Correlational analysis Roshani K. Prematunga BCom, PGDipSci (Statistics), MPH ∗
Centre for Psychiatric Nursing, Faculty of Medicine, Dentistry & Health Sciences, The University of Melbourne, Level 2, 757, Swanston St, Carlton, VIC 3053, Australia
Received 7 October 2011 ; received in revised form 31 January 2012; accepted 14 February 2012
KEYWORDS Summary A common question of interest in nursing research is the relationships
Introduction variable to predict an outcome variable, which
is a statistical technique known as simple linear In nursing research, examining relationships regression. A causal relationship can be estab- between variables are often of interest and corre- lished using regression. However, correlational lation is a common statistical technique employed analysis is only appropriate to explore relation- to perform the task. Correlation refers to how ship between variables and not to infer causal closely two or more variables are related. In relationship. correlational analysis, magnitude and direction of Bivariate, partial and multiple correlational linear relationships between variables would be analyses are common variants of correlational estimated and the statistical significance1 needs to analysis. Bivariate correlation measures the rela- be tested, where the null hypothesis is that there tionship between two variables. Partial correlation is no relationship between the variables and the examines the relationship between two variables alternative hypothesis is that a relationship exists.2 while taking into consideration the effect of a third Once a relationship has been established using variable.3 Multiple correlation examines the cor- correlational analysis, a model can be created relation between the outcome variable and the that allows the researcher to use an explanatory combined effect of other predictors.3 Statistical software has become a common tool to compute ∗ Tel.: +61 3 9035 4210; fax: +61 3 8344 7733. different correlation statistics. The aim of this E-mail address: [email protected] paper is to outline different approaches used to
examine bivariate relationships of interval/ratio,
ordinal and nominal variables4 and demonstrate the Box 1 use of correlational analysis using examples from Aim: To examine whether nurses in Australia published research. participated in clinical decision-making to the extent they desired. A total of 174 registered nurses were selected and interval variables: professional orientation, paramedical orienta- Correlation analysis using interval/ratio tion and perceived decisions were considered. variables Correlation between Professional orientation and Perceived decisions, and Paramedical Correlational analysis in interval/ratio variables orientation and Perceived decisions were prerequisites confirming a linear relationship and examined. no outlier in the data using Scatterplots. Fig. 1A Significance level was set at p < 0.01. The suggests a positive linear relationship between Pearson’s correlation of 0.332 indicates a X and Y, where as X increases, Y also increases. somewhat weak positive relationship between Fig. 1B suggests a negative linear relationship, professional values and perceived decision- where as X increases, Y decreases. Fig. 1C shows making and showed it was highly statistically that the points on the graph are scattered ran- significant (p < 0.001), suggesting nurses hold- domly and depicts no apparent linear relationship. ing professional values participated more No outliers are present in Fig. 1. in perceived decision-making (Table 1). The The most commonly used coefficient to mea- Pearson’s correlation of −0.250 indicates a sure correlation between interval/ratio variables somewhat weak negative relationship between is the Pearson product-moment correlation coeffi- paramedical values and perceived decision- cient (Pearson’s r)2 which provides a measure for making and it was highly statistically significant the magnitude and direction of a linear relation- (p = 0.007), suggesting nurses holding paramed- ship. Values for Pearson’s r vary between −1.00 and ical values participated less in perceived +1.00. A correlation coefficient of zero indicates decision-making (Table 1). no linear relationship between the variables; −1.00 Source: Hoffman et al.6 indicates a perfect negative linear relationship, and +1.00 indicates a perfect positive linear relation- ship. In practice, a perfect relationship between Correlation analysis using ordinal variables is unlikely to be observed. Negative coef- variables ficient indicates a negative linear relationship and positive coefficient indicates a positive linear rela- The most common test statistic used to exam- tionship. A statistical test can be performed to ine the association between ordinal variables is evaluate the strength of a relationship between the Spearman’s rank-order correlation (rs ), also two variables. The null hypothesis is that no lin- referred to as Spearman’s rho [].5 Spearman’s ear relationship exists between the two variables. correlation coefficient is a measure of correlation It is possible for larger samples to show statistical which uses rank order to determine the magnitude significance despite a small correlation coefficient. and direction of relationship between two sets of p-Value is used to assess the statistical significance ranked data. Similar to the Pearson’s r, values for whereas the practical importance can be assessed Spearman’s correlation coefficient vary between with an indicator known as the coefficient of deter- −1.00 and +1.00 and can be interpreted in a similar mination (R2 ). The coefficient of determination is fashion. The strength of the relationship can be given by the square of the Pearson correlation coef- ficient. It describes the proportion of variance in the outcome variable which can be explained by Table 1 Significant correlation between decision- the variance in the predictor variable.5 For exam- making and occupational orientation. ple, if two variables have a correlation coefficient Perceived decisions of 0.6, then the R2 = 0.36. Therefore, 36% of the variation in the outcome variable can be explained 0.332a Professional orientation <0.001b by the variation in the predictor variable. Most sta- tistical packages produce Pearson’s r along with the −0.250a Paramedical orientation 0.007b p-value. An example from published research which illustrates the use of Pearson’s correlation using a Pearson’s correlation. interval variables is provided in Box 1 . b p-Value. Correlational analysis 197
Fig. 1 Scatterplots showing various relationships.
