 # Inferential Statistics

## Presentation on theme: "Inferential Statistics"— Presentation transcript:

Inferential Statistics
Statistical Analysis of Research Data

Statistical Inference
Getting information about a population from a sample How practical are “statistically significant” results? Cost/benefit Crucial difference Client acceptability Public and political acceptability Ethical and legal concerns

Inferential Statistics
These provide a means for drawing conclusions about a population given the data actually obtained from the sample. They allow a researcher to make generalizations to a larger number of individuals, based on information from a limited number of subjects. They are based on: Probability theory Statistical inference Sampling distributions

Inferential Statistics
Probability theory – the basis for decision-making statistical inferences. It refers to a large number of experiences, events or outcomes that will happen in a population in the long run. Likelihood and chance are similar terms. Examples are usually based on tossing a coin and finding heads or tails. Probabilities are statements of likelihood expressed in values from 0 to 1.0. p = the number of outcomes the total possible outcomes

Inferential Statistics
Statistical inference – statistics enable us to judge the probability that our inferences or estimates are close to the truth Sampling distributions – are theoretical distributions developed by mathematicians to organize statistical outcomes from various sample sizes so that we can determine the probability of something happening by chance in the population from which the sample was drawn. They allow us to know the relative frequency or probability of occurrence of any value in the distribution.

Inferential Statistics
Hypothesis testing – 5 basic steps Make a prediction Decide on a statistical test to use Select a significance level and a critical region (region of rejection of the null hypothesis). To do this you must consider two things Whether both ends (tails) of the distribution should be included. How the critical region of a certain size will contribute to Type I or Type II errors.

Critical Region in the Sampling Distribution for a One-Tailed Test

Critical Regions in the Sampling Distribution for a Two-Tailed Test

Levels of Significance
Remember, if a printsout shows a two-tailed test result, and you wanted a one-tailed result, divide the two tailed p value by 2. Example: p = .080 (two-tailed) or p>.05 p = .040 (one-tailed) or p<.05 The first would not be statistically significant, whereas, the second would be statistically significant

Outcomes of Statistical Decision Making

Inferential Statistics
Hypothesis testing cont. Computing the test statistic - The test statistic is not a mean, sd or any form of descriptive data. It is simply a number that can be compared with a set of results predicted by the sampling distribution Compare the test statistic to the sampling distribution (table) and make a decision about the null hypothesis – reject it if the statistic falls in the region of rejection. Consider the power of the test – its probability of detecting a significant difference – parametric tests are more powerful

Degrees of Freedom One sample t-test or paired t-test = N-1
Independent t-test = N-2 Chi-square test = (# rows - 1) x (# columns – 1) ANOVA : df between groups = (# levels or groups – 1) df within groups = (# subjects - # of levels) Correlations = N-2

Levels of Measurement There are four levels or scales of measurement, Each level is classified according to certain characteristics. Data that fall in the first level are limited to certain statistical tests. Choices of statistical tests (and the power of the tests) increase as the levels go up.

Levels of Measurement cont.
Nominal scale – measurement at its weakest – numbers or other symbols are used to classify or partition a class into mutually exclusive subclasses – animals can be classified as dogs, cats, etc. You can test hypotheses regarding distribution among the categories by using the Chi-square test.

Levels of Measurement cont.
Ordinal scale – shows relationships among classes, such as higher than, more difficult than, etc. It allows the attributes of a variable to be ranked in relation to each other. A researcher can test hypotheses using non-parametric statistics of order and ranking.

Levels of Measurement cont.
Interval scale – is similar to the ordinal scale, but the distance between any two numbers is of a known size. The numbers used have absolute values and the interval between each number is considered to be equal. Increasing amounts of a variable are represented by increasing numbers on the scale. The variable is continuous. There is no true zero – where you have none of the variable. All parametric tests can be used with interval data.

Levels of Measurement cont.
Ratio scale – is like the interval scale but it has a true zero point as its origin. Time, length and weight are ratio scales when used alone, but not as a characteristic of a person. Arithmetic, all parametric tests and geometric means can be used with ratio data.

