Introduction
(Slide 1 of 40)

Chapter 18. Crisis Text Line – Crisis Trends: Everything Up to Nonparametric Tests with Ordinal Data

Which Test Is Best?
true
true
You must read each slide, and complete any questions on the slide, in sequence.

By Warren Fass, University of Pittsburgh Bradford, and Susan A. Nolan, Seton Hall University

Crisis Trends (February 2, 2017). Crisis text line. Opens in new window http://crisistrends.org/.

In this activity, we will consider various findings from the Crisis Text Line, a resource that offers crisis counseling via text messages: http://crisistrends.org/. Then you will identify statistical analyses that could be used to support the various conclusions.

Guidelines for choosing the appropriate hypothesis test

A table with books on it.
LDprod/Shutterstock

“help something happened last night.”

“i just feel awful... im in the bathroom at my school crying.”

“I feel completely invisible.”

“Why can't I be happy?!”

These are just some of the texts that kicked off conversations with volunteers trained to help people in crisis. Since 2013, Crisis Text Line has been available for people in the U.S. to seek help when faced with difficult or tragic circumstances. As of February 2017, over 30 million text messages have been exchanged, helping people cope with bullying, substance use, sexual abuse, thoughts of suicide, and many other psychological issues. Just text HELLO to 741 741 if you or a friend ever needs instant help.
Researchers can apply for access to raw data – stripped of all identifying information – from the Crisis Text Line. As their Web site tells us, “The anonymous data collected by text is teaching us when crises are most likely to happen — and helping schools and law enforcement to prepare for them.” Universities, for example, might use trends from these data to plan mental health initiatives on campus.
But anyone can go online to http://crisistrends.org/ and play with the data via a series of interactive graphs. (We’ll call this interactive Web site “Crisis Trends” for short in the scenarios below.) What can you learn be exploring the data? We’ll examine several data-driven trends. For each trend described in this activity, identify the statistical analysis that could have be used to draw a conclusion.

Crisis Trends allows visitors to view the patterns of crisis texting across the U.S. For example, Crisis Trends provides ranks for the frequencies of texting data for each of 18 different psychological issues (e.g., bullying, friend issues, and depression) for all states. Imagine that we determined the rankings of a sample of states in terms of the frequencies of texts about bullying issues for 2016. So, the state with the most texts related to bullying would be ranked one. Imagine that we also determined the rankings of these same states for the frequencies of texts about friend issues, also for 2016. Similarly, the state with the most texts related to friend issues would be ranked one.

Which statistical test could be used to determine if the states’ ranks for bullying and friend issues were related?
Correct! The researchers could have used the Spearman rank-order correlation coefficient because there are only ordinal variables, each state’s bullying rank and friend issues rank. There are no scale or nominal variables. In addition, the question is about the association between two variables.
Actually, that’s not the correct statistical analysis. Let’s walk through the questions on the flowchart in Appendix E to determine what analysis could be used in this case.

In which of the following four categories does this situation fall? Click to see the data again. And click on the flowchart button to see the overview for choosing the best test.

Crisis Trends allows visitors to view the patterns of crisis texting across U.S. For example, Crisis Trends provides ranks for the frequencies of texting data for each of 18 different psychological issues (e.g., bullying, friend issues, and depression) for all states. Imagine that we determined the rankings of a sample of states in terms of the frequencies of texts about bullying issues for 2016. So, the state with the most texts related to bullying would be ranked one. Imagine that we also determined the rankings of these same states for the frequencies of texts about friend issues, also for 2016. Similarly, the state with the most texts related to friend issues would be ranked one.
Correct! There is at least one ordinal variable.
Actually, there is at least one ordinal variable.

We know there is at least one ordinal variable. In which of the following categories – both involving ordinal variables – does this situation fall?
Correct! There are only ordinal variables – bullying rank and friend issues rank for each state. And the question is about whether these two variables are associated.
Actually, there are only ordinal variables – bullying rank and friend issues rank for each state. And the question is about whether these two variables are associated.

Based on the answers to these questions, which statistical test could be used to determine if there was a relationship between the two variables, bullying rank and friend issues rank?
Correct! The researchers could have used the Spearman rank-order correlation coefficient because there are only ordinal variables, each state’s bullying rank and friend issues rank, and the research question is about the association between two variables.
Actually, that’s not the correct statistical analysis. The researchers could have used the Spearman rank-order correlation coefficient because there are only ordinal variables, each state’s bullying rank and friend issues rank, and the research question is about the association between two variables.

Crisis Trends also allows us to look at trends for frequency of texts from any month. For example, let’s say we want to know whether depression-related texts occurred more frequently in December than in January in 2016. Imagine that we take a sample of days from each month and treat each day as a participant; let’s say we look at 11 days, or participants, in January, and 10 days, or participants, in December. Each participant (day) receives a score for the number of texts related to depression occurring during that day.
The bar graph here depicts hypothetical means for samples of days from December and January. These means are based on the actual data for those months – 100 depression-related texts for January and 88 depression-related texts for December.

