of degrees of difference, but not the specific amount of difference. This scale is very common in marketing, satisfaction and attitudinal research. Any questions that ask the respondent to rate something are using ordinal scales. For example,

How would you rate the service of our wait-staff?

Although we would know that respondent X ("very good") thought the service to be better than respondent Y ("good"), we have no idea how much better nor can we even be sure that both respondents have the same understanding of what constitutes "good service" and therefore, whether they really differ in their opinion about its quality.

Likert scales are commonly used in attitudinal measurements. This type of scale uses a five-point scale ranging from strongly agree, agree, neither agree nor disagree, disagree, strongly disagree to rate people's attitudes. Variants of the Likert-scale exist that use any number of points between three and ten, however it is best to give at least four or five choices. Be sure to include all possible responses: sometimes respondents may not have an opinion or may not know the answer, and therefore you should include a "neutral" category or the possibility to check off "undecided/uncertain", "no opinion" or "don't know".

Although some researchers treat them as an interval scale, we do not really know that the distances between answer alternatives are equal. Hence only the mode and median can be calculated, but not the mean. The range and percentile ranking can also be calculated.

Interval scales (definition of interval scale) take the notion of ranking items in order one step further, since the distance between adjacent points on the scale are equal. For instance, the Fahrenheit scale is an interval scale, since each degree is equal but there is no absolute zero point. This means that although we can add and subtract degrees (100° is 10° warmer than 90°), we cannot multiply values or create ratios (100° is not twice as warm as 50°). What is important in determining whether a scale is considered interval or not is the underlying intent regarding the equal intervals: although in an IQ scale, the intervals are not necessarily equal (e.g. the difference between 105 and 110 is not really the same as between 80 and 85), behavioural scientists are willing to assume that most of their measures are interval scales as this allows the calculation of of averages mode, median and mean, the range and standard deviation.

Although Likert scales are really ordinal scales, they are often treated as interval scales. By treating this type of agreement scale or attitudinal measurement as interval, researchers can calculate mean scores which can then be compared. For instance, the level of agreement for men was 3.5 compared to 4.1 for women, or it was 3.3 for first time visitors compared to 2.8 for repeat visitors.

When a scale consists not only of equidistant points but also has a meaningful zero point, then we refer to it as a ratio scale. If we ask respondents their ages, the difference between any two years would always be the same, and ‘zero’ signifies the absence of age or birth. Hence, a 100-year old person is indeed twice as old as a 50-year old one. Sales figures, quantities purchased and market share are all expressed on a ratio scale.

Ratio scales should be used to gather quantitative information, and we see them perhaps most commonly when respondents are asked for their age, income, years of participation, etc. In order to respect the notion of equal distance between adjacent points on the scale, you must make each category the same size. Therefore, if your first category is $0-$19,999, your second category must be $20,000-$39,999. Obviously, categories should never overlap and categories should follow a logical order, most often increasing in size.

Ratio scales are the most sophisticated of scales, since it incorporates all the characteristics of nominal, ordinal and interval scales. As a result, a large number of descriptive calculations are applicable.

Words are extremely powerful and we react instinctively to their underlying meanings. Knowing how to word questions in a neutral yet effective is therefore an art to which many books have been dedicated.

There are a number of common errors in question wording that should be avoided.

1.    Loaded words, or words that stir up immediate positive or negative feelings. When loaded words are used, respondents react more to the word itself than to the issue at hand. Not surprisingly, the estimated speed of the car was much higher for group A than for group B in the following experiment:

Two groups of people (A & B) were shown a short film about a car crash. Group A was asked "How fast was car X going when it smashed into car Y?" while Group B was asked "How fast was car X going when it contacted car Y?"

Many similar experiments have shown the power of words to introduce such a bias.

2.    Loaded response categories, or providing a range of responses that will skew the answers in one direction or another. In the following example:

You have biased the answers towards "increase" since there are two categories that address degrees of increase, compared to only one category for a potential decrease. You need to balance this scale by changing "decreased" to "decreased a lot" and "decreased slightly".

3.    Leading questions, or questions that suggest socially acceptable answers or in some way intimate the viewpoint held by the researcher, can lead respondents to answer in a way that does not reflect their true feelings or thoughts. For instance, in a survey about all-inclusive holidays, the question "How much time did you devote to your child(ren) during your vacation?" will lead people to overestimate the time spent, since to answer "I did not spend any time or little time" would almost be like saying they were not devoted to their offspring. It is very important to find words that do not make assumptions about the respondents or are neutral in nature.