is Kendall’s Tau-a, where each pair of observa-
Box 2 tions is examined and classified as concordant Aim: To examine whether nurses in Australia or discordant.7 Kendall’s Tau coefficient is com- participated in clinical decision-making to the puted based on the ordering of the values and extent they desired. A total of 174 registered extreme values do not alter the estimate of the nurses were selected. Hoffman et al. computed coefficient.8 A large number of concordant pairs correlational analysis on an ordinal variable, would be expected in the sample if the relation- level of appointment and perceived decisions ship between the two variables were positive in to examine linear relationship. the population. A large number of discordant pairs Significance level was set at p < 0.01. The would be expected in the sample if the relationship Spearman’s correlation of 0.338 indicates a between the two variables were negative in the somewhat weak positive relationship between population. Kendall’s Tau is interpreted in a simi- level of appointment and perceived decision- lar fashion to Pearson’s r where the coefficient can making and it was highly statistically significant range between −1.00 and +1.00. The sign of the (p < 0.001), suggesting nurses practicing at a coefficient indicates the direction of the relation- higher level of appointment participated more ship, and its absolute value indicates the strength, in decision-making (Table 2). with larger absolute values indicating stronger rela- Source: Hoffman et al.6 tionships and zero indicating no relationship. A statistical test can be performed to evaluate the strength of a relationship where the null hypoth- assessed using a statistical test, where the null esis is no linear relationship exists between the hypothesis is that there is no linear relationship two variables. Another version of Kendall’s Tau, between the two sets of ranks. Most statistical known as Kendall’s Tau-b is used to measure asso- packages also produce rs and the corresponding ciation based on concordant and discordant pairs p-value. We also see Spearman’s correlation of observations adjusted for the number of ties in employed in examining relationship between ranks, i.e. an observed value occurred more than ordinal and interval variables. An example from once.7 Kendall’s Tau is also employed in exam- published research which illustrates the use of ining relationship between ordinal and interval Spearman’s correlation using ordinal variables is variables. provided in Box 2 . An alternative statistic that can be employed in measuring association between ordinal variables Association with nominal variables
Table 2 Significant correlation between decision- Relationship between nominal variables can be making and level of appointment. assessed by cross-tabulating data in a contingency table9 and computing the Pearson’s Chi-square Perceived decisions statistic (2 ). The computed Chi-square value can 0.338a be tested for statistical significance where the Level of appointment <0.001b null hypothesis is no relationship exists between a Spearman’s correlation. the variables. Most statistical packages produce b p-Value. 2 along with the p-value. The existence of a 198 R.K. Prematunga
Box 3 Box 4 Aim: To examine association between the Aim: To examine the occurrence of epilepsy use of nutritional supplements used among ath- in relation to the etiology of intellectual dis- letes and the motive for using such substances ability in children. for health maintenance. A Chi-square test was performed on nomi- A Chi-square test was performed to deter- nal variables to determine association in the mine association between two nominal vari- occurrence of epilepsy among different cate- ables: supplement used and motive for gories of children with intellectual disability. using such substances, and the strength of To determine the strength of association, this association was estimated by comput- Cramer’s V coefficient was computed. An alpha ing ϕ coefficient. Statistical analyses