The mean numbers of depression texts in January and December.
Source: Number of texter conversations with Depression from Jan & Dec 2016 http://crisistrends.org/

The mean numbers of depression texts in January and December, Image Long Description

The mean numbers of depression texts in January and December. The left column represents the month, and the right column contains the corresponding value.

The mean numbers of depression texts in January and December
Month Mean Number of Depression Texts
January 3.23
December 2.84

What statistical analysis could be used to determine whether there was a significant difference between the numbers of texts for January and December in 2016?














Correct! The researchers could have used an independent-samples t test, because there is one nominal independent variable, month, with two levels or groups: January and December. There is a scale dependent variable, number of texts per day. And participants (days) are only in one of the two groups, so it is a between-groups design.
That’s not the correct statistical analysis. Let’s walk through the questions on the flowchart in Appendix E to determine what analysis could be used in this case.

In which of the following four categories does this situation fall? Click to see the data again. And click on the flowchart button to see the overview for choosing the best test.

Crisis Trends also allows us to look at trends for frequency of texts from any month. For example, let’s say we want to know whether depression-related texts occurred more frequently in December than in January in 2016. Imagine that we take a sample of days from each month and treat each day as a participant; let’s say we look at 11 days, or participants, in January, and 10 days, or participants, in December. Each participant (day) receives a score for the number of texts related to depression occurring during that day. The bar graph here depicts hypothetical means for samples of days from December and January. These means are based on the actual data for those months – 100 depression-related texts for January and 88 depression-related texts for December.
Correct! There is at least one nominal independent variable, and a scale dependent variable.
Actually, there is at least one nominal independent variable and a scale dependent variable.

How many nominal independent variables are there?
Correct! There is one nominal independent variable, month. (The dependent variable, number of texts per day, is scale.)
Actually, there is only one nominal independent variable, month. (The dependent variable, number of texts per day, is scale.)

How many levels does the independent variable have?
Correct! There are two levels or groups, January and December.
Actually, there are two levels or groups, January and December.

How many samples are there?
Correct! There are two samples, one consisting of days in January and one consisting of days in December.
Actually, there are two samples, one consisting of days in January and one consisting of days in December.

What type of design is this?
Correct! This is a between-groups design. Each participant (day) appears in only in one of the groups.
Actually, this is a between-groups design. Each participant (day) appears in only one of the groups.

Based on the answers to these questions, what statistical analysis could be used to determine whether there was a significant difference in the number of texts between the two groups?














Correct! The researchers could have used an independent-samples t test, because there is one nominal independent variable, month, with two levels or groups, January and December. There is a scale dependent variable, number of texts per day. And participants (days) are only in one of the two groups, so it is a between-groups design.
Actually, that’s not the correct statistical analysis. The researchers could have used an independent samples t test because there is one nominal independent variable, month, with two levels or groups, January and December. There is a scale dependent variable, number of texts per day. And participants (days) are only in one of the two groups, so it is a between-groups design.

Crisis Trends also allows us to look at the frequency of texts about different crises, such as depression-related texts, occurring during a given month. Let’s say that we wonder whether the frequency of texts may be related to temperature. Imagine that we take a sample of days from 2015 and treat each day as a participant. Let’s say we look at 60 days, or participants, for 2015. Each participant (day) receives two scores: the number of depression-related texts received that day and the high temperature.

Which statistical test could be used to determine if there was a significant relation between the numbers of texts and temperature?
Correct! The researchers could have used the Pearson correlation coefficient because there are only scale variables, number of depression-related texts and temperature, and the question is about association rather than prediction.
That’s not the correct statistical analysis. Let’s walk through the questions on the flowchart in Appendix E to determine what analysis could be used in this case.

In which of the following four categories does this situation fall? Click to see the data again. And click on the flowchart button to see the overview for choosing the best test.

Crisis Trends also allows us to look at the frequency of texts about different crises, such as depression-related texts, occurring during a given month. Let’s say that we wonder whether the frequency of texts may be related to temperature. Imagine that we take a sample of days from 2015 and treat each day as a participant. Let’s say we look at 60 days, or participants, for 2015. Each participant (day) receives two scores: the number of depression-related texts received that day and the high temperature.
Correct! There are only scale variables, number of depression-related texts and temperature.
Actually, there are only scale variables, number of depression-related texts and temperature.

Is the research question about association or prediction?
Correct! The research question is about association. We want to know if two variables are related, not whether one predicts the other.
Actually, the research question is about association. We want to know if two variables are related, not whether one predicts the other.