4.    Double-barrelled questions are questions that require more than one answer, and therefore should be broken into at least two questions. For instance, "How would you rate the quality of service provided by our restaurant and room-service staff?" does not allow you determine which is being rated, the restaurant staff or the room-service staff, nor whether either or both were used. It would be far better to have separate questions dealing with these issues.

5.    Vague words or phrases can pose a special problem, since sometimes you want to be deliberately vague because you are exploring an issue and don’t want to pre-determine or influence the answers being considered. However, there are some common rules that you can follow:

• Always specify what "you" refers to as in "you, personally" or "you and your family"
• Always give a precise time frame: "How often in the last two months…", "In 1998…(not "last year")", "July and August (instead of "during the summer")"
• Stay away from terms such as "rarely" or "often", if you can since they can be interpreted in many different ways. Use terms such as "once a month", "several times a week" that are more specific.
• When probing for specific quantitative details, don’t tax the respondent’s memory unreasonably. "How many ads for different destinations did you see in the last month?" is an almost impossible question to answer even though the time frame is ‘only’ a month. Since most of us are inundated with advertising messages of all types, you cannot expect a meaningful answer. It would be better to probe for destinations that have stood out or ads that have marked the respondent in some way.

6.    Offensive or threatening questions, even if inadvertent, can lead to a dishonest answer or refusal. This applies even to a blunt question about income as in "What is your income?_______" It is better to rephrase such a question by providing broad categories, for example

It is rare that you really need very precise income information, and for most types of research this type of broad category would be sufficient. This is part of the considerations around "nice to know" information versus information needed to specifically answer the research problem.

7.    Jargon, acronyms and technical language should be avoided in favour of more conversational language and the use of terms that are readily understood by everyone. It should not be taken for granted, for instance, that everyone would be familiar with even relatively common terms such as "baby boomer", "GTA" or "brand image", and it is far better to spell them out: "those born after the Second World War but before 1960", "the cities in the Oshawa-Hamilton corridor" and "the image reflected by the name or symbol used to identify a service provided", even if this may appear to be cumbersome.
Pamela Narins, the Market Research Manager for SPSS, has prepared some excellent pointers on how to "write more effective survey questions". Be sure to check them out!

Probably the most challenging questionnaires to design are for self-administered surveys, since their appearance is critical in motivating respondents. If the font is too small, the instructions confusing, the look unprofessional or cluttered, you can be sure that it will have an immediate impact on both the overall response rate and non-response error.

Even if there is a covering letter to accompany the survey, it is a good idea to reiterate the key points about the importance of the survey, due date and where to return it at the top of the questionnaire, in case it gets separated from the covering letter. It is critical that the first several questions engage the respondent. They should also be relevant to the study itself so that the respondent quickly understands what the survey is about and becomes engaged in is objectives. These questions should be straightforward with relatively few categories of response. Also, respondents should not be asked to agree or disagree with a position at the outset, but rather eased into the survey by asking them what they think is important or what they prefer.

Because of these considerations, it is not a good idea to put demographic questions at the beginning of the questionnaire. Since they are easy to answer and often seen as simple "housekeeping" items, they are much better at the end when respondents are getting tired. The same holds true for any questions considered a bit more sensitive. If placed too early, they can lead the respondent to abandon and it is better that there only be one or two questions at the end that suffer from non-response bias rather than that the response rate as a whole be reduced.

Questions themselves should be grouped in sections, and these sections should have a logical sequence. You want to avoid making the respondent "jump around" mentally, and if necessary, may even want to help him/her "shift gears" by introducing a new section. For instance, "Now we will shift slightly to ask about…" or "We would like to get your opinion on some related areas" can help the respondent travel through the questionnaire or interview.

One other important aspect to watch out for is "position bias". This bias can occur in lists, where the first items are treated differently (often seen as more important) from the last items. It is therefore a good idea to have several versions of the questionnaire, where this is an option, and change the order of the lists in question.