were level of 0.05 was used. Analysis was conducted performed using SPSS.14.0.1. A statistically with computer program SPSS v.13 for Windows. significant association between taking Vitamin A statistically significant association C as a supplement to avoid sickness (2 = 54.67, between the occurrence of epilepsy between p < 0.001), and a somewhat intermediate asso- etiological categories, 2 (3, N = 167) = 30.085, ciation was found by phi correlation coefficient p = 0.001, and a somewhat moderate associa- of 0.324 (p < 0.001). tion was found by the Cramer’s V correlation Source: Petróczi et al.10 coefficient of 0.424 (p = 0.001) (Table 3). Source: Memisevic et al.11
relationship between nominal variables can be
determined using the 2 statistic. However, the Summary strength of the association between the two vari- ables is measured using the Phi coefficient (ϕ)2 or This paper has presented different approaches Cramer’s V. When a contingency table has 2 × 2 used to conduct correlational analysis based on cells, Phi coefficient is used to measure the strength different types of variables. Pearson’s correla- of relationship where the coefficient can range tion is appropriate for interval/ratio variables; between 0 and 1. A large value indicates a strong Spearman’s correlation and Kendall’s Tau appropri- association and a value close to zero indicates a ate for ordinal variables and Pearson’s Chi-square weak association between the two variables. When with Phi coefficient or Cramer’s V appropriate for a contingency table has more than 2 × 2 cells, nominal variables. Correlational analysis is a sta- Cramer’s V is used to measure the strength of rela- tistical technique employed in measuring linear tionship between two nominal level variables. We relationships between variables. A causal relation- also see Pearson’s Chi-square along with Phi coeffi- ship between variables cannot be inferred using cient or Cramer’s V employed in examining rela- correlational analysis, regardless of the magnitude tionship between nominal and ordinal variables. of the correlation observed. Examples from published research illustrating the use of Pearson’s Chi-square, Phi coefficient and Cramer’s V in examining association between two nominal variables are presented in Boxes 3 and 4 . References 1. Fethney J. Statistical and clinical significance, and how to use confidence intervals to help interpret both. Aust Crit Table 3 Occurrence of epilepsy in relation to the Care 2010;23:93—7. etiology of intellectual disability. 2. Polit DF. Statistics and data analysis for nursing research. Etiology Epilepsy 2nd ed. NJ: Pearson; 2010. 3. Sirkin RM. Statistics for the social sciences. 3rd ed. CA: Sage; Without With Total 2006. 4. Fisher MJ, Marshall AP. Understanding descriptive statistics. N % N % N % Aust Crit Care 2009;22:93—7. Down syndrome 34 100.0 0 0.0 34 100.0 5. Coolidge FL. Statistics a gentle introduction. 2nd ed. CA: Other genetic cause 15 68.2 7 31.8 22 100.0 Sage; 2006. 6. Hoffman K, Duffield C, Donoghue J. Barriers to clinical Brain injury 18 51.4 17 48.6 35 100.0 decision-making in nurses in Australia. AJAN 2004;21(3):8. Unknown etiology 66 86.8 10 13.2 76 100.0 7. Pett MA. Nonparametric statistics for health care research: Total 133 79.6 34 20.4 167 100.0 statistics for small samples and unusual distributions. CA: Sage; 1997. Correlational analysis 199
8. Wilcox RR. Fundamentals of modern statistical methods: in athletes’ supplement use for maintenance of health: a substantially improving power and accuracy. New York: retrospective study. Nutr J 2007;6(34). Springer; 2010. 11. Memisevic H, Sinanovic O. Epilepsy in children with intel- 9. Fisher MJ, Marshall AP, Mitchell M. Testing differences in lectual disability in Bosnia and Herzegovina: Effects of sex, proportions. Aust Crit Care 2011;24:133—8. level and etiology of intellectual disability. Res Dev Disabil 10. Petróczi A, Naughton DP, Mazanov J, Holloway A, Bingham 2009;30:1078—83. J. Limited agreement exists between rationale and practice