Based upon the answers to these questions, which statistical test could be used to determine if there was a significant relation between the number of depression-related texts and temperature?
Correct! The researchers should have used the Pearson correlation coefficient because there are only scale variables, number of depression-related texts and temperature, and the research question is about association rather than prediction.
Actually, that’s not the correct statistical analysis. The researchers should have used the Pearson correlation coefficient because there are only scale variables, number of depression-related texts and temperature and the research question is about association rather than prediction.

What is a more common problem among Crisis Trend texters – substance abuse or health issues? The graph below shows the frequency of each crisis over time. Imagine that participants in this study are a sample of months from 2015 and 2016. Each month gets two scores – the number of texts related to substance abuse issues and the number of texts related to health issues.

Source: Number of texter conversations with Substance Abuse & Health Concerns http://crisistrends.org/

The numbers of texter conversations with substance abuse and health concerns over months from 2015 and 2016., Image Long Description

The numbers of texter conversations with substance abuse and health concerns over months from 2015 and 2016. The first column represents the month, the second column shows the corresponding number of texts related to substance abuse, and the last column contains the corresponding number of texts related to health concerns.

The numbers of texter conversations with substance abuse and health concerns over months from 2015 and 2016.
Month Substance Abuse Health Concerns
January 2015 65 96
June 2015 93 100
December 2015 86 49
January 2016 82 45
June 2016 90 56
December 2016 100 60

Which statistical test could be used to determine if there was a significant difference between the number of texts for each month (participant)?
Correct! The researchers could have used a paired-samples t test because there is one nominal independent variable, type of crisis, with two levels or groups: substance abuse and health concerns. There is a scale dependent variable, number of texts. And all participants (months) are in both groups, so it is a within-groups design.
That’s not the correct statistical analysis. Let’s walk through the questions on the flowchart in Appendix E to determine what analysis could be used in this case.

In which of the following four categories does this situation fall? Click to see the data again. And click on the flowchart button to see the overview for choosing the best test.

What is a more common problem among Crisis Trend texters – substance abuse or health issues? The graph below shows the frequency of each crisis over time. Imagine that participants in this study are a sample of months from 2015 and 2016. Each month gets two scores – the number of texts related to substance abuse issues and the number of texts related to health issues.





Correct! There is at least one nominal independent variable and a scale dependent variable.
Actually, there is at least one nominal independent variable and a scale dependent variable.

How many nominal independent variables are there?
Correct! There is one nominal independent variable, type of crisis. (The dependent variable, number of texts, is scale.)
Actually, there is one nominal independent variable, type of crisis. (The dependent variable, number of texts, is scale.)

How many levels does the independent variable have?
Correct! There are two levels or groups, substance abuse and health concerns.
Actually, there are two levels or groups, substance abuse and health concerns.

How many samples are there?
Correct! There are two samples, one consisting of substance abuse texts and one consisting of health concern texts.
Actually, there are two samples, one consisting of substance abuse texts responses and one consisting of health concern texts.

What type of design is this?
Correct! This is a within-groups design. Each participant (month) is in both groups.
Actually, this is a within-groups design. Each participant (month) is in both groups.

Based on the answers to these questions, which statistical test could be used to determine if there was a significant difference between the number of texts for the two groups?
Correct! The researchers could have used a paired-samples t test because there is one nominal independent variable, type of crisis, with two levels or groups: substance abuse and health concerns. There is a scale dependent variable, number of texts. And all participants (months) are in both groups, so it is a within-groups design.
Actually, that’s not the correct statistical analysis. The researchers could have used a paired-samples t test because there is one nominal independent variable, type of crisis, with two levels or groups: substance abuse and health concerns. There is a scale dependent variable, number of texts. And all participants (months) are in both groups, so it is a within-groups design.

Are stress-related issues more common for people living in certain regions of the U.S.? For example, Crisis Trends provides the ranks for stress related-issues for individual states for a specific year. Imagine that we want to look at the stress related-ranks for the year 2015. We then separate the 48 contiguous states into the following groups: states that are east of the Mississippi River and states that are west of the Mississippi River. The separation would result in 26 states east of the Mississippi River and 22 states west of the Mississippi River (the states are considered the participants). We could determine the stress-related ranks of a randomly selected sample of states from each group.

Which statistical test could be used to determine if there was a significant difference between the ranks for the two groups?
Correct! The researchers could have used the Mann-Whitney U test because there is one nominal independent variable, location of state, with two levels or groups, east and west. There is an ordinal dependent variable, stress-related rank. And participants (states) are only in one of the two groups, so it is a between-groups design.
That’s not the correct statistical analysis. Let’s walk through the questions on the flowchart in Appendix E to determine what analysis could be used in this case.

In which of the following four categories does this situation fall? Click to see the data again. And click on the flowchart button to see the overview for choosing the best test.

