Statistical Hypothesis Testing

Dsicuss about the Statistics Without Maths Psychology Interpretation.

The bottom line of any statistical investigation is to draw some inference regarding the underlying probability law of the phenomenon of interest on the basis of available, observable data (Cohen, Manion & Morrison, 2013). According to statistical theory, observations form the foundation on which any assertion, conjecture or research hypothesis could be analysed and inferred upon.

A statistical hypothesis testing method is a tool to evaluate the validity of a conjecture regarding the parameter of the probability distribution which is assumed to be underlying the phenomenon being considered in the hypothesis (Casella & Berger, 2013).. This is done on the basis of a sample statistic. The procedure takes into consideration two contesting hypothesis, namely the null hypothesis and the alternate hypothesis (Levine, 2014). The null hypothesis assumes a position of no difference with respect to the population parameter(s) involved in the conjecture when compared to a specific value or one another. The alternate hypothesis challenges the assertion of the null hypothesis (Zacks, 2014). Drawing on the fact that the process is dependent on observable values and based upon probability theory, the task is then to look at the sample evidence and determine whether there is reasonable doubt on the validity of the assumption. The rule or test which determines whether the null shall be rejected or not, is based upon the sampling distribution of the test statistic (Casella & Berger, 2013).

Consider the following example. It is of interest to test the validity of the claim that babies born to mothers who are regular smokers have lower weight at birth on an average than those born to mothers who do not. Then mean weight at birth for babies born to mothers who are regular smokers could be compared with the mean birth weight of those born to mothers who do not smoke. According to null hypothesis testing method, then mean or average weight at birth would be the statistic of interest (Salkind, 2016). According to central limit theorem, large sample distribution of mean is normal with mean equal to the population mean, µ and variance equal to  where  is the population variance of weights of the babies at birth (Casella & Berger, 2013). Then assuming that specification of labelling parameter asserted by null hypothesis is true, if the probability that observed value of mean falls within the critical region is less than level of significance, α, which is a set probability is determined beforehand, the null is rejected at 100α% level of significance (Lowry, 2014). The following figure depicts the situation when the test fails to reject null.

Understanding Null and Alternate Hypotheses

Figure 1.1.a

Let us refer to the observed of mean weight as t_obs. Had the observed value of the test statistic t_obs, been greater than the critical value t_α, which is the value which marks critical region, then P (T>t_obs | H_{0 is} true) which is defined as the p-value, would be less than the level of significance leading to rejection of null hypothesis (Lehmann & Romano, 2012). Figure 1 however shows the situation when the probability is greater than α.

This shows how statistical hypothesis testing considers how likely a hypothesis is to be true or false. Four cases is likely to be the outcome of a testing procedure as depicted in the following table:

Truth	Null hypothesis is true	Null hypothesis is not true
Decision
Reject Null Hypothesis	Incorrect                            Type I error ( = α)	Correct
Fail to reject null hypothesis	Correct	Incorrect                                        Type II error

Table 1.1.a

This level of significance is the probability of type I error (Wasserman, 2013). The probability of type I error is defined as the probability of rejecting the null hypothesis when it is true. The type II error can be used to compute how powerful the test is. The value (1-probability of type II error) gives the power of the test which is a measure of how good the test is (Park, 2015). The main aim in determining the optimum test usually is to choose the test which has minimum probability of type I error while keeping type II error in check. This is usually best balanced with α=0.05 (Dancey & Reidy, 2007). This is in line with the intuition that rejection null hypothesis would have more severe consequences than acceptance of an alternate which is false.

The theory of hypothesis testing as explained above requires a large volume of data given its assumption being based on large sample distribution. The results are also vulnerable to sampling errors of the observable test statistic that is used to attain the final decision (Foster, Diamond & Jefferies, 2014). Additionally the procedure does not take into account any prior information in its assumptions which may amount to key information getting left out. A key problem thus lies in the interpretation of the results. Statistical significance does not necessarily suggest scientific significance (Wasserman, 2013). Although a small p-value suggests that chance of null being true is low, does not exactly indicate that chance of the alternative being true is high. Therefore p-value being less than significance level leads to rejection of the null but not the acceptance of the alternative hypothesis (Casella & Berger, 2013). This means that even if the null is failed to be rejected, it is simply owing to the fact that there is not enough evidence to accept is given the sample used to arrive at the conclusion. However there is no guarantee that it is really so (Levine, 2014).

Type I and Type II Errors

A confidence interval of a parameter is a range of values within which the population value of the parameter is expected to lie with a specific degree of certainty or confidence (Salkind, 2016). The confidence interval estimates as opposed to the point estimates of the parameter estimates a set of values that the parameter value may assume. These values are dispersed around the point estimate and lie within a margin of error, which depends on the size of the sample used for estimation and the required degree of certainty. Confidence intervals are thus by definition provide more information about the estimate, specifically the degree of uncertainty that the estimate has (Levine, 2014). This has far greater implications practically and the measure has thus gained in popularity in recent times.

