Hello Nicole,

Step1. Defining variables. So, the first thing you have to think about is your design. Your research purpose brings up a causal effect where education in prison has an outcome in "lower recidivism rates". Your research question is actually beautiful, as usually prisons become more of a reproduction of the concept of hell, instead of recovering people, but back to the main topic.

After underlying your variable as education in prison and recidivism rates, you have to know also a couple more things: 1. how many prison scenarios are you going to compare. 2. how are you going to measure reicidivism.

Step2. Defining groups. For problem 1. you need one control group, which might be prisoners that are not going through any type of educational program - or that are going through a standard or common ground scenario. Besides the control group, you want at least one more group that is going to be under an educational program. That would give you two groups. You might add a third if you want to compare multiple educational strategies and your criteria on picking additional groups depends on what is of your interest. 

Step3. knowing what you are measuring. Here you have to figure out how are you going to measure reicidivism. That's what is going to give you a hint on which statistical procedure you are going to use. Why it matters so bad? Because statistical procedures depend on the level of the variable like if it is categorical, ordinal, intervalar or ratio. So you have to think what kind of response you are expecting.

"Question 1. I have been out of prison on my most previous charge for more than 3 years? " - this seems to be a yes/no question. Should be coded as 0/1

on the other questions 1, 2, 3 and 4, you think of using a rating scale with a forced response without a mid-term, which is really good to avoid a neutral subpopulation. But you might have a problem here. The layout of answers doesn't look like corresponding that much to what you are asking (e.g."1: I have been out of prison on my most previous charge for more than 3 years", that's not something you can agree within a rating scale, that's a no/yes answer model). So let's go back in to making the right answers on scaling what you want to measure.

1st. You need to define the group towards which the person belongs. "DId you participate of educational programs while in prison on your most previous charge?" - that would be yes/no again

2nd. You want to measure recidivism. So you might actually want to ask "Did you get back into jail again after prison?" - or some similar question; That can be a yes/no,  but you might want to ask how many times the person got back to jail or explore other issues related to reicidivism.

3rd, you are asking at question 2 if the person was capable of keeping the same job since he/she got out of prison. That's still a yes/no (1/0) question. You might add a new question "For how long have you been able to keep the same job". Explore the topic.

4th."I support myself?" - here you mean actually supporting economically or emotionally? Economically can be yes/no. If you mean emotionally, it can be a rating scale. Still here you can use also a more complex measure, like an affective scale such as Positive Affect Negative Affect Scale (PANAS) or similar measures.

5th. Here you are asking the person itself how much he/she believes in the efficacy of educational programs in jail. Then here you will be able to use a different measurement level. you might really use a rating scale as you proposed, or you might check on Visual-Analogic Scale (VAS), which can be a 10cm printed line in one end with "I don't believe in the efficacy at all" to "I think it is extremely effective" - please name the extremes the way you think it is better. 

The point is, here, looking at your questions, simplifying: for your binary "yes/no" questions, a Chi-square test can suffice. Still, I think a logistic regression would give a much more interesting information as it gives you information on probability (like, "participants who went through prison without educational programs are 2.23 times more likely to commit a crime again, etc".

For questions with a rating scale or VAS, you can use methods such as t-test and ANOVA. T-Test would suffice if you define only two groups (didn't go through educational procedures x went through educational procedures). A one-way ANOVA would be your solution if you have more than 2 educational scenarios. Still, that's true if you compose scores (summing up items). If you work with them isolated, you would have to work with contingency tables and chi-square again. Another option might be using other non-parametric tests, but all of this is going to depend on the rest of your research design.

Advice: make more questions. Your list still is too small. Tha would also help giving you more hints on how would you do the data analysis.

Hello Nicole,

Step1. Defining variables. So, the first thing you have to think about is your design. Your research purpose brings up a causal effect where education in prison has an outcome in "lower recidivism rates". Your research question is actually beautiful, as usually prisons become more of a reproduction of the concept of hell, instead of recovering people, but back to the main topic.

After underlying your variable as education in prison and recidivism rates, you have to know also a couple more things: 1. how many prison scenarios are you going to compare. 2. how are you going to measure reicidivism.

Step2. Defining groups. For problem 1. you need one control group, which might be prisoners that are not going through any type of educational program - or that are going through a standard or common ground scenario. Besides the control group, you want at least one more group that is going to be under an educational program. That would give you two groups. You might add a third if you want to compare multiple educational strategies and your criteria on picking additional groups depends on what is of your interest. 

Step3. knowing what you are measuring. Here you have to figure out how are you going to measure reicidivism. That's what is going to give you a hint on which statistical procedure you are going to use. Why it matters so bad? Because statistical procedures depend on the level of the variable like if it is categorical, ordinal, intervalar or ratio. So you have to think what kind of response you are expecting.