Are stress-related issues more common for people living in certain regions of the U.S.? For example, Crisis Trends provides the ranks for stress related-issues for individual states for a specific year. Imagine that we want to look at the stress related-ranks for the year 2015. We then separate the 48 contiguous states into the following groups: states that are east of the Mississippi River and states that are west of the Mississippi River. The separation would result in 26 states east of the Mississippi River and 22 states west of the Mississippi River (the states are considered the participants). We could determine the stress-related ranks of a randomly selected sample of states from each group.
Correct! There is at least one ordinal variable.
Actually, there is at least one ordinal variable.

We know there is at least one ordinal variable. In which of the following categories – both involving ordinal variables – does this situation fall?



Correct! There is one nominal independent variable, location of state, and one ordinal dependent variable, stress-related ranks.
Actually, there is one nominal independent variable, location of state, and one ordinal dependent variable, stress-related ranks.

What type of design is this?
Correct! This is a between-groups design. Each participant (state) is located in only one of the groups (regions).
Actually, this is a between-groups design. Each participant (state) appears in only one of the groups (regions).

How many groups (regions) are being studied?
Correct! There are two groups (regions) – east and west of the Mississippi.
Actually, there are two groups (regions) – east and west of the Mississippi.

Based upon the answers to these questions, what statistical analysis could be used to determine if there was a significant difference between the ranks for the two groups?
Correct! The researchers could have used the Mann-Whitney U test because there is one nominal independent variable, location of state, with two levels or groups (regions): east and west. There is an ordinal dependent variable, stress-related rank. And participants (states) are only in one of the two groups (regions), so it is a between-groups design.
Actually, that’s not the correct statistical analysis. The researchers could have used the Mann-Whitney U test because there is one nominal independent variable, location of state, with two levels or groups (regions): east and west. There is an ordinal dependent variable, stress-related rank. And participants (states) are only in one of the two groups (regions), so it is a between-groups design.

Let’s look at one more set of data from Crisis Trends. Are there differences in the numbers of suicidal-thought texts based on time of the year? Imagine that we take a sample of days from each month and treat each day as a participant. Let’s say we look at 20 days (or participants) in January, 20 days (or participants) in June, and 20 days (or participants) in December. Each day (participant) would receive a score for the number of texts sent that day related to suicidal-thoughts.

Which statistical test could be used to determine if there were significant differences in the number of texts sent during each of the three months?
Correct! The researchers could have used a one-way within-groups ANOVA, because there is one nominal independent variable, month, with three levels or groups: January, June, and December. There is a scale dependent variable, number of suicidal-thoughts texts. And days (participants) are only in one of the three groups, so it is a between-groups design.
Actually, that’s not the correct statistical analysis. Let’s walk through the questions on the flowchart in Appendix E to determine what analysis could be used in this case.

In which of the following four categories does this situation fall? Click to see the data again. And click on the flowchart button to see the overview for choosing the best test.

Let’s look at one more set of data from Crisis Trends. Are there differences in the number of suicidal-thought texts based on time of the year? Imagine that we take a sample of days from each month and treat each day as a participant. Let’s say we look at 20 days (or participants) in January, and 20 days (or participants) in June, and 20 days (or participants) in December. Each day (participant) would receive a score for the number of texts sent that day related to suicidal-thoughts.
Correct! There is at least one nominal independent variable and a scale dependent variable.
Actually, there is at least one nominal independent variable and a scale dependent variable.

How many nominal independent variables are there?
Correct! There is one nominal independent variable, month. (The dependent variable, number of texts, is scale.)
Actually, there is one nominal independent variable, type of crisis. (The dependent variable, number of texts, is scale.)

How many nominal independent variables are there?
Correct! There are three levels or groups: January, June, and December.
Actually, there are there are three levels or groups: January, June, and December.

What type of design is this?
Correct! Each participant (day) appears in only one of the three groups, so it is a between-groups design.
Actually, each participant (day) appears in only one of the three groups, so it is a between-groups design.

Based on the answers to these questions, what statistical analysis could be used to determine if there were significant differences in the number of texts sent during each of the three months?
Correct! The researchers could have used a one-way within-groups ANOVA. There is one nominal independent variable, month, with three levels or groups: January, June, and December. There is a scale dependent variable, number of suicidal-thoughts texts. And days (participants) are only in one of the three groups, so it is a between-groups design.
Actually, that’s not the correct statistical analysis. The researchers could have used a one-way within-groups ANOVA. There is one nominal independent variable, month, with three levels or groups: January, June, and December. There is a scale dependent variable, number of suicidal-thoughts texts. And days (participants) are only in one of the three groups, so it is a between-groups design.

Congratulations! You have completed the activity and gained some good experience in choosing the best hypothesis test.