The upper and lower bounds of the confidence interval, in general, are therefore defined as:

Upper bound = estimate + error margin

Lower bound = estimate – error margin

The error margin is computed using the formula:

The exact form of the intervals may however differ as per the type of estimates and thus the underlying distribution, the confidence level as reflected by the confidence coefficient (1-α) and the kind of sampling method that had been employed for estimation purposes. A 95% confidence intervals imply that that the confidence of the true parameter value to lie in the specified range of values is 95%. This does not mean that the probability of the true value to lie in the interval is 0.95. It means that 95% of the time that the estimation process has been generated, the true value has been found to lie in the specified interval (Migon, Gamerman & Louzada, 2014). A specific interval may either have the true population value or may not, but there is no in between. The rationale driving the theory is that, although two confidence intervals with two different sample of same length would be different, but repeatedly generating those using different samples of same size drawn randomly would exhibit that proportion of the times the intervals contain the true parameter value is equal to the confidence coefficient 1-α (Salkind, 2016).

Considering the example of estimating the mean of a normal distribution where the population standard deviation is known. The mean of random variables following normal distribution is known to be .The critical value for a 95% confidence interval would then be the z-statistic with 1-α = 0.95 that is, α = 0.05. The standard deviation of the mean would then be   (Lehmann & Romano, 2012).

Confidence Intervals and Estimation

Clearly the confidence interval could then be considered as a complement of the significance level of a two-tailed hypothesis test for significance of mean. Confidence intervals could then also be used to test hypotheses (Reichardt & Gollob, 2016).The confidence interval would not contain the hypothesised value of the mean under null hypothesis if the test statistic were to be significantly different from the value (Dancey & Reidy, 2007). Thus the null hypothesis would be rejected.

The above example explains confidence intervals when the population standard deviation is known, however when the population value of standard error is unknown, the t-interval confidence intervals are used where the critical value is computed on basis of t-statistic and the standard error on basis of sample standard error (Cohen, Manion & Morrison, 2013).

A confidence interval with a specific level of confidence which is narrower is the one which is more preferred. The width of a confidence interval with a specific confidence coefficient depends on the standard error and hence on the size of the sample used to generate the confidence interval (Kalinowski, 2010). The width of the confidence interval increases with decrease in size of the sample that has been used to compute the confidence interval and vice versa (Dancey & Reidy, 2007). This is owing to the decrease in sampling error as the size of sample increases. Thus a 95% confidence interval with sample size, n=1000 would be narrower than one which has n=500.There for a confidence interval based on larger sample size will always be better. Again, a 99% confidence interval would be narrower than a 95% confidence interval (Migon, Gamerman & Louzada, 2014). This is due to the fact that increasing the confidence level would decrease the critical value and hence decrease the width. Lastly, a greater standard deviation of the sample would increase standard error and hence the width (Ott & Longnecker, 2015).  

The data relates to 253 participants from a cross-sectional survey of community volunteers in Australia. The relevant data for this part consists of four different sets of attributes for each participant, namely, age, sex, employment status and number of days that the participant had engaged in volunteering. Each participant has been assigned a unique ID, ranging from 1 to 253 for identification purposes. The age variable has an interval scale assuming discrete value of age in years, the “sex” variable has a nominal scale with two categories (Male = “1” and Female = “2”), the “EMPLOY” variable representing employment status is also categorical with two categories (Employed= “1” and Unemployed= “0”) and finally the “DAY_VOL” variable which represents number of days volunteered which is also interval scale assuming discrete values of the count of days volunteered.

Complementarity of Hypothesis Testing and Confidence Intervals

The chosen hypothesis out of the two options provided for this paper states that, “The male participants of the survey are older, on average, relative to the female participants”.

The formal definition of hypothesis is thus stated as follows:

H₀: The age of men are same as that of the women participants (Null)

H₁: The men are older than the women participants. (Alternate)

The following analysis makes use of the theory of null hypothesis testing. The data variables relevant for the chosen hypothesis are “Age” and “Sex”. It is seen that there are 75 individuals who are female and 178 who are male. The data can thus be viewed upon two sets of ages, one for females, having length of 75 data points and the other for males, having a length of 178 data points.

The frequency distribution of the ages of the females seem to be skewed right where as the length of right tail for males seem to be of much lesser degree than the females (Argyrous, 2011). This suggests that more females under the median age are involved rather than those who are over median age whereas the age distribution is comparatively even in case of the males (Salkind, 2016).

The minimum age among male participants was found to be 23 years and maximum was 73 years. The mean of age for the males is found to be 45.4 and the standard error is 1.02.Due to the fact that population variance of age distribution of the males was unknown, the confidence interval had been computed using t-intervals (George & Mallery, 2016). 

Then the computed 95% confidence interval for the mean age of the males was found to be (43.00, 44.99).