Instructions should clearly stand out. Too often, it is the questions themselves that are bolded or italicized, rather than the instructions. Remember that respondents will always read the questions, but unless their attention is really drawn to the instructions, they are likely to skip these. This could lead to response error, particularly if you are asking them to skip certain questions based on their response.

Be sure to also consider coding of the questionnaire to save time and confusion during the data entry stage. Whatever rationale you chose for coding, it should be discreet (i.e. not distract the respondent in a self-administered questionnaire) and consistent throughout.

Pamela Narins, the Market Research Manager for SPSS, has written some excellent "guidelines for creating better questionnaires" that you should definitely check out!

No matter how experienced you are in developing questionnaires or how "routine" the survey might be considered, it is always important to pretest your instrument before it is printed and fielded. Start by reading it aloud – the flow of words should sound conversational/read smoothly and communicate clearly. Instructions should not leave any doubt about what the respondent is supposed to do. Check each question carefully to ensure that it will indeed provide the information you need, and that the meaning of every word is clear. Go back and revise your survey!

Next, give the questionnaire to a small group of people, who preferably know little or nothing about the research itself. Ask them to read the questionnaire to see if they, too, can clearly understand what is required and whether the flow makes sense to them. Go back and revise your survey!

And finally, select a small number of people from your sampling frame, if at all possible, and test your questionnaire with them (even if your questionnaire turns out to be "pretty good", you should not include these respondents in your final sample). Look at the frequency distribution: if there are too many "neutral", "don’t know" or "don’t remember" responses, you need to revise the questions themselves. If the questions that require a written response look too "squished", provide more generous spacing.

You should ask the respondents to also comment on the questionnaire itself, and on whether you should perhaps be asking some additional questions relevant to the research problem as stated in the introduction. Since people are often reluctant to admit that they might have difficulty responding, ask them whether they believe other people would have difficulty and which questions in particular might pose problems? You will elicit many useful comments, so: Go back and revise your survey!

And remember that you will have to go through the same steps for each language used for the survey, and that you will need to check for cultural and regional differences as well in these cases.

Pamela Narins, the Market Research Manager for SPSS, has provided "13 important tips to help you pretest your surveys" that you should definitely check out!

In probability sampling, the sample is selected in such a way that each unit within the population or universe has a known chance of being selected. It is this concept of "known chance" that allows for the statistical projection of characteristics based on the sample to the population.

Most estimates tend to cluster around the true population or universe mean. When plotted on a graph, these means form what is called the normal or bell curve. This theoretical distribution allows for the calculation of the probability of a certain event occurring (e.g. the likelihood that an activity studied will be undertaken by people over 65 years old, if those are the variables being studied).

There are three main types of probability or random sampling that we will review more closely:

• (Simple) Random
• Stratified
• Cluster

In non-probability sampling, the sample is selected in such a way that the chance of being selected of each unit within the population or universe is unknown. Indeed, the selection of the subjects is arbitrary or subjective, since the researcher relies on his/her experience and judgement. As a result, there are no statistical techniques that allow for the measurement of sampling error, and therefore it is not appropriate to project the sample characteristics to the population.

In spite of this significant shortcoming, non-probability sampling is very popular in hospitality and tourism research for quantitative research. Almost all qualitative research methods rely on non-probability sampling techniques.

There are three main types of non-probability sampling that we will review more closely:

• Judgment Sampling
• Quota Sampling
• Convenience Sampling

Measures of central tendency describe the location of the center of a frequency distribution. There are three different measures of central tendency: the mode, median and mean.

The mode is simply the value of the observation that occurs most frequently. It is useful when you want the prevailing or most popular characteristic or quality. In a survey of adults aged 18 or older, the question "What is your age?" was answered as follows:

What is your age?

Hence the mode is 46, since 47 respondents provided this answer, more than any other category. Since there is only one mode in this distribution, it is referred to as ‘unimodal’. If a distribution has two modes (or two values that have the same amount of responses that are also the highest), it is referred to as ‘bimodal’.

The median is the middle observation, where half the respondents have provided smaller values, and half larger ones. It is calculated by arranging all observations from lowest to highest score and counting to the middle value. In our example above, the median is 47. The cumulative percentage will tell you at a glance where the median falls. Since the median is not as sensitive as the mean to extreme values, it is used most commonly in cases where you are dealing with ‘outliers’ or extreme values in the distribution that would skew your data in some way. The median is also useful when dealing with ordinal data and you are most concerned with a typical score.

