• No se han encontrado resultados

Sistema Gestor de Base de Datos (SGBD)

CAPÍTULO 2 TENDENCIAS Y TECNOLOGÍAS A UTILIZAR

2.2 Lenguajes y Herramientas a utilizar para la creación y funcionamiento de la Capa de

2.2.3 Sistema Gestor de Base de Datos (SGBD)

Th is approach to design involves creating groups of research participants that diff er on one variable, and then statistically evaluating them to observe if these groups also diff er signifi cantly on the behavior of interest. Th e goal of the research may be either to fi nd out if one variable is simply related to another (post facto study), or to establish if one variable actually infl uences another (true experiment). With either goal, the question being asked using this method is whether or not the groups diff er, as opposed to the previous correlational design that questioned on whether the scores were related to a single group.

If the groups are formed based on the amount of one variable that the participants currently possess (e.g., age, sex, height) and assigning them to the appropriate group, then it is a post facto design. If there is a signifi cant group diff erence on the behavior performance, then the interpretation may still be that the group diff erence variable and behavior are related without knowing the reason for it. Furthermore, the information obtained from a post facto group-diff erence study is similar to that obtained from the correlational relationship post facto study described earlier.

Th e statistical evaluation for “signifi cance” may not be based on a correlation coeffi cient, but may use procedures like t-test or analysis of variance (ANOVA). Th ese two techniques allow a researcher to calculate the probability of obtaining the observed diff erences in the mean values (assuming random sampling), if the populations are not diff erent. In other words, it is possible that the samples have diff er-ent means when their populations do not have diff erer-ent means. Sampling variability can certainly lead to this situation. Random samples may not necessarily match the population accurately, and hence, two samples can easily diff er when their populations do not. However, if the observed groups have diff erent mean values that have a very low probability (≤0.05) of coming from equal populations, that is, diff ering owing to sampling error only, then it is possible to conclude that the group variable being studied and the behavior are truly related in the population, not just for the sample studied.

Th is is similar to the result from a post facto relationship question evaluated with a correlation coef-fi cient described in the previous section. Th e legitimate interpretation of a post facto study may be the same, irrespective of whether the researcher evaluates the result as a relationship question with a

correlation coeffi cient, or as a group diff erence question with a test for signifi cant diff erences between the means. If the more powerful interpretation that a variable actually infl uences the behavior is required, then the researcher may need to conduct a true experiment.*

To obtain the cause-and-eff ect information, a research design where only the group diff erence variable could lead to the observed diff erence in the group performance is required. Th is research would begin by creating two or more groups that do not initially diff er on the group diff erence variable, or anything else that might infl uence the performance on the behavior variable; for example, research participants do not decide which group to join, the top or lowest performers are not placed in “groups,” and existing intact groups are not used. Instead, equal groups are actively formed by the researcher, and controls are imposed to keep unwanted factors from infl uencing the behavior performance. Experimental controls are then imposed to make sure that the groups are treated equally throughout the experiments. Th e only factor that is allowed to diff er between the groups is the amount of the group diff erence variable that the participants experience. Th us, the true experiment starts with equal groups and imposes diff erences on the groups to observe whether a second set of diff erences is obtained.

In this way, it is possible to determine whether the imposed group diff erence actually infl uences the performance, because all the alternate logical possibilities for why the groups diff er on the behavior of interest are eliminated. In practice, the equal groups are formed either by randomly assigning an exist-ing pool of research participants into equal groups, or by selectexist-ing several equal random samples from a large population of research participants. In either procedure, the groups are formed so that the groups are equal on all factors, known and unknown, which have any relationship or potential infl uence on the behavior performance. Th e researcher then imposes the research variable diff erence on the groups, and later measures the individuals and compares the group means on the behavior performance.

As discussed earlier, random sampling or random assignment might have assigned people to groups in such a way that it failed to produce exact equality. Th us, the researcher needs to know if the resulting group diff erences are greater than the initial inequality that the random chance might have produced.

Th is is easily evaluated using a test for statistical signifi cance. If the statistic value of the test has a prob-ability of 0.05, then the sampling variprob-ability only may have a 5/100 chance of producing the group-mean diff erence as large as the one found. Again, for any observed result that has a probability of being pro-duced by sampling luck alone, which is as small as or smaller than 5/100, one may conclude that the dif-ference may be from something other than this unlikely source and is “statistically signifi cant.” In this case, the researcher may conclude that the reason for the groups to have diff erent behavior performance means is that the imposed group diff erence variable created these performance diff erences, and, if these performance diff erences are imposed on other groups, then one may expect to reliably fi nd similar performance diff erences.