The minimum age among female participants was found to be 20 years and the maximum age was found to be 74 years. The mean age of the females was estimated to be 41.23. The standard error was found to be 1.55. The 95% confidence interval for the mean age of females was computed in the same way as the males due to same reasons. The confidence interval of the mean age of females was computed to be (39.70, 42.75).

Therefore it is indicated that the age for females are in a range of values which are slightly lower than those of the males.

The test for significance of the hypothesis can then be framed as follows:

Benefits and Limitations of Hypothesis Testing and Confidence Intervals

H₀: mean age for males= mean age for females (Null)

H₁: mean age for males > mean age for females (Alternate)

The task is then to compare the mean values of two groups of variables, namely the men and the women. The population variances are unknown and therefore the one tailed t-test for significance is used to test the hypothesis. The level of significance assumed is 0.05.The p-value is observed to be 0.026 which is less than the level of significance assumed. The null hypothesis is thus rejected at 0.05 or 5% level of significance. Therefore the test indicates that participating men are indeed older on an average than the women who are participating (George & Mallery, 2016). 

Approaching the problem through confidence intervals, the confidence interval of the difference of mean age of women from the men should be considered. The null hypothesis states that there is no difference between the mean ages of the two groups, or in other words, the difference between the mean is equal to zero. The null is therefore rejected if the confidence interval does not contain 0 (Salkind, 2016).Then the 95% confidence interval of the t-interval variety has been computed since the population variance of the statistic is unknown. The estimate of difference between the means has been found to be equal to 4.13. The confidence interval for the estimated difference in means was found to be (0.500, 7.844). Therefore the null is rejected at 5% level of significance. The analysis thus concludes that the hypothesis, “The male participants of the survey are similar, on average, to the female participants” is rejected at 5% level of significance in favour of the conjecture that age of the males are higher than the females.

References

Argyrous, G. (2011). Statistics for research: with a guide to SPSS. Sage Publications.

Casella, G., & Berger, R. L. (2013). Statistical inference. Australia: Brooks/Cole Cengage Learning.

Cohen, L., Manion, L., & Morrison, K. (2013). Research methods in education. Routledge.

Cronk, B. C. (2017). How to use SPSS®: A step-by-step guide to analysis and interpretation. Routledge.

Dancey, C. P., & Reidy, J. (2007). Statistics without maths for psychology. Pearson Education.

Foster, L., Diamond, I., & Jefferies, J. (2014). Beginning statistics: An introduction for social scientists. Sage.

George, D., & Mallery, P. (2016). IBM SPSS Statistics 23 step by step: A simple guide and reference. Routledge.

Greenland, S., Senn, S. J., Rothman, K. J., Carlin, J. B., Poole, C., Goodman, S. N., & Altman, D. G. (2016). Statistical tests, P values, confidence intervals, and power: a guide to misinterpretations. European journal of epidemiology, 31(4), 337-350.

Hinton, P. R., McMurray, I., & Brownlow, C. (2014). SPSS explained. Routledge.

Kalinowski, P. (2010). Understanding confidence intervals (CIs) and effect size estimation. APS Observer, 23(4).

Leech, N. L., Barrett, K. C., & Morgan, G. A. (2014). IBM SPSS for intermediate statistics: Use and interpretation. Routledge.

Lehmann, E. L., & Casella, G. (2013). Theory of point estimation. Springer Science & Business Media.

Lehmann, E. L., & Romano, J. P. (2012). Testing statistical hypotheses. Springer Science & Business Media.

Levine, G. (2014). Introductory statistics for psychology: the logic and the methods. Academic Press.

Lowry, R. (2014). “Concepts and applications of inferential statistics.”

McCormick, K., Salcedo, J., Peck, J., & Wheeler, A. (2017). SPSS Statistics for data analysis and visualization. John Wiley & Sons.

Migon, H. S., Gamerman, D., & Louzada, F. (2014). Statistical inference: an integrated approach. CRC press.

Ott, R. L., & Longnecker, M. T. (2015). An introduction to statistical methods and data analysis. Nelson Education.

Park, H. M. (2015). Hypothesis testing and statistical power of a test.

Reichardt, C. S., & Gollob, H. F. (2016). When confidence intervals should be used instead of statistical tests, and vice versa. What If There Were No Significance Tests?: Classic Edition, 231.

Salkind, N. J. (2016). Statistics for people who (think they) hate statistics. Sage Publications.

Sekaran, U., & Bougie, R. (2016). Research methods for business: A skill building approach. John Wiley & Sons.

Tressoldi, P. E., & Utts, J. (2015). Statistical guidelines for empirical studies. Parapsychology. A Handbook for the 21st Century, 83-93.

Wasserman, L. (2013). All of statistics: a concise course in statistical inference. Springer Science & Business Media.

Zacks, S. (2014). Parametric statistical inference: basic theory and modern approaches (Vol. 4). Elsevier.