The mean is also known as the ‘arithmetic average’ and is symbolized by ‘X’. The formula for calculating the mean is ∑ X/n (The Greek letter sigma ∑ is the symbol for sum) This means that you total all responses (X) and then divide them by the total number of observations (n). In our example, you would have to multiply all the values (actual ages or ‘x’) by the number of respondents or frequencies (‘f’) for each (∑ x • f = X or. 18 • 17 + 19 • 14 + etc.) and then divide the total by the 1500 respondents who participated in the survey. The result is 47.04. Since the computer will follow the same steps, you must be sure that the values are real and not just codes for categories. For example, the computer would calculate a mean of 3.7367 for the same information as the frequency above but recoded into age categories, based on the assumption that the values under x are 1 to 6:

What is your age?

Another useful calculation is the range, the calculation of the spread of the numerical data. It is calculated by subtracting the lowest value (in our example 18) from the highest value (or 86) to give us a total of 68. This is particularly useful when dealing with rating scores, for instance, where you would like to determine how close people are in agreement or alternatively, how wide the discrepancies are. 

When you want to know how respondents answered on two or more questions at the same time, you will need to run a cross-tabulation. In order to do so, you must first determine which is your independent variable, and which your dependent variable, since the first is traditionally used as column headings and the latter are found in the row.

Independent variables explain or predict a response or an outcome, which is the dependent variable under study. As a basic rule, demographic information is usually considered independent, since characteristics such as gender, age, education etc. will normally determine the responses we make. If the variables being studied are not demographic, then the independent variable is determined by the study’s objectives. For instance, if the objective is to determine whether the level of satisfaction with the past holiday at a destination influences the likelihood of return, then level of satisfaction is our independent variable and the likelihood to return the dependent one.

This is the typical output of a simple cross-tabulation (of education levels and overall satisfaction with a holiday) as produced by SPSS, when we also ask that column percentage be calculated. Note that the title gives the two variables with the dependent one first separated by *. When producing this information in a table, we would reword it to read "Overall holiday satisfaction by highest level of education completed" ( see Table 1), removing all extraneous information and leaving it as a statement, not a question.

Overall satisfaction with your holiday * What is the highest level of education completed Crosstabulation

Obviously, you would not be able to use this table as is in a report. It requires ‘cleaning’. Your first consideration would be whether you want to keep all of the categories in your independent and dependent variable. This depends, of course, on what you are trying to illustrate and the responses in each cell. First of all, very few people have less than a high school degree, and we could therefore collapse the first two categories into ‘high school or less’. But that still leaves us with five categories or more detail than we would probably need. So we could collapse the categories ‘graduated from technical or vocational school’ and ‘some college/university’ into ‘some advanced education’ and the last two into ‘graduated from university or more’. Similarly, we notice that the level of satisfaction with the holiday is very high. Indeed, any rows with less than 5% of respondents in cells should be collapsed. At the very least we should only have one category ‘not at all or nor very satisfied’. This collapsing of categories is knows as recoding and is a way of changing existing variables or creating new variables based on existing data as explained by John Urbik, the Technical Marketing Specialist for SPSS.

The resultant cross-tabulation would look like this:

overall satisfaction with holiday * highest level of education Crosstabulation

We can now proceed to present this information in a more pleasing table format by giving it the appropriate table number and title, indicating the total number of respondents who answered this question, and cleaning the table, as follows:

Table 1: Overall holiday satisfaction by highest level of education

n=1243

Graphically, we would follow very similar rules: the graph is numbered (Figure 1) with the same title and the number of respondents indicated; the independent variable identifies the columns since we want to compare the satisfaction level of each of the three education categories. It is the column percentage that is used for comparison purposes. The type of graph below is called a clustered bar chart.

The chi-square (pronounced ‘kai’) distribution is the most commonly used method of comparing proportions. Its symbolized by the Greek letter chi or  χ²). This test makes it possible to determine whether the difference exists between two groups and their preference or likelihood of doing something is real or just a chance occurrence. In other words, it determines whether a relationship or association exists between being in one of the two groups and the behaviour or characteristic under study. If in a survey of 692 respondents we asked whether or not they are interested attending attractions and events that deal with history and heritage during their vacation, and we wanted to determine whether there is a difference in how men and women respond to this question, we could calculate a chi-square.