"Question 1. I have been out of prison on my most previous charge for more than 3 years? " - this seems to be a yes/no question. Should be coded as 0/1

on the other questions 1, 2, 3 and 4, you think of using a rating scale with a forced response without a mid-term, which is really good to avoid a neutral subpopulation. But you might have a problem here. The layout of answers doesn't look like corresponding that much to what you are asking (e.g."1: I have been out of prison on my most previous charge for more than 3 years", that's not something you can agree within a rating scale, that's a no/yes answer model). So let's go back in to making the right answers on scaling what you want to measure.

1st. You need to define the group towards which the person belongs. "DId you participate of educational programs while in prison on your most previous charge?" - that would be yes/no again

2nd. You want to measure recidivism. So you might actually want to ask "Did you get back into jail again after prison?" - or some similar question; That can be a yes/no,  but you might want to ask how many times the person got back to jail or explore other issues related to reicidivism.

3rd, you are asking at question 2 if the person was capable of keeping the same job since he/she got out of prison. That's still a yes/no (1/0) question. You might add a new question "For how long have you been able to keep the same job". Explore the topic.

4th."I support myself?" - here you mean actually supporting economically or emotionally? Economically can be yes/no. If you mean emotionally, it can be a rating scale. Still here you can use also a more complex measure, like an affective scale such as Positive Affect Negative Affect Scale (PANAS) or similar measures.

5th. Here you are asking the person itself how much he/she believes in the efficacy of educational programs in jail. Then here you will be able to use a different measurement level. you might really use a rating scale as you proposed, or you might check on Visual-Analogic Scale (VAS), which can be a 10cm printed line in one end with "I don't believe in the efficacy at all" to "I think it is extremely effective" - please name the extremes the way you think it is better. 

The point is, here, looking at your questions, simplifying: for your binary "yes/no" questions, a Chi-square test can suffice. Still, I think a logistic regression would give a much more interesting information as it gives you information on probability (like, "participants who went through prison without educational programs are 2.23 times more likely to commit a crime again, etc".

For questions with a rating scale or VAS, you can use methods such as t-test and ANOVA. T-Test would suffice if you define only two groups (didn't go through educational procedures x went through educational procedures). A one-way ANOVA would be your solution if you have more than 2 educational scenarios. Still, that's true if you compose scores (summing up items). If you work with them isolated, you would have to work with contingency tables and chi-square again. Another option might be using other non-parametric tests, but all of this is going to depend on the rest of your research design.

Advice: make more questions. Your list still is too small. Tha would also help giving you more hints on how would you do the data analysis.

ANOVA  will be most suitable statistical technique

My 2 pence (possibly not worth considering after Pedros comprehensive answer):

Forget testing.

Better try to estimate sensible effects and relations.

Your response seems to be a proportion. This calls for a logistic regression model.

The answers to the questions would be your predictors. There are no models for ordinal predictors (like Likert scaled data). This will strongly limit your possibilities. You could treat the predictors as (unordered) factors, but then you will ignore all the information provided by the ordered structure of answers.

If I had to plan such a study I would try to get at leadt interval-scaled predictors. In your case I would let the people place a mark on a line and then measure the distance to the start of the line (relative to the length of the line) and use this as "degree of aggreement".

Example:

Q: I support myself:

                   _____________________________________

              '                           '                           '

       not at all                 so-so                absolutely

When the mark is drawn at "not at all", the rel. distance is 0, at "so-so" it would be 0.5 and at "absolutely" it would be 1.0. All positions inbetween are also possible.

If required, the resulting values (L) are bounded in [0,1] and they can be transformed to the whole real line by a logit-transformation:

X = log(L/(1-L))

possibly after squeezing the values a bit so that L=0 and L=1 will not occur (these would give infinite logit values).

With these predictors you can set up a logistic regression model making the best use of all information.

I recommend the use of "path analysis"... or structural equation modeling (SEM)....

During survey creation one issue to try to correct for will be selection bias. That is, perhaps prisoners who opt to take advantage of education opportunities while incarcerated are different from those who don't EVEN BEFORE they get the education. You may need to include questions in the survey to determine whether the education programs were voluntary and/or do some investigations with the prison staff to find out how prisoners are selected for the programs.

It's a great question and I wish you success.

Nicole, I agree with  Danilo that structural equation modeling would be an approach whereby you would see the interrelationship among your variables. This relates to what Gary said that prisoners who choose to become educated in prison may have differences in their personality that could confound the association of education and recidivism. Education and job skill training prior to prison and during prison would be considerations. Education and job skills prior to prison could mediate education and its affect on recidivism. SEM could be set to test the degree of associations among variables and possible confounders such as prior offenses, neighborhood type, socioeconomic status, positive social support...etc. On the other hand, logistic regression could work if you consider recidivism (yes) as compared to (no) as the outcome. The main independents could be acquiring job skills or academic education  and adjustment variable would be all the previously mentioned factors that could influence either wanting job skills or education and/or recidivism. Another approach is using a mixed method design where open-ended interviews inform the quantitative design. In other words, concepts gained by finding out what motivates prisoners to receive an education when in prison would help you design a study to measure association.