4.5.5 Examples

As an example of a group diff erence of true experiment versus a group diff erence of post facto study, consider an investigation to determine whether unusual attitude training infl uences the pilot perfor-mance in recovering from an uncommanded 135 degree roll. Researcher A investigates this by locating 30 pilots in his company, who have had unusual attitude training within the past 6 months and who volunteered for such a study. He compares their simulator performance with that of a group of 30 pilots from the company, who have never had such training and have expressed no interest in participating in the study. A statistical comparison of the performance of the two groups in recovering from the

*Although it is possible to conduct a true experiment as a relationship question evaluated with a correlation coeffi cient, this is very rare in practice. True experiments producing information on one variable and actually infl uencing the performance on another, are almost always conducted as a question of group diff erences and evaluated for statistical signifi cance with some factors other than correlation coeffi cient.

uncommanded 135 degree roll indicated the mean performances for pilots who were or were not trained in unusual attitude recover, which were 69.6 and 52.8, respectively. Th ese means do diff er respectively with t(38) = 3.45, p = 0.009.

With such a design, one can conclude that the performance means for the populations of trained and untrained pilots do diff er in the indicated direction. Th e chance of obtaining nonrepresentative samples with such diff erent means (from populations without mean diff erences) is less than 1 in 100. However, as this is a post facto study, it is impossible to know whether the training or other pilot characteristics are responsible for the diff erence in the means.

As Researcher A used a post facto study—that is, did not start with equal groups and did not impose the group diff erence variable (i.e., having or not having unusual attitude training) on the groups—there are many possible reasons that trained group performed better. For example, the more skilled pilots sought out such training and thus, could perform any fl ight test better because of their inherent skill, not because of the training. Allowing the pilots to self-select the training created groups that diff er in ways other than the training variable under study.

It is, of course, also possible that the attitude training is the real active ingredient leading to the roll-recovery performance, but this cannot be investigated using Researcher A’s study. It is only possible to know that seeking and obtaining attitude training is related to better roll recovery. Is it because better pilots seek such training, or because such training produces increased skill? It is impossible to know.

Is this diff erence in interpretations relevant? If one is selecting pilots to hire, perhaps not. One cannot simply hire those who have obtained such training, and think that they will (based on group averages) be more skilled. If one is trying to decide whether to provide unusual attitude training for a company’s pilots and the cost of such training is expensive, then one would want to know if such training actually leads to (causes) improved skill in pilots in general. If the relationship between attitude training and performance is owing to the fact that only highly skilled pilots have historically sought out such train-ing, then providing such training to all may be a waste of time and money.

On the other hand, Researcher B has a better design for this research. Sixty pilots are identifi ed in the company, who have not had unusual attitude training. Th ey are randomly assigned to one of the two equal groups, either to a group that is given such training or to a group that gets an equal amount of additional standard training. Again, the mean performance of the two groups are observed to diff er signifi cantly with p = 0.003.

Th is research provides much better information from the signifi cant diff erence. It is now possible to conclude that the training produced the performance diff erence and would reliably produce improved performance if imposed on all of the company’s pilots. Th e pilot’s average performance on unusual atti-tude recovery would be better because of the training. Th e extent of improvement could be indicated by looking at our eff ect-size index. If eta squared equaled to 0.15, then we can conclude that the training leads to 15% of the variability among pilots on the performance being measured.

Oft en, these questions on group diff erence are addressed with a research design involving more than two groups in the same study. For example, a researcher might randomly assign research participants to one of the three groups and then impose a diff erent amount of training or a diff erent type of training on each group. One could then use a statistical analysis called ANOVA to observe whether the three amounts or types diff er in their infl uence on performance. Th is is a very typical design and analysis in behavioral science studies. Such research can be either a true experiment (as described earlier) or a post facto study. Th e question of signifi cance is answered with an F statistic, rather than the t in a two-group study, but eta squared is still used to indicate the amount or size of the treatment eff ect.

For example, unusual attitude recovery was evaluated with three random samples of pilots using a normal attitude indicator, a two-dimensional outside-in heads-up display (HUD), or a three-dimen-sional HUD. Th e mean times to recovery were 16.3, 12.4, and 9.8 s, respectively. Th e means did diff er signifi cantly with a one-way ANOVA, F(2, 27) = 4.54, p < 0.01. An eta squared value of 0.37 indicated that 37% of the pilot variability in attitude recovery is owing to the type of display used. One can conclude that the three methods would produce diff erences among the pilots in general, because the

probability of fi nding such large sample diff erences just from random assignment eff ects, rather than training eff ects, is less than 1 in 100. Further, the display eff ects produced 37% of the individual pilot variability in time to recover. Th e ANOVA established that the variance among the means was from the display eff ects, and not from the random assignment diff erences regarding who was assigned to which group. Th is ANOVA statistical procedure is very typical for the analysis of data from research designs involving multiple groups.

Documento similar