Tests of Significance Parametric tests of significance – used if there are at least 30 observations, the population can be assumed to be normally distributed, variables are at least in an interval scale Z tests are used with samples over 30. There are four kinds (two samples or two categories) t-tests are used when samples are 30 or less. Single sample t-test (one sample) Independent t-test (two samples) Paired t-test (two categories

Tests of Significance Non-parametric tests of significance – small numbers, can’t assume a normal distribution, or measurement not interval Chi-square – requires only nominal data – allows researcher to determine whether frequencies that have been obtained in research differ from those that would have been expected – use a X2 sampling distribution Chi-square goodness of Fit Chi-Square test of independence

Tests of Significance Mann Whitney U – an alternate to the independent t-test – must have at least ordinal data. It counts the comparative ranks of scores in two samples (from highest to lowest) The null hypothesis is that the two samples are randomly distributed. Use U sampling distribution tables for small sample sizes (1-8) and medium sample sizes (9-20) and the Z test for large samples

Tests of Significance Wilcoxin Matched Pairs (signed rank test) – is an alternate to the paired t-test. It is used for repeated measures on the same individual. It requires a measurement between ordinal and interval scales – the scores must have some real meaning. Use a T table. If the T is less than or equal to the T in the table, the null hypothesis is rejected.

Measures of Association
Parametric Measures of Association – These answer the question, “within a given population, is there a relationship between one variable and another variable?” A measure of association can exist only if data can be logically paired. It can be tested for significance. Correlation – answers the question, “What is the degree of relationship between “x” and “y” – Use Pearson Product Moment Correlation (Pearson r ) –see next slide

Measures of Association The Pearson Correlation Coefficient (Pearson r)
The r examines the relationship between two quantitative sets of scores. The r varies from – 1.00 to The r is not a proportion and cannot be multiplied by 100 to get a percentage. To think of the r as a percentage, it needs to be converted to the “Coefficient of Determination” or R2 . An r of .50 is 25% better than an r of 0.00

Measures of Association
Non-parametric tests for association Correlation The Spearman Rank Order Correlation (Rs)– “To what extent and how strongly are two variables related?” Phi coefficient – it can be used with nominal data, but should have ordinal data Kendall’s Q – can be used with nominal data

Prediction Parametric Prediction – using a correlation, if you know score “x”, you can predict score “y” for one person – Use regression analysis Simple linear regression – allows the prediction from one variable to another – you must have at least interval level data Multiple linear regression – this allows the prediction of one variable from several other variables. The dependent variable must be on the interval scale

Prediction Non-parametric Prediction – measures the extent to which you can reduce the error in predicting the dependent variable as a consequence of having some knowledge of the independent variable such as, predicting income [DV] by education [IV] Kendall’s Tau – used with ordinal data and ranking - is better than the Gamma because it takes ties into account Gamma - used with ordinal data to predict the rank of one variable by knowing rank on another variable Lambda – can be used with nominal data – knowledge of the IV allows one to make a better prediction of the DV than if you had no knowledge at all

Parametric Multiple Comparisons
The analysis of variance (ANOVA) is probably the most commonly encountered multiple comparison test. It compares observed values with expected values in trying to discover whether the means of several populations are equal. It compares two estimates of the population variance. One estimate is based on variance within each sample – within groups. The other is based on variation across samples – between groups. The between group variance is the explained variance (due to the treatment) and the variation within each group is the unexplained variance (the error variance).

Parametric Multiple Comparisons
ANOVA cont. The ratio of the explained scores to the unexplained scores gives the F statistic. If the variance between the groups is larger, giving an F ratio greater than 1, it may be significant depending upon the degrees of freedom. If the F ratio is approximately 1, it means that the null hypothesis is supported and there was no significant difference between the groups.

Parametric Multiple Comparisons
ANOVA cont. If the null hypothesis is rejected, then one would be interested in determining which groups showed a significant difference. The best way to check this is to conduct a post hoc test such as the Tukey, Bonferrioni, or Scheffe. (SPSS will do this for you if you click on Post-hoc and check the test desired.Check on descriptives while you still in ANOVA, and the program will also give you the mean for each group)

Parametric Multiple Comparisons
Two-Way Analysis of Variance Classifies participants in two-ways Results answer three questions Two main effects An interaction effect

Non-parametric Multiple Comparison
Kruskal-Wallis Test – an alternative to the one-way ANOVA. The scores are ranked and the analyses compare the mean rank in each group. It determines if there is a difference between groups. McNemar Test – an adaptation of the Chi-square that is used with repeated measures at the nominal level. Friedman Test –an alternative to the repeated ANOVA. Two or more measurements are taken from the same subjects. It answers the questions as to whether the measurement changes over time.