 χ² determines the differences between the observed (f_o) and expected frequencies (f_e). The observed frequencies are the actual survey results, whereas the expected frequencies refer to the hypothetical distribution based on the overall proportions between the two characteristics if the two groups are alike. For example, if we have the following survey results:

Then we can calculate our expected frequencies (f_e) based on the proportion of respondents who said ‘yes’ versus ‘no’. It can also be calculated for each cell by the row total with the column total divided by the grand total (e.g. 254 x 294 : 692 = 108).

This second table, where no relationship exists between the interest in attending history and heritage attractions and events and gender, also represents the null hypothesis or H_o. (Therefore, if a study says that it "fails to reject the null hypothesis", it means that no relationship was found to exist between the variables under study.)

Hence, the calculation is as follows:

The critical value for a level of significance of .05 (or 95% level of confidence, the normal level in this type of research) is 3.841. This means that you are confident that 95% of the distribution falls below this critical value. Since our result is above this value, we can:

• Reject the null hypothesis that no difference exists between interest in attending historical attractions and events and gender (in other words, there is a difference between genders); and
• Conclude that the differences in the groups are statistically significant (or not due to chance)

You will not need to memorize all the critical values since computer programs such as SPSS will not only calculate the  χ² values for you, but will also give you the precise level of observed significance (known as p value), which in our case is .039. If this level of significance is above the standard .05 level of statistical significance, you are dealing with a statistically significant relationship.

A correlation is used to estimate the relationship between two characteristics. If we are dealing with two ordinal or one ordinal and one numerical (interval or ratio) characteristic, then the correct correlation to use is the Spearman rank correlation, named after the statistician, also known as rho or r_s. Computer software programs such as SPSS will execute the tedious task of calculating Spearman’s rho very easily.

Correlations range from –1 to +1. At these extremes, the correlation between the two characteristics is perfect, although it is negative or inverse in the first instance. A perfect correlation is one where the two characteristics increase or decrease by the same amount. This is a linear relationship as illustrated below.

A correlation coefficient of 0 therefore refers to a situation where no relationship exists between the two characteristics. In other words, changes in one characteristic cannot be explained by the changes occurring in the second characteristic. A common way of interpreting the strength of a correlation is as follows:

However, in most tourism and hospitality related research, corrections between ± .26 to ± .5 are generally considered to be quite high, with strong or very strong correlations rarely found.

The variation in one characteristic can be predicted if we know the value of correlation, since we can calculate the coefficient of determination or r².

If we have a correlation of  ±.5, then  ±.5² = .25. We can therefore conclude that 25% of the variation in one characteristics can be predicted by the value of the second measure.

Let us take a look at an example to make these interpretations clearer. Let us assume we wish to determine whether pleasure travellers who are interested in staying in first class hotels are motivated because they wish to indulge in luxury or because of some other motivation, such as the identification of luxury hotels with big modern cities. By running a Spearman correlation on these three ordinal variables that asked respondents to rate the importance of each, we obtain the following output:

**Correlation is significant at the .01 level (2-tailed).

If we look at the first pair of variables (‘big modern cities’ by ‘first class hotel’) we note that the correlation coefficient is .229 or ‘weak’, while the statistical significance is very high (.000 is so small that the number itself is cut off) Indeed, as the footnote says, the results are significant at the .01 level. This can also be interpreted as "the level of confidence is 99%"(1 - .01 = .99). Finally, the output informs us that 1472 of our survey respondents answered both questions.

Similarly, the second pair of variables (‘indulging in luxury’ by ‘first class hotel’) show us a strong correlation (indeed, unusually high for this type of research at .615) that is statistically significant and where 1468 respondents answered both questions.

We can therefore conclude that ‘indulging in luxury’ is strongly correlated with the choice of first class hotel accommodation and that 38% (r² = .615 x .615) of the variance in first class hotels is determined by this factor. Furthermore, this finding is statistically significant, which means we can reject the null hypothesis that there is no relationship between the importance attributed to staying in first class hotels and indulging in luxury. We can further conclude that while ‘big modern cities’ are associated with luxury hotels, that correlation is weak with only 5% (r² = .229 x .229) of the variance in luxury hotels explained by it.